1MAC Layer Abstraction for Simulation ScalabilityImprovements in Large-scale Sensor Networks

Tian He†, Brain Blum∗, Yvan Pointurier∗, Chenyang Lu‡, John A. Stankovic∗, Sang Son∗∗Department of Computer Science, University of Virginia

†Department of Computer Science and Engineering, University of Minnesota‡Department of Computer Science and Engineering, Washington University in St. Louis

Abstract— It is risky to implement a large-scale wireless sen-sor system without predicating its behavior beforehand throughsimulation, an indispensable step toward the final system. Sincein wireless simulation, over 90% simulation time is spent onthe characterization of the MAC layer behaviors, abstraction atthe MAC layer becomes the key to the scalability of wirelesssensor network simulations. In this work, we employ an abstractsimulation mode, established by online data collection, to describethe MAC behaviors efficiently. Observing that the abstract sim-ulation mode becomes inaccurate when network traffic changessignificantly, we adopt a hybrid approach. We switch between theabstract simulation mode and full simulation mode in responseto the traffic changes. Statistics are collected online under fullsimulation mode to update the abstract simulation mode to beused. In evaluation, we demonstrate that our approach is 9 timesfaster than the original simulation with a loss of accuracy less than10% in the end-to-end delay and 1% in the number of messagessuccessfully transmitted across the network.

I. INTRODUCTION

With the advance of sensor network research, numerousprotocols have been proposed recently. Validation and analysisof these protocols are very important for system designers tomake right decisions to meet application requirements. Onemethod available for validation and analysis is mathematicalmodeling. Unfortunately work in mathematics is limited dueto the complex, dynamic, and unpredictable nature of thesesystems as they scale to thousands of devices. One can alsoexplore new protocols for sensor networks directly with thephysical devices (e.g., mica2). Problems occur again with thecomplexity and scale of these networks. The sensor devicesthat exist today provide a limited subset of the functionalityand scale. Due to the practical limitations of mathematicalanalysis or physical network implementations, simulation is animportant tool for researchers in this field, at least during theinitial stage of the system design. Aside from just providing ameans of representing models where it was previously impos-sible or just not practical, simulation also allows fast evolutionof protocols without the significant effort of reworking a math-ematical analysis or reloading code onto thousands of physicalmotes. The need for simulators capable of modeling these large,complex, and seemingly limitless networks grows every day.Although many simulators have been built to date (ns-2 [8],GloMoSim [20], SSF [18], SensorSim [16], TOSSIM [10],SENSE [3], J-Sim [6], ATEMU [15], SENS [19], TOSSF [14]),

the scalability of these tools is not satisfactory. For examplens-2, currently the most popular network simulator, suffers anorder of magnitude performance penalty by invoking the Tclinterpreter during a simulation run and occupies approximatelygigabytes memory to simulate thousands of nodes. This limita-tion in scalability is a problem as networks being tested growand simulators are left with the task of modeling the incrediblycomplex behavior of these large systems.

One method of improving the scalability of network simula-tors is through model abstraction. The goal of model abstractionis to sufficiently reduce the complexity of a model, withoutsuffering (too great) a loss in accuracy. With this thought inmind, our work is an attempt to study the feasibility of usingabstraction as a solution to the problem of scalability. Specifi-cally, we are targeting to MAC layer abstraction, because basedon our empirical profiling results, over 90% simulation time isspent on the characterization of the MAC layer behaviors.

The originality of our approach lies in three aspects: First,to the best of our knowledge, all prior work on data collectionfor abstraction occurs off-line prior to simulation. In contrast,our abstraction is established on-line by analyzing the datacollected from a full simulation mode. This approach providesa high fidelity than off-line static abstractions. Second, wedynamically switch the granularity of the simulation basedon the feedback from the traffic rate monitor. This approachembraces the accuracy gain from the detailed simulation andspeedup from high-level abstraction. Third, our approach isindependent of upper layer implementations. Different networkprotocols and applications can run on top of our implementationseamlessly. we demonstrate that our approach is 9 times fasterthan the original simulation with a loss of accuracy less than10% in the end-to-end delay and 1% in the number of messagessuccessfully transmitted across the network.

The rest of paper is organized as follows. In section II,we discuss research to date that has pursued similar goals.Specifically we look at different solutions to scalability andidentify the uniqueness of our solution. In section III we discussthe rationale behind our abstraction and the basic ideas behindour solution. Section IV provides detail on our implementationwithin the GloMoSim framework and section V follows withresults and analysis. Section VI inspects the unresolved issuesthat needs to be addressed in the future work. We conclude insection VII.

II. STATE OF THE ART

In search of a solution to augmenting scalability in wirelessnetwork simulators, we looked at prior work in two generalareas of network simulation research. The first, achievingsimulation scalability, gave us insight into prior strategy andpotential solutions to help guide us towards our ultimate attemptat a MAC layer abstraction. The second, simulation validation,helped us understand the important aspects and techniques tovalidate our ultimate solution. Prior research in both of theseareas are discussed in this section.

A. Scalability

To solve the scalability problem, modern research on net-work simulation has taken two major approaches: parallel sim-ulation and simulation abstraction. The first, parallel simulation,has been implemented in architectures such as DaSSF [18]andGloMoSim [20] and has proven adequate in speeding upsimulations on a limited scale. [11], [13] and [20] discuss theapplication and results of implementing parallel and distributedsimulation across several processes to achieve speedup. Thebiggest advantage of parallel simulation is that speedup isgained without sacrificing the accuracy and granularity of thesimulation. This research on parallel simulation environmentsand techniques is complementary to our work.

The second approach, called model abstraction, is the focusof our work. [1], [4] and [17] outline fundamental goals andtradeoffs to consider when simulating a network layer abstrac-tion. Additionally these papers provide important insights intounderstanding the effects of model abstraction and testing forpossible simulator invalidation. [5] looks at two abstractiontechniques: centralized computation and abstract packet dis-tribution. Centralized computation saves memory consumptionand time by centrally computing protocol states to reduce theworkload and complexity of performing these computationsfor every simulated node. Abstract packet distribution avoidslink-by-link packet transmission by scheduling packets directlyat the receiving end. This second technique is similar to ourwork in that it eliminates potentially unnecessary details duringpacket transfer. In our work we implement a similar, but lessaggressive technique that results in a more detailed and hencemore accurate abstraction. Hybrid simulation, as implementedin SensorSim [16], is another approach to achieving speedupthrough abstraction. This approach utilized data collected froma real system to apply statistical data to simulation models. Ourapproach is different from this technique in the sense that weuse adaptive simulation granularity from fine detailed packet-level simulation and apply it to coarse abstract-level simulation.[1] uses a fluid-based approach to speedup simulation. In thissolution, speedup is achieved by coarsening the representationof network traffic from a packet level granularity to a flow levelgranularity where closely related packets are substituted with asingle packet. [12] abstracts details of an 802.11 MAC Protocolby substituting probabilistic packet arrival data and statisticalpacket arrival times to remove the details of the 802.11 MAC

protocol contention phase. This research most closely resemblesour own with the exception that data collection occurs off-line.This limitation prevents researchers from applying statisticaldata to situations that have not been previously encountered andmakes dynamic traffic patterns difficult to model. Our researchaddresses this limitation by performing online data collectionby monitoring traffic patterns to determine when simulatedabstraction is appropriate.

B. Validation

Aside from the work designing and implementing our ab-straction, we also studied model validation to better understandthe overall effect and validity of our research. Validation testssuch as the chi-square goodness of fit test [9] can be usedto test whether the simulated model reflects the real situation.[7] and [4] discuss issues and techniques for model validationand analysis. Although various methods and techniques areproposed and discussed in these papers, our primary focus ison comparing the simulated results of our abstraction with theresults of the high granularity simulation [12]. Since differ-ent networking scenarios have different impacts on simulatedresults, a limited set of case comparisons are not enough toclaim the validation of any model. Rigorous model validation,specifically as it applies to abstraction, is left as an area forfuture research.

III. METHODOLOGY OVERVIEW

Accurately modeling wireless networks on the scale ofhundreds of thousands of nodes has been and will remaina challenging problem for research. Despite the effects ofMoore’s law it remains impractical to handle simulations oflarge-scale networks over reasonable periods of simulation timeto collect precise experimental data.

A. Design Goals

The array of simulators developed to date provide varyinglevels of detail, accuracy, scalability, modularity (flexibility),etc. Unfortunately the more detailed a simulation becomes,the less capable it is of scaling to large complex networks.It is crucial to decide what level of detail can be sacrificedin exchange for simulation speed and scalability. Since in thewireless scenario, nearly 90 ∼ 95% of simulation time is spentmodeling details of the MAC and below, abstracting these layerscan provide a large speedup while preserving the accuracyof upper-layer behavior. The higher level protocols are idealfor our abstraction because they are not concerned with MAClayer detail so long as the packet transmission delay remainsconsistent with that of a detailed model’s behavior. Morespecifically, this work attempts to solve the speedup problemby abstracting out MAC layer implementation detail for theMAC protocols (e.g.MACAW [2]) in GloMoSim [20]. Whilethe specific abstraction we implement is under GloMoSim, ourconcept of MAC layer abstraction is generic and can be appliedacross other simulators.

2

MMSN: Multi-Frequency Media Access Control forWireless Sensor Networks

Gang Zhou, Chengdu Huang†, Ting Yan, Tian He‡, John A. Stankovic and Tarek F. Abdelzaher†

Department of Computer Science, University of Virginia, Charlottesville 22903{gzhou,ch4pp,ty4k,tianhe,stankovic,zaher}@cs.virginia.edu

Abstract— Multi-frequency media access control has beenwell understood in general wireless ad hoc networks, whilein wireless sensor networks, researchers still focus on singlefrequency solutions. In wireless sensor networks, each deviceis typically equipped with a single radio transceiver and ap-plications adopt much smaller packet sizes compared to thosein general wireless ad hoc networks. Hence, the multi-frequencyMAC protocols proposed for general wireless ad hoc networksare not suitable for wireless sensor network applications, whichwe further demonstrate through our simulation experiments.In this paper, we propose MMSN, which takes advantage ofmulti-frequency availability while, at the same time, takes intoaccount the restrictions in wireless sensor networks. In MMSN,four frequency assignment options are provided to meet differentapplication requirements. A scalable media access is designedwith efficient broadcast support. Also, an optimal non-uniformbackoff algorithm is derived and its lightweight approximationis implemented in MMSN, which significantly reduces congestionin the time synchronized media access design. Through extensiveexperiments, MMSN exhibits prominent ability to utilize paralleltransmission among neighboring nodes. It also achieves increasedenergy efficiency when multiple physical frequencies are avail-able.

I. INTRODUCTION

As an emerging technology, wireless sensor networks(WSNs) have a wide range of potential applications [1] [2] [3],including environment monitoring, smart buildings, medicalcare, industry and military applications. Being an essentialpart of the communication stack, media access control (MAC)has received intense research attention, and a number ofsolutions have been proposed [4] [5] [6] [7] [8] [9]. Whilethese solutions work well when one physical frequency isused, parallel data transmission when multiple frequencies areavailable is not considered. On one hand, the radio bandwidthin WSNs is very limited, 19.2Kbps in MICA2 [10] and250Kbps in MICAz [11] and Telos [12]. On the other hand,the current WSN hardware, such as MICAz and Telos that useCC2420 radio [13], already provide multiple frequencies [10].So it is imperative to design multi-frequency MAC protocolsin wireless sensor networks to take full advantage of paralleltransmission to improve network throughput.

In the state-of-the-art research, a significant number ofmulti-frequency MAC protocols have been proposed, for wire-less networks in general. However, these protocols are not

†Chengdu Huang and Tarek F. Abdelzaher are now with University ofIllinois at Urbana-Champaign.

‡Tian He is now with University of Minnesota.

suitable for typical WSN applications. First, to save energyandreduce product cost, each sensor device is usually equippedwith a single radio transceiver. This single transceiver cannot transmit and receive at the same time, nor can it func-tion on different frequencies simultaneously. This restrictedhardware is quite different from more powerful hardwareassumed in other wireless systems. For example, protocols[14] [15] are designed for frequency hopping spread spectrum(FHSS) wireless cards, and protocol [16] assumes the busy-tone functionality on the hardware. In protocols [17] [18] [19][20], the hardware is assumed to have the ability to listento multiple frequencies at the same time. Second, the networkbandwidth in WSNs is very limited and the MAC layer packetsize is very small, 30∼50 bytes, compared to 512+ bytesused in general wireless ad hoc networks. Due to the smalldata packet size, the RTS/CTS control packets in IEEE 802.11[21] no longer constitute a small overhead that can be ignored.So protocols [22] [23] [24] [25] [26] that are based on IEEE802.11, and protocols [27] [28] [29] [14] that use RTS/CTS forfrequency negotiation are not suitable for WSN applications,even though they exhibit good performance in general wirelessad hoc networks.

In this paper, we propose MMSN, abbreviation for Multi-frequency Media access control for wireless Sensor Networks.MMSN takes full advantage of multiple frequencies and isespecially designed to meet WSN requirements. The detailedMMSN design is presented from two aspects:frequency as-signmentand media access, and its performance is evaluatedthrough extensive simulation. The main contributions of thiswork can be summarized as follows:

• To the best of our knowledge, the MMSN protocol is thefirst multi-frequency MAC protocol especially designedfor WSNs, in which each device is equipped with a singletransceiver and the MAC layer packet size is very small.

• Instead of using pair-wise RTS/CTS frequency negotia-tion [27] [28] [29] [14], we propose lightweight frequencyassignment, which takes advantage of the static propertyof many deployed wireless sensor networks [30] [31][32] [33]. Even though pair-wise frequency negotiationis efficient when devices are highly mobile, it involvesunnecessary overhead and is too costly when applied tostatic WSN applications.This paper gives a complete study of tradeoffs amongphysical frequency requirements, potential conflict re-

duction and communication overhead, during frequencyassignment. Four optional frequency assignment schemesare proposed for MMSN, which exhibit distinguishedadvantages in different scenarios.

• We develop new toggle transmission and toggle snoop-ing techniques to enable the single transceiver sensordevice to achieve scalable performance, avoiding the non-scalable “one control channel + multiple data channels”design [34]. Also, MMSN has efficient broadcast support,which either is not addressed in [27] or is implementedby repeated link-layer retransmission of broadcast packetsenqueued by higher layers in [22].Moreover, through strict theoretical analysis, an opti-mal non-uniform backoff algorithm is derived and itslightweight approximation is implemented in MMSN.Compared with a uniform backoff algorithm, this non-uniform scheme significantly reduces potential conflictsamong neighboring nodes.

The rest of this paper is organized as follows: In Section II,we present the motivation of this work. In Section III, the de-sign details of MMSN are explained. In Section IV, extensiveexperiments are conducted to evaluate MMSN’s performance.Finally, in Section V, we give the conclusions and point outfuture work.

II. MOTIVATION

To obtain a better understanding of the cost that RTS/CTScontrol packets incur in general wireless ad hoc networksversus WSNs, we choose a typical multi-frequency MAC pro-tocol, the MMAC [27] protocol, proposed for general wirelessad hoc networks, as a case study. In MMAC, periodicallytransmitted beacons divide time into fixed-length beacon inter-vals. At the beginning of each beacon interval, there is a smallwindow called the ATIM window, in which the nodes thathave packets to send negotiate frequencies with destinationnodes. After the ATIM window, nodes that have successfullynegotiated frequencies with their destinations can send outdata packets using the IEEE 802.11 protocol, i.e. exchangingRTS/CTS before sending out DATA packets. We implementMMAC in GloMoSim [35], a scalable discrete-event simulatordeveloped by UCLA, and observe the performance. We adoptthe same experiment set up as in [27]: 100 nodes are randomlyplaced in a 500m×500m terrain. The transmission range ofeach node is 250m. Each node has 3 physical frequencies.Forty nodes are randomly chosen to be sources, and 40 nodesare randomly chosen to be destinations. Source nodes generateCBR traffic to destinations with a rate of 10 packets persecond. Figure 1 plots the aggregate MAC throughput of thenetwork with different packet sizes.

As can be observed in Figure 1, when the packet size islarge, the MMAC protocol with 3 frequencies and a beaconinterval of 100ms (the default configuration suggested in [27])impressively enhances the aggregate MAC throughput by afactor of nearly 2 over IEEE 802.11. This result is consistentwith that presented in [27]. However, the performance ofboth MMAC and IEEE 802.11 degrades when the packet

size reduces. This is because the overhead of using RTS/CTSpackets becomes more prominent when the data packet sizeis small. Moreover, the performance improvement of MMACover IEEE 802.11 diminishes when the packet size becomessmaller. When the packet size is as small as 32 bytes, IEEE802.11 has even a slightly higher throughput than MMAC.The reason is when the packet size reduces, more packetscould be sent in a beacon interval. However, since nodesgenerally can not switch frequency during a beacon interval,the bandwidth wasted is more severe compared to the casewhen the packet size is large. Changing the length of thebeacon interval could be beneficial, but the effect is two-sided. While lengthening the beacon interval can mitigate theoverhead of having a fixed period of frequency negotiation,it deteriorates the bandwidth caused by the requirement thatnodes have to stick to the frequency they have negotiated withsome of their neighbors. In Figure 1, we also plot the caseswith different beacon intervals. We can see that while usingashorter beacon interval (50ms) helps to some extent, MMACwith 3 frequencies still can not even outperform IEEE 802.11with a single frequency, when the packet size is as small as64 or 32 bytes. The main observation we make here is thatwhile MMAC is a good multi-frequency MAC protocol forgeneral wireless ad hoc networks where packets usually havelarge sizes, it is not suitable for WSNs where packets are muchsmaller.

0

200

400

600

800

1000

1200

1400

512 256 128 64 32Agg

rega

te M

AC

Thr

ough

put (

Kbp

s)

Packet Size (bytes)

Packet arrival rate per flow is 10 packets/sec

802.11MMAC, Beacon Interval = 50msMMAC, Beacon Interval = 100msMMAC, Beacon Interval = 150ms

Fig. 1. Effect of Packet Size on MMAC

III. MMSN PROTOCOL

This section presents the MMSN multi-frequency MACprotocol. MMSN is especially designed for WSNs, whichis composed of hundreds of simple devices geographicallydispersed in an ad hoc network over a large geographic area.Each device is equipped with a single transceiver and thepacket size is very small, 30∼50 bytes. The MMSN protocolconsists of two aspects: frequency assignment and mediaaccess. The frequency assignment is used to assign differentfrequencies if enough frequencies exist, or evenly allocateavailable frequencies if there are more neighbors than availablefrequencies, to nodes that have potential communication con-flicts. MMSN allows users to choose 1 of 4 available frequencyassignment strategies. In media access design, nodes that have

Gradient-Driven Target Acquisition in Mobile

Wireless Sensor Networks

Qingquan Zhang1, Gerald Sobelman1, Tian He2

1Department of Electrical and Computer Engineering, University of Minnesota, Twin

cities, USA2Department of Computer Science and Engineering, University of Minnesota, Twin

cities, USA

{zhan0511,sobelman}@umn.edu

{tianhe}@cs.umn.edu

Abstract. Navigation of mobile wireless sensor networks and fast target

acquisition without a map are two challenging problems in search and

rescue applications. In this paper, we propose and evaluate a novel Gra-

dient Driven method, called GraDrive. Our approach integrates per-node

prediction with global collaborative prediction to estimate the position of

a stationary target and to direct mobile nodes towards the target along

the shortest path. We demonstrate that a high accuracy in localization

can be achieved much faster than other random work models without

any assistance from stationary sensor networks. We evaluate our model

through a light-intensity matching experiment in MicaZ motes, which

indicates that our model works well in a wireless sensor network envi-

ronment. Through simulation, we demonstrate almost a 40% reduction

in the target acquisition time, compared to a random walk model, while

obtaining less than 2 unit error in target position estimation.

Key words: Wireless Sensor Network, Navigation, Localization, Prob-

abilistic Model, Rescue

1 Introduction

Wireless sensor networks have gained extensive attention in many applicationssuch as tracking, differentiated surveillance, and environment monitoring [1–3].Moreover, the hybrid systems of mobile objects (e.g. Robots) and sensor net-works create new frontiers for civilian and military applications, such as searchand rescue missions in which the background environments are inaccessible tohumans. A heterogeneous searching team consisting of robots and a wireless sen-sor network has greater advantage, considering its distributed computation andnavigation capability achieved through the cooperation of embedded wirelesssensor networks.

Although the applications of mobile sensor networks keep diversifying, sev-eral underlying capabilities remain fundamental and critical. In this work, we

2 Qingquan Zhang1, Gerald Sobelman1, Tian He2

focus on the target acquisition – finding the locations of stationary targets us-ing mobile sensor nodes. The challenging problem we address in this work is to

navigate a team of mobile sensor nodes toward the stationary targets fast and ac-

curately while consuming the least amount if energy and other resources. In thisemerging research arena, most research groups employed static wireless sensornetworks to navigate the mobile sensor nodes. Tan in [4] used distributed staticsensor networks to collect the data and execute local calculations to generatea path for a mobile sensor network to move toward the goal. Although the in-network calculation implemented in that project was quite efficient in creatingthe shortest routing path, the additional requirement of a stationary distributedsensor network sets a barrier for rescue applications, because of the high costto cover a large geographic area with a large number of sensors. Other researchgroups [5] proposed gradient methods in which the mobile wireless sensor nodesmove toward the gradient direction assuming that targets carried the most inten-sive strength of interested signals. However, in all of their implementations, theassistance of a stationary wireless sensor network was assumed to be available ingenerating a local signal distribution map. A probabilistic navigation algorithmis presented in [6], where a discrete probability distribution of vertex is intro-duced to point to the moving direction. This algorithm computes the utilitiesfor every state and then picks the actions that yield a path toward the goal withmaximum expected utility. The shortcoming of this method is that it requiresthe arrival of a mobile sensor node to localize the target position and significantcommunication overhead is introduced by the iteration process.

2 Contribution

In this paper, we propose to compensate those deficiencies by incorporating aprediction model of real-time processes into a mobile sensor network sensing andnavigation architecture. We are interested in the mutually beneficial collabora-tion of the algorithms described above but seek to reduce the costs and providefaster target localization. The novelty of our approach is the seamless integration

of a per-node prediction model with a global prediction model. The per-node pre-diction model guarantees that a mobile node can acquire the position of a targetalone, while the global prediction significantly reduces the navigation overheadand time, if collaboration among the nodes is available. Specifically, the maincontributions of our prediction models are:

– Our model provides more meaningful description of individual sensor read-ings in term of accuracy and confidence.

– Our model works with a single mobile sensor node as well as a swarm ofmobile sensor nodes. In the latter case, the sensor nodes have the ability toshare local information in order to draw a global picture, which helps eachsensor node to acquire the target along a significantly shorter path.

– The in-network prediction algorithm enables faster yet accurate target po-sition acquisition: sensor nodes would be required to reach the target only

Efficiency Centric Communication Model forWireless Sensor Networks

Qing Cao†, Tian He‡, Lei Fang, Tarek Abdelzaher†, John Stankovic, Sang SonDepartment of Computer Science, University of Virginia, Charlottesville VA 22904, USA

Abstract— Recent studies on radio reality provided strongevidence that radio links between low-power sensor devices areextremely unreliable. In this paper, we study how to improveenergy efficiency for reliable communication using such unreli-able links. We identify an optimal bound on energy efficiencyfor reliable communication, and propose a new communicationmodel in the link layer that asymptotically approaches thisbound. This new model indicates a better path metric comparedto previous path metrics, and we validate this by establishinga routing infrastructure based on this metric, which indeedachieves a higher energy efficiency compared to other state-of-the-art approaches. We present results from a systematicanalysis, simulations and prototype experiments based on theMicaZ platform. The results give us fundamental insights oncommunication efficiency over unreliable links.

I. INTRODUCTION

The evolution of radio models has consistently shaped theupper layers of the communication stack in wireless sensornetworks. Early communication stacks were usually designedbased on over-simplified radio assumptions, such as the unitdisk graph model. Recently, this model has been repeatedlychallenged by empirical measurements. For example, studiesin [31], [26], [8] suggested that wireless links are irregularand unreliable. These studies also observed highly differentpacket delivery ratios for the same link in reverse directions.The design of communication stacks must take into accountthese radio layer realities.

Motivated by these observations, we focus on how to min-imize overhead while providing reliable end-to-end commu-nication over these unreliable links. Reliable communicationis important for sensor networks despite the wide use ofaggregation techniques, because realistic applications oftenuse alarms and other one-time event notifications that needto be communicated reliably. The successful delivery of suchinformation has a direct effect on the overall performance ofthe system.

Formally, we model an unreliable link between nodes Aand B as (p, q), where p represents the packet delivery ratiofrom A to B, and q represents the packet delivery ratiofrom B to A. When p and q are less than 100%, we mustuse a packet recovery mechanism, such as retransmissionsand acknowledgements, to achieve reliability. The problem ofminimizing overhead, therefore, is equivalent to minimizingthe additional traffic compared to the ideal scenario where

† Qing Cao and Tarek Abdelzaher are now with the University of Illinois,Urbana-Champaign.‡ Tian He is now with the University of Minnesota.

every link is perfect. To model this traffic, we introduce anew parameter, Energy Per Bit, or EPB, to characterize theenergy efficiency aspect of communication. EPB representsthe average energy consumption for each delivered bit from thesource to the destination. EPB is decided by several factors,such as the link layer packet recovery mechanism, the routinglayer path selection, the relative positions of the source andthe destination, and the network topology, among others. Someof these factors, such as the relative positioning of the sourceand the destination, are unique to a particular transport task.Therefore, we do not aim to optimize EPB across differenttransport tasks. Rather, we focus on optimizing EPB for agiven transport task, for example, node A sends 1000 bytes toB through multiple hops.

The optimization presented in this paper is a joint optimiza-tion process in two layers: in the link layer, it optimizes howlost packets are detected; in the routing layer, it optimizeshow paths are selected. More specifically, we present twocorresponding techniques: 1) the lazy packet loss detectionin the link layer and 2) the use of a stream based path metricin the routing layer. The first optimization technique appliesto a particular chosen path, while the second one applies tothe path selection process. These two optimization techniquesare unified by their consistency: we first analyze the firstoptimization technique and demonstrate its effectiveness inreducing EPB given a particular path. Based on the analysis,we distill a path metric that features a stream nature, interactswith the commonly used spanning tree routing structure,and leads to the second optimization technique. Therefore,by jointly applying these two techniques, we further reduceEPB, as validated by an extensive evaluation.

Our optimization techniques have applications in a widevariety of scenarios. One of them is surveillance [27], [3].In this type of applications, sensed data, such as temperature,light or pressure readings, are periodically sampled and relayedto a central data collection node, referred to as the basestation. Normally, the data reporting rate is relatively low, andtiming requirements are not strict. Since data are generatedperiodically, traffic naturally exhibits a streaming nature (i.e.,a data flow to the base via multiple hops for a long periodof time). Our optimization techniques can therefore be readilyapplied to such applications.

The rest of this paper is organized as follows. Section IIpresents the lazy lost packet recovery mechanism. We thenderive a general stream path metric based on analysis ofthis mechanism. Section III presents a systematic performance

evaluation of the mechanism from two aspects: the end-to-end delay and the buffer requirements. Section IV integratesthe stream metric into the routing layer design, leading tothe second optimization technique. We compare its perfor-mance to two state-of-the-art protocols that take into accountthe unreliable nature of sensor network communication. Wedemonstrate that, by using the joint optimization, our protocolstack achieves a considerably better EPB value than eitherof them. At last, we outline related work in Section V andconclude this paper with Section VI.

II. LAZY LOST PACKET RECOVERY

We now describe the first optimization technique, calledlazy lost packet recovery. This section is organized into threeparts. First, we describe different link quality metrics basedon empirical experiments. Next, we present the design of theoptimization algorithm, and demonstrate that it approaches theoptimal efficiency bound. At last, we discuss applications ofreliable packet delivery in sensor networks.

A. Link Quality Metrics Overview

Link quality metrics are used to classify and select links.Prior measurements on link quality reveal that different linkshave considerably different packet reception properties [30].One popular model is to treat a link as a bi-directionalpacket reception probability vector (p, q). However, receptionprobability is not the only representation of link quality.Recently, another metric, LQI , or Link Quality Indicator, wasdefined by IEEE standard 802.15.4 [1], and was implementedon the Chipcon CC2420 radio component [7]. The CC2420radio has been used on MicaZ and Telos nodes. Anotherinteresting link quality indicator is RSSI (Received SignalStrength Indicator), also implemented on the CC2420 radio.In this section, we demonstrate that LQI is closely correlatedwith packet reception probability, and either of them can beused as the underlying metric of our optimization model. Onthe other hand, our experiments demonstrate that RSSI isnot a satisfactory indicator of link quality, and should not beadopted for link classification.

The implementation of LQI , as defined in [1], was that it“may be implemented using receiver ED (energy detection),a signal-to-noise ratio estimation, or a combination of thesemethods.” The CC2420 radio module implemented LQI basedon a sampling of the error rate for the first eight symbols ofeach incoming packet. This sampling generates a correlationvalue in the range of [50, 110], followed by a linear conversionof this value to a range of [0, 255], which is the value providedto the user.

RSSI is not detailed in the 802.15.4 standard, but it isprovided by the CC2420 radio module together with LQI .RSSI is also based on eight symbols, but instead of usingthe error rate, it uses the average energy level to calculate itsvalue.

We carried out a series of experiments with both indoor andoutdoor environments during different periods of the day. Due

to space limitations, we only present one representative set ofdata in Figure 1.

-100-90-80-70-60-50-40-30-20-10

0 10 20 30 40 50 60 70 80 90

100 110 120

2 4 6 8 10 12 14 16 18 20 22 24 26 28 30 32 34 36 38 40 42 44 46 48Pac

ket R

ecep

tion

Per

cent

/LQ

I Raw

Val

ue/R

SS

I Raw

Val

ue

Experiment Round

Received Packet PercentReceived Packet LQI (averaged per round)

Received Packet RSSI (averaged per round)

Fig. 1. Correlation between the packet receiving ratio and LQI of CC2420radio at receiver side. At each distance, six rounds of packets are sent fromthe sender to the receiver.

In this particular experiment, we use a pair of MicaZ nodes,one as the sender and the other as the receiver. We vary thedistance from the receiver to the sender from 5ft to 40ft, insteps of 5ft. At each distance, the receiver sends six rounds ofpackets, with 100 packets in each round. The packet receptionratio, the average LQI and the average RSSI are plottedfor comparison. Both LQI and RSSI are calculated in everyround.

We have three observations. First, we observe a strongcorrelation between the averaged LQI values and packet re-ception probabilities at the receiver. Statistical analysis showsthat the Pearson’s correlation coefficient is 0.90 between thesetwo variables. Of course, there are still some inconsistenciesobserved, especially when the received signal is weak. Theseinconsistencies explain why the Pearson’ s coefficient is not1.0. Nevertheless, the observed correlation is still quite inter-esting, since LQI is calculated only from those packets thatare received, whereas the packet reception probability takesinto account those packets that are dropped. This correlationimplies that LQI is a good measurable indicator of the packetreception probability. Second, at each distance, the packetreception probability has a narrow range of variation that isless than 20%. Our additional experiments also exhibit thesame property. This observation is consistent with results inthe literature [26]. We, therefore, consider the packet receptionprobability as a relatively stable parameter to classify links.Third, observe that there is a much smaller correlation betweenRSSI and the packet reception probability. The Pearson’s cor-relation coefficient is only 0.56 between the packet receptionprobability and the RSSI value. Furthermore, observe thatwhen the signal is weak, even though there is a considerablevariation in the packet loss rate, RSSI does not changeappreciably. Therefore, we do not recommend using RSSIas a reliable link quality indicator.

Impact of Radio Irregularity on Wireless Sensor Networks

Gang Zhou, Tian He, Sudha Krishnamurthy, John A. Stankovic

Department of Computer Science University of Virginia, Charlottesville, VA 22903

{gz5d, tianhe, skrish, stankovic}@cs.virginia.edu

ABSTRACT In this paper, we investigate the impact of radio irregularity on the communication performance in wireless sensor networks. Radio irregularity is a common phenomenon which arises from multiple factors, such as variance in RF sending power and different path losses depending on the direction of propagation. From our experiments, we discover that the variance in received signal strength is largely random; however, it exhibits a continuous change with incremental changes in direction. With empirical data obtained from the MICA2 platform, we establish a radio model for simulation, called the Radio Irregularity Model (RIM). This model is the first to bridge the discrepancy between spherical radio models used by simulators and the physical reality of radio signals. With this model, we are able to analyze the impact of radio irregularity on some of the well-known MAC and routing protocols. Our results show that radio irregularity has a significant impact on routing protocols, but a relatively small impact on MAC protocols. Finally, we propose six solutions to deal with radio irregularity. We evaluate two of them in detail. The results obtained from both the simulation and a running testbed demonstrate that our solutions greatly improve communication performance in the presence of radio irregularity.

Categories and Subject Descriptors C.2.1 [Computer Communication Network]: Network Architecture and Design; I.6 [Computing Methodologies]: Simulation and Modeling

General Terms Design, algorithms, measurement, performance, experimentation

Keywords Sensor networks, wireless communication, radio irregularity, sending power, path loss, link asymmetry, packet loss

1. INTRODUCTION Radio irregularity is a common and non-negligible phenomenon in wireless sensor networks. It results in irregularity in radio range and variations in packet loss in different directions, and is considered as an essential reason for asymmetric links as viewed by upper layers in the protocol stack. Several empirical studies [4][9][23][28] on the Berkeley mote platform have shown that the radio range varies significantly in different directions and the percentage of asymmetric links in a system varies depending on the average distance between nodes. The impact of radio irregularity on protocol performance can be investigated through a running system. However, few researchers have actually pursued this direction, because of two reasons: First, the complexity and cost of performance evaluations on a running system escalate, when sensor networks scale up to thousands or more nodes. Second, repeatable results of radio performance are extremely hard to obtain from uncontrolled environments, hence leading to difficulties in system tuning and performance evaluation. As a result, simulation techniques are used as an efficient alterative to evaluate protocol performance. Unfortunately, most existing simulations don’t take radio irregularity, a common phenomenon in wireless communication, into account. The spherical radio patterns assumed by simulators such as [27] may not approximate real radio properties well enough and hence may lead to an inaccurate estimation of application performance. Several researchers [4][9][23][28] have already shown extensive evidence of radio irregularity in wireless communication. Their main focus is to observe and quantify such phenomena. This paper is distinguished from the previous ones for the initiative in bridging the gap between spherical radio models used by simulators and the physical reality of radio signals. We first verify the presence of radio irregularity using empirical data obtained from the MICA2 platform. The results demonstrate that the radio pattern is largely random; however, it exhibits a continuous change with incremental changes in direction. Based on experimental data, a radio model for simulations, called the Radio Irregularity Model (RIM), is formulated. RIM takes into account both the non-isotropic1 properties of the propagation media and the heterogeneous properties of devices.

1 Exhibiting properties with different values when measured along

axes in different directions.

Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. MobiSYS’04, June 6-9, 2004, Boston, Massachusetts, USA. Copyright 2004 ACM 1-58113-000-0/00/0004…$5.00.

With the help of the RIM model, we explore the impact of radio irregularity on MAC and routing performance. Among the protocols we evaluate, we find that radio irregularity has a significant impact on the routing protocols, but a relatively small impact on the MAC protocols. We also find that location-based routing protocols, such as Geographic Forwarding (GF) [17] perform worse in the presence of radio irregularity than on-demand protocols, such as AODV [21] and DSR [14]. We propose several potential solutions to deal with radio irregularity in wireless sensor networks. We evaluate two of them through a simulation and a running system, respectively. Our results illustrate that our solutions do succeed in alleviating the performance penalties due to radio irregularity. The rest of this paper is organized as follows: we briefly analyze the causes and impact of radio irregularity in Section 2. In Section 3, we describe experimental data collected from the MICA2 platform and make some general conclusions about radio irregularity. Based on these conclusions, we propose the RIM radio model in Section 4. We then use the RIM model in simulations to analyze the impact of radio irregularity on MAC and network layer protocols in Section 5 and Section 6, respectively. Solutions to deal with radio irregularity are proposed and evaluated in Section 7. Finally, we conclude the paper in Section 8.

2. ANALYSIS OF RADIO IRREGULARITY In this section, we first identify the causes of radio irregularity, and then briefly discuss the impact of irregularity on the different protocol layers.

2.1 Causes of Radio Irregularity Radio irregularity is caused by two categories of factors: devices and the propagation media. Device properties include the antenna type (directional or omni-directional), the sending power, antenna gains (at both the transmitter and receiver), receiver sensitivity, receiver threshold and the Signal-Noise Ratio (SNR). Media properties include the media type, the background noise and some other environmental factors, such as the temperature and obstacles within the propagation media. In general, the radio irregularity is caused by the non-isotropic properties of the propagation media and the heterogeneous properties of devices. Among all these factors, we focus on the non-isotropic path losses and the differences in signal sending power, which are commonly regarded as the key causes of radio irregularity.

• Non-isotropic Path Losses: The variance in the signal path loss is one of the major causes of radio irregularity. When a signal propagates within a medium, it may be reflected, diffracted, and scattered [22]. Reflection occurs when an electromagnetic signal encounters an object, such as a building, that is larger than the signal’s wavelength. Diffraction occurs when the signal encounters an irregular surface, such as a stone with sharp edges. Scattering occurs when the medium through which the electromagnetic wave propagates contains a large number of objects smaller than the signal wavelength. The medium is normally different in different directions. Consequently, radio propagation exhibits non-isotropic patterns in most environments.

Another significant reason for non-isotropic path loss is hardware calibration. A node may not have the same antenna gain along all propagation directions, possibly due to hardware manufacturing. Hence, the non-isotropic antenna gain of each node also contributes to the non-isotropic path loss.

• Heterogeneous Sending Powers: Sensor devices may transmit RF signal at different sending powers, even though they are the same kind of devices. This difference may arise from some random factors during the manufacture of sensor devices. In addition, after the sensor devices are deployed, the batteries of different sensor devices deplete at different rates, due to different workloads and different environments in which they are deployed. Heterogeneous sending powers result in variable communication ranges, and cause non-isotropic connectivity.

2.2 Impact of Radio Irregularity Radio irregularity is a non-negligible phenomenon in wireless communication. It’s an essential reason for asymmetric radio interference and asymmetric links in upper layers. It can directly or indirectly affect many aspects of upper layer performance. Asymmetric radio interference between neighboring nodes affects the correctness of MAC layer functions. For example, in the presence of radio irregularity, a node might not be able to successfully reserve the wireless channel through RTS and CTS handshaking, because those neighboring nodes of the receiver, which cannot hear the CTS control packet, might disrupt the receiving node. This impacts the delivery ratios of data frames at the MAC layer. Radio irregularity can also affect the performance and even correctness of networking protocols such as [12] [14] [16] [17] [21]. For example, link asymmetry is one of the ways in which radio irregularity manifests itself at the higher layer. Link asymmetry has an adverse impact on protocols that use path-reversal techniques to establish an end-to-end connection. Actually, the impact of radio irregularity is not only confined to the MAC and routing layers, radio irregularity also influences other protocols, such as the localization, sensing converge and topology control protocols. Localization protocols such as DV-Hop [20] and Centroid [2] assume a spherical radio range. The study in [10] shows that the performance of such protocols degrades when the radio range becomes irregular. The sensing coverage scheme in [26] assumes that sensing and communication ranges are spherical. In the presence of radio irregularity, they might not be able to guarantee full coverage and blind points would occur. The topology-control scheme in GAF [25] builds a communication mesh based on the assumption of a spherical range. This might lead to the network partition in the presence of a non-spherical range. We note that some other topology-control protocols, such as ASCENT [3] and Span [5] don’t depend on such an assumption, however, performance evaluations of those protocols considering radio irregularity are desired. Due to space limitation, in the rest of paper, we only focus on the impact of radio irregularity on MAC and routing performance and leave the rest as future work.

Range-Free Localization

Radu Stoleru1, Tian He2 and John A. Stankovic3

1 Department of Computer Science, University of Virginia,stoleru@cs.virginia.edu

2 Department of Computer Science and Engineering, University of Minnesotatianhe@cs.umn.edu

3 Department of Computer Science, University of Virginia,stankovic@cs.virginia.edu

1 Introduction

Advances in micro-electro-mechanical systems have triggered an enormousinterest in wireless sensor networks (WSN). WSN are formed by large numbersof densely deployed nodes enabled with sensing and actuating capabilities.These nodes have very limited processing and memory capabilities, limitedenergy resources and it is envisioned that they will be mass produced, toreduce costs.

Several challenging problems exist in wireless sensor networks. Amongthese is how to obtain location information for sensor nodes and events presentin the network. From this perspective, we categorize the localization problemas: node localization, target localization and location service. Node localiza-tion is the process of determining the coordinates of the sensor nodes in theWSN. Target localization is the process of obtaining the coordinates of anevent or a target present in the sensor network. The location of a target canbe obtained either passively (the nodes sense the target) or actively, whenthe target cooperates and communicates with the sensor network. A locationservice acts as a repository that can be used to answer questions like “whereis entity X?”. In the remaining part of this chapter we focus on the nodelocalization problem in WSN.

Node localization is a complicated and important problem for wirelesssensor networks (WSN). The aspects of this problem that have challenged theresearch community can be summarized as follows:

• Assumptions - The node localization problem remains a difficult chal-lenge to be solved practically. To make the problem practically tractable,its complexity had to be reduced, by making simplifying assumptions. Asa result, many localization schemes proposed solutions that are based onassumptions that do not always hold or are not practical. Examples of such

2 Radu Stoleru, Tian He and John A. Stankovic

assumptions are: circular radio range, symmetric radio connectivity, addi-tional hardware (e.g., ultrasonic), lack of obstructions, lack of line-of-sight,no multipath and flat terrain.

• Localization Protocol Design - The problem of localization in WSNis further complicated by the large number of parameters that need tobe considered when designing a localization system for a particular WSNdeployment. Among these parameters are: the deployment method for thesensor network; the existence of a line-of-sight between sensor nodes anda remote, central point; the time required by the localization scheme; thepresence of reference points (anchors) in the network, and the density;the cost for localization, represented by additional hardware (form fac-tor) and energy expenditure (messages exchanged or time necessary forlocalization).

• Cost/Accuracy trade-off - Due to the mostly static nature of manyWSN, obtaining the location information by each sensor node is often aone time or rare event. Adding hardware to each sensor node, to assistin the localization, is a costly solution, and, so far, has been ruled outfrom real system deployments. For example, GPS is a typical high-endsolution, which requires sophisticated hardware to achieve high resolutiontime synchronization with satellites. The constraints on power and cost fortiny sensor nodes and the need for a line of sight from a sensor node to fouror more satellites preclude this as a viable solution. Other solutions requireper node devices that can perform ranging among neighboring nodes. Thedifficulties of these approaches are two-fold. First, under constraints ofform factor and power supply, the effective ranges of such devices are verylimited. For example the effective range of an ultrasonic transducer is onthe order of a few meters, when the sender and receiver are not facingeach other. Second, since most sensor nodes are static, i.e., the location isnot expected to change, it is not cost-effective to equip these sensors withspecial circuitry just for a one-time localization.

• Performance Evaluation - The problem of localization in wireless sen-sor networks has been studied and evaluated predominantly in simulators.Due to the severe hardware constraints imposed on wireless sensor nodes,real system implementations of the proposed simulated solutions have notproduced encouraging results. Solutions that use the most tempting meansof evaluating relative distances between sensor nodes - RF signal strength,have largely failed in practice, due to the unreliable nature and irregularpattern of the radio communication. Localization schemes that are basedon the receive signal strength indicator (RSSI) have been, however, inten-sively studied in simulators.

• Security - Since localization is a critical factor in WSN, attacks on it canrender the sensor network ineffective. To date, very little work has beendone on creating robust and secure localization schemes. A few notableexceptions are [15] [14] [17] [16] [5].

RESTORE: A Real-Time Event Correlation andStorage Service for Sensor Networks

Sudha KrishnamurthyDeutsche Telekom Laboratories

Berlin, Germanysudha.krishnamurthy@telekom.de

Tian HeDepartment of Computer Science

University of MinnesotaMinneapolis, MN

tianhe@cs.umn.edu

Gang Zhou, John A. Stankovic, Sang H. SonDepartment of Computer Science

University of VirginiaCharlottesville, VA

{gzhou,stankovic,son}@cs.virginia.edu

Abstract— This paper describes the design of RESTORE,which is a framework for providing in-network event correla tionand storage service for sensor environments. RESTORE uses adata-centric approach in which it partitions a sensor network intozones and maps every event to a zone. The sensor nodes in a zonemake use of their cooperative storage resources and redundancyto improve information availability and energy efficiency. RE-STORE is useful for temporarily buffering information in se nsorenvironments that have intermittent connectivity to a basestation.RESTORE may also be used as the underlying framework foran event notification service that publishes events directly from asensor network to the subscribers. The contributions of this paperinclude a) an event taxonomy for correlating sensor events withinthe network, and b) mechanisms to organize information usingthe collaborative resources, in order to enable real-time eventcorrelation within a sensor network. We also present preliminarysimulation results that examine how the collaboration within azone in RESTORE impacts the energy consumption, informationavailability, and message overhead.

I. I NTRODUCTION

Wireless sensor networks are primarily large-scale, de-centralized information systems. An information system forsensors needs to have the ability to store the stream ofobservations generated locally by the sensors and more im-portantly, correlate the raw observations from different partsof the network to generate events that are meaningful to theend users. There have been several efforts to correlate theinformation gathered by the deployed sensors and generatehigher-level inferences in the context of different applications,such as surveillance [1], environmental monitoring [2], [3],and structural health monitoring [4]. However, in a majority ofthese efforts, the storage and correlation of sensor informationis performed at the edge of the network in more powerfuldevices that serve as base stations. This approach is certainlythe most viable option when the devices in the networkare primitive and the infrastructure allows easy and uninter-rupted access to the powerful edge devices. However, recenttechnological advances are making it possible to integratemulti-modal sensing capabilities, resulting in network devicesthat are smarter and capable of fine-grained classification.Should this technological trend continue, performing eventcorrelation and storage within the sensor network should beanincreasingly viable option and benefit several applications. Inthis paper, we motivate the need for an in-network service that

buffers the observations from the sensors for a limited periodof time and uses that as a basis for correlating events withinthe network. We also present an architecture that makes useof the inherent redundancy and collaborative nature of sensornetworks to realize such a service.

A. The Case for In-Network Correlation and Storage

The use of external base stations for storage and correlationhas some drawbacks. Unlike the tiny sensor devices that can beleft unattended in remote environments, the base stations areoften more power-consuming and less unobtrusive. Leavingthese base stations unattended in remote sites may not be afeasible option, especially in harsh and unfriendly environ-ments where stealthiness is important, such as in battlefieldsurveillance applications. This problem can be addressed byusing long-range relays to transmit the sensor observations toremote base stations. However, this long-range communicationmay not always be reliable. End-to-end connectivity maybe intermittent or periodic and may be disrupted due toseveral factors, such as the weather. Moreover, such long-rangecommunication may be more vulnerable to interception.

A better alternative in such environments, where commu-nication may be disrupted, is to buffer the observations ofthe sensors within the sensor network for limited periods oftime and correlate the observations within the network. Eventnotifications generated as a result of the correlation can thenbe directly delivered to the end subscribers. The data storedin the network may also be uploaded in batches, wheneverthe base stations are connected. Such an approach reduces thedependence on external base stations and minimizes the use oflong-range communication. Recent efforts, such as the Zigbeegateway working group [5], delay-tolerant architectures [6],and TinyREST [7] are addressing the interoperability issuesinvolved in integrating the sensor networks more closely withthe Internet and traditional networks. We think that an in-network event correlation and storage service would vastlyimprove the utility of sensor networks and provide impetusand motivation for such efforts that attempt to bridge the gapbetween sensor networks and traditional networks.

Sensor networks are typically perceived as challenged en-vironments [6]. So a valid question is whether the deployedsensor devices are capable of realizing an in-network storage

and correlation service. Recent technological trends havebeenencouraging in this regard. If the sensor devices are low-enddevices, like motes, each of which has a storage capacityof only 512 KB [8], building an in-network storage schemewill require cooperative caching schemes. However, if themotes are deployed along with devices that have larger storagecapacity, such as PDAs and Stargate processors [9], thensuch a heterogeneous collection of devices would make itincreasingly feasible to provide long-term, in-network storage.

In order to implement an in-network event correlationservice, the nodes in the network should at the minimumbe able to order events in space and time. Mote-like devicespossess this ability and this allows them to perform simplecorrelations pertaining to a single object, such as tracking themotion of a single car with the help of magnetic sensors. How-ever, if multiple cars of identical size appear in the networkconcurrently and their trajectories intersect, this basicabilityis insufficient to distinguish between the different objects. Toperform event correlation within the network in such scenarios,we would need nodes with the ability to perform finer-grainedclassification. Recent advances in nodes with multi-modalsensing is promising in this regard. For example, the integratedRFID-sensor systems from SkyeTek [10] combine a MicaDotnode [11] with an RFID reader. This combination of sensorsand RFID provides access to a richer source of data andallows different instances of the same kind of event to bedistinguished, making it possible to perform more fine-grainedevent correlation within the network.

B. Overview of RESTORE

In this paper, we describe an in-network event-correlationand storage service for sensor networks, called RESTORE.The design principle behind RESTORE is that no single nodein a zone has enough resources to store and correlate informa-tion about all the events in the network. So RESTORE makesuse of the inherent redundancy and cooperative resources ofa sensor network to store and correlate events within thenetwork. RESTORE organizes the nodes in the sensor networkinto zones. Each zone is responsible for storing informationabout one or more events. A zone functions like a cooperativecache and is the smallest unit of collaborative storage inthe sensor network. Each zone has a storage manager thatis responsible for coordinating the storage and event corre-lation within its zone. Zone members collaborate to achieveenergy efficiency and fault tolerance. The two main goals ofRESTORE are the following:

Real-time event correlation: RESTORE enables a sensornetwork to publish event notifications directly to the sub-scribers, without any intermediate processing at base stations.In many applications, the subscribers may be interested in be-ing notified about composite events instead of every individualevent. The cooperative resources of the members within a zonecan be used to record observations from sensors across thenetwork in a decentralized manner to form a more compositeview of the events in the network. As new observations arrive,they can be correlated with relevant portions of the stored data

in real-time. While the correlation happens as new data arrives,the event notifications are generated according to the frequencyand specifications of the subscribers.

Short-term, reliable storage: RESTORE not only mini-mizes the need for long-range communication, it also providesreliability. In environments where connectivity to an externalpersistent storage is intermittent, if the observations are di-rectly transmitted to the base station and are lost in the processdue to unreliable channels, the base station may not be ableto generate some event notifications. In such an environment,RESTORE enables a sensor network to store information andcorrelate events within the network. If the event notificationsto the subscribers are lost, they can be retransmitted, becausethe observations used to generate the notifications are storedin the network for a period of time. Whenever a base stationis available, the individual zone managers can upload thestored data from their zones and reclaim the storage resources.RESTORE also allows the stored data to be uploaded to mobilebase stations, which is useful for applications, such as, wildlifetracking[12].

The RESTORE architecture described in this paper assumesa network consisting of homogeneous low-end devices, likethe motes. However, if nodes with more powerful storagecapabilities, such as PDAs, are also present, RESTORE wouldtake advantage of such nodes by favoring them as storagemanagers. RESTORE also takes advantage of RFID-enabledobjects, if they are present, and uses the tags to distinguishbetween different instances of an object and perform fine-grained event correlation.

The rest of the paper is organized as follows. In Section II,we describe the use of an in-network correlation and storageservice for a few application scenarios. In Section III, wediscuss the architecture of RESTORE and describe how itpartitions the network into zones, in order to provide energy-efficient storage. In Section IV, we present an event taxonomyand describe how RESTORE uses the zones to store and cor-relate information for events characterized by this taxonomy.In Section V, we present preliminary performance results. InSection VI, we compare RESTORE with other sensor networkstorage schemes. In Section VII, we present our conclusions.

II. A PPLICATIONS

We now illustrate the use of an in-network event correlationand storage service, like RESTORE, in some typical sensorapplications.

• Consider an application that makes use of integratedRFID-sensor nodes in parking meters to monitor parkingviolations. Most vehicles these days have RFID-basedtoll tags or license plates. The sensors keep track ofthe parking duration and the RFID reader identifies thevehicle. The sensor nodes are grouped into regions andcollectively store the violations in that region. The storeddata can be logged to an external persistent storage eitherperiodically or when the number of violations in a regionexceeds a certain threshold. Alternatively, a cop can drivearound with a PDA and retrieve the stored data from

RID: Radio Interference Detection in WirelessSensor Networks

Gang Zhou, Tian He, John A. Stankovic,Tarek AbdelzaherDepartment of Computer Science

University of Virginia, Charlottesville 22903{gzhou,tianhe,stankovic,zaher}@cs.virginia.edu

Abstract— In wireless sensor networks, many protocols assumethat if node A is able to interfere with node B’s packet reception,node B is within node A’s communication range. It is alsoassumed that if node B is within node A’s communication range,node A is able to interfere with node B’s packet receptionfrom any transmitter. While these assumptions may be usefulin protocol design, they are not valid, according to the realexperiments we conducted in MICA2 platform. For a strong linkthat has a high packet delivery ratio, the interference range isobserved smaller than the communication range, while for a weaklink that has a low packet delivery ratio, the interference rangeis larger than the communication range. So using communicationrange information alone is not enough to design real collision-free media access control protocols. This paper presents a radiointerference detection protocol (RID) and its variation (RID-B)to detect run-time radio interference relations among nodes. Theinterference detection results are used to design real collision-freeTDMA protocols. With extensive simulations in GlomoSim, andwith sensor network application scenarios, we observe that theTDMA which uses the interference detection results has 100%packet delivery ratio, while the traditional TDMA has packet lossup to 60%, in heavy load. In addition to the scheduling-basedTDMA protocols, we also explore the application of interferencedetection on contention-based MAC protocols.

I. INTRODUCTION

Wireless sensor networks (WSN) is an emerging technologythat has a wide range of potential applications [1], includingenvironment monitoring, smart houses, remote medical sys-tems, sheep shepherding, and intrusion detection. Recent work[2][3][4][5][6] found that radio communication in wirelesssensor networks (WSN) differs significantly from traditionalMobile Ad-Hoc Networks (MANET). For example, when nodeC’s signal can interfere with node A’s signal, preventing A’ssignal from being received at node B, it’s usually assumedthat node B must be within node C’s communication rangeand there is communication connectivity from C to B. Wename this assumption the interference-connectivity assump-tion, which is widely used to design collision-free MediaAccess Control (MAC) protocols [7][8].

However, from our experiments on the MICA2 platform, wefind that this interference-connectivity assumption is not valid.In actuality, a node can interfere with another node even if itis beyond its communication range. In our experiments weshow that when the receiver is within, but close to the edge ofthe transmitter’s radio range, (referred to as a weak link), thetransmitter’s signal is easily interfered with at the receiver byanother node which has no connectivity with the receiver. Such

experiments are repeated several times and it is always foundthat the interference-connectivity assumption is violated.

The interference-connectivity assumption assumes that in-terference always comes from connectivity. Another assump-tion, the connectivity-interference assumption, assumes thatconnectivity always leads to interference. The connectivity-interference assumption was firstly addressed in [9] and isdescribed as follows: when two transmitters, A and B, bothhave connectivity to a receiver C and transmit simultaneously,a collision occurs, the data is corrupted and neither packet isreceived correctly. Our experiments confirm the observation in[9] that the connectivity-interference assumption is not alwaysmaintained. In other words in spite of the logical interference,one packet may be received while the other is corrupted.We observe that this assumption is usually violated in thecase of strong links, that is, when the transmitter is close tothe receiver and has a very strong signal that dominates theinterfering signal.

Without these assumptions between connectivity and in-terference, it’s extremely challenging to design collision-freeMAC protocols. In this paper, the idea of radio interference de-tection at run-time is put forth, for the first time, to design realcollision free MAC protocols that don’t depend on these non-realistic assumptions. Our solutions obtain the interferencerelations among nodes to assist in achieving real collision-freepacket delivery. The design of the radio interference detectionprotocol, RID, is presented. A lightweight version, the RID-Basic (RID-B), is also presented. Extensive simulations usingsensor network scenarios are conducted to compare the per-formance of TDMA and TDMA-RID-B (TDMA with RID-Bsupport). The performance evaluation shows that traditionalTDMA can have up to 60% packet loss in heavy-loadednetworks, while TDMA-RID-B can maintain 100% packetdelivery ratio.

The rest of this paper is organized as follows: in Sec-tion II, experimental observations of radio interferences in theMACA2 platform are presented. Then in Section III, the radiointerference detection protocol, RID, and its variation, RID-B, are explained. In Section IV, extensive performance eval-uations of TDMA and TDMA-RID-B, in which the TDMAscheduling is based on RID-B’s interference detection result,are analyzed. In addition, it is also explained how to use RIDon contention-based MAC protocols. In Section VI, relatedwork is analyzed, and finally in Section VII, conclusions are

given and future work is pointed out.

II. EXPERIMENTS ON RADIO INTERFERENCEIn this section, we present experimental results collected

in an outdoor environment with MICA2 motes. From theseexperiments, it is confirmed that radio interferences are ubiq-uitous phenomena. It is also observed that for a weak link thathas a low packet delivery ratio, the interference range is largerthan the communication range, while for a strong link that hasa high packet delivery ratio, the interference range is smallerthan the communication range.

A. Experimental Setup

For each experiment, three MICA2 motes are used. OneMICA2 mote is used as the transmitter, and another MICA2mote is used as the receiver, and the third one is used asthe jammer, whose transmission is synchronized with that ofthe transmitter to generate possible interference. The carriersensing and backoff operations in the MAC layer are disabledto ensure packets are simultaneously sent out by the jammerand the transmitter. The radio interferences are reflected bythe changing packet delivery ratios.

Mover1

r2

TR J

(a) A Weak Link From T to R

Mover1

r2

TR J

(b) A Strong Link From T to R

Fig. 1. Empirical Experiment Setting

All experiments are conducted in an open parking lot lateat night, in order to separate possible influence from peopleand moving cars on the radio interference measurements. Inthe experiment, as Figure 1 illustrates, the transmitter T andreceiver R are fixed in positions, and jammer J moves alongthe line determined by the positions of the transmitter andthe receiver. The jammer is tried at different positions alongthe line to observe different degrees of interferences, andinterference observations in different directions are measuredas well.

In the experiments, two kinds of links are used, strong linksand weak links. The setting of a weak link case is presentedin Figure 1(a). The receiver R is put on the communicationedge of transmitter T , i.e. the fan area [3], resulting in a weaklink that only has 80% packet delivery ratio. The distancebetween the transmitter and the receiver is measured to be16.2 feet. In Figure 1(b), the setting of the strong link case isshown. The receiver R is not put on the edge of transmitter T ’scommunication range. Instead, it is put close to the transmitter,to get a strong link that has a stable 100% packet deliveryratio. The distance between the transmitter and the receiveris measured to be 8.5 feet. Similarly, the jammer is puton different positions in one direction to observe differentinterferences, and also interferences from different directionsare measured.

All the experiments are repeated several times and consis-tent results are obtained.

B. Interference in one Direction

In this experiment, the jammer’s interference is measuredin one direction, on both a strong link and a weak link. In theweak link case, the packet delivery ratio is 80% when thereis no interference, and the distance between transmitter T andreceiver R (Figure 1(a)) is 16.2 feet. The experimental resultsof this observation are illustrated in Figure 2. As Figure 2(a)shows, when the distance between jammer J and receiver Rincreases, more packets from transmitter T are able to gothrough the channel and be correctly received by the receiver.On the other hand, less and less packets from the jammer arecorrectly received by the receiver. This is because when thejammer moves further away from the receiver, its own signalgets weaker when it arrives at the receiver, thus becomingless capable of interfering with the transmitter’s signal. Onthe contrary, the transmitter’s signal makes the signal of thejammer harder to receive.

In the case of a strong link that has 100% packet deliveryratio, similar phenomena are observed. The packet deliveryratio of the transmitter increases from 0% to 100%, with theincrease of the distance between the jammer and the receiverfrom 2.92 feet to 19 feet. On the other hand, the packetdelivery ratio of the jammer decreases from 99.2% to 0%.

In addition, both Figure 2(a) and (b) illustrate that whenthe transmitter and the jammer have similar distances tothe receiver, the total communication throughput of the link,including packets from both the jammer and the transmitter,goes down. This is because when the transmitter’s distance tothe receiver is similar to that of the jammer, their signals are atsimilar power levels, resulting in higher probability that bothof them get corrupted.

C. Interference in Different Directions

Besides the interferences in one direction, interferences indifferent directions are also measured, and Figure 3 showsthe experimental results. As Figure 3 shows, neither the radiointerference pattern nor the radio communication pattern isspherical, which is consistent with the result in [2].

Models and Solutions for Radio Irregularity in

Wireless Sensor Networks

GANG ZHOU, TIAN HE, SUDHA KRISHNAMURTHY, JOHN A. STANKOVIC

Department of Computer Science

University of Virginia

In this paper, we investigate the impact of radio irregularity on wireless sensor networks. Radio

irregularity is a common phenomenon which arises from multiple factors, such as variance in

RF sending power and different path losses depending on the direction of propagation. From our

experiments, we discover that the variance in received signal strength is largely random; however, it

exhibits a continuous change with incremental changes in direction. With empirical data obtained

from the MICA2 and MICAZ platforms, we establish a radio model for simulation, called the Radio

Irregularity Model (RIM). This model is the first to bridge the discrepancy between sphericalradio models used by simulators and the physical reality of radio signals. With this model,we investigate the impact of radio irregularity on several upper layer protocols, including MAC,routing, localization and topology control. Our results show that radio irregularity has a relativelylarger impact on the routing layer than the MAC layer. It also shows that radio irregularity leads tolarger localization errors and makes it harder to maintain communication connectivity in topology

control. To deal with these issues, we present eight solutions to deal with radio irregularity. Weevaluate three of them in detail. The results obtained from both the simulations and a running

testbed demonstrate that our solutions greatly improve system performance in the presence ofradio irregularity.

Categories and Subject Descriptors: C.2.1 [Computer Communication Network]: NetworkArchitecture and Design; I.6 [Computer Methodologies]: Simulation and Modeling

General Terms: Design, Algorithms, Measurement, Performance, Experimentation

Additional Key Words and Phrases: Sensor networks, wireless communication, radio irregularity,sending power, path loss, link asymmetry, packet loss, localization, topology control

1. INTRODUCTION

Radio irregularity is a common and non-negligible phenomenon in wireless sensornetworks. It results in irregularity in radio range and variations in packet loss indifferent directions, and is considered as an essential reason for asymmetric linksas viewed by upper layers in the protocol stack. Several empirical studies [Gane-san et al. 2002][Woo et al. 2003][Zhao and Govindan 2003][Cerpa et al. 2003] onthe Berkeley mote platform have shown that the radio range varies significantlyin different directions and the percentage of asymmetric links in a system variesdepending on the average distance between nodes.

The impact of radio irregularity on protocol performance can be investigatedthrough a running system. However, few researchers have actually pursued thisdirection, because of two reasons: First, the complexity and cost of performanceevaluations on a running system escalate, when sensor networks scale up to thou-sands or more nodes. Second, repeatable results of radio performance are extremelyhard to obtain from uncontrolled environments, hence leading to difficulties in sys-tem tuning and performance evaluation. As a result, simulation techniques are used

as an efficient alterative to evaluate protocol performance. Unfortunately, most ex-isting simulations do not take radio irregularity, a common phenomenon in wirelesscommunication, into account. The spherical radio patterns assumed by simulatorssuch as [Zeng et al. 1998] may not approximate real radio properties well enoughand hence may lead to an inaccurate estimation of application performance.

Several researchers [Ganesan et al. 2002][Woo et al. 2003][Zhao and Govindan2003][Cerpa et al. 2003] have already shown extensive evidence of radio irregular-ity in wireless communication. Their main focus is to observe and quantify suchphenomena. This paper is distinguished from the previous ones for the initiative inbridging the gap between spherical radio models used by simulators and the phys-ical reality of radio signals. We first verify the presence of radio irregularity usingempirical data obtained from MICA2 and MICAZ platforms. The results demon-strate that the radio pattern is largely random; however, it exhibits a continuouschange with incremental changes in direction. Based on experimental data, a radiomodel for simulations, called the Radio Irregularity Model (RIM), is formulated.RIM takes into account both the anisotropic properties of the propagation mediaand the heterogeneous properties of devices.

With the help of the RIM model, we explore the impact of radio irregularity onMAC, routing, localization and topology control performance. Among the proto-cols we evaluate, we find that radio irregularity has a significant impact on rout-ing, localization and topology control protocols, but a relatively small impact onthe MAC protocols. We also find that location-based routing protocols, such asGeographic Forwarding (GF) [Karp 2000] perform worse in the presence of radioirregularity than on-demand protocols, such as AODV [Perkins and Royer 1999]and DSR [Johnson and Maltz 1996]. We propose several potential solutions todeal with radio irregularity in wireless sensor networks. We evaluate the Symmet-ric Geographic Forwarding solution in simulation, and implement the AsymmetryDetection Service as well as the Bounded Distance Forwarding solution in runningsystems with 27∼60 MICA2 devices. Our results illustrate that our solutions dosucceed in alleviating the performance penalties due to radio irregularity.

The rest of this paper is organized as follows: we briefly analyze the causes andimpact of radio irregularity in Section 2. In Section 3, we describe experimentaldata collected from the Berkeley mote platform and make some general conclusionsabout radio irregularity. Based on these conclusions, we propose the RIM radiomodel in Section 4. We then use the RIM model in simulations to analyze theimpact of radio irregularity on MAC protocols in Section 5, routing protocols inSection 6, localization protocols in Section 7 and topology control protocols inSection 8. Solutions to deal with radio irregularity are proposed and evaluated inSection 9. Finally, we conclude the paper in Section 10.

2. ANALYSIS OF RADIO IRREGULARITY

In this section, we first identify the causes of radio irregularity, and then brieflydiscuss the impact of irregularity on the different protocol layers.

2.1 Causes of Radio Irregularity

Radio irregularity is caused by two categories of factors: devices and the propa-gation media. Device properties include the antenna type (directional or omni-

RAP: A Real-Time Communication Architecture for Large-Scale Wireless Sensor Networks

Chenyang Lu Brian M. Blum Tarek F. Abdelzaher John A. Stankovic Tian He

Department of Computer Science University of Virginia

Charlottesville, VA 22903 {chenyang, bmb5v, zaher, stankovic, th7c}@cs.virginia.edu

Abstract

Large-scale wireless sensor networks represent a new generation of real-time embedded systems with significantly different communication constraints from traditional networked systems. This paper presents RAP, a new real-time communication architecture for large-scale sensor networks. RAP provides convenient, high-level query and event services for distributed mi-cro-sensing applications. Novel location-addressed communication models are supported by a scalable and light-weight network stack. We present and evalu-ate a new packet scheduling policy called velocity monotonic scheduling that inherently accounts for both time and distance constraints. We show that this policy is particularly suitable for communication scheduling in sensor networks in which a large number of wireless devices are seamlessly integrated into a physical space to perform real-time monitoring and control. Detailed simulations of representative sensor network environ-ments demonstrate that RAP significantly reduces the end-to-end deadline miss ratio in the sensor network.

1. Introduction

With the advances in MEMS devices and embedded processors and radios, it will soon be feasible to deploy large-scale sensor networks to perform distributed mi-cro-sensing and control of physical environments [11]. For example, a surveillance system may use a large network of acoustic sensors to detect and track vehicles in a security area. Similarly, biometric sensors can be deployed in airports to detect harmful bio-agents and issue alarms to command and control centers during potential bio-attacks. These smart sensors and actuators are equipped with low-power processors and short-range radio transceivers [9]. They will automatically form multi-hop ad hoc networks to communicate both among themselves and to remote base stations (e.g., PDA’s).

Because distributed micro-sensing involves direct in-teraction with a physical environment, data communi-cation in sensor networks often has timing constraints in the form of end-to-end deadlines. Surveillance may require the position of an intruder be reported to a command center within 15 sec so that pursuing actions can be initiated in time. Data in a system may have different deadlines due to different validity intervals. The validity intervals (and hence, update deadlines) of the locations of different intruders such as pedestrians and motor vehicles may depend on their movement speeds. For example, locations of tanks have shorter update deadlines than those of pedestrians. Similarly, the location of an intruder should have a shorter update deadline than the temperature measurement of a region because the former can change faster than the latter. Sensor network protocols should support real-time communication by minimizing the packet deadline miss ratio, i.e., the percentage of packets that meet their end-to-end deadlines.

While sensor networks share the notion of timing con-straints with more traditional embedded systems, they differ in two respects. First, individual sensors are typi-cally very small in size and resource capacity. Hence, the philosophy of sensor networks relies on resource dedication rather than sharing. In other words, individ-ual sensor devices and nodes are likely to be dedicated for individual tasks, thereby eliminating much of the need for sophisticated CPU scheduling in a multitask environment.

Second, it is envisioned that sensor nodes will operate in groups, since individual nodes are too limited and unreliable to perform useful activities from the applica-tion’s perspective. Group activities require coordina-tion and communication among member nodes. Sens-ing results of groups need to be sent back to base sta-tions through multi-hop communication. Thus, the main schedulable resource becomes the wireless com-

munication channel. Progress of user-level activities and their ability to meet end-end deadlines are there-fore determined by scheduling of the communication medium rather than scheduling of the processor. To-wards that end, new real-time communication architectures are required for ad hoc wireless environ-ments.

Multi-hop wireless communication scheduling differs from CPU scheduling in that it has an inherent notion of distance. In sensor networks, the distance is deter-mined by the physical locations of source and destina-tion. These locations impose distance constraints on messages, in addition to time constraints, calling for communication scheduling policies that are cognizant of both time and space.

The first contribution of this paper is RAP, a real-time communication architecture for large-scale wireless sensor networks. RAP provides a set of convenient, high-level query and event services to real-time dis-tributed micro-sensing applications. Query and event services are based on novel location-addressed com-munication models supported by a scalable and light-weight network stack.

The second contribution of RAP is a novel Velocity Monotonic Scheduling (VMS) policy suitable for packet scheduling in sensor networks. VMS is based on a notion of packet requested velocity. Each packet is expected to make its end-to-end deadline if it can move toward the destination at its requested velocity, which reflects its local urgency. Compared with non-prioritized packet scheduling, VMS improves the dead-line miss ratios of sensor networks by giving higher priority to packets with higher requested velocities. VMS can outperform deadline-based packet scheduling because velocity more accurately reflects the local ur-gency at each hop when packets with the same deadline have different distances to their destinations. Assuming that each sensor knows its own location (using GPS or other location services [8]), the requested velocity can be determined locally. This property enables VMS to scale well in large-scale sensor networks.

The final contribution of this paper is a detailed simula-tion study of the real-time performance of several rout-ing protocols and packet scheduling algorithms in a typical sensor network scenario. Our simulation ex-periments demonstrate that, for sensors far away from their base station, RAP reduces the deadline miss ratio from 90.0% to 17.9%, compared to existing wireless communication protocols (DSR over 802.11b). To our best knowledge, ours is the first detailed performance study on deadline issues in multi-hop wireless sensor network settings under overload conditions.

In the following sections, we discuss the key character-istics of sensor networks, present the design of RAP, report a set of simulation experiments with sensor net-work configurations, and conclude the paper by sum-marizing our key results and future work.

2. Real-time Communication in Sensor Networks

In this section, we describe the characteristics of sensor networks and communication models on sensor net-works. This analysis serves as a basis for our design of real-time communication protocols.

Sensor networks are an instance of mobile ad hoc net-works (MANET) [10] that have recently attracted a lot of interest and visibility due to their flexibility and the feasibility of their deployment at low costs. In general, mobile ad hoc networks depend on peer-to-peer com-munication protocols that do not require a fixed infra-structure such as centralized servers and access points. Sensor networks are different from their traditional ad hoc wireless counterparts (composed of laptops and PDA’s) in that they have a larger scale, higher density, smaller devices, and a tighter interaction with a physi-cal environment. Energy conservation is critical in sen-sor networks because of their long lifetime and form-factor constraints, which preclude the use of large bat-teries or power supplies.

In most envisioned sensor network applications, a large number of sensors are deployed in an area and a small number of more powerful nodes (such as PDA’s with Internet connections) form possibly mobile interface stations which serve as the entry points to the sensor network. In the following, we shall call such interface stations, base stations. A user may query the physical environment through such base stations. Alternatively she may also register for an event. The occurrence of the event will automatically trigger a specified query. A query can specify timing requirements including rate, start time, duration, and end-to-end deadlines. For example, a user can register for a virus_found event in a rectangular area with coordinates (10,10,20,20), and specify a query on the event to report the density of the detected virus. If a virus is found, the density of the viruses should be reported to the base station from where they are found every 1.5 sec for a duration of 30 min. Every reading should reach the base station within an end-to-end deadline of 5 sec.

Communication in a sensor network can be divided into two categories: local coordination and sensor-base communication. Before sending information to the base station, sensors within the local area coordinate among

Feedback Control Scheduling in Distributed Real-Time Systems*

John A. Stankovic, Tian He, Tarek Abdelzaher, Mike Marley, Gang Tao, Sang Son and Chenyang LuDepartment of Computer Science, University of Virginia

Charlottesville, VA 22903e-mail: { stankovic, tianhe, zaher, son, chenyang }@cs.virginia.edu, gt9s@virginia.edu

* This work was supported, in part by NSF grants CCR-0098269, the MURI award N00014-01-1-0576 from ONR, and DAPRPA ITOoffice under the NEST project (grant number F336615-01-C-1905).

AbstractDistributed soft real-time systems are becoming in-

creasingly unpredictable due to several important factorssuch as the increasing use of commercial-off-the-shelfcomponents, the trend towards open systems, and the pro-liferation of data-driven applications whose executionparameters vary significantly with input data. Such sys-tems are less amenable to traditional worst-case real-timeanalysis. Instead, system-wide feedback control is neededto meet performance requirements. In this paper, we ex-tend our previous work on developing software controlalgorithms based on a theory of feedback control to dis-tributed systems. Our approach makes three importantcontributions. First, it allows the designer for a distrib-uted real-time application to specify the desired temporalbehavior of system adaptation, such as the speed of con-vergence to desired performance upon load or resourcechanges. This is in contrast to specifying only steady-statemetrics, e.g., deadline miss ratio. Second, unlike QoS op-timization approaches, our solution meets performanceguarantees without accurate knowledge of task executionparameters - a key advantage in an unpredictable envi-ronment. Third, in contrast to ad hoc algorithms based onintuition and testing, our solution has a basis in the theoryand practice of feedback control scheduling. Performanceevaluation reveals that the solution not only has excellentsteady state behavior, but also meets stability, overshoot,and settling time requirements. We also show that the so-lution outperforms several other algorithms available inthe literature.

1. IntroductionMany soft real-time systems such as smart spaces, fi-

nancial markets, processing audio and video, and theworld wide web are not amenable to traditional worst-casereal-time analysis. Most of these systems are distributed.They operate in open environments where both load andavailable resources are difficult to predict. Monitoring andfeedback control are needed to meet performance con-

straints. Several difficulties are observed in meeting per-formance constraints in these systems. One main difficultylies in their data-dependent resource requirements, whichcannot be predicted without interpreting input data. Forexample, the execution time of an information server (aweb or database server) heavily depends on the content ofrequests, such as the particular web page requested. Asecond major challenge is that these systems have highlyuncertain arrival workloads; it is not clear how many userswill request some resource in the www or how many usersmight walk into a smart space. A third challenge involvesthe complex interactions among many distributed sites,often across an environment with poor or unpredictabletiming behavior. Consequently, developing certain typesof future real-time systems will involve techniques formodeling the unpredictability of the environment, han-dling imprecise or incomplete knowledge, reacting tooverload and unexpected failures (i.e., those not expressedby design-time failure hypotheses), and achieving the re-quired performance levels and temporal behavior.

We envision a trend in the theory of real-time com-puting that aims at providing performance guaranteeswithout the requirement of fine-grained task executionmodels such as those involving individual task executiontimes. Instead, we shall see the emergence of coarse-grained models that describe the aggregate behavior of thesystem. Coarse-grained models are easier to obtain andthey need not be accurately computed. These models aremore appropriate for system analysis in the presence ofuncertainty regarding load and resources. In this paper, weexplore one such model based on difference equations.Unlike the more familiar queuing-theory models of aggre-gate behavior, difference equation models do not makeassumptions regarding the statistics of the load arrivalprocess. Thus, while both types of models describe queu-ing dynamics, difference equation models are independentof load assumptions and consequently more suitable forsystems where load statistics are difficult to obtain orwhere the load does not follow a distribution that is easyto handle analytically. The latter is the case, for example,with web traffic, which cannot be modeled by a Poisson

distribution. Differential equation models include inputload as a measured variable. Hence, they are particularlysuitable for feedback control architectures in which inputload is measured at run-time.

In this paper, we show that our solution outperformsseveral other algorithms available in the literature. Someof our key performance results are the ability to adapt tounpredictable environments and better transient and steadystate response. In addition to improvements in perform-ance, our solution has a basis in the theory and practice offeedback control scheduling. This is in contrast to themore common ad hoc algorithms based on intuition andtesting where it is very difficult to characterize the aggre-gate performance of the system and where major over-loads and/or anomalous behavior can occur since thealgorithms are not developed to avoid these problems.

2. DFCS Framework OverviewEarly research on real-time computing was concerned

with guaranteeing avoidance of undesirable effects such asoverload and deadline misses. Solutions were predicatedon knowing worst-case resource requirements a priori. Incontrast, in highly uncertain environments, the main con-cern is to design adaptation capabilities that handle un-certain effects dynamically and in an analyticallypredictable manner. To address this issue, we propose aframework called Distributed Feedback Control Real-timeScheduling (DFCS). The framework is based on feedbackcontrol that incrementally corrects system performance toachieve its target in the absence of initial load and re-source assumptions. One main performance metric of sucha system is the quality of performance-convergence to thedesired level. In our framework, the desired convergenceattributes may be specified and enforced using mecha-nisms borrowed from control theory. These mechanismsare very robust and have been applied successfully fordecades in physical process-control systems that are oftennon-linear and subject to random external disturbances.Before establishing our DFCS framework, we give anoverview of the software system being controlled and de-scribe the feedback-control mechanism involved.

We assume that there are N computing nodes con-nected via a network. Tasks arrive at nodes in unknownpatterns. Each task is served by a periodically invokedschedulable entity (such as a thread) with each instancehaving a soft deadline equal to its period. The periodicityconstraint is motivated by the requirements of real-timeapplications such as process control and streaming media.It is also motivated by recent trends in real-time operatingsystem design, such as temporal isolation of independentapplications. Note that temporal isolation is usuallyachieved using operating system constructs such as capac-ity reserves, hierarchical schedulers, and resource shares,all of which rely on periodic scheduling in the kernel.

We abstract a typical adaptive system by two sets of

performance metrics. The primary set represents metricsto be maintained at specified levels, for example, the op-timal utilization of a server, or the desired altitude of anairplane. The secondary set represents negotiable metricssuch as service quality. The objective of adaptation is toincur minimum degradation in secondary metrics whilemaintaining the primary metrics at their desired values. Torepresent multiple levels of degradation in secondary met-rics, we assume that each task has several service levels ofdifferent quality. For example, a task can execute forvarying amounts of time with the quality of the result im-proving with greater execution time.

The goal of our DFCS architecture is to maintain theprimary performance metrics around their targets. A keyintuition that affects the architecture of the feedback loopsis that the dynamics of a distributed system manifestthemselves on two different time-scales. Fast dynamicsare observed on individual nodes. These dynamics arisefrom local load changes due to individual task arrivals andterminations. Slower dynamics are observed globally onthe scale of the entire system. These dynamics arise fromchanges in aggregate load distribution. Hence, our feed-back architecture necessarily involves two sets of loops,local and distributed ones, each tuned to the dynamics ofthe corresponding scale.

Each node in the distributed system (Figure 1) has alocal feedback control system (LFC) and a distributedfeedback control system (DFC). The distributed feedbackcontroller is responsible for maintaining the appropriateQoS balance between nodes. The local feedback controlleris responsible for tracking the global QoS set point set bydistributed controller and ensuring that tasks that are ad-mitted to this node have a minimum miss ratio (MR) andthe node remains fully utilized. It is important to note thatthese two types of controllers form the main parts of thedistributed scheduling in the system, but they are not theentire algorithm. In this paper we develop the distributedcontroller, rely on a local controller (developed earlier andpresented in a previous paper [10][11][12]), and then em-bed these two controllers into the complete distributedscheduling algorithm (see section 3.7). We test the com-plete algorithm.

Now consider a few more details about the control-lers’ structure. The distributed controller commands thelocal controller via the QoS set point. The entire controlsystem of the local node becomes an actuator for the dis-tributed controller to control the state of one local node.The local controller manipulates its actuators in the LFCto achieve the target QoS set point. While the frameworkis general enough to accommodate different definitions ofQoS, in this paper, we let the primary performance metricbe the deadline miss ratio (MR). Since performance met-rics of admitted tasks can be trivially satisfied if the ad-mitted task set is empty, it is especially important toquantify the loss of QoS due to task rejection to avoidtrivial solutions. For this reason, we use two different miss

Achieving Realistic Sensing Coveragein Wireless Sensor Networks

Joengmin Hwang, Tian He, Yongdae KimDepartment of Computer Science, University of Minnesota, Minneapolis

1 Introduction

Despite the well-known fact that sensing patterns in realityare highly irregular, researchers continue to develop protocolswith simplifying assumptions about the sensing. For example,a circular 0/1 sensing model [2] is widely used in most existingsimulators and analysis. While this model provides high-levelguidelines, it could cause wrong estimation of system perfor-mance in the real world.

Our answer to this issue is a sensing area modeling tech-nique, which obtains the coverage of sensor nodes through eventtraining. The main idea is using discrete controlled events toidentify the sensing coverage based on event detection resultsby individual sensor nodes. A key architectural advantage ofthis approach is a lightweight design in sensor node with mini-mal overhead. Besides communication, each sensor node onlyneeds to support a simple detection function (with optional timesynchronization requirement). We evaluate our model using aphysical experiment on a testbed consisting of 40 MicaZ motes.Our evaluation results shows we can identify the sensing areawith low error rate with simple training method.

2 Sensing Area Modeling

In this section, we introduce the design of our system at thearchitectural level.

2.1 Assumptions

We assume that we can generate controlled events withknown time and location. This can be done through two ap-proaches. First, events can be generated by using real targets.For example, one or multiple robots can move along predefinedtraces to activate PIR motion sensors in the field. Other events,such as vibration and sound, can be generated similarly. Sec-ond, the controlled events can be injected using special devicessuch as infrared radiation generator, Spotlight [1] and our pro-totype system described. Since the methods to generate con-trolled events are diversified, we intentionally describe our ap-proach conceptually independent of the concrete method used.The targeted application scenarios are 1) to identify the cover-age of motion sensors with a room, 2) to discover blind spots ina surveillance area, and 3) to compensate the irregularity of tinyproximity sensors used for paper-edge detection in printers.

2.2 Main Idea

The main idea of our training-based modeling approach is togenerate controlled events in the space where the sensor nodesare deployed. An event could be, for example, the presenceof an object in an area or a light spot projected on a plate ofsensors. Formally, an event can be defined as a detectable phe-nomenon e(t, p) that occurs at time t and at location p ∈ A ⊂R

k

(k = 1,2,3). Without loss of generality, we use k = 2 in the restof the paper. An event is said to be controlled if we can enforce

a relationship between the time t and location p. In other words,a set of controlled events can be described as the event locationsover the discrete time: G : R → R

2, where G(t) = pt = (xt ,yt)where t ∈ {t1,t2, ...,tn}.

It consists of two major parts: an event generator G and aset of sensor nodes ni (i ∈ N). The event generator G couldbe a single device or multiple distributed devices that can gen-erate a sequence of controlled events e(t, p) with known spa-tiotemporal correlation G(t) = p(xt ,yt). We define Si(t, p) asthe detection function of node ni. If node ni can detect evente(t, p), Si(t, p) = 1, otherwise Si(t, p) = 0. In case of detection,sensor nodes store the timestamp t locally. By the end of train-ing, a sensor computes the location of all events it detects byinputting the timestamps into G(t). Therefore, a set of times-tamps Ti = {t i

1,ti2, . . . ,t i

n} stored in node ni can be converted to

a set of locations Pi = {pi1, pi

2, . . . , pin} within the sensing area.

The location set Pi can be directly used to describe the sensingarea of node ni.

2.3 Design of Event Generator G(t)Since the overhead and accuracy of the sensing modeling

is largely determined by G(t), it is important to consider sev-eral solutions to optimize G(t) under different system configu-rations.

2.3.1 Regular G(t)To illustrate the basic functionality of an event generator, we

start with a simple sensor system where the sensing area of anode is a line segment. We shall find out the portion of the lineincluded in the sensing ranges of sensor node n1 and n2. Toachieve this, the event generator creates discrete point eventsalong this line [0,L] with constant speed v with same intervalD. Formally, G(t) = t · v where t = kD/v and 0 ≤ k ≤ L/D.For example, a sensor node n1 collects a set of six timestampsT1 = {t1,t2, . . . ,t6} at which the events are detected. From func-tion G, the actual locations of events are converted to a set oflocations P1 = {t1v,t2v, . . . ,t6v}. The sensing coverage of sen-sor n1 can be defined as the line segment that covers P1. Forthe sensor n2, it reports timestamps T2 = {t4,t5,t6,t7} and thesensing coverage of sensor n2 is defined as the line segmentthat covers P2 = {t4v,t5v,t6v,t7v}. The intersection of T1 andT2, T1 ∩ T2 = {t4,t5,t6} indicates the coverage of two sensorsis overlapped. The regular training can be generalized to thecase when the events occur in a plane by dividing the plane intoseveral lines with a certain interval.

In addition to the progressive scanning, the G(t) function ofthe regular training can generate events with an arbitrary se-quence as long as every point in the area is covered. Also, aslong as we can match the event and its position we can use anytraining method for G(t) which includes unsupervised event.

1

Figure 1. System Setup

1 1.5 2 2.5 3 3.5 40

0.05

0.1

0.15

0.2

0.25

0.3

0.35

interval

co

ve

rag

e e

rro

r

fp, var=1fp, var=2fp, var=3fn, var=1fn, var=2fn, var=3

Figure 2. Errors in regular G(t) withvarying interval and irregularity

1 1.5 2 2.5 3 3.5 40

0.05

0.1

0.15

0.2

0.25

0.3

0.35

last level interval

co

ve

rag

e e

rro

r

fp, var=1fp, var=2fp, var=3fn, var=1fn, var=2fn, var=3

Figure 3. Errors in hierarchical G(t)with varying interval and irregularity

Algorithm 1 Hierarchical G(t) process

Output: Pi: The sensing area of ni.1: G(t) starts with level-1 events e(t, p) (The number of level-

1 event is decided by the minimum sensing area)2: Node ni reports Si(t, p) for all level-1 events3: repeat4: for all level-k adjacent pair e(tm, pm) and e(tn, pn) do5: if any node detects only one of events && no event

generated at positionpm+pn

2before then

6: Generate a level-(k +1) event at positionpm+pn

27: end if8: end for9: k = k +1

10: until (k = Maximum Level)11: Pi is a set of positions p where Si(t, p) = 1

2.3.2 Hierarchical G(t)Hierarchical G(t) is motivated by the observation that the

boundary area of a sensing coverage requires more detail thanthe area in the middle of coverage. With the hierarchical G(t),we can reduce the number of events required to obtain the sameaccuracy as regular G(t). A level-1 event divides the area intofour sub-areas, and level-2 events divide the area into 16 sub-areas. In general, level-i events divide an area into 4i sub-areas.If an event is a level-i event, it is also a level- j event, wherej ≥ i. Two events are said to be adjacent(or a pair) if theyare neighboring each other vertically, horizontally or diagonally(e.g., an event could have maximal eight adjacent events). Twoadjacent events are said to be a boundary pair if only one of twoadjacent events is within a sensing range of some node. form aboundary pair). The event in the boundary pair is called bound-ary event. The main idea of Hierarchical G(t) is to recursivelygenerate new events in the middle of boundary pairs. It worksin a way similar to the binary search within a two-dimensionalspace. We describe the step by step operation of HierarchicalG(t) in Algorithm 1.

3 System Implementation

We design and implement a complete version of our sys-tem which includes regular and hierarchical training on theTinyOS/Mote platform. The NesC language is used to programthe motes and Java is used to build the regular and hierarchicalgenerators. The compiled image of a full mote implementation

occupies 14,500 bytes of code memory and 605 bytes of datamemory. For each event we generate, we assign a unique ID.By using these IDs, we eliminate the need for time synchroniza-tion. We use an oracle algorithm that assumes the knowledge ofthe sensing area of the nodes. Basically, this algorithm activatesa sensor node (e.g., through projecting light to a sensor), if thecontrolled event e(t, p) is within the sensing area of the node.We want to emphasize that the oracle algorithm and generatedground truth is used only for the purpose of evaluation. Thisknowledge is not used in any part of our proposed algorithm.Figure 1 shows the implementation setting. After each run, thetraining results are visualized on the board and compared withthe ground truth.

In experiments, we investigate the impact of training interval(resolution) and sensing irregularity in both training methods.We divide error into two types: (1) false positive f p: the areameasured as a part of sensing coverage but is not a part of realsensing coverage, or (2) false negative f n: the area which ismeasured as not in the sensing coverage but is a part of realsensing coverage. Figure 2 and 3 shows the coverage error inregular training and hierarchical training respectively. The de-gree of irregularity of sensing coverage is denoted by var wherehigher value means more severe irregularity. Our result showsthat with a small training interval, we can achieve very precisecoverage modeling, f p is almost 0% and f n is at 1% to 8%and even if we increase the interval, f p is still almost 0%. Theperformance of hierarchical training is almost similar to the reg-ular training, which means we can do an efficient training whilesaving the cost of event generation.

4 Conclusion

This paper intends to draw attention to the sensing irregular-ity issue known but largely ignored by many designers. We con-tribute to this area by designing and implementing two training-based methods that accurately identify the sensing patterns ofnodes using controlled events. Our design has been fully im-plemented and evaluated in a test-bed consisting of 40 MicaZmotes. We hope this work motivates our community to seri-ously consider the reality issues existed in the sensor networks.

5 References[1] R. Stoleru, T. He, J. A. Stankovic, and D. Luebke. High-Accuarcy, Low-

Cost Localization System for Wireless Sensor Networks. In Sensys’05,November 2005.

[2] T. Yan, T. He, and J. A. Stankovic. Differentiated Surveillance Service forSensor Networks. In SenSys’03, November 2003.

2

Lightweight Detection and Classification for WirelessSensor Networks in Realistic Environments

Lin Gu 1 Dong Jia 2 Pascal Vicaire 1 Ting Yan 1 Liqian Luo 1

Ajay Tirumala 3 Qing Cao 1 Tian He 1

John A. Stankovic 1 Tarek Abdelzaher 1 Bruce H. Krogh 2

{lingu, pv9f, ty4k, ll4p, qc3h, th7c, stankovic, zaher}@cs.virginia.edu{djia, krogh}@ece.cmu.edu tirulama@uiuc.edu

ABSTRACTA wide variety of sensors have been incorporated into a spectrum ofwireless sensor network (WSN) platforms, providing flexible sens-ing capability over a large number of low-power and inexpensivenodes. Traditional signal processing algorithms, however, oftenprove too complex for energy-and-cost-effective WSN nodes. Thisstudy explores how to design efficient sensing and classification al-gorithms that achieve reliable sensing performance on energy-and-cost-effective hardware without special powerful nodes in a contin-uously changing physical environment. We present the detectionand classification system in a cutting-edge surveillance sensor net-work, which classifies vehicles, persons, and persons carrying fer-rous objects, and tracks these targets with a maximum error in ve-locity of 15%. Considering the demanding requirements and strictresource constraints, we design a hierarchical classification archi-tecture that naturally distributes sensing and computation tasks atdifferent levels of the system. Such a distribution allows multiplesensors to collaborate on a sensor node, and the detection and clas-sification results to be continuously refined at different levels ofthe WSN. This design enables reliable detection and classificationwithout involving high-complexity computation, reduces networktraffic, and emphasizes resilience and adaptation to the realisticenvironment. We evaluate the system with performance data col-lected from outdoor experiments and field assessments. Based onthe experience acquired and lessons learned when developing thissystem, we abstract common issues and introduce several guide-lines which can direct future development of detection and classifi-cation solutions based on WSNs.

Categories and Subject DescriptorsC.2.1 [Computer-Communication Networks]: Network Archi-tecture and Design; C.3 [Computer System Organization]: Spe-cial Purpose And Application-Based Systems—Real-Time and em-bedded systems; C.4 [Performance of Systems]: Design Studies

Permission to make digital or hard copies of all or part of this work forpersonal or classroom use is granted without fee provided that copies arenot made or distributed for profit or commercial advantage and that copiesbear this notice and the full citation on the first page. To copy otherwise, torepublish, to post on servers or to redistribute to lists, requires prior specificpermission and/or a fee.SenSys’05, November 2–4, 2005, San Diego, California, USA.Copyright 2005 ACM 1-59593-054-X/05/0011 ...$5.00.

General TermsDesign, Experimentation, Measurement, Performance

KeywordsClassification, Wireless Sensor Networks, VigilNet

1. INTRODUCTIONSensing is a fundamental function in wireless sensor networks.

Researchers have built WSN platforms with a wide spectrum ofsensors, ranging from simple thermistors to micropower impulseradars [4, 13, 14]. They provide flexible sensing capability witha large number of low-power and inexpensive sensor nodes. Non-trivial as it is, the selection and integration of sensors on a WSNplatform is often a manageable task given a certain amount of en-gineering effort. The situation is, however, completely differentabove the physical sensor and computing hardware layer – the ac-quisition and processing of sensor data impose great challenges onWSN design because of strict resource constraints.

Cost-effectiveness being an important objective, WSN design-ers often choose mass produced commercial off the shelf (COTS)sensors when designing a sensor network system. Moreover, a sen-sor node must be energy efficient. As a result, the raw sensor datais often of low-quality – they are not always reliable, not alwaysrepeatable, usually not self-calibrated, and often not shielded to en-vironment and circuit board noise. Obviously, it is necessary touse signal processing algorithms to filter, process, and abstract sen-sor data with software to provide precise, reliable, and easy-to-useinformation to applications.

Traditional signal processing algorithms, however, often provetoo complex to implement on inexpensive sensor network hardwarewithout digital signal processing co-processors. For example, thepopular Berkeley Mica series has an 8-bit micro-processor runningat 7.3827MHz, no hardware floating-point support, and only 4KBdata memory. Though recent versions of MicaZ and Telos motes

1Department of Computer Science, University of Virginia.2Department of Electrical and Computer Engineering, CarnegieMellon University.3Department of Computer Science, University of Illinois atUrbana-Champaign4This work is supported in part by NSF grant CCR-0098269 andCCR-0325197, the MURI award N00014-01-1-0576 from ONR,and the DAPRPA IXO offices under the NEST project (grant num-ber F336615-01-C-1905).

employ a bigger data memory, we expect the growth of computa-tional resources on WSN platforms to be rather slow because ofthe emphasis on low power consumption, low cost, and small formfactor. Generally, resource constraints will continue to represent thereality of energy-and-cost-effective embedded systems. This strictresource limitation makes it very difficult to execute Fast FourierTransformation (FFT) and other signal processing algorithms withmoderate or high time/space complexity. Also, the stringent en-ergy budget favors simple and quick algorithms over complex al-gorithms that require prolonged execution time.

While the computation/energy resources are limited, the appli-cation requirement is not. Specifically, the development of recentsurveillance WSNs requires the network to provide functionalitieswell beyond sensing and routing. Such surveillance WSNs are de-signed to detect and report certain classes of events of interest.When such an event happens, the WSN needs to detect it quickly,classify it into one category (e.g., person, vehicle), and compute itsattributes (e.g., location, velocity).

Designing such surveillance WSNs is a research challenge. Be-sides the obviously severe resource constraints, the following fac-tors also contribute to the difficulty of the task.• To provide sensing coverage for a relatively large area, the net-

work is usually comprised of a large number of densely de-ployed nodes. This imposes a challenge on efficient data prop-agation and reliable operation.

• The detection, classification, and reporting must be performedin a timely manner. It is usually required that the network com-plete the detection and classification before the target travelsout of the field so that the system can respond to the event. As aresult, offline-style processing performed by base stations withglobal and relatively “complete” data is often not feasible inthis context.

• To perform quality signal processing, the sensors often need tosample at a high sampling rate, stressing resource utilization.The sensing data is bursty and in large quantity.

• Surveillance networks are often deployed on rough terrains fora long period of time. Hence, it must be adaptive to the realistic,ever-changing environment.

Given the numerous technical challenges, important research ques-tions are: Can we construct a reliable surveillance WSN that meetsthe requirements within the strict resource constraints? What per-formance will such a system achieve? This study attempts to an-swer these questions by presenting the detection and classifica-tion system in VigilNet [10], which is a recently deployed surveil-lance WSN detecting and classifying vehicles, persons and personswith ferrous objects. Specifically, this paper explores the designchoices involved in constructing an efficient detection and classifi-cation system that achieves reliable performance on a network ofenergy-and-cost-effective sensor nodes, analyzes the performance,and proposes a set of guidelines for future designs of WSNs in asimilar design context.

It is worth clarifying that advanced signal processing mathemat-ics and algorithms are not the emphasis of this paper. Instead, thispaper focuses on the system design issues involved in creating areliable and realistic classification system for a surveillance WSNusing homogeneously low-end sensor nodes, as well as evaluationof the effectiveness of these designs. To the authors’ best knowl-edge, there has not been a large-scale deployment of such a sophis-ticated surveillance network without using special powerful nodes.Hence, our study is focused on answering this challenge: Withoutenhancing any individual nodes’ capability and cost, can a networkof distributed sensor nodes provide advanced functions and work

reliably in realistic environments? We believe that the experienceacquired and lessons learned in constructing such a system, and theanalysis of the trade-offs and design decisions in it, will benefit theresearch in this area, and help transform the research potential ofWSNs into real-world technology and market success.

The paper is organized as follows. Section 2 presents back-ground information and surveys related work. Section 3 gives anoverview of the VigilNet surveillance system. Section 4 presentsthe design of the hierarchical classification architecture. Systemlevel evaluation is shown in Section 5. Section 6 discusses sev-eral guidelines for designing a large-scale WSN for detection andclassification tasks. Finally, Section 7 concludes the paper.

2. BACKGROUNDFocusing on VigilNet’s hardware platform, we present a brief

overview of the sensing subsystem – sensors and their supportingcircuitry – on a sensor node. The sensing subsystem is the hard-ware foundation on which classification systems are constructed.We also survey the related work in the area of detection and classi-fication WSN systems.

2.1 Overview of the sensing subsystemVigilNet uses the ExScal motes as sensor nodes. Based on the

Mica2 [3] mote design, the ExScal mote, shown in Fig. 1, is de-signed by CrossBow Inc. and Ohio State University for large-scalesurveillance WSNs [8]. The major difference between the ExScalmote and the Berkeley Mica2 mote is that the former integratesa magnetometer (Honeywell HMC1052[2]), a microphone, and 4PIR sensors on the same circuit board as the processor’s. After thefirst prototype ExScal motes were delivered in March 2004, Cross-Bow released several versions with various improvements through-out the year of 2004.

Several correlated factors contribute to the complexity of thesensing circuitry. First, applications require a long sensing dis-tance, which implies a finer granularity for the sensor readings.Second, as a general purpose platform, designers hope to choosesensors with a wide measuring range. Third, the wider measuringrange combined with a finer granularity maps to more numeric val-ues which, however, have to all fall into the representation capabil-ity of the A/D converter (ADC), I/O bus, and CPU I/O port width.Finally, as the sensors on the sensor board grow in both number andsophistication, the support circuitry may need to support better fil-tering, handle more advanced signaling protocols, employ a fasteror wider bus, provide wider functionality (such as waking up thesensor node), or build better shielding to avoid cross-talk amongvarious components. These factors make the design of the sens-ing subsystem a significant engineering effort involving numerousdesign choices which often depend on the application domain.

Figure 1: ExScal mote

To solve the aforementionedrange and granularity problems forthe magnetometer, the ExScal moteincludes circuitry that allows theapplication program to adjust theinput signal to be amplified. Toprovide a quality signal for acous-tic processing, the microphone cir-cuitry incorporates a high-pass fil-ter and a low-pass filter. Both theinput adjustment and filtering arecontrollable by the processor, withan I2C bus connecting the proces-

sor and sensor components.

ATPC: Adaptive Transmission Power Controlfor Wireless Sensor Networks

Shan Lin, Jingbin Zhang, Gang Zhou, Lin Gu, Tian He†, and John A. Stankovic

Department of Computer Science, University of Virginia†Department of Computer Science and Engineering, University of Minnesota

{shanlin,jz7q,gzhou,lingu}@cs.virginia.edu, tianhe@cs.umn.edu, stankovic@cs.virginia.edu

AbstractExtensive empirical studies presented in this paper con-

firm that the quality of radio communication between lowpower sensor devices varies significantly with time and envi-ronment. This phenomenon indicates that the previous topol-ogy control solutions, which use static transmission power,transmission range, and link quality, might not be effectivein the physical world. To address this issue, online trans-mission power control that adapts to external changes is nec-essary. This paper presents ATPC, a lightweight algorithmof Adaptive Transmission Power Control for wireless sen-sor networks. In ATPC, each node builds a model for eachof its neighbors, describing the correlation between trans-mission power and link quality. With this model, we em-ploy a feedback-based transmission power control algorithmto dynamically maintain individual link quality over time.The intellectual contribution of this work lies in a novel pair-wise transmission power control, which is significantly dif-ferent from existing node-level or network-level power con-trol methods. Also different from most existing simulationwork, the ATPC design is guided by extensive field experi-ments of link quality dynamics at various locations and overa long period of time. The results from the real-world exper-iments demonstrate that 1) with pairwise adjustment, ATPCachieves more energy savings with a finer tuning capabilityand 2) with online control, ATPC is robust even with envi-ronmental changes over time.

Categories and Subject DescriptorsC.2.1 [Computer-Communication Networks]: Net-

work Architecture and Design—wireless communication

General TermsAlgorithms, Design, Experimentation, Measurement,

Performance

Permission to make digital or hard copies of all or part of this work for personal orclassroom use is granted without fee provided that copies are not made or distributedfor profit or commercial advantage and that copies bear this notice and the full citationon the first page. To copy otherwise, to republish, to post on servers or to redistributeto lists, requires prior specific permission and/or a fee.SenSys’06, November 1–3, 2006, Boulder, Colorado, USA.Copyright 2006 ACM 1-59593-343-3/06/0011 ...$5.00

KeywordsAdaptive, Feedback, Link Quality, Wireless Sensor Net-

work, Transmission Power Control

1 IntroductionWith the integration of sensing and communication abil-

ities in tiny devices, wireless sensor networks are widelydeployed in a variety of environments, supporting militarysurveillance [1] [24], emergency response [41], and scien-tific exploration [36]. The in-situ impact from these en-vironments, together with energy constraints of the nodes,makes reliable and efficient wireless communication a chal-lenging task. Under a constrained energy supply, reliabilityand efficiency are often at odds with each other. Reliabil-ity can be improved t by transmitting packets at the maxi-mum transmission power [13] [38], but this situation intro-duces unnecessarily high energy consumption. To providesystem designers with the ability to dynamically control thetransmission power, popularly used radio hardware such asCC1000 [6] and CC2420 [7] offers a register to specify thetransmission power level during runtime. It is desirable tospecify the minimum transmission power level that achievesthe required communication reliability for the sake of savingpower and increasing the system lifetime.

Although theoretical study and simulation provide a valu-able and solid foundation, solutions found by such effortsmay not be effective in real running systems. Simplified as-sumptions can be found in these studies, for example, statictransmission power, static transmission range, and static linkquality. These studies do not consider the spatial-temporalimpact on wireless communication. In this paper, we presentsystematic studies on these impacts. There are a number ofempirical studies on communication reality conducted withreal sensor devices [43] [40] [44] [4] [29] [20]. Their resultssuggest that for a specified transmission power and commu-nication distance, the received signal power varies and thelink quality is unstable. But they do not focus on a system-atic study on the radio and link dynamics in the context ofdifferent transmission power settings. Our extensive exper-iments with MICAz [8] confirm the observations presentedin previous work. We also go further and explore the radioand link dynamics when different transmission power levelsare applied. Our experimental results identify that link qual-ity changes differently according to spatial-temporal factorsin a real wireless sensor network. To address this issue, we

design a pairwise transmission power control. Our empiri-cal study also reveals that it is feasible to choose a minimaland environment-adapting transmission power level to savepower, while guaranteeing specified link quality at the sametime.

To achieve the optimal transmission power consumptionfor specified link qualities, we propose ATPC, an adaptivetransmission power control algorithm for wireless sensornetworks. The result of applying ATPC is that every nodeknows the proper transmission power level to use for each ofits neighbors, and every node maintains good link qualitieswith its neighbors by dynamically adjusting the transmis-sion power through on-demand feedback packets. Uniquely,ATPC adopts a feedback-based and pairwise transmissionpower control. By collecting the link quality history, ATPCbuilds a model for each neighbor of the node. This modelrepresents an in-situ correlation between transmission powerlevels and link qualities. With such a model, ATPC tunesthe transmission power according to monitored link qualitychanges. The changes of transmission power level reflectchanges in the surrounding environment. ATPC supportspacket-level transmission power control at runtime for MACand upper layer protocols. For example, routing protocolswith transmission power as a metric [33] [35] [12] [9] [5]can make use of ATPC by choosing the route with optimalpower consumption to forward packets.

The topic of transmission power control is not new, butour approach is quite unique. In state-of-art research, manytransmission power control solutions use a single transmis-sion power for the whole network, not making full use ofthe configurable transmission power provided by radio hard-ware to reduce energy consumption. We refer to this group asnetwork-level solutions, and typical examples in this groupare [27] [25] [2] [18] [31]. Also, some other work takes theconfigurable transmission powers into consideration. Theyeither assume that each node chooses a single transmissionpower for all the neighbors [2] [18] [19] [28] [37] [17][26] [30] [22], which we refer to as node-level solutions, ornodes use different transmission powers for different neigh-bors [23] [42] [3], which we call neighbor-level solutions.While these solutions provide a solid foundation for our re-search, ATPC goes further to support packet-level transmis-sion power control in a pairwise manner.

Also, most existing real wireless sensor network systemsuse a network-level transmission power for each node, suchas in [13] [38]. These coarse-level power controls lead tohigh energy consumption. The authors of [34] present a valu-able study about the impact of variable transmission poweron link quality. Through our empirical experiments withthe MICAz platform, it is observed that different transmis-sion powers are needed to achieve the same link quality overtime. This leads to our feedback-based transmission powercontrol design, which is not addressed in [34]. Also, the au-thors of [34] use a fixed number of transmission powers (13levels), which fixes the maximum accuracy for power tun-ing. The ATPC we propose chooses different transmissionpower levels based on the dynamics of link quality, and italso allows for better tuning accuracy and more energy sav-ings. Our approach essentially represents a good tradeoff

between accuracy and cost, a finer control at each node inexchange for less energy consumption when transmitting thepackets.

In this work, we invest a fair amount of effort to obtainempirical results from three different sites and over a rea-sonably long time period. These results give practical guid-ance to the overarching design of ATPC. We demonstratethat ATPC greatly extends the system lifetime by choosinga proper transmission power for each packet transmission,without jeopardizing the quality of data delivery. In our3-day experiment with 43 MICAz motes, ATPC achievesabove a 98% end-to-end Packet Reception Ratio in natu-ral environment through fair and rainy days. The solu-tions without online tuning can barely deliver half of pack-ets. Compared to other solutions, ATPC also significantlysaves transmission power. With equivalent communicationperformance, ATPC only consumes 53.6% of the transmis-sion energy of the maximum transmission power solutionand 78.8% of the transmission energy of the network-leveltransmission power solution. More specifically, the contri-butions of our work lie in two aspects.

• Our systematic study and experiments reveal the spa-tiotemporal impacts on wireless communication andidentify the relationship between dynamics of link qual-ity and transmission power control.

• With run-time pairwise transmission power control, weachieve high packet delivery ratio successfully withsmall energy consumption under realistic scenarios.

The rest of this paper is organized as follows: the motiva-tion of this work is presented in Section 2. In Section 3, thedesign of ATPC is stated. In Section 4, ATPC is evaluatedin real world experiments. The state of the art is analyzedin Section 5. In Section 6, conclusions are given and futurework is pointed out.

2 MotivationRadio communication quality between low power sen-

sor devices is affected by spatial and temporal factors. Thespatial factors include the surrounding environment, such asterrain and the distance between the transmitter and the re-ceiver. Temporal factors include surrounding environmen-tal changes in general, such as weather conditions. In thissection, we present experimental results for investigation ofthese impacts. We note that previous empirical studies oncommunication reality [43] [4] [44] [10] [29] [20] suggestthat for a specified transmission power, fixed communicationdistance, and antenna direction, the received signal powerand the link quality vary. But they do not focus on a sys-tematic study of the radio and link dynamics when differ-ent transmission powers are considered. We conducted thesemeasurements, and we are the first to study systematicallythe spatial and temporal impacts on the correlation betweentransmission power and Received Signal Strength Indicator(RSSI)/ Link Quality Indicator (LQI) [15]. Both RSSI andLQI are useful link metrics provided by CC2420 [7]. RSSIis a measurement of signal power which is averaged over 8symbol periods of each incoming packet. LQI is a measure-ment of the “chip error rate” [7] which is also implementedbased on samples of the error rate for the first eight symbols

StarDust: A Flexible Architecture for Passive Localizationin Wireless Sensor Networks ∗

Radu Stoleru, Pascal Vicaire, Tian He†, John A. StankovicDepartment of Computer Science, University of Virginia

†Department of Computer Science and Engineering, University of Minnesota

{stoleru, pv9f}@cs.virginia.edu, tianhe@cs.umn.edu, stankovic@cs.virginia.edu

ABSTRACTThe problem of localization in wireless sensor networks wherenodes do not use ranging hardware, remains a challeng-ing problem, when considering the required location accu-racy, energy expenditure and the duration of the localiza-tion phase. In this paper we propose a framework, calledStarDust, for wireless sensor network localization based onpassive optical components. In the StarDust framework,sensor nodes are equipped with optical retro-reflectors. Anaerial device projects light towards the deployed sensor net-work, and records an image of the reflected light. An imageprocessing algorithm is developed for obtaining the locationsof sensor nodes. For matching a node ID to a location wepropose a constraint-based label relaxation algorithm. Wepropose and develop localization techniques based on fourtypes of constraints: node color, neighbor information, de-ployment time for a node and deployment location for anode. We evaluate the performance of a localization systembased on our framework by localizing a network of 26 sen-sor nodes deployed in a 120 × 60ft2 area. The localizationaccuracy ranges from 2ft to 5ft while the localization timeranges from 10 milliseconds to 2 minutes.

Categories and Subject Descriptors: C.2.4 [ComputerCommunications Networks]: Distributed Systems; C.3 [Spe-cial Purpose and Application Based Systems]: Real-timeand embedded systems

General Terms: Algorithms, Measurement, Performance,Design, ExperimentationKeywords: Wireless Sensor Networks, Localization

1. INTRODUCTION

∗This work was supported by the DARPA IXO office (grantF336616-01-C-1905), by the ARO grant W911NF-06-1-0204and by the NSF grants CCR-0098269, CNS-0626616, CNS-0626614. We would like to thank the anonymous reviewersand our shepherd, Nirupama Bulusu, for their valuable feed-back.

Permission to make digital or hard copies of all or part of this work forpersonal or classroom use is granted without fee provided that copies arenot made or distributed for profit or commercial advantage and that copiesbear this notice and the full citation on the first page. To copy otherwise, torepublish, to post on servers or to redistribute to lists, requires prior specificpermission and/or a fee.SenSys’06,November 1–3, 2006, Boulder, Colorado, USA.Copyright 2006 ACM 1-59593-343-3/06/0011 ...$5.00.

Wireless Sensor Networks (WSN) have been envisionedto revolutionize the way humans perceive and interact withthe surrounding environment. One vision is to embed tinysensor devices in outdoor environments, by aerial deploy-ments from unmanned air vehicles. The sensor nodes forma network and collaborate (to compensate for the extremelyscarce resources available to each of them: computationalpower, memory size, communication capabilities) to accom-plish the mission. Through collaboration, redundancy andfault tolerance, the WSN is then able to achieve unprece-dented sensing capabilities.

A major step forward has been accomplished by devel-oping systems for several domains: military surveillance [1][2] [3], habitat monitoring [4] and structural monitoring [5].Even after these successes, several research problems remainopen. Among these open problems is sensor node localiza-tion, i.e., how to find the physical position of each sensornode. Despite the attention the localization problem inWSN has received, no universally acceptable solution hasbeen developed. There are several reasons for this. On onehand, localization schemes that use ranging are typicallyhigh end solutions. GPS ranging hardware consumes en-ergy, it is relatively expensive (if high accuracy is required)and poses form factor challenges that move us away fromthe vision of dust size sensor nodes. Ultrasound has a shortrange and is highly directional. Solutions that use the radiotransceiver for ranging either have not produced encouragingresults (if the received signal strength indicator is used) orare sensitive to environment (e.g., multipath). On the otherhand, localization schemes that only use the connectivityinformation for inferring location information are character-ized by low accuracies: ≈ 10ft in controlled environments,40− 50ft in realistic ones.

To address these challenges, we propose a framework forWSN localization, called StarDust, in which the complexityassociated with the node localization is completely removedfrom the sensor node. The basic principle of the frame-work is localization through passivity: each sensor nodeis equipped with a corner-cube retro-reflector and possi-bly an optical filter (a coloring device). An aerial vehicleprojects light onto the deployment area and records imagescontaining retro-reflected light beams (they appear as lu-minous spots). Through image processing techniques, thelocations of the retro-reflectors (i.e., sensor nodes) is deter-mined. For inferring the identity of the sensor node presentat a particular location, the StarDust framework develops aconstraint-based node ID relaxation algorithm.

The main contributions of our work are the following. We

propose a novel framework for node localization in WSNsthat is very promising and allows for many future exten-sions and more accurate results. We propose a constraint-based label relaxation algorithm for mapping node IDs tothe locations, and four constraints (node, connectivity, timeand space), which are building blocks for very accurate andvery fast localization systems. We develop a sensor nodehardware prototype, called a SensorBall. We evaluate theperformance of a localization system for which we obtainlocation accuracies of 2 − 5ft with a localization durationranging from 10 milliseconds to 2 minutes. We investigatethe range of a system built on our framework by consider-ing realities of physical phenomena that occurs during lightpropagation through the atmosphere.

The rest of the paper is structured as follows. Section 2is an overview of the state of art. The design of the Star-Dust framework is presented in Section 3. One implemen-tation and its performance evaluation are in Sections 4 and5, followed by a suite of system optimization techniques, inSection 6. In Section 7 we present our conclusions.

2. RELATED WORKWe present the prior work in localization in two major

categories: the range-based, and the range-free schemes.The range-based localization techniques have been de-

signed to use either more expensive hardware (and hencehigher accuracy) or just the radio transceiver. Ranging tech-niques dependent on hardware are the time-of-flight (ToF)and the time-difference-of-arrival (TDoA). Solutions thatuse the radio are based on the received signal strength indi-cator (RSSI) and more recently on radio interferometry.

The ToF localization technique that is most widely used isthe GPS. GPS is a costly solution for a high accuracy local-ization of a large scale sensor network. AHLoS [6] employsa TDoA ranging technique that requires extensive hardwareand solves relatively large nonlinear systems of equations.The Cricket location-support system (TDoA) [7] can achievea location granularity of tens of inches with highly direc-tional and short range ultrasound transceivers. In [2] thelocation of a sniper is determined in an urban terrain, byusing the TDoA between an acoustic wave and a radio bea-con. The PushPin project [8] uses the TDoA between ul-trasound pulses and light flashes for node localization. TheRADAR system [9] uses the RSSI to build a map of signalstrengths as emitted by a set of beacon nodes. A mobilenode is located by the best match, in the signal strengthspace, with a previously acquired signature. In MAL [10],a mobile node assists in measuring the distances (acting asconstraints) between nodes until a rigid graph is generated.The localization problem is formulated as an on-line state es-timation in a nonlinear dynamic system [11]. A cooperativeranging that attempts to achieve a global positioning fromdistributed local optimizations is proposed in [12]. A veryrecent, remarkable, localization technique is based on radiointerferometry, RIPS [13], which utilizes two transmitters tocreate an interfering signal. The frequencies of the emittersare very close to each other, thus the interfering signal willhave a low frequency envelope that can be easily measured.The ranging technique performs very well. The long timerequired for localization and multi-path environments posesignificant challenges.

Real environments create additional challenges for therange based localization schemes. These have been empha-

Figure 1: Stars, Planets or Smart Dust?

sized by several studies [14] [15] [16]. To address these chal-lenges, and others (hardware cost, the energy expenditure,the form factor, the small range, localization time), severalrange-free localization schemes have been proposed. Sensornodes use primarily connectivity information for inferringproximity to a set of anchors. In the Centroid localizationscheme [17], a sensor node localizes to the centroid of itsproximate beacon nodes. In APIT [18] each node decidesits position based on the possibility of being inside or out-side of a triangle formed by any three beacons within node’scommunication range. The Gradient algorithm [19], lever-ages the knowledge about the network density to infer theaverage one hop length. This, in turn, can be transformedinto distances to nodes with known locations. DV-Hop [20]uses the hop by hop propagation capability of the networkto forward distances to landmarks. More recently, severallocalization schemes that exploit the sensing capabilities ofsensor nodes, have been proposed. Spotlight [21] creates wellcontrolled (in time and space) events in the network whilethe sensor nodes detect and timestamp this events. Fromthe spatio-temporal knowledge for the created events andthe temporal information provided by sensor nodes, nodes’spatial information can be obtained. In a similar manner,the Lighthouse system [22] uses a parallel light beam, thatis emitted by an anchor which rotates with a certain period.A sensor node detects the light beam for a period of time,which is dependent on the distance between it and the lightemitting device.

Many of the above localization solutions target specificsets of requirements and are useful for specific applications.StarDust differs in that it addresses a particular demandingset of requirements that are not yet solved well. StarDust ismeant for localizing air dropped nodes where node passive-ness, high accuracy, low cost, small form factor and rapid lo-calization are all required. Many military applications havesuch requirements.

3. STARDUST SYSTEM DESIGNThe design of the StarDust system (and its name) was

inspired by the similarity between a deployed sensor net-work, in which sensor nodes indicate their presence by emit-ting light, and the Universe consisting of luminous and il-luminated objects: stars, galaxies, planets, etc. We depictthis similarity in Figure 1, which shows the duality betweenstars, galaxies and sensor nodes equipped with light emittingcapabilities.

The main difficulty when applying the above ideas to thereal world is the complexity of the hardware that needs tobe put on a sensor node so that the emitted light can bedetected from thousands of feet. The energy expenditure for

SIGF: A Family of Configurable, Secure Routing Protocols forWireless Sensor Networks

Anthony D. Wood Lei FangJohn A. Stankovic

Department of Computer ScienceUniversity of Virginia

{wood—leifang—stankovic}@cs.virginia.edu

Tian HeDept. of Computer Science and Engineering

University of Minnesotatianhe@cs.umn.edu

AbstractAs sensor networks are deployed in adversarial envi-ronments and used for critical applications such as bat-tlefield surveillance and medical monitoring, securityweaknesses become a big concern. The severe resourceconstraints of WSNs give rise to the need for resourcebound security solutions.

In this paper we present SIGF (Secure Implicit Ge-ographic Forwarding), a configurable secure routingprotocol family for wireless sensor networks that pro-vides “good enough” security and high performance.By avoiding or limiting shared state, the protocols pre-vent many common attacks against routing, and containothers to the local neighborhood.

SIGF makes explicit the tradeoff between securityprovided and state which must be stored and main-tained. It comprises three protocols, each forming abasis for the next: SIGF-0 keeps no state, but pro-vides probabilistic defenses; SIGF-1 uses local his-tory and reputation to avoid attackers; and SIGF-2 usesneighborhood-shared state to provide stronger securityguarantees.

Our performance evaluation shows that SIGF achieveshigh packet delivery ratios with low overhead and end-to-end delay. We evaluate the security of SIGF pro-tocols under various security attacks and show thatit effectively contains the damage from compromisednodes and defends against black hole, selective for-warding, Sybil, and some denial of service attacks.

Permission to make digital or hard copies of all or part of this work forpersonal or classroom use is granted without fee provided that copies arenot made or distributed for profit or commercial advantage and that copiesbear this notice and the full citation on the first page. To copy otherwise, torepublish, to post on servers or to redistribute to lists, requires prior specificpermission and/or a fee.In submission to SASN ’06Copyright c© 2006 ACM . . . $5.00.

1. IntroductionSecurity is critical for many wireless sensor networkapplications such as battlefield surveillance, medicalmonitoring, and emergency response. However, manysecurity mechanisms developed for the Internet or ad-hoc networks cannot be applied directly to wirelesssensor networks (WSNs) due to their limited resourcesin computation, memory, communication bandwidth,and energy.

The severe resource constraints of WSNs give rise tothe need for resource bound security solutions. Thereare at least two interesting aspects of this concept.First, individual security mechanisms must be efficientin memory, computation, energy and bandwidth. Forexample, certain cryptographic schemes are inappro-priate because ciphertext message expansion results incostly memory, bandwidth and energy use. Second, theresource consumption of all security mechanisms in-stalled together at a node must not exceed the amountof resources allocated for security and they cannot de-grade performance to an unacceptable level during nor-mal operation nor when an attack is underway.

It is not possible in today’s state of the art to includestrong security mechanisms for each of the services at anode such as medium access control, routing, localiza-tion, time synchronization, power management, sens-ing, and group management. Consequently, even if asecure (to a wide variety of attacks) routing protocolis implemented, it may suffer from low efficiency andwould not protect against attacks on the other services.

Our approach for resource bound security is to haveminimal active security protection. This results in veryhigh performance and minimal resource consumptionwhen no attacks are underway. Then upon detectingan attack or if the system designers expect increasedthreats, the appropriate security mechanism is acti-vated. The result is not 100% security protection—butgood enough security, activated at the right time. This

general approach makes it possible to have high per-formance and to react to current security attacks, andis even more evolvable to new attacks than approachesthat fix a set of solutions into a node.

In this paper we present Secure Implicit GeographicForwarding (SIGF), a family of configurable securerouting protocols that follow this general approach. Fora complete WSN solution similar families of protocolswould be required for each of the other services.

SIGF is based on IGF [1], a nondeterministic Net-work/MAC hybrid routing protocol that is completelystateless. This allows it to handle network dynamics ef-fortlessly, and intrinsically limits the effects of a com-promised node to a local area. There are no routing ta-bles to corrupt, since forwarding decisions are madeas late as possible—when a packet is ready to transmit.Nevertheless, it is susceptible in the local neighborhoodto a simple CTS rushing attack.

SIGF comprises three protocols which extend IGFand populate the gap between pure statelessness andtraditional shared-state security. SIGF-0 keeps no state,but uses nondeterminism and candidate sampling toachieve high packet delivery ratios probabilistically.SIGF-1 keeps local state, building reputations for itsneighbors to aid in next-hop selection. SIGF-2 usesstate shared with neighbors to provide the strongest de-fense against attack, yet at the greatest cost. Each proto-col encompasses the features of the previous, layeringadditional mechanisms to defend against more sophis-ticated attacks. The layered family of protocols allowsa network to activate only the protections currently nec-essary, and to change to stronger ones only if they arewarranted.

We evaluate the performance of each protocol bysimulating with no attacks, and with black hole, selec-tive forwarding, Sybil, and denial of service attacks.We show that each protocol represents a tradeoff be-tween state and security, and that despite keeping nostate, SIGF-0 performs well.

We make several contributions in this work. First,we show that even with no security countermeasures,the base protocol IGF has desirable attack contain-ment properties, but nevertheless falls to several at-tacks that completely disrupt communication in a localneighborhood. Then we present the design and evalu-ation of SIGF, a secure routing protocol family builtupon IGF. We show that the stateless SIGF-0 protocolmaintains 45% packet delivery ratio (PDR) under blackhole attack from a single node, and that the reputation-based SIGF-1 achieves 83% PDR under Sybil attack.To the best of our knowledge, SIGF is the first config-urable routing protocol for WSNs that makes explicitthe tradeoffs between resources required and security

provided, and enables resource bound security that isboth efficient and effective.

It is possible that some WSNs require much strongersecurity than what our dynamic approach offers. How-ever, no perfect solution exists—nor is likely to exist onseverely resource constrained devices. Our approach,as exemplified by the SIGF routing protocols presentedin this paper, can significantly improve security (notmake it absolute), allow operation in the presence ofattacks, and support a high performance system.

The rest of this paper is organized as follows. Inthe next section we briefly review our foundationalrouting protocol, IGF. Then Section 3 describes sys-tem assumptions and routing attacks on IGF. Section 4presents SIGF, our secure routing protocol family. InSection 5 we present our experiments and a detailedevaluation of the protocols under various attacks. Fi-nally, we discuss related work in Section 6 and thenconclude.

2. Implicit Geographic ForwardingOur foundational routing protocol is Implicit Geo-graphic Forwarding (IGF), which is completely state-less, without dependence on knowledge of the net-work topology or the presence/absence of any othernodes [1]. It makes nondeterministic routing decisions,implicitly allowing receiving nodes to determine apacket’s next-hop at transmission time. IGF couplesthe routing and MAC components into a single inte-grated Network/MAC protocol. It identifies the bestforwarding candidate during MAC-layer handshakingat the instant a packet is sent.

O D60S

Figure 1. Forwarding area for message sender S.

Figure 1 presents an example topology, where sourcenode S transmits a message toward D. Routing is inte-grated with the RTS/CTS hand-shake of MACA/802.11[9] DCF MAC protocols. Overhead from the small con-trol messages is acceptable for the stateless benefitsthey provide, and are negligible in WSNs that carrymoderate or large packet loads of aggregated data.

The communication handshake for this exampletopology is shown in Figure 2. It begins when thesender S’s NAV timer is zero and it carrier senses an

2

International Conference on Distributed Computing Systems ICDCS 2003

SPEED: A Stateless Protocol for Real-Time Communication

in Sensor Networks

Tian Hea John A Stankovica Chenyang Lub Tarek Abdelzahera

a Department of Computer Science b Department of Computer Science & Engineering University of Virginia Washington University in St Louis {tianhe, stankovic, zaher}@cs.virginia.edu lu@cs.wustl.edu

Abstract In this paper, we present a real-time communication

protocol for sensor networks, called SPEED. The protocol provides three types of real-time communication services, namely, real-time unicast, real-time area-multicast and real-time area-anycast. SPEED is specifically tailored to be a stateless, localized algorithm with minimal control over-head. End-to-end soft real-time communication is achieved by maintaining a desired delivery speed across the sensor network through a novel combination of feedback control and non-deterministic geographic forwarding. SPEED is a highly efficient and scalable protocol for sensor networks where the resources of each node are scarce. Theoretical analysis, simulation experiments and a real implementation on Berkeley motes are provided to validate our claims.

1. Introduction Many exciting results have been recently developed for

large-scale sensor networks. These networks can form the basis for many types of smart environments such as smart hospitals, battlefields, earthquake response systems, and learning environments. While these potential applications remain diverse, one commonality they all share is the need for an efficient and robust routing protocol.

The main function of sensor networks is data delivery. We distinguish three types of communication patterns asso-ciated with the delivery of data in such networks. First, it is often the case that one part of a network detects some activ-ity that needs to be reported to a remote base station. This type of communication is called unicast. Alternatively, a base station may issue a command or query to an area in the sensor networks. For example, it may ask all sensors in the region of a damaged nuclear plant to report radiation read-ings, or command all lights in a given area to turn on. This type of communication motivates a different routing service where one end-point of the route may be an area rather than an individual node. We call this area-multicast. Finally, since sensors often measure highly redundant information, in some situations it may be sufficient to have any node in an area respond. We call a routing service that provides such capability, area-anycast. SPEED provides the afore-

mentioned three types of communication services. Since sensor networks deal with real world, it is often

necessary for communication to meet real-time constraints. In surveillance systems, for example, communication de-lays within sensing and actuating loops directly affect the quality of tracking. To date, few results exist for sensor networks that adequately address real-time requirements. In this paper we develop a protocol SPEED that supports soft real-time communication based on feedback control and stateless algorithms for large-scale sensor networks. We evaluate SPEED via simulation using GloMoSim [15] and compare it to five other ad hoc routing protocols: DSR [5], AODV [10], GF [13] and two scaled down versions of SPEED. The performance results show that SPEED 1) re-duces the number of packets that miss their end-to-end deadlines, 2) reacts to transient congestion in the most sta-ble manner, and 3) efficiently handles voids [6] with mini-mal control overhead. We also implement SPEED on the Berkeley motes [4]. The results show that SPEED helps balance the traffic load to increase the system lifetime.

2. State of the Art Several routing protocols have been developed for ad

hoc wireless networks. Sensor networks can be regarded as a sub-category of such networks, but with a number of dif-ferent requirements.

In sensor networks, location is more important than a specific node’s ID. For example, tracking applications only care where a target is located, not the ID of the reporting node. In sensor networks, such location-awareness is neces-sary to make the sensor data meaningful. Therefore, it is natural to utilize location-aware routing. A set of location based routing algorithms have been proposed. Finn [2] pro-posed a greedy geographic forwarding protocol with limited flooding to circumvent the voids inside the network. GPSR [6] by Karp and Kung use perimeter forwarding to get around voids. Geographic distance routing (GEDIR) [13] guarantees loop-free delivery in a collision-free network. LAR [7] by Young-Bae Ko improves the efficiency of the on-demand routing algorithms by restricting routing packet flooding in a specified “request zone.”

SPEED also utilizes geographic location to make local-ized routing decisions. The difference is that SPEED is de-signed to handle congestion and provide a soft real-time communication service, which are not the main goals of previous location-based routing protocols. Moreover, SPEED provides an alternative solution to handle voids other than approaches based on planar graph traversal [6] and limited flooding [2].

Several real-time protocols have been proposed for sensor networks. SWAN [1] uses feedback information from the MAC layer to regulate the transmission rate of non-real-time TCP traffic in order to sustain real-time UDP traffic. RAP [9] uses velocity monotonic scheduling to pri-oritize real-time traffic and enforces such prioritization through a differentiated MAC Layer. Woo and Culler [14] proposed an adaptive MAC layer rate control to achieve fairness among nodes with different distances to the base station. All of these algorithms work well by locally de-grading a certain portion of the traffic. However, this kind of local MAC layer adaptation cannot handle long-term congestion where routing assistance is necessary to divert traffic away from any hotspot. SPEED provides a combina-tion of MAC layer and network layer adaptation that effec-tively deals with such issues. To the best of our knowledge, no routing algorithm has been specifically designed to pro-vide soft real-time guarantees for sensor networks.

Reactive routing algorithms such as AODV [10] and DSR [5] maintain routing information for a small subset of possible destinations, namely those currently in use. If no route is available for a new destination, a route discovery process is invoked. Route discovery broadcasts can lead to significant delays in a sensor network with a large network diameter. This limitation makes on-demand algorithms less suitable for real-time applications.

3. Design Goals Our design is inspired by the observation that unlike

wired networks, where the delay is independent of the physical distance between the source and destination, in multi-hop wireless sensor networks, the end-to-end delay depends on not only single hop delay, but also on the dis-tance a packet travels. In view of this, the key design goal of the SPEED algorithm is to support a soft real-time com-munication service with a desired delivery speed across the sensor network, so that end-to-end delay is proportional to the distance between the source and destination. It should be noted that delivery speed refers to the approaching rate along a straight line from the source toward the destination. Unless the packet is routed exactly along that straight line, delivery speed is smaller than the actual speed of the packet in the network. For example, if the packet is routed in the opposite direction from the destination, its speed is negative. Our algorithm ensures that this condition never occurs.

Upon this soft real-time delivery service, SPEED pro-vides three types of real-time communication services,

namely, real-time unicast, real-time area-multicast and real-time area-anycast, for sensor networks. In doing so, SPEED satisfies the following design objectives.

1. Stateless Architecture. The physical limitations of sen-

sor networks, such as large scale, high failure rate, and constrained memory capacity necessitate a stateless ap-proach. SPEED only maintains immediate neighbor in-formation. It doesn’t require a routing table as in DSDV [11] nor per-destination states as in AODV [10]. Thus, its memory requirements are minimal.

2. Soft Real-Time. Sensor networks are commonly used to monitor and control the physical world. SPEED provides a uniform delivery speed across the sensor network to meet the requirement of real-time applications such as disaster and emergency surveillance in sensor networks.

3. Minimum MAC Layer Support. SPEED doesn’t re-quire real-time or QoS aware MAC support. The feed-back control scheme employed in SPEED allows it to be compatible with all existing best effort MAC layers.

4. QoS Routing and Congestion Management. Most re-active routing protocols can find routes that avoid net-work hot spots during the route acquisition phase. Such protocols work well when traffic patterns don’t fluctuate during a session. However, these protocols (e.g. [5]) are less successful when congestion patterns change rapidly compared to the session lifetime. When a route becomes congested, such protocols either suffer a delay or initiate another round of route discovery. As a solution, SPEED uses a novel backpressure re-routing scheme to re-route packets around large-delay links with minimum control overhead.

5. Traffic Load Balancing. In sensor networks, the band-width and energy are scarce resources compared to a wired network. Because of this, it is valuable to utilize several simultaneous paths to carry packets from the source to the destination. SPEED uses non-deterministic forwarding to balance each flow among multiple concur-rent routes.

6. Localized Behavior. Pure localized algorithms are those in which any action invoked by a node should not affect the system as a whole. In algorithms such as AODV, DSR and TORA, this is not the case. In these protocols a node uses flooding to discover new paths. In sensor net-works where thousands of nodes communicate with each other, broadcast storms may result in significant power consumption and possibly a network meltdown. To avoid that, all distributed operations in SPEED are localized to achieve high scalability.

7. Void Avoidance. In some scenarios, pure greedy geo-graphic forwarding may fail to find a greedy path to the destination, even when one actually exists. SPEED han-dles the void the same way as it handles congested areas and guarantees that if there is a greedy route between the source and destination, it will discover it.

IEEE TRANSACTIONS ON PARALLEL AND DISTRIBUTED SYSTEMS, TPDS-0195-1003.R2 1

A Spatiotemporal Communication Protocol for Wireless Sensor Networks

Tian He, Member, IEEE, John A. Stankovic, Fellow, IEEE, Chenyang Lu, Member, IEEE and Tarek F. Abdelzaher, Member, IEEE

Abstract— In this paper, we present a spatiotemporal communication protocol for sensor networks, called SPEED. SPEED is specifically tailored to be a localized algorithm with minimal control overhead. End-to-end soft real-time communication is achieved by maintaining a desired delivery speed across the sensor network through a novel combination of feedback control and non-deterministic geographic forward-ing. SPEED is a highly efficient and scalable protocol for sensor networks where the resources of each node are scarce. Theoretical analysis, simulation experiments and a real implementation on Berkeley motes are provided to validate the claims.

Index Terms— Wireless sensor networks, routing, real-time, spatiotemporal, networking.

—————————— � ——————————

1 INTRODUCTION

ANY sensor network applications, such as battle-field surveillance and earthquake response systems, are designed to interact with fast changing events in

the real world. It is often necessary for the underlying communication infrastructure to meet real-time constraints [9][16][4]. In surveillance systems [8], for example, com-munication delays within sensing and actuating loops di-rectly affect the quality of tracking. Our spatiotemporal communication protocol, called SPEED, is inspired by fol-lowing observation: In wired networks, the end-to-end de-lay is independent of the physical distance between the source and destination. While in multi-hop wireless sensor networks, since communication is physically bounded, the end-to-end delay depends not only on single hop delay (time-constraints), but also on the distance a packet travels (spatial-constraints). In view of this, the key design goal of this work is to support a spatiotemporal communication service with a desired delivery speed across the sensor net-work, so that end-to-end delay is proportional to the dis-tance between the source and destination. We deem this service as one type of soft real-time communication, be-cause it achieves a predicable end-to-end communication delay under given distance (spatial) constraints. The key contribution of SPEED is achieving the spatio-temporal requirements through a novel combination of a time-aware feedback control mechanism and a spatial-aware non-deterministic geographic forwarding scheme. We evaluate SPEED using GloMoSim [22]. The perform-ance results show that SPEED 1) reduces the number of

packets that miss their end-to-end deadlines, 2) reacts to transient congestion in the most stable manner, and 3) effi-ciently handles voids [12] with minimal control overhead. To demonstrate its applicability, we also implement SPEED on the Berkeley motes [3]. The results show that SPEED helps balance the traffic load to increase the system lifetime.

2 STATE OF THE ART Several routing protocols have been developed for ad hoc wireless networks. Sensor networks can be regarded as a sub-category of such networks, but with a number of dif-ferent requirements. In sensor networks, location is more important than a spe-cific node’s ID. For example, tracking applications only care where a target is located, not the ID of the reporting node. In sensor networks, such spatial-awareness [7] is necessary to make the sensor data meaningful. Therefore, it is natural to utilize spatial-aware routing. A set of location based routing algorithms have been proposed. Finn [5] proposed a greedy geographic forwarding protocol with limited flooding to circumvent the voids inside the net-work. GPSR [12] by Karp and Kung use perimeter for-warding to get around voids. Similarly, GOAFR [14] com-bines the greedy routing with the adaptive face routing to provide an asymptotically optimal path to the destination. Geographic distance routing (GEDIR) [20] guarantees loop-free delivery in a collision-free network. LAR [13] by Young-Bae Ko improves the efficiency of the on-demand routing algorithms by restricting packet flooding in a speci-fied “request zone”. Basagni, et. al. propose a distance routing algorithm [2] for mobility (DREAM), in which each node periodically updates its location information to other nodes. An updating rate is set according to a distance effect in order to reduce the number of control packets. Recently, Huang [4] et. al. proposed Mobicast protocol extended geo-cast by providing just-in-time information dissemination to nodes in a mobile delivery zone.

0195-1003/05/$20.00 © 2005 IEEE

————————————————

• Tian He is with Department of Computer Science, University of Vir-ginia,Charlottesville, 22904. E-mail: tianhe@cs.virginia.edu

• John A. Stankovic is with Department of Computer Science, University of Virginia, Charlottesville, 22904. E-mail: stankovic@cs.virginia.edu

• Chenyang Lu is with Department of Computer Science & Engineering, Washington University in St Louis. E-mail: lu@cs.wustl.edu

• Tarek F. Adelzaher is with Department of Computer Science, University of Virginia, Charlottesville, 22904. E-mail: zaher@cs.virginia.edu

Manuscript received Oct.29, 2003;, revised Sept 2,2004; accepted Jan 6,2005

M

2 IEEE TRANSACTIONS ON PARALLEL AND DISTRIBUTED SYSTEMS, TPDS-0195-1003.R2

SPEED also utilizes geographic location to make localized routing decisions. The key difference is that SPEED takes timely delivery into account and is designed to be the first spatiotemporal-aware communication protocol for sensor networks. Moreover, SPEED provides an alternative solu-tion to handle voids other than approaches based on planar graph traversal [12] [14] and limited flooding [5]. Reactive routing algorithms such as AODV [17] and DSR [10] maintain routing information for a small subset of pos-sible destinations, namely those currently in use. If no route is available for a new destination, a route discovery process is invoked. Route discovery broadcasts can lead to significant delays in a sensor network with a large network diameter. This limitation makes on-demand algorithms less suitable for real-time applications. Several real-time protocols have been proposed for ad hoc and sensor networks. SWAN [1] uses feedback information from the MAC layer to regulate the transmission rate of non-real-time TCP traffic in order to sustain real-time UDP traffic. RAP [16] uses velocity monotonic scheduling to prioritize real-time traffic and enforces such prioritization through a differentiated MAC Layer. V. Kanodia etc. [11] proposed a service differentiation for delay-sensitive traffic by prioritizing 802.11. Woo and Culler [21] proposed an adaptive MAC layer rate control to achieve fairness among nodes with different distances to the base station. All of these algorithms work well by locally degrading a certain portion of the traffic. However, this kind of local MAC layer adaptation cannot handle long-term congestion where routing assistance is necessary to divert traffic away from any hotspot. SPEED provides a combination of MAC layer and network layer adaptation that effectively deals with such issues. To the best of our knowledge, no routing algo-rithm has been specifically designed to provide soft real-time guarantees under spatiotemporal constraints for sen-sor networks.

3 DESIGN GOALS The key design goal of the SPEED algorithm is to support a spatiotemporal communication service with a desired de-livery speed across the sensor network, so that end-to-end delay is proportional to the distance between the source and destination. It should be noted that delivery speed refers to the approaching rate along a straight line from the source toward the destination. Unless the packet is routed exactly along that straight line, delivery speed is smaller than the actual speed of the packet in the network. For ex-ample, if the packet is routed in the opposite direction from the destination, its speed is negative. Our algorithm en-sures that this condition never occurs. More specifically, SPEED satisfies the following design objectives. Soft Real-Time: We define the soft real-time guarantee provided by SPEED as delay guarantee per unit delivery distance (speed guarantee). Under this guarantee, we can obtain a predictable end-to-end communication delay un-der given spatial (distance) constraints before hand. Con-sequently, applications can make admission control to de-liver packets that are able to meet end-to-end deadlines.

Minimal State Architecture: The physical limitations of sensor networks, such as large scale, high failure rate, and constrained memory capacity necessitate a minimal state approach. SPEED only maintains immediate neighbor in-formation. It does not require a routing table as in DSDV [18] nor per-destination states as in AODV [17]. Thus, its memory requirements are minimal. Minimum MAC Layer Support: SPEED does not require real-time MAC support. The feedback control scheme em-ployed in SPEED allows all existing best effort MAC layers. QoS Routing and Congestion Management: Most reactive routing protocols can find routes that avoid network hot spots during the route acquisition phase. Such protocols work well when traffic patterns do not fluctuate during a session. However, these protocols (e.g. [10]) are less suc-cessful when congestion patterns change rapidly compared to the session lifetime. When a route becomes congested, such protocols either suffer a delay or initiate another round of route discovery. As a solution, SPEED uses a novel backpressure re-routing scheme to re-route packets around large-delay links with minimum control overhead. Traffic Load Balancing: In sensor networks, the bandwidth and energy are scarce resources compared to a wired net-work. Because of this, it is valuable to utilize several simul-taneous paths to carry packets from the source to the desti-nation. SPEED uses non-deterministic forwarding to bal-ance each flow among multiple concurrent routes. Localized Behavior: Pure localized algorithms are those in which any action invoked by a node should not affect the system as a whole. In algorithms such as AODV, DSR and TORA, this is not the case. In these protocols, a node uses flooding to discover new paths. In sensor networks where thousands of nodes communicate with each other, broad-cast storms may result in significant power consumption and possibly a network meltdown. To avoid that, all dis-tributed operations in SPEED are localized to achieve high scalability. Void Avoidance: In some scenarios, pure greedy geo-graphic forwarding may fail to find a greedy path to the destination, even when one actually exists. SPEED handles the void the same way as it handles congested areas and guarantees that if there is a greedy route between the source and destination, it will discover it. Note, while SPEED does not use routing tables, SPEED does utilize location information to carry out routing. Thus, we assume that each node is location-aware [7].

4 SPEED PROTOCOL SPEED maintains a desired delivery speed across sensor networks with a two-tier adaptation included for diverting traffic at the networking layer and locally regulating pack-ets sent to the MAC layer. It consists of the following com-ponents:

• An API • A delay estimation scheme • A neighbor beacon exchange scheme

• A Non-deterministic Geographic Forwarding algo-rithm (NGF)

A High-Accuracy, Low-Cost Localization System for Wireless Sensor Networks

Radu Stoleru, Tian He, John A. Stankovic, David Luebke Department of Computer Science

University of Virginia, Charlottesville, VA 22903 {stoleru, tianhe, stankovic, luebke}@cs.virginia.edu

ABSTRACT The problem of localization of wireless sensor nodes has long been regarded as very difficult to solve, when considering the realities of real world environments. In this paper, we formally describe, design, implement and evaluate a novel localization system, called Spotlight. Our system uses the spatio-temporal properties of well controlled events in the network (e.g., light), to obtain the locations of sensor nodes. We demonstrate that a high accuracy in localization can be achieved without the aid of expensive hardware on the sensor nodes, as required by other localization systems. We evaluate the performance of our system in deployments of Mica2 and XSM motes. Through performance evaluations of a real system deployed outdoors, we obtain a 20cm localization error. A sensor network, with any number of nodes, deployed in a 2500m2 area, can be localized in under 10 minutes, using a device that costs less than $1000. To the best of our knowledge, this is the first report of a sub-meter localization error, obtained in an outdoor environment, without equipping the wireless sensor nodes with specialized ranging hardware.

Categories and Subject Descriptors C.2.4 [Computer-Communications Networks]: Distributed Systems; C.3 [Special-Purpose and Application-Based Systems]: Real-Time and embedded systems.

General Terms Algorithms, Measurement, Performance, Design, Experimentation

Keywords Wireless Sensor Network, Localization, Event Distribution, Laser

1. INTRODUCTION Recently, wireless sensor network systems have been used in many promising applications including military surveillance, habitat monitoring, wildlife tracking etc. [12] [22] [33] [36]. While many middleware services, to support these applications, have been designed and implemented successfully, localization - finding the position of sensor nodes - remains one of the most difficult research challenges to be solved practically. Since most emerging applications based on networked sensor nodes require location awareness to assist their operations, such as annotating sensed data with location context, it is an indispensable requirement for a sensor node to be able to find its own location.

Many approaches have been proposed in the literature [4] [6] [13] [14] [19] [20] [21] [23] [27] [28], however it is still not clear how these solutions can be practically and economically deployed. An on-board GPS [23] is a typical high-end solution, which requires sophisticated hardware to achieve high resolution time synchronization with satellites. The constraints on power and cost for tiny sensor nodes preclude this as a viable solution. Other solutions require per node devices that can perform ranging among neighboring nodes. The difficulties of these approaches are two-fold. First, under constraints of form factor and power supply, the effective ranges of such devices are very limited. For example the effective range of the ultrasonic transducers used in the Cricket system is less than 2 meters when the sender and receiver are not facing each other [26]. Second, since most sensor nodes are static, i.e. the location is not expected to change, it is not cost-effective to equip these sensors with special circuitry just for a one-time localization. To overcome these limitations, many range-free localization schemes have been proposed. Most of these schemes estimate the location of sensor nodes by exploiting the radio connectivity information among neighboring nodes. These approaches eliminate the need of high-cost specialized hardware, at the cost of a less accurate localization. In addition, the radio propagation characteristics vary over time and are environment dependent, thus imposing high calibration costs for the range-free localizations schemes. With such limitations in mind, this paper addresses the following research challenge: How to reconcile the need for high accuracy in location estimation with the cost to achieve it. Our answer to this challenge is a localization system called Spotlight. This system employs an asymmetric architecture, in which sensor nodes do not need any additional hardware, other than what they currently have. All the sophisticated hardware and computation reside on a single Spotlight device. The Spotlight device uses a steerable laser light source, illuminating the sensor nodes placed within a known terrain. We demonstrate that this localization is much more accurate (i.e., tens of centimeters) than the range-based localization schemes and that it has a much longer effective range (i.e., thousands of meters) than the solutions based on ultra-sound/acoustic ranging. At the same time, since only a single sophisticated device is needed to localize the whole network, the amortized cost is much smaller than the cost to add hardware components to the individual sensors.

2. RELATED WORK In this section, we discuss prior work in localization in two major categories: the range-based localization schemes (which use either expensive, per node, ranging devices for high accuracy, or less accurate ranging solutions, as the Received Signal Strength Indicator (RSSI)), and the range-free schemes, which use only connectivity information (hop-by-hop) as an indication of proximity among the nodes.

Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. SenSys’05, November 2–4, 2005, San Diego, California, USA. Copyright 2005 ACM 1-59593-054-X/05/0011...$5.00.

The localization problem is a fundamental research problem in many domains. In the field of robotics, it has been studied extensively [9] [10]. The reported localization errors are on the order of tens of centimeters, when using specialized ranging hardware, i.e. laser range finder or ultrasound. Due to the high cost and non-negligible form factor of the ranging hardware, these solutions can not be simply applied to sensor networks.

The RSSI has been an attractive solution for estimating the distance between the sender and the receiver. The RADAR system [2] uses the RSSI to build a centralized repository of signal strengths at various positions with respect to a set of beacon nodes. The location of a mobile user is estimated within a few meters. In a similar approach, MoteTrack [17] distributes the reference RSSI values to the beacon nodes.

Solutions that use RSSI and do not require beacon nodes have also been proposed [5] [14] [24] [26] [29]. They all share the idea of using a mobile beacon. The sensor nodes that receive the beacons, apply different algorithms for inferring their location. In [29], Sichitiu proposes a solution in which the nodes that receive the beacon construct, based on the RSSI value, a constraint on their position estimate. In [26], Priyantha et al. propose MAL, a localization method in which a mobile node (moving strategically) assists in measuring distances between node pairs, until the constraints on distances generate a rigid graph. In [24], Pathirana et al. formulate the localization problem as an on-line estimation in a nonlinear dynamic system and proposes a Robust Extended Kalman Filter for solving it. Elnahrawy [8] provides strong evidence of inherent limitations of localization accuracy using RSSI, in indoor environments.

A more precise ranging technique uses the time difference between a radio signal and an acoustic wave, to obtain pair wise distances between sensor nodes. This approach produces smaller localization errors, at the cost of additional hardware. The Cricket location-support system [25] can achieve a location granularity of tens of centimeters with short range ultrasound transceivers. AHLoS, proposed by Savvides et al. [27], employs Time of Arrival (ToA) ranging techniques that require extensive hardware and solving relatively large nonlinear systems of equations. A similar ToA technique is employed in [3].

In [30], Simon et al. implement a distributed system (using acoustic ranging) which locates a sniper in an urban terrain. Acoustic ranging for localization is also used by Kwon et al. [15]. The reported errors in localization vary from 2.2m to 9.5m, depending on the type (centralized vs. distributed) of the Least Square Scaling algorithm used.

For wireless sensor networks ranging is a difficult option. The hardware cost, the energy expenditure, the form factor, the small range, all are difficult compromises, and it is hard to envision cheap, unreliable and resource-constraint devices make use of range-based localization solutions. However, the high localization accuracy, achievable by these schemes is very desirable.

To overcome the challenges posed by the range-based localization schemes, when applied to sensor networks, a different approach has been proposed and evaluated in the past. This approach is called range-free and it attempts to obtain location information from the proximity to a set of known beacon nodes.

Bulusu et al. propose in [4] a localization scheme, called Centroid, in which each node localizes itself to the centroid of its proximate beacon nodes. In [13], He et al. propose APIT, a scheme in which each node decides its position based on the possibility of being inside or outside of a triangle formed by any three beacon nodes heard by the node. The Global Coordinate System [20],

developed at MIT, uses apriori knowledge of the node density in the network, to estimate the average hop distance. The DV-* family of localization schemes [21], uses the hop count from known beacon nodes to the nodes in the network to infer the distance. The majority of range-free localization schemes have been evaluated in simulations, or controlled environments. Several studies [11] [32] [34] have emphasized the challenges that real environments pose. Langendoen and Reijers present a detailed, comparative study of several localization schemes in [16].

To the best of our knowledge, Spotlight is the first range-free localization scheme that works very well in an outdoor environment. Our system requires a line of sight between a single device and the sensor nodes, and the map of the terrain where the sensor field is located. The Spotlight system has a long effective range (1000’s meters) and does not require any infrastructure or additional hardware for sensor nodes. The Spotlight system combines the advantages and does not suffer from the disadvantages of the two localization classes.

3. SPOTLIGHT SYSTEM DESIGN The main idea of the Spotlight localization system is to generate controlled events in the field where the sensor nodes were deployed. An event could be, for example, the presence of light in an area. Using the time when an event is perceived by a sensor node and the spatio-temporal properties of the generated events, spatial information (i.e. location) regarding the sensor node can be inferred.

Figure 1. Localization of a sensor network using the

Spotlight system We envision, and depict in Figure 1, a sensor network

deployment and localization scenario as follows: wireless sensor nodes are randomly deployed from an unmanned aerial vehicle. After deployment, the sensor nodes self-organize into a network and execute a time-synchronization protocol. An aerial vehicle (e.g. helicopter), equipped with a device, called Spotlight, flies over the network and generates light events. The sensor nodes detect the events and report back to the Spotlight device, through a base station, the timestamps when the events were detected. The Spotlight device computes the location of the sensor nodes.

During the design of our Spotlight system, we made the following assumptions: - the sensor network to be localized is connected and a

middleware, able to forward data from the sensor nodes to the Spotlight device, is present.

- the aerial vehicle has a very good knowledge about its position and orientation (6 parameters: 3 translation and 3 rigid-body rotation) and that is possesses the map of the field where the network was deployed.

- a powerful Spotlight device is available and it is able to generate

Spotlight: Low-cost Asymmetric Localization System for Networked Sensor Nodes

University of Virginia

Tian He, Radu Stoleru, John A. Stankovic

{ Tianhe,rs6bd,Stankovic}@cs.virginia.edu Many middleware services for sensor network applications have been developed successfully. Localization - finding the position of sensor nodes - remains one of the most difficult research challenges to be solved economically and practically. Our answer to this challenge is a localization system called Spotlight. This system employs an asymmetric architecture, in which field motes do not need any additional hardware than what they currently have. All sophisticated hardware and computation reside on a single Spotlight device. We demonstrate that this localization is much more accurate than the range-based localization schemes and that it has a much longer effective range than the solutions based on ultrasound/acoustic ranging. Meanwhile, since only a single sophisticated device is needed to localize the whole network, the amortized cost will be much smaller than the cost to add hardware components to individual sensors. The general idea of the Spotlight technique is to use a sophisticated spotlight device to generate controlled events that assist the localization of sensor nodes. An event could be, for example, the presence of light in an area. Using the time when an event is perceived by a sensor node the sensor node can be inferred. We envision, and depict in Figure 1, a sensor network deployment and localization scenario as follows: wireless sensor nodes are randomly deployed from a helicopter. A device (e.g. helicopter) flies over the network and generates light events over predefine traces. The sensor nodes detect the events and report to a base station the times when the events were detected. The base station computes the location of the sensor nodes.

Figure 1: Localization of a sensor network using the Spotlight system

Formally, suppose that space A contains all sensor nodes Ni, and each node Ni is located at Pi(x, y, z). To determine Pi(x, y, z), a spotlight localization system needs to support three major functions as shown in Figure 2, namely a Detection Function D(e), an Event Distribution Function E(t) and a Localization Function L(t). The Detection Function D(e) is supported by the sensor nodes. It is a sensing module to determine whether an external event happens or not. The Localization Function L(t) and the Event Distribution Function E(t) are supported by a spotlight device. The Localization Function is an aggregation algorithm, which calculates the intersection of multiple sets of points. The Event Distribution Function E(t) describes the locations of the

events (e.g. light spots) within the space A over a certain period of time. We implement multiple realizations of the E(t) function including Point Scan, Line Scan and Space Cover. Since the E(t) function is only realized at the spotlight device, its complexity is transparent to sensor nodes.

Figure 2: Spotlight system architecture

With the support of the three functions, the localization scheme proceeds as follows: 1. A spotlight device distributes events into the space A over a period of time. 2. During the event distribution, sensor nodes record the time sequence T = {t1, t2 ... tn} when they detect the event. 3. After the event distribution, each sensor node sends the detection time sequences to the spotlight device. 4. The spotlight device estimates the location of sensor nodes, using the time sequence T and the known E(t) function.

Figure 3: Spotlight System implementation As shown in Figure 3, during this demonstration, we use a NEC MultiSync MT1030 projector connected to a DELL laptop. By projecting light to a vertical Veltex board, we distribute well-controlled events to the Mica2 motes attached to the board. Specifically, we will demonstrate the localization process of the point scan, line scan and area cover. We will also provide insights into how this localization scheme can be successfully applied to realistic outdoor environments. © Copyright by Tian He, Radu Stoleru and John A. Stankovic All Rights Reserved

Localization System for Outdoor Wireless Sensor NetworksRadu Stoleru, Tian He, John A. Stankovic

Department of Computer Science, University of Virginia {stoleru, tianhe, stankovic}@cs.virginia.edu

1. INTRODUCTION We demonstrate a localization system, called Spotlight [1], for Wireless Sensor Networks (WSN). The system, developed at the University of Virginia, has been tested in an outdoor, realistic, environment and has shown a very high accuracy (e.g., tens of centimeters). Our contribution to the area of localization in WSNs is two-fold: develop a novel scheme for localizing sensor nodes, and implement the system on real sensor nodes and common-off-the-shelf (COTS) hardware.

2. SYSTEM DESCRIPTION The Spotlight localization system employs an asymmetric architecture, in that all the complexity associated with the localization is embedded in a single device, called the Spotlight device. In our system, no hardware changes are necessary to the existing line of Mica sensor boards.

The Spotlight device is a sophisticated, powerful device, capable of producing controlled events (e.g., light events) in the deployment area of a sensor network. A visual depiction of the Spotlight device, is a helicopter equipped with a search light.

The sensor network is time synchronized. Its sensor nodes execute an event detection algorithm, and report back to the Spotlight device, through a base station, the time stamps of the detected events. Using its knowledge of the distribution (in space and time) of events, and the time when a sensor node detected an event, the Spotlight device is able to compute the node’s location.

3. DEMONSTRATION For our indoor demonstration, we use the following hardware: a computerized telescope mount, three very low power (less than 1mW) diode lasers, able to produce circular and linear beams, an LCD projector, a laptop and a sensor network of 10 MicaZ motes. The setup is shown in Figure 1.

The sensor nodes are attached to a vertically positioned Veltex board (attached to an easel). The Spotlight device, comprised of the telescope mount and laptop is positioned 3-4 meters away from the sensor network. A Localization GUI, running on the laptop, controls the orientation of the computerized telescope mount and the communication with the sensor network, through a base station. The Localization GUI displays, through the LCD projector, an aerial image (map) of a real outdoor environment, called deployment area. During the demonstration, the system goes through the following phases: - Time synchronization phase, to synchronize the clocks of all sensor nodes and the PC. This phase executes only once, at the

beginning of the demo. The following phases execute in a loop. - Event Generation phase, during which the laser will scan the deployment area. We demonstrate two different Event Generation schemes: a Point Scan and a Line Scan. In the Point Scan scheme, the deployment area is scanned by a laser with a circular beam. In the Line Scan scheme, two lasers, with linear beams, scan the deployment area, on the horizontal and vertical axes. One run of the demo will exhibit one of the two Event Generation schemes. - Reporting phase, initiated by the Localization GUI. In this phase all sensor nodes report back to the Localization GUI, the time stamps of the detected events. - Localization phase in which the Localization GUI computes and displays on the map the position of sensor nodes. - System reset phase, initiated from the Localization GUI, in which all sensor nodes clear the set (i.e. empty set) of timestamps for the observed events.

Figure 1. Indoor demonstration setup.

The error in localization of the system is represented by the distance between the position of the photo-sensor, on the MTS310 sensor board, and the position of a sensor node, shown as a red dot through the LCD projector.

4. REFERENCES [1] R. Stoleru, T. He, J. A. Stankovic, “High-Accuracy, Low-Cost Localization System for Wireless Sensor Networks”, in International Conference on Embedded Networked Sensor Systems (SenSys), 2005

Achieving Generic Asymmetric Sensing Architecture

Yu Gu and Tian He{yugu,tianhe}@cs.umn.edu

Department of Computer Science and EngineeringUniversity of Minnesota, Twin Cities

1 Introduction

As a key approach to achieve energy efficiency in sensor net-works, sensing coverage has been studied extensively in the lit-erature. Researchers have designed many coverage protocolsto provide various kinds of service guarantees on the networklifetime, coverage ratio and detection delay. While the state-of-the-art is encouraging, we believe there are some aspectsthat need further investigation. First, currently different sens-ing coverage algorithms focus on different service guarantees(e.g., coverage vs. detection delay). Any single design is notgeneral enough to meet a wide range of sensing requirementsunder different operating scenarios. Second, lifetime extensionin most algorithms is essentially achieved through coordinationamong neighboring nodes. The local node density, therefore,imposes a theoretical upper bound on the system lifetime. Sucha bound can be surpassed through global scheduling. However,the overhead of global scheduling would increase significantlyif the coordination among the nodes goes beyond the boundaryof neighborhood.

To address these two issues simultaneously, in this abstract,we propose a Unified Sensing Coverage Architecture, calleduSense, which features three novel ideas: Asymmetric Archi-tecture, Generic Switching and Global Scheduling. We pro-pose asymmetric architecture based on the conceptual separa-tion of switching from scheduling. Switching is efficiently sup-ported in sensor nodes, while scheduling is done in a separatedcomputational entity, where multiple scheduling algorithms aresupported. Such asymmetric architecture enables us to designsophisticated coverage algorithms in an unconstrained designspace, represent such intelligence with a lightweight algorithmimplemented in sensor nodes and achieve a fast and efficientchange of coverage algorithms by disseminating only severalparameters. As far as we know, the proposed generic switch-ing is the first generic and lightweight switching algorithm forresource-limited sensor nodes. As an instance, we propose atwo-level global coverage algorithm, called uScan. At the firstlevel, coverage is scheduled to activate different portions of anarea. At the second level, sets of nodes are selected to coveractive portions. Interestingly, we show that it is possible to ob-tain optimal set-cover results in linear time if the layout of areassatisfies certain conditions. The global scheduling, which co-ordinates nodes within a large area, could lead to significantlybetter performance compared with localized solutions. Our ini-tial evaluation results indicate that uSense is a promising ar-chitecture to support flexible and efficient coverage in sensornetworks.

2 uSense Architecture

The uSense design consists of two parts as shown in Figure 1.The switching algorithm is implemented in the sensor nodes,

Sensor i

Sensor j

Sensor k

Algorithm 1

Algorithm 2

Algorithm n

…..

Param

eter Translating

Generic

Switching

Algorithm

Parameters

Connectivity

Sensor NetworkComputational

Entity

Switching Side Scheduling Side

Fig. 1. Overview of uSense Architecture

which takes only two scheduling parameters as input: workingschedule bits: S - an infinite binary string in which 1 denotes theactive state and 0 denotes the inactive state, and switching rate:R - defines the rate of toggling between states. The compu-tational entity is responsible for generating the scheduling pa-rameters. It takes the schedules decided by sensing coveragealgorithms and translates them into the parameters used by theswitching algorithm. In order to efficiently disseminate thesetwo parameters to the sensor nodes, we propose to express theschedule bits, which is usually periodic and following a certainpattern, in the form of regular expression. To further reducecommunication costs, we use a parameterized schedule, whichis a binary regular expression that changes with the node’s lo-cation and enables us to disseminate the schedules through onelimited flooding instead of unicasting to every single node.

3 Global Scheduling Algorithms

Conceptually, uSense can support many existing coveragealgorithms. Due to its asymmetric architecture, it is especiallyfriendly to the global scheduling algorithms. Since a globalscheduling allows many more nodes to activate in turn ratherthan the localized ones that only schedule the nodes withinneighborhood, it leads to a significant energy savings. In thissection, we describe a global scheduling algorithm called uS-can, which is one of sensing coverage algorithms in Figure 1that are supported by the uSense architecture.

The outputs of uScan are the scheduling bits S and switch-ing rate R for individual nodes. uScan is a two-level schedulealgorithm, which works as follows: Suppose we provide sens-ing coverage to a given area using uScan. First, uScan dividesthe area into small regions, and decides the working schedulesfor these regions. This level of scheduling is conceptually in-dependent of the deployment of the nodes. At the second-level,we assign nodes to cover the active regions at different timeintervals, using a set-cover technique. By combining the first-level schedule and the set-cover assignment, we can decide theschedule bits S for individual nodes.

1

Fig. 2. uSense System Setup

0

0.2

0.4

0.6

0.8

1

150 300 450 600 750 900 1050 1200 1350

CD

F

Detection Delay

Line ScanSystolic Scan

Fig. 3. Detection Delay of Line and Sys-tolic Scan for Static Targets

0

2

4

6

8

10

1 1.5 2 2.5 3 3.5 4

Syst

em H

alf

Lif

e

Node Density

Ideal Upper Bound

0

2

4

6

8

10

1 1.5 2 2.5 3 3.5 4

Syst

em H

alf

Lif

e

Node Density

Ideal Upper BoundDiffSurv

0

2

4

6

8

10

1 1.5 2 2.5 3 3.5 4

Syst

em H

alf

Lif

e

Node Density

Ideal Upper BoundDiffSurvuSense

0

2

4

6

8

10

1 1.5 2 2.5 3 3.5 4

Syst

em H

alf

Lif

e

Node Density

Ideal Upper BoundDiffSurvuSense

All Working

Fig. 4. System Half Life vs. Node Den-sities

3.1 Level I: Tile Scheduling

In uScan, we partition an area under surveillance into somesmall regions of the same shape, a process called tessellation.These small regions are called tiles, which can be regular trian-gles, rectangles or regular hexagons in a 2-D space. The size oftiles is set to be smaller than the minimum target size, so that atarget is detected as long as a portion of a tile is covered. As areminder, nodes within a sensor network only support a genericswitching algorithm, which has neither the concept of tiles northe partition information of the tiles. All the complex logic re-sides in the computational entity. In this section, we discuss twomethods for the tile-level scheduling. They differ in the energyconsumption rate and the detection delay.

• Line Scan: Instead of trying to cover all tiles, we onlycover a column/row of tiles in a certain interval of timeduring one round of scan. The covered columns/rows areincreasing or decreasing consecutively.

• Systolic Scan: Systolic Scan emulates the cardiac cyclesof a beating heart. Over the area under surveillance, wesense the tiles from the inner layer to the outer layer con-tinuously.

Both line scan and systolic scan specify only the set of tilesneed to be activated (covered) at a given point of time. The taskof covering each tile set is accomplished by the second-levelnode scheduling, which is described in the next section.

3.2 Level II: Node Scheduling

Tile-level scheduling determines the set of active tiles TSi

at the time interval i. In this section, we describe how we cantranslate a known tile schedule into a corresponding node sched-ule bits S, which can be interpreted directly by a generic switch-ing algorithm.

The main idea of our approach is to find the optimal set ofnodes which could cover all the tiles that need to be active attime interval i. Before node scheduling, we first map physicalnode coverage into the coverage bipartite graph according tothe coverage relationship. Then we divide node scheduling intotwo steps. First, for a tile set T Si, we keep identifying 1-coverset with minimal number of nodes, until the size of 1-cover setis above a certain threshold. Secondly, we create schedules fornodes such that each identified 1-cover set provides coverage toT Si in a round-robin fashion.

The generic Minimum Set Cover problem has been provenNP-Hard [1]. Fortunately, we find line scan coverage is a spe-cial case of the generic set cover problem, because a node onlyneed to cover a continues segment of tiles. By mapping thecoverage bipartite graph into a directed acyclic graph with fol-lowing rules:

1. Map N tiles in TSi into N vertex V = {v1, ...,vN} and addone extra vertex vN+1.

2. If a node covers a set of tiles {Ti, ...,Ti+n}, we create ndirectional edges (vi,v j) where v j = vi+1, ...,vi+n+1. Eachedge has a unit cost.

We reduce the tile set cover problem to the problem of find-ing out the shortest paths from v1 to vN+1, with the overall run-time of O(|V |)+ O(|E|).

We note that the proposed polynomial algorithm does notapply to generic tile scheduling, where a tile set does not forma continuous curve or where a node can cover multiple seg-ments of a tile set simultaneously. In these cases, we adopt agreedy set-cover method by choosing the node that covers themost number of tiles first.

In order to support line scan or systolic scan in a 2-D space,we need to identify cover sets for the whole area(not just for asingle tile set). Thus a node may need to cover multiple tile setsT Si. To effectively handle the cases where we have to selectcover sets for multiple T S, we designed an algorithm that eachnode maintains a counter SC which records how many timesit has been selected into a unique cover sets. While selectingcover sets for each tile set T S, instead of solely consider thenumber of nodes in the cover sets, the algorithm calculates theminimum cover set among the nodes whose SC counter valuesare as small as possible.

After obtaining cover sets for every tile set T Si, we buildthe final schedule of node according to all the cover sets it’s be-longed to, which can be executed directly by our generic switch-ing algorithm.

4 Implementation and Evaluation

We have implemented a complete version of uSense onBerkeley TinyOS/Mote platform, using 30 MicaZ motes asshown in Figure 2. The results showed that uSense is alightweight and efficient architecture. To reveal the system per-formance at scale, we have conducted some initial large scalesimulations with 10,000-node. Under full coverage mode, wedemonstrated that our global scheduling algorithms provide sig-nificant energy savings over previous protocols such as Diff-Surv [2] under metrics such as Half-life, Coverage Overtimeand Node Energy Consumption. In the future, we plan to inves-tigate the performance under partial coverage mode at scale aswell, with additional metrics such as Detection Delay for StaticTargets and Worst-Case Breach (WCB) for Mobile Targets.

5 References[1] V. T. Paschos. A Survey of Approximately Optimal Solutions to Some

Covering and Packing Problems. In ACM Computing Surveys, June 1997.

[2] T. Yan, T. He, and J. A. Stankovic. Differentiated Surveillance Service for

Sensor Networks. In SenSys’03, November 2003.

2

1

uCast: Unified Connectionless Multicast for EnergyEfficient Content Distribution in Sensor Networks

Qing Cao, Tian He, Member, IEEE, and Tarek Abdelzaher, Member, IEEE

Abstract— In this paper, we present uCast, a novel multicastprotocol for energy efficient content distribution in sensor net-works. We design uCast to support a large number of multicastsessions, especially when the number of destinations in a sessionis small. In uCast, we do not keep any state information relevantto ongoing multicast deliveries at intermediate nodes. Rather,we directly encode the multicast information in the packetheaders, and parse these headers at intermediate nodes using thescoreboard algorithm as proposed in this paper. We demonstratethat 1) uCast is powerful enough to support multiple addressingand unicast routing schemes and 2) uCast is robust, efficient andscalable in the face of changes in network topology, such as thoseintroduced by energy conservation protocols. We systematicallyevaluate the performance of uCast through simulations, compareit with other state-of-the-art protocols, and collect preliminarydata from a running system based on the Berkeley motesplatform.

Index Terms— Sensor networks, multicast, content distribu-tion.

I. INTRODUCTION

Recent research work has observed the unique challengesof sensor networks from a variety of aspects. One challengeis that sensor nodes are severely limited by their computation,storage and communication power. In this paper, we studythe implications of such limitations on multicast protocols.In particular, we address the problem of designing protocolsto support a large number of small group multicasts, wherethe number of destinations in a single session is limited. Thisproblem is unique, as we will show later, in that no previousprotocol can be easily adapted to solve it.

There are many applications of small group multicast insensor networks. For example, in a typical directory service,such as the protocol described in [16], each node period-ically updates a small set of other nodes (named directoryservers in [16]) with its location. Therefore, one node needsto multicast information to several destination nodes, whichform a small group. Furthermore, when multiple nodes usethe directory service, they will generate many small groupmulticast sessions. Another common example involves data-centric storage (DCS) [21], [24]. One key component of someDCS protocols is the use of Geographic Hash Tables (GHT).GHT hashes keys, usually the names of data or events, intogeographic coordinates. It then stores the values at the node

Qing Cao and Tarek Abdelzaher are with the Department of ComputerScience, University of Illinois at Urbana Champaign, 201 North GoodwinAvenue, Urbana, IL 61801. Email: {qcao2,zaher}@cs.uiuc.edu.

Tian He is with the Department of Computer Science and Engineering,University of Minnesota, 200 Union Street SE, Minneapolis, MN 55455.Email: tianhe@cs.umn.edu.

that is geographically nearest to the hash value of the key.Usually, the data storage protocol suggests that the key-valuepairs should be stored at multiple locations for robustness.Therefore, such protocol naturally requires a small groupmulticast session for each storage process, and many multicastsessions for storing a large amount of data.

However, providing small group multicast service in sen-sor networks is complicated by several unique challenges.One challenge is that sensor nodes are extremely resourceconstrained. Consequently, sensor networks generally employcertain energy conservation protocols, which usually allowindividual nodes to switch between sleep and wake states.Therefore, the topology of sensor networks will change, whichposes unique requirements on the multicast service. Conse-quently, any protocol that relies on certain multicast structuresto keep state information, such as multicast trees, must adaptthese structures to topology changes. However, so far, noprevious multicast protocol has addressed this issue properly.

Second, there are many unicast routing protocols for sensornetworks. In fact, there is no consensus on which one isthe best, and the choice unicast protocol usually dependson the particular application. For example, when geograph-ical location information is readily available for each node,several well-known routing protocols that take advantage ofgeographical information are probably the best [4], [12], [14].On the other hand, when there is no geographical informationavailable, protocols based on certain topology encodings arepreferred [5], [20], [18]. Because of the wide range of unicastchoices, it is undesirable to design multicast protocols thatmake assumptions regarding any particular unicast, since thatwill limit the applicability of the multicast service. On theother hand, it is also not smart for the multicast to providerouting service from scratch, since doing so will lead to con-siderable functional overlapping with the unicast protocol. Noprotocol so far has addressed both problems simultaneously, infact, most protocols provide their services totally independentof available unicast protocols, or provide unicast as a specialsub-service.

To address these two challenges, we present a multicast pro-tocol that is both powerful (supporting multiple unicast routingprotocols by using a unified interface) and robust (tolerant oftopological changes by providing a connectionless service).We call this protocol unified connectionless multicast, oruCast. To the best of our knowledge, uCast is the first protocolspecifically designed to support a large number of small groupmulticast sessions in sensor networks. We now give a moredetailed explanation on the two features of uCast.

The first feature is that uCast can support multiple unicast

2

routing protocols through a unified interface. In this sense,uCast is a modular extension to the underlying unicast layer.In fact, it can extend any unicast routing protocol as long asthis unicast can export a common comparison interface, whichallows a comparison operation between two nodes for the samedestination node to determine which one is better, using somenotion of cost. We implemented uCast on top of three unicastrouting protocols with different addressing schemes to proveour point.

The second feature is that uCast is tolerant of topologychanges caused by energy saving protocols. To achieve this,uCast does not keep any multicast-specific state at interme-diate nodes. Instead, uCast dynamically decides the multi-cast delivery path at each intermediate node, based only onlocal topology information and the comparison interface asdiscussed earlier. Since local information is much easier to re-construct on topological changes compared to a superimposedglobal multicast overlay, uCast is much more adaptable forunpredictable changes in network connectivity than previousmulticast protocols.

The rest of the paper is organized as follows. We outlinerelated work in Section II. We discuss the details of uCast inSection III. In Sections IV and V, we report simulations basedand real platform based results. At last, we provide furtherdiscussions and conclusions in Section VI.

II. RELATED WORK

A. MulticastMulticast is a classical topic in networking. Interestingly,

we find only a few multicast protocols designed for sensornetworks. Some may assume that work in ad-hoc networks oreven the Internet can be easily customized for sensor networks.This is not the case because of the unique challenges ofsensor networks. In the following, we survey the work onmulticast from three aspects, namely, multicast work for sensornetworks, ad-hoc networks and the Internet.

There has been some work on multicast for sensor networksfor various purposes. One category is called geocast, whichconsiders the scenario that multicast destinations are locatedwithin a bounded geographical area. Representative workincludes [13], [17]. Another multicast category is calledspatiotemporal multicast, or mobicast [10]. Mobicast featuresa moving zone of multicast destinations. The goal is to deliverpackets just in time to this zone for tracking purposes. Anothercategory [1] studies multicast in data caching and placement. Itfocuses on using multicast trees for asynchronously updateddata deliveries. Yet another is called TTDD [30], which isoptimized for mobile sinks. It uses a grid structure, coupledwith localized flooding to track mobile sinks. All these proto-cols do not consider the effect of topology changes introducedby energy conservation protocols, nor are they designed tohandle the small group multicast scenarios. Further, none ofthese protocols takes into account the compatibility issue withunicast protocols. Therefore, they are usually implementedindependently with the unicast routing protocols that arealready available, or provide unicast as a special case [13],[17]. Both approaches are likely to introduce functionality

redundancy, if there is already a unicast layer implemented.Since the memory size of current sensor network nodes isextremely limited (4K bytes on Mica2/MicaZ), it is notsmart to tolerate any functional redundancy between differentrouting services. From this point, these previous multicastprotocols are significantly different from uCast, especiallyon the compatibility with multiple unicast routing protocols.Therefore, we do not compare uCast with them in this paper.

There are many multicast protocols for ad-hoc networks.Representative protocols include multicast-tree based (Multi-cast AODV [22]), mesh based (CAMP [7]), or group based(ODMRP [15]). However, these protocols can not be easilyapplied to sensor networks because they all rely on preestab-lished overlays. These overlays are associated with consider-able signaling costs, therefore, they are usually too expensiveto reconstruct in the face of frequent topology changes, suchas those introduced by energy conservation protocols. Further,since they are usually designed for mobile nodes that are muchmore powerful than sensor nodes, such as laptops, they areusually too heavy-weighted to implement in sensor networks.

At last, we also find many multicast protocols for IPnetworks in the literature. Representative protocols includeIGMP [6], Xcast [3], [2], [25] and DVMRP [23]. Amongthese protocols, Xcast [3], [2], [25] is the most relevant toour work in that similar to ours, it encodes the destination listinto the packet headers. However, our work is considerablydifferent from Xcast in two aspects. First, Xcast relies on therouting tables at intermediate routers to decide the packet flow.However, we do not assume any particular routing structure,such as the routing table. Second, Xcast can only work witha single unicast routing protocol. Therefore, if the underlyingrouting protocol modifies the structure of the routing table,Xcast has to be modified as well. Following this principle,it is impractical to build a single multicast layer for wirelesssensor networks using Xcast. We overcome this problem bydesigning uCast on top of the common comparison interfaceexported by the underlying unicast layer. This design choiceessentially decouples uCast from any underlying unicast rout-ing details and leads to a generalized and flexible service thatis significantly different from Xcast.

Based on this survey, we consider uCast as a necessarycomplement for previous protocols. Primarily, our work istargeted at the small group multicast scenarios. Conceptually,the application domain of uCast is shown in Figure 11.

As shown in Figure 1, as the number of members for aparticular multicast session increases, or the traffic per sessionincreases, the average cost per member decreases for con-nection based protocols, because the signaling cost becomesless significant. This implies connection based protocols aremore suitable for long-term large scale multicast. However,when the number of members is small and the traffic is low,the corresponding application domain can be characterized byspontaneous, short-term content delivery requests within smallgroups. The expected cost per member will increase due tothe signaling cost for the connection based multicast service.

1The curve shown is only conceptual and helps the understanding of theapplication domain of uCast. It does not reflect any quantitative results.

uScan: A Lightweight Two-Tier Global Sensing Coverage Design

Yu Gu and Tian HeDepartment of Computer Science and Engineering, University of Minnesota

{yugu,tianhe}@cs.umn.edu

1 Introduction

Wireless Sensor Networks (WSNs), consisting of thousandsof low-cost sensor nodes, have been used in many applica-tion domains such as military surveillance [1], habitat monitor-ing [4] and scientific exploration [6]. Limited power suppliesand difficulties in harvesting ambient energy make energy con-servation a critical issue to address. As one of solutions, energy-efficient sensing coverage extends system lifetime by leveragingon the redundant deployment of sensor nodes. Existing algo-rithms [5, 7, 8, 9] are designed to be well distributed and local-ized with solid performance gains. While the state-of-the-art isencouraging, we believe there are some aspects that need furtherinvestigation. In most algorithms, extending system lifetime isachieved essentially through coordination among neighboringnodes. The local node density, therefore, imposes a theoreti-cal upper bound on the system lifetime, if a continuous sensingcoverage or a partial coverage is required. Such a bound can besurpassed through global scheduling. However, the overhead ofglobal scheduling would increase significantly if the coordina-tion among the nodes goes beyond the neighborhood.

To address these issues, we introduce a two-tier globalscheduling method, called uScan. At the first level, coverageis scheduled to activate different portions of an area. We pro-pose an optimal scheduling algorithm to minimize area breach.At the second level, sets of nodes are selected to cover activeportions. Interestingly, we show that it is possible to obtainoptimal set-cover results in linear time if the layout of areas sat-isfies certain conditions. We have implemented and evaluatedour design on the Berkeley TinyOS/Mote platform [2], using 30MicaZ motes. The results indicate that uScan is a lightweightsolution with significant energy savings, compared with local-ized solutions.

2 uScan Design

uScan is a two-level schedule algorithm, which works as fol-lows: Suppose we provide sensing coverage to a given area us-ing uScan. First, uScan divides the area into small regions, anddecides the working schedules for these regions. This level ofscheduling is conceptually independent of the deployment ofthe nodes. At the second-level, we assign nodes to cover theactive regions at different time intervals, using a set-cover tech-nique. By combining the first-level schedule and the set-coverassignment, we can decide the working schedule of individualnodes.

2.1 Assumption

We assume that nodes are time-synchronized and their lo-cations are known. These are common assumptions for manysensor network applications [1, 4, 6]. Accuracy of time syn-chronization and localization do not need to be precise, becauseclock drift can be resolved by slightly extending the active du-

ration and localization error can addressed using the methodproposed in [8]. For the clarity of the protocol description inthe rest of the paper, we refer the sensing area of a node as acircle with a nominal radius r centered at the location of thenode. However, our design works under irregular sensing areasas long as nodes are aware of their sensing areas.

2.2 Definition of the Node Schedule

In essence, a sensing coverage algorithm decides the work-ing schedule of individual sensor nodes. Specifically, uScandescribes the behavior of a node using two parameters, namelythe schedule bits S and switching rate R.

• Schedule bits S: It is an infinite binary string in which1 denotes the active state and 0 denotes the inactive state.Since the sensing coverage schedule is usually periodic,follows a certain pattern. Therefore, we can express S witha regular expression. For example, (0010)∗ can be used todenote a repeated off-off-active-off schedule.

• Switching rate R: It defines the rate of toggling betweenstates. For example, a switching rate of 0.5HZ requires anode to read one bit from the schedule S every 2 seconds(when consecutive bits have the same value, there is noneed to change power state of the node physically).

2.3 Level I: Tile Scheduling

In uScan, we partition an area under surveillance into somesmall regions of the same shape, a process called tessellation.These small regions are called tiles, which can be regular trian-gles, rectangles or regular hexagons in a 2-D space. The size oftiles is set to be smaller than the minimum target size, so that atarget is detected as long as a portion of a tile is covered. In thissection, we discuss two methods for the tile-level scheduling.They differ in the energy consumption rate and the detectiondelay.

• Line Scan: Instead of trying to cover all tiles, we onlycover a column/row of tiles in a certain interval of timeduring one round of scan. The covered columns/rows areincreasing or decreasing consecutively.

• Systolic Scan: Systolic Scan emulates the cardiac cyclesof a beating heart. Over the area under surveillance, wesense the tiles from the inner layer to the outer layer con-tinuously.

Both line scan and systolic scan specify only the set of tilesneed to be activated (covered) at a given point of time. The taskof covering each tile set is accomplished by the second-levelnode scheduling, which is described in the next section.

2.4 Level II: Node Scheduling

Tile-level scheduling determines the set of active tiles TSi

at the time interval i. In this section, we describe how we can

1

Fig. 1. Test-bed Setup

0

0.2

0.4

0.6

0.8

1

150 300 450 600 750 900 1050 1200 1350

CD

F

Detection Delay

Line ScanSystolic Scan

Fig. 2. Detection Delay

0

2

4

6

8

10

1 1.5 2 2.5 3 3.5 4

Sy

stem

Hal

f L

ife

Node Density

Ideal Upper Bound

0

2

4

6

8

10

1 1.5 2 2.5 3 3.5 4

Sy

stem

Hal

f L

ife

Node Density

Ideal Upper BoundDiffSurv

0

2

4

6

8

10

1 1.5 2 2.5 3 3.5 4

Sy

stem

Hal

f L

ife

Node Density

Ideal Upper BoundDiffSurvuSense

0

2

4

6

8

10

1 1.5 2 2.5 3 3.5 4

Sy

stem

Hal

f L

ife

Node Density

Ideal Upper BoundDiffSurvuSense

All Working

Fig. 3. System Half Life vs. Densities

translate a known tile schedule into a corresponding node sched-ule bits S, which can be interpreted directly by a generic switch-ing algorithm.

The main idea of our approach is to find the optimal set ofnodes which could cover all the tiles that need to be active attime interval i. Before node scheduling, we first map physicalnode coverage into the coverage bipartite graph according tothe coverage relationship. Then we divide node scheduling intotwo steps. First, for a tile set T Si, we keep identifying 1-coverset with minimal number of nodes, until the size of 1-cover setis above a certain threshold. Secondly, we create schedules fornodes such that each identified 1-cover set provides coverage toT Si in a round-robin fashion.

The generic Minimum Set Cover problem has been provenNP-Hard [3]. Fortunately, we find line scan coverage is a spe-cial case of the generic set cover problem, because a node onlyneeds to cover a continuous segment of tiles. By mapping thecoverage bipartite graph into a directed acyclic graph with fol-lowing rules:

1. Map N tiles in TSi into N vertex V = {v1, ...,vN} and addone extra vertex vN+1.

2. If a node covers a set of tiles {Ti, ...,Ti+n}, we create ndirectional edges (vi,v j) where v j = vi+1, ...,vi+n+1. Eachedge has a unit cost.

We reduce the tile set cover problem to the problem of find-ing out the shortest paths from v1 to vN+1, with the overall run-time of O(|V |)+ O(|E|).

We note that the proposed polynomial algorithm does notapply to generic tile scheduling, where a tile set does not forma continuous curve or where a node can cover multiple seg-ments of a tile set simultaneously. In these cases, we adopt agreedy set-cover method by choosing the node that covers themost number of tiles first.

In order to support line scan or systolic scan in a 2-D space,we need to identify cover sets for the whole area (not just for asingle tile set). Thus a node may need to cover multiple tile setsT Si. To effectively handle the cases where we have to selectcover sets for multiple T S, we designed an algorithm that eachnode maintains a counter SC which records how many timesit has been selected into a unique cover sets. While selectingcover sets for each tile set T S, instead of solely consider thenumber of nodes in the cover sets, the algorithm calculates theminimum cover set among the nodes whose SC counter valuesare as small as possible.

After obtaining cover sets for every tile set T Si, we buildthe final schedule of node according to all the cover sets it’s be-longed to, which can be executed directly by our generic switch-ing algorithm.

3 Implementation and Evaluation

We have implemented a complete version of uScan onBerkeley TinyOS/Mote platform, using 30 MicaZ motes asshown in Figure 1. The compiled image of a full implemen-tation occupies 21,040 bytes of code memory and 907 bytesof data memory. The results showed that uScan is a lightweightand efficient coverage design. To reveal the system performanceat scale, we have conducted some initial large scale simulationswith 10,000-node. Under full coverage mode as shown in Fig-ure 3, we demonstrated that our global scheduling algorithmsprovide significant energy savings over previous protocols suchas DiffSurv [8] under metrics such as Half-life, Coverage Over-time and Node Energy Consumption. In the future, we planto investigate the performance under partial coverage mode atscale as well, with additional metrics such as Detection Delayfor Static Targets and Worst-Case Breach (WCB) for MobileTargets.

4 Conclusion

The major contribution of this work is a two-level globalscheduling algorithm called uScan. In the first level, we pro-pose two tile-level scheduling algorithm. In the second level,we propose a linear algorithm to address the set-cover problemwhen the layout of tiles satisfies certain conditions. We evaluateour architecture with a network of 30 MicaZ motes, an exten-sive simulation with 10,000 nodes, as well as theoretical anal-ysis. We believe our work has successfully provided flexibilityand efficiency for the sensor network coverage problem.

5 References[1] A. Arora and et al. A Wireless Sensor Network for Target Detection,

Classification, and Tracking. Computer Networks (Elsevier), 2004.

[2] J. Hill, R. Szewczyk, A. Woo, S. Hollar, D. E. Culler, and K. S. J. Pister.System Architecture Directions for Networked Sensors. In ASPLOS’00,2000.

[3] V. T. Paschos. A Survey of Approximately Optimal Solutions to SomeCovering and Packing Problems. In ACM Computing Surveys, June 1997.

[4] R. Szewczyk, A. Mainwaring, J. Anderson, and D. Culler. An Analysis ofa Large Scale Habit Monitoring Application. In SenSys’04, 2004.

[5] D. Tian and N. Georganas. A Node Scheduling Scheme for Energy Con-servation in Large Wireless Sensor Networks. Wireless Communications

and Mobile Computing Journal, May 2003.

[6] G. Tolle, J. Polastre, R. Szewczyk, N. Turner, K. Tu, S. Burgess, D. Gay,P. Buonadonna, W. Hong, T. Dawson, and D. Culler. A Macroscope in theRedwoods. In Sensys’05, November 2005.

[7] X. Wang, G. Xing, Y. Zhang, C. Lu, R. Pless, and C. Gill. IntegratedCoverage and Connectivity Configuration in Wireless Sensor Networks. InSensys’03, November 2003.

[8] T. Yan, T. He, and J. A. Stankovic. Differentiated Surveillance Service forSensor Networks. In SenSys’03, November 2003.

[9] F. Ye, G. Zhong, S. Lu, and L. Zhang. PEAS: A Robust Energy Conserv-ing Protocol for Long-lived Sensor Networks. In Proc. of International

Conference on Distributed Computing Systems (ICDCS), May 2003.

2

Achieving Long-Term Surveillance in VigilNetTian He†, Pascal Vicaire, Ting Yan, Qing Cao‡, Gang Zhou, Lin Gu,Liqian Luo‡, Radu Stoleru, John A. Stankovic, Tarek F. Abdelzaher‡

Department of Computer ScienceUniversity of Virginia, Charlottesville 22903

Abstract— Energy efficiency is a fundamental issue for out-door sensor network systems. This paper presents the designand implementation of multi-dimensional power managementstrategies in VigilNet, a major recent effort to support long-term surveillance using power-constrained sensor devices. Weintegrate a novel tripwire service with an effective sentry andduty cycle scheduling in order to increase the system lifetime,collaboratively. Through extensive system implementation, wedemonstrate the feasibility to achieve high surveillance perfor-mance and energy efficiency, simultaneously. We invest a fairamount of effort to evaluate our architecture with a networkof 200 XSM motes in an outdoor environment, an extensivesimulation with 10,000 nodes, as well as an analytical probabilisticmodel. These evaluations demonstrate the effectiveness of ourintegrated approach and identify many interesting lessons andguidelines, useful for the future development of energy-efficientsensor systems.

I. INTRODUCTION

VigilNet is a recent major effort to support long-termmilitary surveillance, using large-scale micro-sensor networks.Besides requirements of accurate target tracking and classifi-cation [1], one of the key design goals of VigilNet is to achievelong-term surveillance in a realistic mission deployment. Dueto the small form factor and low-cost requirements, sensordevices such as the XSM motes [2] are normally equippedwith limited power sources (e.g., two AA batteries). Moreover,because of the hostile environment and a large number ofnodes deployed, currently it is not operationally and economi-cally feasible to replace the power source without introducingenormous effort and elements of risk to the military personnel.In addition, the static nature of the nodes in the field preventsthe scavenging of the power from ambient motion or vibra-tion [3][4]. The small form factor and possible lack of theline of sight (e.g., deployment in the forest) make it difficultto harvest solar power. On the other hand, a 3∼6 month systemlife span is essential to guarantee the effectiveness of normalmilitary operations, which necessitates a 12∼24 fold extensionof the normal lifetime of active sensor nodes. Consequently, itis critical to investigate practical approaches of spending thepower budget effectively.

Many solutions have been proposed for energy effi-ciency at various levels of the system architecture, rang-ing from the hardware design [5][2], coverage [6][7][8][9]MAC [10][11][12], routing [13][14][15], data dissemina-tion [16], data gathering [17][18], data aggregation [19][20],

†Tian He is now with University of Minnesota tianhe@cs.umn.edu‡Qing Cao, Liqian Luo and Tarek Abdelzaher are now with University of

Illinois, Urbana-Champaign

data caching [21], topology management [22], clustering [23],placement [24] [25] to energy-aware applications [26][27].Instead of focusing on a single protocol, our answer to energyefficiency is an integrated multi-dimensional power manage-ment system. Our contributions, presented in this paper, areidentified in the following aspects: 1) Our design is vali-dated through an extensive system implementation: VigilNet– a large-scale sensor network system delivered to militaryagencies. 2) VigilNet takes a systematic approach, and theenergy efficiency is not narrowly accounted for within a singleprotocol. We propose a novel tripwire service, integrated withan effective sentry and duty cycle scheduling to increase thesystem lifetime, collaboratively. 3) Tradeoffs are investigatedto meet requirements of both surveillance performance and thenetwork lifetime. We present a complete system with 40,000lines of code, running on motes, that achieves performance andenergy efficiency simultaneously. 4) We devote considerableeffort to evaluate the system with 200 XSM motes in anoutdoor environment and an extensive simulation of 10,000nodes, in order to identify a set of useful lessons and guidelinesfor future research.

The remainder of the paper is organized as follows: Sec-tion II categorizes power management features for differentapplication scenarios. Section III describes the power manage-ment requirements in VigilNet. Section IV introduces threepower management strategies utilized in VigilNet namely,sentry service, tripwire service and duty cycle scheduling. Sec-tion V describes the integrated power management architecturein VigilNet. Section VI briefly discusses some additionalenergy efficient techniques applied in VigilNet. Section VIIaddresses the tradeoff between energy efficiency and networkperformance. Section VIII details the system implementation.Section IX provides the evaluation of a network of 200 XSMmotes as well as an extensive hybrid simulation with 10,000nodes. Finally, Section X concludes the paper.

II. BACKGROUND

Power management is by no means a stand-alone researchissue. It can be dramatically affected by the underlying systemconfiguration and by the application requirements. These in-clude the form factor [28], hardware capability [5], possibilityof energy scavenging [4][29], network/sensing topology anddensity [6], link quality [30], event patterns, node mobility,availability and accuracy of time synchronization [31], real-time requirements and the nature of the applications [26]. Atthe hardware level, multi-level sleep modes in the low power

microcontroller [5] enable software to control the rate of powerdissipation. Fine-grained power control [2] allows applicationsto activate hardware modules incrementally. Radio wakeupcircuits [32] achieve passive vigilance with a minimal powerdraw. Energy scavenging [3] is also possible for some ap-plication scenarios, where ambient energy can be harvested.Sensing coverage schemes [6][7] exploit redundancy in thenode deployment to activate only a subset of nodes. Thecoordinated scheduling of the sensor duty cycle [33] increasesthe probability of detection and reduces the detection delaywith a minimal power consumption. Communication protocolsturn off the radio when a node is not the intended receiver [12].Though many individual solutions are proposed, few realsystems actually achieve power efficiency comprehensively,which makes the integrated approach in VigilNet novel andpractically useful. Considering the diversity of the differentapproaches, we categorize power management strategies inthe context of two types of systems: sampling systems andsurveillance systems.

A. Power Management in Sampling Systems

Great Duck Island [26] and Structural Monitoring [27] aretypical sampling systems, which are deployed as distributedlarge-scale data acquisition instruments. Power managementstrategies in these systems normally make use of the followingtechniques:

• Predefi ned sampling schedules: Most environmentalphenomena, such as temperature, exist ubiquitously overspace and continuously over time. The static nature ofthese phenomena makes it sufficient to construct thedata profile by sampling the environment within discretetime and space. Nodes can conserve energy by turningthemselves off, according to a predefined schedule.

• Synchronized and coordinated operations: Once thesampling interval is defined a priori, nodes can commu-nicate in a synchronized fashion. With a precise time syn-chronization [31], a receiver can turn on the radio moduleright before the message payload arrives. Consequently,we can avoid low-power listening over radio [10] during anon-active period. In addition, with the knowledge aboutthe sending rate of individual nodes, we are able toestimate the link quality without control messages [34].

• Data aggregation and compression: Since channel me-dia access is costly, especially when the receiver is ina deep-sleep state [10], it is beneficial to send out oneaggregate containing multiple sensing readings [19][20].In addition, due to the value locality of the sensed data,we can compress the total number of bits to be sent overthe air. Since both aggregation and compression needto buffer a relatively large number of readings, whichintroduces a certain delay, they are not quite suitable fortime-critical surveillance systems. However, they matchmost sampling systems very well.

B. Power Management in Surveillance System

On the other hand, operations in surveillance sys-tems [35][36][37] [38], such as VigilNet, are event-driven innature. In surveillance systems, we are more interested in thedata profile between inception and conclusion of the transientevents. These systems should remain dormant in the absenceof the events of interests, and switch to an active state to obtainhigh fidelity in detection. Normally, the surveillance systemsimprove the system lifetime through the following approaches:

• Coverage control: Surveillance systems are normallydeployed with a high density (For instance, the defaultconfiguration of VigilNet [38] has 28 nodes per nominalradio range (30 m)) for the sake of robustness in detectionand fine-grained sensing during tracking. We can increasethe system lifetime by activating only a subset of nodesat a given point of time, waiting for potential targets.

• Duty cycle scheduling: The duration of transient eventswithin the area of surveillance is normally non-negligible.By coordinating nodes’ sleep schedules, we can conserveenergy without noticeably reducing the chance of de-tection. Duty cycle scheduling is different from samplescheduling in the sense that duty cycle scheduling isat the micro-scale (milliseconds vs. minutes) and it isstrongly affected by the dynamics of the events (e.g.,target velocity).

• Incremental activation: The sampling systems are nor-mally designed for data logging. At each sample instance,all sensors should be activated to obtain a complete dataprofile. In contrast, surveillance systems are designedto detect transient events of interest. It is sufficient toactivate only a subset of sensors for the initial detection.After the initial detection, we can activate other sensorsto achieve a higher sensing fidelity and to performclassification.

III. POWER MANAGEMENT REQUIREMENTS IN VIGILNET

Our power management strategies are motivated by a typicalmilitary surveillance application. The mission objective ofsuch a system is to conduct remote, persistent, clandestinesurveillance to a certain geographic region to acquire andverify enemy capabilities and transmit summarized intelli-gence worldwide in a near-real time manner. Several systemrequirements affect our power management design withinVigilNet:

• Continuous surveillance: Due to the dynamic/transientnature of the event, VigilNet is required to providecontinuous surveillance. This requirement significantlyaffects the overall architecture of power managementstrategies and the degree of energy conservation VigilNetcan achieve.

• Real-time: As a real-time online system for target track-ing, VigilNet is required to cope with fast changingevents in a responsive manner. The delays introducedby the power management directly affect the maximumtarget speed our VigilNet can track. It is an essential

VigilNet: An Integrated Sensor Network System forEnergy-Efficient Surveillance

Tian He, Sudha Krishnamurthy, John A. Stankovic, Tarek Abdelzaher, Liqian Luo,

Radu Stoleru, Ting Yan, Lin Gu, Gang Zhou

Department of Computer Science

University of Virginia, Charlottesville, VA 22903

{tianhe, skrish, stankvoic, zaher, ll4p, rs6bd, ty4k, lg6e, gz5d}@cs.virginia.edu

and Jonathan Hui, Bruce Krogh

Department of Electrical and Computer Engineering

Carnegie-Mellon University, Pittsburgh, PA 15213

jwhui@cs.berkeley.edu, krogh@ece.cmu.edu

This paper describes one of the major efforts in the sensor network community to build an in-

tegrated sensor network system for surveillance missions. The focus of this effort is to acquire

and verify information about enemy capabilities and positions of hostile targets. Such missions

often involve a high element of risk for human personnel and require a high degree of stealthiness.

Hence, the ability to deploy unmanned surveillance missions, by using wireless sensor networks, is

of great practical importance for the military. Because of the energy constraints of sensor devices,

such systems necessitate an energy-aware design to ensure the longevity of surveillance missions.

Solutions proposed recently for this type of system show promising results through simulations.

However, the simplified assumptions they make about the system in the simulator often do not

hold well in practice and energy consumption is narrowly accounted for within a single protocol.

In this paper, we describe the design and implementation of a complete running system, called

VigilNet, for energy-efficient surveillance. The VigilNet allows a group of cooperating sensor

devices to detect and track the positions of moving vehicles in an energy-efficient and stealthy

manner. We evaluate middleware and system performance extensively on a network of 70 MICA2

motes. Our results show that our surveillance strategy is adaptable and achieves a significant

extension of network lifetime. Finally, we share lessons learned in building such an integrated

sensor system.

Categories and Subject Descriptors: C.2.1 [Computer Communication Networks]: Network

Architecture and Design

General Terms: Design, Performance, Experimentation, Measurement

Additional Key Words and Phrases: Sensor networks, Energy conservation, Tracking, Wireless

1. MOTIVATION

One of the key advantages of wireless sensor networks (WSN) is their ability tobridge the gap between the physical and logical worlds, by gathering certain usefulinformation from the physical world and communicating that information to morepowerful logical devices that can process it. If the ability of the WSN is suitablyharnessed, it is envisioned that WSNs can reduce or eliminate the need for humaninvolvement in information gathering in certain civilian and military applications.In the near future, sensor devices will be produced in large quantities at a very low

cost and densely deployed to improve robustness and reliability. They can be minia-turized into a cubic millimeter package (e.g., smart dust [Kahn et al. 1999]) in orderto be stealthy in a hostile environment. Cost and size considerations imply thatthe resources available to individual nodes are severely limited. We believe, how-ever, that limited processor bandwidth and memory are temporary constraints insensor networks. They will disappear with fast developing fabrication techniques.The energy constraints on the other hand are more fundamental. According toR.A. Powers [Powers 1995], battery capacity only doubles in 35 years. Energyconstraints are unlikely to be solved in the near future with the slow progress inbattery capacity and energy scavenging. Moreover, the untended nature of sensornodes and the hazardous sensing environment preclude manual battery replace-ment. For these reasons, energy awareness becomes the key research challenge forsensor network protocol design. Several researchers have addressed energy conser-vation recently. Most of them focus on particular protocols and investigate whethertheir energy conservation goal can be achieved. To the best of our knowledge, noneof them investigate energy-conservation for a running system as whole. Normallythey evaluate their approach through simulations. Simulation approaches tend tomake simplified assumptions that often do not hold well in practice and they aresubject to incompleteness. For example, in [Yan et al. 2003; Wang et al. 2003; Yeet al. 2003], several sensing coverage schemes are proposed for energy conservation.None of them consider energy consumption in activities other than sensing.

In this paper, we describe our effort that involves system design and implementa-tion of VigilNet on a MICA2 platform with 70 MICA2 motes. The primary goal ofthe VigilNet is to support the ability to track the position of moving targets in anenergy-efficient and stealthy manner. Our experimental results show that the prob-ability of false alarms observed reaches zero when aggregation is achieved amongmore than 3 member motes. The experimental results we obtained also show thatwith 5% of deployed motes serving as sentries and the non-sentries operating ata 4% duty cycle, our algorithm extends the lifetime of a sensor network by up to900%.

The main contributions of this paper are: 1) the design and implementation of anintegrated system with energy-awareness as the main design principle across a wholeset of middleware services, 2) mechanisms for dynamic control, which allow tradeoffsbetween energy-efficiency and system performance by adjusting the sensitivity ofthe system, and 3) a physical implementation and extensive field evaluation thatreveal the practical issues that are hard to capture in simulation.

The remainder of this paper is organized as follows. Section 2 describes therequirements of a typical ground surveillance application. In Section 3, we describethe system setup and hardware components. In Section 4, we provide an overviewof VigilNet design. In Section 5, we elaborate on the individual components of thesystem. In Section 6, we discuss the VigilNet implementation issues. We presentexperimental results in Section 7, and summarize the lessons learned from ourexperience in Section 8. We present related work in Section 9, Finally we concludein Section 10 and discuss some future work in Section 11.

Walking GPS: A Practical Solution for Localization in Manually Deployed Wireless Sensor Networks

Radu Stoleru, Tian He, John A. Stankovic Department of Computer Science

University of Virginia {stoleru, tianhe, stankovic@cs.virginia.edu}

Abstract In this paper we present the design, implementation

and evaluation of a simple, practical and cost effective localization solution, called Walking GPS, that can be used in real, manual deployments of WSN. We evaluate our localization solution exclusively in real deployments of MICA2 and XSM motes. Our experiments show that 100% of the deployed motes localize (i.e,. have a location position) and that the average localization errors are within 1 to 2 meters, due mainly to the limitations of the existing commercial GPS devices. 1. Introduction

Wireless Sensor Networks (WSN) have generated a lot

of interest recently. The interest, initially in the academic community, was due to the challenging research problems posed by these devices with very limited resources (e.g. memory, processing power, radio bandwidth, energy). More recently, due to the tremendous potential of WSN in military, civil and industrial applications, real deployments of these networks have become imminent.

Many elegant and clever solutions have been proposed and evaluated in simulators and real system deployments, for several of the problems present in the WSN. Among these problems are energy conservation, efficient data placement and aggregation, programming paradigms and topology control. However, despite the attention it has received, accurate localization is a problem that remains unsolved in a real, ad-hoc deployment, without sophisticated, expensive hardware. In this paper we present a solution to this problem, when manual deployment is an option.

In many applications it is envisioned that WSN will be deployed from Unmanned Aerial Vehicles. In the meantime, manual deployments have been prevalent and the employed localization solutions have used some variant of associating the sensor node ID with prior knowledge of that ID’s position in the field.

We propose in this paper a solution, called Walking GPS, in which the deployer (either person or vehicle) carries a GPS device that periodically broadcasts its location. The sensor nodes being deployed, infer their position from the location broadcast by the GPS device.

The main contribution of this paper is that we present the design, implementation and the real world evaluation of a solution for the localization of wireless sensor nodes that are deployed manually in the field. Our solution is simple, cost effective and has very little overhead. Despite the simplicity of the idea, many system design and implementation issues had to be addressed in order to make the solution work and be efficient. We further hypothesize that the lessons learned from our experience can be extended to aerial deployments, for an initial, coarse localization.

The rest of the paper is structured as follows. In the second section we present related work, emphasizing research that focused on real system implementations. We present the system design and architecture of our Walking GPS localization solution in section three and the implementation of the system in section four. We present our extensive experimental results in section five and conclude in section six by summarizing the main contributions and propose future extensions of our solution. 2. Related Work

Recent developments, driven by both industrial and academic research, continuously expand the applicability of sensor networks at an unprecedented momentum. The inherent interaction between sensor networks and the physical world makes location-awareness one of the essential services for many emerging applications such as location-based directory service (e.g., GHT) [1] and entity tracking [2]. In response, many algorithms have been proposed to address the localization problem in sensor networks.

Among these solutions, some of them are designed under certain assumptions and mostly evaluated in

simulation environments. For example, the Amorphous positioning algorithm proposed in [3] uses offline hop-distance estimations and multilateration to estimate nodes’ locations, assuming an isotropic RF radio. The APIT positioning algorithm [4] is a scheme in which a node infers its position based on the possibility of being inside or outside of a triangle formed by any three anchors . In this scheme, more powerful anchors, that cover the entire deployment area, are required. In the Received Signal Strength Indicator (RSSI) techniques (RADAR [5] and SpotOn [6]), either theoretical or empirical models are used to translate signal strength into distance estimates. These approaches show promising results in simulation and controlled laboratory environment. However, in practice, many empirical studies [7] [8] [9] demonstrate that in most environments, RF radio is not isotropic and there is, actually, no connection between the signal strength degradation and the distance an RF signal travels.

Another set of solutions use Time of Arrival (TOA) [10] and Time Difference of Arrival (TDOA) [1] [11] [12] techniques to obtain pair-wise distances. These techniques demonstrate high accuracy in localization in real deployment. However, they require extensive infrastructure support, such as GPS in [10], high per-node cost as in Cricket [11] [12] and AHLos [13]. In addition, the limited range of ultrasound (normally 7 – 10 meters) used in TDOA, imposes the requirement of a high-density anchor nodes, which makes the overall system cost for localization prohibitively high for large scale sensor networks.

Recently, mobile robots have been utilized as an effective instrument to localize sensor nodes [14] [15]. Although the ideas are similar to those in this paper, due to the significant discrepancy in the deployment configuration, their designs currently require a fine-tuning of parameters, in order to achieve a high accuracy. This tuning may prove to be problematic for a non-expert.

In view of various limitations exhibited by current localization schemes, we aim to find a practical solution that not only provides high localization accuracy in a running system, but also requires very low system cost. 3. System Architecture

In this section, we present the architecture of our system, and the design decisions we encountered. We support our decisions with background material and a description of the internals of our software components.

An alternative to the Walking GPS localization scheme is enabling each sensor node with GPS capabilities. This monolithic solution is both expensive and inefficient.

In the Walking GPS architecture, however, the system is decoupled into two software components: the GPS Mote and the Sensor Mote. A UML deployment diagram is shown in Figure 1.

Sensor Mote

Localization

GPS Mote

GPS

Figure 1. Software components for the Walking GPS scheme

The GPS Mote runs on a MICA2 mote. The mote is

connected to a GPS device, and outputs its location information at periodic intervals. The Sensor Mote component runs on all sensor nodes in the network. This component receives the location information broadcast by the GPS Mote and infers its position from the packets received.

This architecture enabled us to push all complexity derived from the interaction with the GPS device to a single node, the GPS Mote, and to significantly reduce the size of the code and data memory used on the sensor node. Through this decoupling, a single GPS Mote is sufficient for the localization of an entire sensor network, and the costs are thus reduced.

A relatively simple design for the GPS Mote would have been to periodically broadcast the actual GPS location received from the GPS device. A GPS location is represented by a latitude and a longitude, which are angular measures from the Equator to North or South, and Prime Meridian to East or West, respectively. Due to the relatively small size of a sensor network (hundreds to a few thousand meters), the use of global (i.e. GPS) coordinates is very inefficient. The inefficiency stems from the size of the packets used for passing location information – a significant portion of the location is likely to be the same for all sensor nodes – as well as from the computational costs encountered when aggregating data, e.g., triangulation of several GPS coordinates for positioning a target.

In order to reduce the overhead incurred when exchanging data containing global GPS coordinates (the GPS coordinates take 11 bytes out of 29 bytes, which is the payload size of a TinyOS packet), we decided to use a local, Cartesian, coordinate system. This local coordinate system of reference, which uses linear units, is better suited for WSN, than a global coordinate system.

A local coordinate system is built from a global system, that uses GPS coordinates, in the following way: the local system of reference has an origin (called a Reference Point) specified in terms of global coordinates (GPS coordinates). The distance between this Reference Point (with coordinates λ1 and ϕ1) and another point, with a GPS location specified by λ2 and ϕ2 , can be computed as follows [16]:

( )( ) ( )( )221

221Distance λλϕϕ −∗+−∗= lonlat FF (1)

1An Overview of Data Aggregation Architecture forReal-Time Tracking with Sensor Networks

Tian He†, Lin Gu∗, Liqian Luo‡, Ting Yan∗, John A. Stankovic∗, Sang H. Son∗∗Department of Computer Science, University of Virginia

†Department of Computer Science and Engineering, University of Minnesota‡Department of Computer Science, University of Illinois, Urbana-Champaign

Abstract— Since sensor nodes normally have limited re-sources in terms of energy, bandwidth and computation capabil-ity, efficiency is a key design goal in sensor network research.As one of techniques to achieve efficiency, data aggregationhas been extensively investigated in recent literature. Previousresearch on data aggregation has demonstrated its effectivenessin reducing traffic, easing congestion and decreasing the energyconsumption. However few are actually designed for a real-world application and implemented in a running system. Thispaper describes our design and implementation of a physicaltracking system, using an aggressive data aggregation archi-tecture as one of building blocks. This architecture can begenerally applied to other sensor systems, where communicationefficiency is a paramount concern and networking resources arelimited.1

I. INTRODUCTION

Traditional military surveillance systems, using long-range cameras/radars, are effective in the open terrains wheredirect line-of-sight is available. While in the urban areas andforests, the effectiveness of these solutions is affected by theobstacles such as tall buildings and trees. To address thisissue, the military starts to use wireless sensor networks asan effective surveillance instrument to deal with the non-line-of-sight (NLOS) situations, because that sensor nodes can bedeployed anywhere around an environment to provide ubiq-uitous surveillance. Due to the stealthiness requirement of themilitary surveillance systems, a tiny form factor is essential.Consequently, sensor nodes have very limited resources,suffering the bandwidth, energy and memory constraints thatlimit the amount of information that could be transferred.These factors are coupled with unpredictable traffic patternsand dynamic network topologies, making the task of datadelivery for such networks difficult. Theoretically, for a givenenergy budget available in the network, the total amount ofbits that can be transmitted is limited. It is desirable to havethe capability to deliver more information with the sameamount of bits over the air. It is often the case that thedata represented by these bits is redundant. For example, aseries of sensing readings with same values can be conciselydescribed by an average with a zero standard deviation. As

1This work was supported in part by the DAPRPA IXO offices underthe NEST project (grant number F336615-01-C-1905), the MURI awardN00014-01-1-0576 from ONR and NSF grant CCR-0098269.

one type of data aggregation techniques, this parameterizeddescription of the data distribution can effectively reduce theamount of data transmitted. Since data aggregation can re-duce transmissions while still distributing information aboutthe events of interests, it is deemed as a very effective resortto balance the communication needs and energy constraints.

This paper addresses the research challenges related tothe data aggregation technique in real-time surveillance ap-plications. We identify the fundamental tradeoffs that canbalance the performance within a three-dimensional designspace: namely the timeliness, the energy consumption and theinformation availability. Ideally, we desire to deliver enoughinformation in real-time with minimal energy consumption.Unfortunately, these performance goals are often at oddswith each other. For example, a swift delivery prevents anode accumulating sufficient data for energy-efficient dataaggregation. It is an interesting research problem to identifya performance surface within this three-dimensional space,so that a system designer can make guided decisions to tradeoff among the energy, time and data availability, accordingto the application requirements and system configurations.

To demonstrate our approach to achieve this goal, weemploy a typical sensor tracking system, called VigilNet,as a case study. We introduce the data aggregation architec-ture designed, implemented and integrated in VigilNet andidentify the tradeoffs VigilNet provides. Since VigilNet is atypical sensor network system, we believe our studies canrender insights for the system designers of similar systems.

The contribution of this work lies in the following aspects:1) Unlike the previous approaches that mainly focus on thesimulation study, we demonstrate how various data aggrega-tion techniques can be designed and implemented practicallyin a real world application. 2) We reveal the impact of dataaggregation on the quality of surveillance, the timeliness andthe related overheads. Such an analysis can guide systemdesigners to flexibly change the system configurations inorder to accommodate various kinds of operation scenarios.3) Compared with the solutions without data aggregation,we demonstrate analytically that our approach significantlyreduces the amounts of energy consumed.

II. RELATED WORK

In this paper, we focus on the data aggregation techniquesapplicable to the wireless sensor tracking systems. We divideour discussion into two categories: data aggregation tech-niques and sensor tracking systems in general.

A. Data Aggregation Approach

Data aggregation techniques have been widely used inwireless sensor networks. In [13], Intanagonwiwat proposesseveral basic forms of data aggregation methods, including1) the Center at the Nearest Source method (CNS), wherethe source nearest to the destination aggregates the data fromother nodes; 2) The shortest Path Trees (SPT) method, wheredata aggregation happens at the intermediate nodes withina shortest path tree rooted at the sink; and 3) the greedyIncremental Trees (GIT) method, where an aggregation treeis constructed by connecting each destination sequentially tothe existing tree via a shortest path. GIT assumes a com-plete knowledge of global topology information; thereforeit provides more opportunities for data aggregation. TAGprovides a hierarchical data aggregation scheme at a datacollection phase. Using an acquisitional query processor fordata collection, TinyDB [17] optimizes the query processto aggregate data with a low energy overhead. DirectedDiffusion [14], as a popular data-centric architecture fordata acquisition, can be augmented by aggregating dataalong the reinforced paths from the sources to the sinks.Another type of data aggregation techniques focuses onthe data placement of caching services. Bhattacharya etal. [1] investigate the optimal placement of caches betweenmultiple sensor sources and sinks. These caches aggregatethe updates from the source nodes and distribute data toleaf sink nodes with minimal requested rates. Since allaforementioned approaches are designed for the systems withno stringent time requirement, the designers of these systemsnaturally treat the timeliness as a less important issue relatedto the data aggregation. The closest research related to thiswork is the AIDA protocol [8]. AIDA takes the timelydelivery of messages as well as the protocol overhead intoaccount to adaptively adjust aggregation strategies in accor-dance with assessed traffic conditions and expected sensornetwork requirements. Through simulations, it demonstratesthe feasibility to reduce the energy consumption and theend-to-end communication delay simultaneously. Our workpresented here differs from aforementioned approaches inseveral aspects: First, this work deals with the real-time issuealong with the data aggregation. Second, this work introducesnot only the usage of multiple aggregation techniques, butalso how these techniques can be intergraded within a tieredarchitecture. Third, our work is not a simulation study.Instead it is designed for a realistic tracking application witha running implementation.

B. Research on Sensor-based Surveillance and Tracking

Traditional surveillance systems are widely used fordecades. Due to space constraints, we only name a fewdirectly related ones. The ASDE system [3] and SecurePerimeter Area Network (SPAN) [4], normally use long-range cameras/radars with a 360-degree view of an area asthe instruments for the detection. While these infrastructureprotection systems are effective, they are subject to severallimitations: First, they can not be emplaced swiftly withoutinfrastructure, which makes such system not suitable forspontaneous military deployment in the remote areas. Sec-ond, the large form factor of these systems makes them easilydetectable and evadable. Third, the number of surveillancepoints is small, which makes systems vulnerable to attack.To overcome these limitations, sensor networks is pursuedrecently by [2], [10], [11], [20], [19] as a more efficientmechanism to accomplish the remote unmanned surveillancemissions. Feng et al. [20] design a surveillance and trackingsystem using a distributed Bayesian estimation technique.Brook et al. [2] propose a distributed surveillance systembased on extended Kalman filter techniques. These solutionsprovide very nice features to improve the surveillance per-formance in one aspect or another, however ignoring otherperformance goals. For example, some systems provide highperformance at cost of the system lifetime - a critical per-formance metric for long-term surveillance. The differencebetween our proposed work and aforementioned approachesis that we adopt our solution in a multiple-dimensional designspace, where we consider not only the tracking performance,communication efficiency through data aggregation, but alsothe timeliness in delivery. This requires a balanced andflexible system design. Another highlight of our approach isthe system implementation, which addresses many practicalissues hard to capture in the simulated tracking scenarios.

III. DATA AGGREGATION REQUIREMENTS IN VIGILNET

The VigilNet is an online surveillance system, whichconsists of hundreds of tiny sensor nodes. These nodesdetect, track and classify incoming targets in a timely andenergy efficient manner. The final results are reported toa remote back-end via a long-haul bandwidth-constrainedsatellite links. In the current hardware platform, each nodeis equipped with three types of sensors. Magnetic sensorscan detect the changes in the magnetic field caused by themovement of ferrous objects. Motion sensors can detect thechanges in the infrared radiation caused by warm objectssuch as personnel. The acoustic sensors detect sound waves.Target signatures can be identified in VigilNet by samplingaforementioned three sensors at the particular rates (e.g.10bits data at 1000HZ from acoustic sensors). Suppose 100sensors participate in a tracking process, VigilNet couldgenerate 1Mbit data in one second by using the acousticsensor alone. The total amount of data to be transmittedmultiplies with the number of hops over the network. Giventhe fact that the current long-haul satellite link only provides

2

ACM Transaction on Embedded Computing System

1

AIDA: Adaptive Application Independent Data Aggregation in Wireless Sensor Networks1 Tian He Brian M. Blum John A Stankovic Tarek Abdelzaher

Department of Computer Science University of Virginia

{Tianhe,Bmb5v,Stankovic,Zaher@cs.virginia.edu

Abstract Sensor networks, a novel paradigm in distributed wireless communication technology, have been proposed

for use in various applications including military surveillance and environmental monitoring. These systems could deploy heterogeneous collections of sensors capable of observing and reporting on various dynamic properties of their surroundings in a time sensitive manner. Such systems suffer bandwidth, energy, and throughput constraints that limit the quantity of information transferred from end to end. These factors cou-pled with unpredictable traffic patterns and dynamic network topologies make the task of designing optimal protocols for such networks difficult. Mechanisms to perform data centric aggregation utilizing application specific knowledge provide a means to augmenting throughput, but have limitations due to their lack of adap-tation and reliance on application specific decisions. We therefore propose a novel aggregation scheme that adaptively performs application independent data aggregation in a time sensitive manner. Our work isolates aggregation decisions into a module that resides between the network and the data link layer and does not require any modifications to the currently existing MAC and network layer protocols. We take advantage of queuing delay and the broadcast nature of wireless communication to concatenate network units into an ag-gregate using a novel adaptive feedback scheme to schedule the delivery of this aggregate to the MAC layer for transmission. In our evaluation we show that end-to-end transmission delay is reduced by as much as 80% under heavy traffic loads. Additionally, we show as much as a 50% reduction in transmission energy con-sumption with an overall negative header overhead. Theoretical analysis, simulation, and a test-bed imple-mentation on Berkeley’s MICA motes are provided to validate our claims.

1. Introduction

Wireless Sensor Networks have emerged as a new information-gathering paradigm based on the collabora-tive effort of a large number of sensing nodes. In such networks, nodes deployed in a remote environment must self-configure without any a priori information about the network topology or global view. Nodes will act in response to environmental events and relay collected and possibly aggregated information through the formed multi-hop wireless network in accordance with desired system functionality. The inherently dynamic and distributed behavior of these networks, coupled with inherent physical limitations such as small instruction and data memory, constrained energy resources, short communication radii, and a low bandwidth medium in which to communicate, make developing communication protocols difficult.

Research on hardware for such devices has taken place at Berkeley [14][32][34] and various other research institutions [26] throughout the world. Using such hardware as a basis for development, the software architec-ture and communication stack residing on these devices are built taking into consideration the prolific research in the areas of ad-hoc networking [10][15][17][20], data aggregation [16][21][28], cluster formation [27], dis-tributed services [22], group formation [6], channel contention [3][5][7][19], and power conservation [4][12]. Work targeted to these devices include research in query processing (e.g. TinyDB [25]), and aggregation (e.g. TAG [24]). Work on the utility of such innovative technologies has unearthed potential applications includ-ing, event tracking [1], environmental monitoring, disaster relief, and search and rescue.

In this work, we address the problems of low bandwidth and energy limitations inherent to sensor devices. These networks’ ever-changing and unpredictable state demands a self-configuring, adaptive solution. We

1 This work was supported in part by NSF grant CCR-0098269, the MURI award N00014-01-1-0576 from

ONR, and the DAPRPA IXO offices under the NEST project (grant number F336615-01-C-1905).

ACM Transaction on Embedded Computing System

2

develop a novel adaptive application independent data aggregation (AIDA) component that fits seamlessly into current sensor network communication stack. Our goal is to maximize utilization of the communication chan-nel (single frequency) with energy savings coming as an ancillary benefit. With significant costs incurred from channel contention, packet header overhead, and data padding for fixed sized packets, this work abates such costs by employing varying degrees of data aggregation at forwarding nodes in accordance with current local traffic patterns.

Data aggregation techniques have been extensively investigated in recent literature. Our work, as a novel data aggregation approach, distinguishes itself from current state of the art solutions in three respects. First, prior Application Dependent Data Aggregation (ADDA shown in Figure 1b) relies on application layer infor-mation and must have a bi-directional interface, and therefore dependence with, the data centric routing protocol implemented. AIDA isolates aggregation decisions from application specifics by performing adaptive aggregation in an intermediate layer that resides between the traditional data-link and network layer protocols (Figure 1.a). This component is generalized enough to be utilized over a wide range of applications (data types) without incurring the costs of rewriting components to support application-specific logics. Second, no prior work in data aggregation adapts itself to the traffic situation in a time sensitive manner. AIDA takes the timely delivery of messages as well as protocol overhead into account to adaptively adjust aggregation strategies in accordance with assessed traffic conditions and expected sensor network requirements. Simulation results show that AIDA can adapt to varying traffic situations and dramatically reduce network congestion and trans-mission energy consumption. Third, previous data aggregation schemes (e.g., data centric routing [16]) per-form in-network processing to reduce the amount of application data transmitted. These in-network processes (e.g. averaging) can achieve higher degrees of aggregation; however data are less available to the application (e.g. standard deviation of the data set can not obtained from the average). In contrast, AIDA performs loss-less aggregation allowing the upper layer to decide whether information compression is appropriate at the time. Very important, our design enables AIDA to remain complementary to other data aggregation strategies (Figure 1.c) while providing significant timesaving benefits in the lower layers of the communication stack.

Figure 1: Architectural Designs

This paper attempts to address the aforementioned problems through a novel adaptive time sensitive data aggregation component. As an introduction to sensor networks, and to provide a more in depth discussion of the type of research taking place within this field, we begin section 2 with a discussion of related and ongoing work. Section 3 addresses the need for adaptation, data aggregation, and real-time data delivery. Section 4 then presents specific details about our protocol. Sections 5 and 6 describe our simulation environment, the type of experiments run, and a discussion of the results we obtain in both simulation and in the Berkeley MICA test-bed. Finally, we conclude in Section 7.

2. Leveraging Previous Work

Efforts to maximize channel utilization have been spread across various layers of the sensor network communication stack. Starting at the MAC layer, these include attempts to minimize collisions through con-tention-based mechanisms designed for a lossy wireless medium. Such work includes 802.11 [3], MACA [19], MACAW [5], FAMA[7], S-MAC[31], and Multi-Hop Scheduling [18], to name a few. All of these solutions

Range-Free Localization Schemes for Large Scale Sensor Networks1

Tian He, Chengdu Huang, Brian M. Blum, John A. Stankovic, Tarek Abdelzaher

Department of Computer Science, University of Virginia

151 Engineer's Way, P.O. Box 400740, Charlottesville, Virginia 22904-4740, USA

{TianHe, ch4pp, bmb5v, Stankovic, Zaher}@cs.virginia.edu

ABSTRACT Wireless Sensor Networks have been proposed for a multitude of location-dependent applications. For such systems, the cost and limitations of the hardware on sensing nodes prevent the use of range-based localization schemes that depend on absolute point-to-point distance estimates. Because coarse accuracy is sufficient for most sensor network applications, solutions in range-free localization are being pursued as a cost-effective alternative to more expensive range-based approaches. In this paper, we present APIT, a novel localization algorithm that is range-free. We show that our APIT scheme performs best when an irregular radio pattern and random node placement are considered, and low communication overhead is desired. We compare our work via extensive simulation, with three state-of-the-art range-free localization schemes to identify the preferable system configurations of each. In addition, we study the effect of location error on routing and tracking performance. We show that routing performance and tracking accuracy are not significantly affected by localization error when the error is less than 0.4 times the communication radio radius.

Categories and Subject Descriptors C.2. [Computer Communication Networks]: Network Protocols

General Terms Algorithms, Performance, Design

Keywords Localization, Positioning, Location discovery, Sensor Networks1

1

1. INTRODUCTION Wireless Sensor Networks (WSNs) distinguish themselves from

other traditional wireless or wired networks through sensor and actuator based interaction with the environment. Such networks have been proposed for various applications including search and rescue, disaster relief, target tracking, and smart environments. The inherent characteristics of these sensor networks make a node’s location an important part of their state. For such networks, location is being used (i) to identify the location at which sensor readings originate, (for example, identifying a target’s position during tracking, providing the location of an earthquake survivor buried underneath rubble during disaster relief operations) (ii) in novel communication protocols that route to geographical areas instead of IDs ([13], [16], [17], [19], [35]), and (iii) when providing other location based services ( such as sensing coverage [36] and location directory service [20] to provide medical information about a nearby patient in a smart hospital). In addition to the applications and protocols mentioned, continued research in WSNs will serve to invent and identify many additional protocols and applications, many of which will likely depend on location aware sensing devices.

Many localization algorithms for sensor networks have been proposed to provide per-node location information. With regard to the mechanisms used for estimating location, we divide these localization protocols into two categories: range-based and range-free. The former is defined by protocols that use absolute point-to-point distance estimates (range) or angle estimates for calculating location. The latter makes no assumption about the availability or validity of such information. Because of the hardware limitations of WSN devices, solutions in range-free localization are being pursued as a cost-effective alternative to more expensive range-based approaches.

This paper makes three major contributions to the localization problem in WSNs. First, we propose a novel range-free algorithm, called APIT, with enhanced performance under realistic system configurations. Second, though many different protocols [3][22][26] have been proposed to solve the localization problem in a range-free context, no prior work has been done to compare them in realistic settings. This paper is the first to provide a realistic and detailed quantitative comparison of existing range-free algorithms to determine the system configurations under which each is optimized. We perform such a study to serve as a guide for future research. Third, no attempt has previously been made to broadly study the impact of location error on various location-dependent applications and protocols. This paper

This work was supported in part by NSF grant CCR-0098269, the MURI award N00014-01-1-0576 from ONR, and the DAPRPA IXO office under the NEST project (grant number F336615-01-C-1905). Permission to make digital or hard copies of all or part of this work for personal or classroom use is granted without fee provided that copies are not made or distributed for profit or commercial advantage and that copies bear this notice and the full citation on the first page. To copy otherwise, or republish, to post on servers or to redistribute to lists, requires prior specific permission and/or a fee. MobiCom ’03, September 14-19, 2003, San Diego, California, USA Copyright 2003 ACM 1-58113-753-2/03/0009…$5.00.

provides insight into the effect of localization accuracy on application performance degradation and identifies bounds on the estimation error tolerated by applications.

The remainder of the paper is organized as follows: Section 2 discusses previous work in localization for sensor networks. Section 3 describes APIT. Section 4 gives brief descriptions of three other state-of-the-art range-free protocols to which we compare our work. Section 5 describes our simulation. Section 6 follows with a detailed performance comparison of the four range-free localization algorithms described. Section 7 further investigates the impact of localization error on various location-dependent applications and protocols such as routing and target tracking. Finally, we conclude in Section 8.

2. STATE OF THE ART Many existing systems and protocols attempt to solve the

problem of determining a node’s location within its environment. The approaches taken to solve this localization problem differ in the assumptions that they make about their respective network and device capabilities. These include assumptions about device hardware, signal propagation models, timing and energy requirements, network makeup (homogeneous vs. heterogeneous), the nature of the environment (indoor vs. outdoor), node or beacon density, time synchronization of devices, communication costs, error requirements, and device mobility. In this section, we discuss prior work in localization with regard to these network characteristics, device restrictions, and application requirements. We divide our discussion into two subsections where we present both range-based and range-free solutions..

2.1 Range-Based Localization Schemes Time of Arrival (TOA) technology is commonly used as a

means of obtaining range information via signal propagation time. The most basic localization system to use TOA techniques is GPS [33]. GPS systems require expensive and energy-consuming electronics to precisely synchronize with a satellite’s clock. With hardware limitations and the inherent energy constraints of sensor network devices, GPS and other TOA technology present a costly solution for localization in wireless sensor networks.

The Time Difference of Arrival (TDOA) technique for ranging (estimating the distance between two communicating nodes) has been widely proposed as a necessary ingredient in localization solutions for wireless sensor networks. While many infrastructure-based systems have been proposed that use TDOA [1][12][28], additional work such as AHLos ([30][31]) has employed such technology in infrastructure-free sensor networks. Like TOA technology, TDOA also relies on extensive hardware that is expensive and energy consuming, making it less suitable for low-power sensor network devices. In addition, TDOA techniques using ultrasound require dense deployment (numerous anchors distributed uniformly) as ultrasound signals usually only propagate 20-30 feet.

To augment and complement TDOA and TOA technologies, an Angle of Arrival (AOA) technique has been proposed that allows nodes to estimate and map relative angles between neighbors [27]. Similar to TOA and TDOA, AOA estimates require additional hardware too expensive to be used in large scale sensor networks.

Received Signal Strength Indicator (RSSI) technology such as RADAR [1] and SpotOn [15] has been proposed for hardware-constrained systems. In RSSI techniques, either theoretical or empirical models are used to translate signal strength into distance estimates. For RF systems [1][15], problems such as multi-path fading, background interference, and irregular signal propagation characteristics (shown in an empirical study of this technology [10]) make range estimates inaccurate. Work to mitigate such errors such as robust range estimation ([11]), two-phase refinement positioning ([29], [31]), and parameter calibration ([34]) have been proposed to take advantage of averaging, smoothing, and alternate hybrid techniques to reduce error to within some acceptable limit. While solutions based on RSSI have demonstrated efficacy in simulation and in a controlled laboratory environment, the premise that distance can be determined based on signal strength, propagation patterns, and fading models remains questionable, creating a demand for alternate localization solutions that work independent of this assumption.

2.2 Range-Free Localization Schemes In sensor networks and other distributed systems, errors can

often be masked through fault tolerance, redundancy, aggregation, or by other means. Depending on the behavior and requirements of protocols using location information, varying granularities of error may be appropriate from system to system. Acknowledging that the cost of hardware required by range-based solutions may be inappropriate in relation to the required location precision, researchers have sought alternate range-free solutions to the localization problem in sensor networks.

In [3], a heterogeneous network containing powerful nodes with established location information is considered. In this work, anchors beacon their position to neighbors that keep an account of all received beacons. Using this proximity information, a simple centroid model is applied to estimate the listening nodes’ location. We refer to this protocol as the Centroid algorithm.

An alternate solution, DV-HOP [26] assumes a heterogeneous network consisting of sensing nodes and anchors. Instead of single hop broadcasts, anchors flood their location throughout the network maintaining a running hop-count at each node along the way. Nodes calculate their position based on the received anchor locations, the hop-count from the corresponding anchor, and the average-distance per hop; a value obtained through anchor communication. Like DV-Hop, an Amorphous Positioning algorithm proposed in [22] uses offline hop-distance estimations, improving location estimates through neighbor information exchange.

These range-free techniques are described in more depth in section 4, and are used in our analysis for comparison with our work.

3. APIT LOCALIZATION SCHEME In this section, we describe our novel area-based range-free

localization scheme, which we call APIT. APIT requires a heterogeneous network of sensing devices where a small percentage of these devices (percentages vary depending on network and node density) are equipped with high-powered transmitters and location information obtained via GPS or some

Range-Free Localization and Its Impact on Large Scale

Sensor Networks

Tian He, Chengdu Huang, Brian M. Blum, John A. Stankovic, Tarek Abdelzaher

ABSTRACT With the proliferation of location dependent applications in sensor

networks, location awareness becomes an essential capability of

sensor nodes. Because coarse accuracy is sufficient for most sensor

network applications, solutions in range-free localization are being

pursued as a cost-effective alternative to more expensive range-

based approaches. In this paper, we present APIT, a novel

localization algorithm that is range-free. We show that our APIT

scheme performs best when an irregular radio pattern and random

node placement are considered, and low communication overhead

is desired. We compare our work via extensive simulation, with

three state-of-the-art range-free localization schemes to identify the

preferable system configurations of each. In addition, we provide

insight into the impact of localization accuracy on various location

dependent applications and suggestions on improving their

performance in the presence of such inaccuracy.

1. INTRODUCTION

Sensor networks have been proposed for various

applications including search and rescue, disaster relief,

target tracking, and smart environments. The inherent

characteristics of these sensor networks make a node’s

location an important part of their state. For such networks,

location is being used to identify the location at which sensor

readings originate, (for example, identifying a target’s

position during tracking, providing the location of an

earthquake survivor buried underneath rubble). It is also

used in communication protocols that route to geographical

areas instead of IDs ([18][19][21][37]), and in other location

based services, such as sensing coverage [38] and location

directory service [22]. In addition to the applications and

protocols mentioned, continued research in WSNs will serve

to invent and identify many additional protocols and

applications, many of which will likely depend on location

aware sensing devices.

Many localization algorithms for sensor networks have

been proposed to provide per-node location information.

With regard to the mechanisms used for estimating location,

we divide these localization protocols into two categories:

range-based and range-free. The former is defined by

protocols that use absolute point-to-point distance estimates

(range) or angle estimates for calculating location. The latter

makes no assumption about the availability or validity of

such information. Because of the hardware limitations of

WSN devices, solutions in range-free localization are being

pursued as a cost-effective alternative to more expensive

range-based approaches.

This paper makes three major contributions to the

localization problem in WSNs. First, we propose a novel

range-free algorithm, called APIT, with enhanced

performance under realistic system configurations. Second,

though many different protocols [4][24][28] have been

proposed to solve the localization problem in a range-free

context, no prior work has been done to compare them in

realistic settings. This paper is the first to provide a realistic

and detailed quantitative comparison of existing range-free

algorithms to determine the system configurations under

which each is optimized. We perform such a study to serve

as a guide for future research. Third, no attempt has

previously been made to broadly study the impact of location

error on various location-dependent applications and

protocols. This paper provides insight into the effect of

localization accuracy on applications and suggestions on

how to improve their performance in the presence of such

inaccuracy.

The remainder of the paper is organized as follows:

Section 2 discusses previous work in localization for sensor

networks. Section 3 describes APIT. Section 4 gives brief

descriptions of three other state-of-the-art range-free

protocols to which we compare our work. Section 5

describes our simulation. Section 6 follows with a detailed

performance comparison of the four range-free localization

algorithms described. Section 7 further investigates the

impact of localization error on various location-dependent

applications and protocols such as routing and target

tracking. Finally, we discuss future work in Section 8 and

conclude in Section 9.

2. STATE OF THE ART

Many existing systems and protocols attempt to solve the

problem of determining a node’s location within its

environment. The approaches taken to solve this localization

problem differ in the assumptions that they make about their

respective network and device capabilities. These include

assumptions about device hardware, signal propagation

models, timing and energy requirements, network makeup

(homogeneous vs. heterogeneous), the nature of the

environment (indoor vs. outdoor), node or beacon density,

time synchronization of devices, communication costs, error

requirements, and device mobility. In this section, we

discuss prior work in localization with regard to these

characteristics. We divide our discussion into two

subsections where we present both range-based and range-

free solutions.

2.1 Range-Based Localization Schemes

Time of Arrival (TOA) technology is commonly used as a

means of obtaining range information via signal propagation

time. The most basic localization system to use TOA

techniques is GPS [35]. GPS systems require expensive and

energy-consuming electronics to precisely synchronize with

a satellite’s clock. With hardware limitations and the

inherent energy constraints of sensor network devices, GPS

and other TOA technology present a costly solution for

localization in wireless sensor networks.

The Time Difference of Arrival (TDOA) technique for

ranging (estimating the distance between two

communicating nodes) has been widely proposed as a

necessary ingredient in localization solutions for wireless

sensor networks. While many infrastructure-based systems

have been proposed that use TDOA [1][13][30], additional

work such as AHLos ([32][33]) has employed such

technology in infrastructure-free sensor networks. Like TOA

technology, TDOA also relies on extensive hardware that is

expensive and energy consuming, making it less suitable for

low-power sensor network devices. In addition, TDOA

techniques using ultrasound require dense deployment

(numerous anchors distributed uniformly) as ultrasound

signals usually only propagate 20-30 feet.

To augment and complement TDOA and TOA

technologies, an Angle of Arrival (AOA) technique has been

proposed that allows nodes to estimate and map relative

angles between neighbors [29]. Similar to TOA and TDOA,

AOA estimates require additional hardware too expensive to

be used in large scale sensor networks.

Received Signal Strength Indicator (RSSI) technology

such as RADAR [1] and SpotOn [17] has been proposed for

hardware-constrained systems. In RSSI techniques, either

theoretical or empirical models are used to translate signal

strength into distance estimates. For RF systems [1][17],

problems such as multi-path fading, background interference,

and irregular signal propagation characteristics (shown in an

empirical study of this technology [11]) make range

estimates inaccurate. Work to mitigate such errors such as

robust range estimation ([12]), two-phase refinement

positioning ([31], [33]), and parameter calibration ([36])

have been proposed to take advantage of averaging,

smoothing, and alternate hybrid techniques to reduce error to

within some acceptable limit. While solutions based on

RSSI have demonstrated efficacy in simulation and in a

controlled laboratory environment, the premise that distance

can be determined based on signal strength, propagation

patterns, and fading models remains questionable, creating a

demand for alternate localization solutions that work

independent of this assumption.

2.2 Range-Free Localization Schemes

In sensor networks and other distributed systems, errors

can often be masked through fault tolerance, redundancy,

Feedback Control of Data Aggregation in Sensor Networks ∗

Tarek Abdelzaher, Tian He, John StankovicDepartment of Computer Science, University of Virginia, Charlottesville, VA 22903

{zaher,tianhe,stankovic}@cs.virginia.edu

AbstractSensor networks have recently emerged as a new paradigm for

distributed sensing and actuation. This paper describes funda-mental performance trade-offs in sensor networks and the utility ofsimple feedback control mechanisms for distributed performanceoptimization. A data communication and aggregation frameworkis presented that manipulates the degree of data aggregation tomaintain specified acceptable latency bounds on data delivery whileattempting to minimize energy consumption. An analytic model isconstructed to describe the relationships between timeliness, en-ergy, and the degree of aggregation, as well as to quantify con-straints that stem from real-time requirements. Feedback controlis used to adapt the degree of data aggregation dynamically inresponse to network load conditions while meeting applicationdeadlines. The results illustrate the usefulness of feedback con-trol in the sensor network domain.

1. Introduction

The work reported in this paper is motivated by the rapid emer-gence of sensor networks [4] as a new paradigm for writing dis-tributed applications. These networks are composed of a largenumber of small wireless sensor devices, each equipped with lim-ited processing, communication, and storage capacity. Sensor net-works are especially useful in applications involving a poorly ac-cessible, dangerous, or unfriendly environment, where it is diffi-cult to provide a fixed monitoring infrastructure. Instead, a myriadof wireless sensor devices can be deployed (e.g., by air-droppingfrom a UAV) for remote monitoring and surveillance purposes.Such air-dropped networks are called ad hoc sensor networks todistinguish them from other types of sensor networks where nodesare laid out in some fixed predetermined pattern. Ad hoc wire-less sensor networks present the most challenge to the researchcommunity due to their inherent lack of structure. Example ap-plications include habitat monitoring, defense, border control, and

∗The work reported in this paper was supported in part by theNational Science Foundation under grants CCR-0093144, ANI-0105873, EHS-0208769, and MURI N00014-01-1-0576.

emergency response systems.

This paper introduces fundamental research challenges presentedby ad hoc wireless sensor networks from a feedback control per-spective. These challenges lie in optimally reconciling the fun-damental performance trade-offs that underlie network operation.Hence, the purpose of feedback control in this paper is not to con-trol the dynamics of an external environment, but rather to con-trol network performance itself. At a high level, performance ofa sensor network can be viewed as a point in a three-dimensionalspace. These dimensions are (i) timeliness, (ii) energy consump-tion, and (iii) information output. It is desired to minimize energyconsumption and maximize information output while maintainingtimeliness. These requirements are mutually at odds; communicat-ing more information takes more time and consumes more energy.

The nature of the trade-off among the basic sensor network per-formance requirements depends on current network input, whichis the amount of sensory data infused into the network. For ex-ample, at low network load, timeliness can be easily achieved to-gether with the other requirements. However, at a higher load, adecision has to be made between timeliness of delivery and theamount of deliverable information. Feedback control loops areneeded to trade-off these performance requirements dynamicallyin a distributed fashion in response to current network conditions,essentially solving a distributed constrained optimization problem.This paper describes an instance of such a feedback control archi-tecture, and derives some results in real-time computing that helpquantify the constraints imposed on optimization.

The performance trade-offs mentioned above are fundamen-tally inherent to ad hoc sensor networks because they invariablyarise from the main goal of such networks, namely the collectionof sensory data. The most important output of a sensor network isthe information it provides to external observers. One of the mostlimited resources in an ad hoc network is battery capacity. Thisis partly because advances in battery capacity have developed at aslower rate than advances in processing and communication band-width. Moreover, since the network is typically deployed in re-mote or harsh environments, changing batteries is quite expensiveif not infeasible. Hence, maximizing battery lifetime by conserv-ing energy is a predominant concern.

Omni-directional communication is the most energy-consumingoperation in a sensor network due to the high degree of signal at-tenuation and the multipath phenomena that occur when wirelesssensors are placed on the ground. Directional communication re-mains a big challenge since it requires sensors to know to a highdegree of accuracy both their own position and orientation, as wellas that of their neighbors. The fundamentally high cost of com-

munication results in an important trade-off between the amountof information that the network delivers and its lifetime. A goodcompromise to conserve battery capacity is to perform appropri-ate aggregation on collected data to reduce the amount of networkcommunication without much reduction in the information deliv-ered.

Timeliness of delivery is a fundamental performance concernin sensor networks because such networks must react to externalphenomena in real-time. An unbounded delay in the loop is un-acceptable. From an application’s perspective, discovering an in-truding target too late is not useful for producing an effective re-sponse. Timeliness is generally at odds with energy consumption.If it is possible to delay delivery until more data can be aggregated,overhead can be saved and delivery energy can be reduced. Whilelimited aggregation (or batching) may actually improve the over-all timeliness by reducing total traffic, additional aggregation willimpair timing performance due the introduced aggregation delay.The break-even point depends on the amount of data currently gen-erated, which is a dynamic quantity that depends on activities inthe environment. The timeliness-energy trade-off therefore opensa realm of opportunity for feedback control research in the sensornetwork domain.

An important consideration in the design of a feedback per-formance control framework for sensor networks is to quantifythe fundamental constraints within which each sensor node op-erates to solve the global performance optimization problem. Inthe three-dimensional trade-off space introduced earlier, the basicconstraints are on energy, time, and information content. To ad-dress the timing constraint, in this paper, we describe importantrecent results in real-time computing theory that quantify the abil-ity of the network to communicate data in real-time. We relateglobal timing requirements to the local amount of traffic that canbe processed by each node.

As a specific instance of performance control in sensor net-works, this paper describes a data communication and aggrega-tion framework that manipulates the degree of data aggregationto maintain specified acceptable latency bounds on data deliverywhile attempting to minimize energy consumption. An analyticmodel is constructed to describe the relationships between timeli-ness, energy, and the degree of aggregation, as well as to quantifyconstraints that stem from real-time requirements. Feedback con-trol is used to adapt the degree of data aggregation dynamicallyin response to network load conditions while meeting applicationdeadlines.

The rest of this paper is organized as follows. Section 2 de-scribes the problem statement and the general architecture of ourservice, which is based on two types of data aggregation; lossyand lossless. Section 3 derives an expression for real-time sys-tem capacity that quantifies the amount of information that can bedelivered through the network by the deadlines. This bound is afundamental design constraint that must be enforced by the feed-back control architecture. Section 4 investigates feedback controlof lossy aggregation. It describes the conditions under which sys-tem capacity is maximized, and describes a feedback scheme thatoptimizes capacity subject to time constraints by adjusting the de-gree of aggregation. Section 5 describes local optimization usingfeedback control of lossless aggregation. Section 6 presents a brief

performance evaluation. The paper concludes with final observa-tions and open questions in Section 7.

2. Problem Statement and Architecture

We consider a real-time sensor network where sensory measure-ments should be delivered to their destinations within specifiedtime constraints. Data is divided into multiple classes. Each classis associated with a bound on delivery time. For example, mo-tion sensor measurements might have to be delivered within 3 sec-onds to allow real-time tracking of moving targets. In contrast,temperature measurements could be delivered within 30 seconds,since they exhibit slower dynamics. It is desired to deliver all dataat the minimum energy cost while satisfying all time constraints.Since the environment is dynamic, the amount of data generatedat any time is unpredictable and can vary considerably from timeto time. We assume that some sensors report their measurementsperiodically at all times, while others become active only whentriggered by environmental events. For example, flurries of activ-ity in the monitored environment may generate a burst of motionsensor readings. These sensors will be silent when the environ-ment is quiescent. Since the network load is dynamic, overloadmay occur which can significantly increase communication delay,possibly making it infeasible to deliver data in time. A feedbackmechanism is needed to control network delay such that time con-straints are met.

The main actuator “knob” that can be manipulated in our sys-tem is the degree of data aggregation. In contention-based MediumAccess Control (MAC) communication protocols, packing datainto larger units reduces the chances of packet collisions, hencereducing energy expenditure and improving delay.1 Two differenttypes of aggregation are possible; namely, lossless aggregation andlossy aggregation. Lossless aggregation refers to concatenating in-dividual data items into larger packets, thus amortizing per-packetprotocol overhead. In this case, no data is lost. The approachis especially effective in sensor networks where individual sensorreadings are small in size, leaving much room for concatenation.Another type of overhead that can be amortized is the local hand-shake performed ahead of per-hop data transmission to reserve thechannel. This handshake is common to contention-based wirelessMAC protocols such as 802.11.

Lossless aggregation is effective if the load on the system isnot excessive. If the total communication load approaches systemcapacity, the amount of communicated data must be forcibly re-duced. We call the latter case, lossy aggregation. This technique isalso useful for energy saving, even when the system is not heavilyloaded. The best example of lossy aggregation is the averagingof sensor values. Averaging is a natural choice in many applica-tions. For example, a user may need to know only of the averagetemperature in a region, as opposed the individual readings of allsensors. Similarly, it may be enough to report only the averageestimated location of a target, as opposed to the exact locationsof all triggered motion sensors. Lossy aggregation can be eitherspatial or temporal. In the former, data is averaged from multiplesensors, while in the latter, data from the same sensor is averaged

1Due to the difficulty in synchronizing clocks across all nodes in asensor network, slot-based communication protocols are less prac-tical than contention-based ones.

A Communication Architecture and Programming Abstractions for Real-TimeEmbedded Sensor Networks

�

T. Abdelzaher, J. Stankovic, S. Son, B. Blum, T. He, A. WoodDepartment of Computer Science

University of VirginiaCharlottesville, VA 22904

Chenyang LuDepartment of Computer Science

University of Washington, St. LouisSt. Louis, MO 63130

AbstractData distribution in embedded real-time sensor networks re-quires new protocols and programming environments thatachieve time-sensitive message delivery and provide useful ab-stractions to the application programmer. Attainment of thesegoals requires changes to multiple layers of the communica-tion protocol stack. In this paper, we review a protocol suitedeveloped by the authors for data communication in embed-ded sensor networks. It takes into account time constraints andexports attribute-based connections that are tightly integratedwith properties of the monitored environment. A programminglanguage is described that allows external physical objects tobe represented as first class abstractions in the computing sys-tem. The language facilitates writing monitoring applications.The system was implemented on a prototypical sensor networkbased on MICA motes.

Keywords: sensor networks, programming paradigms, track-ing, QoS, distributed systems.

1 Introduction

Ad hoc wireless sensor networks, made possible by advances incommunication technology and hardware miniaturization [11],raise the need for a new suite of communication protocols andnew programming abstractions for distributed deeply embed-ded computing. Such sensor networks are especially usefulwhen an inhospitable, poorly accessible, or delicate environ-ment prevents the installation of needed computing infrastruc-ture. An example could be the site of a natural disaster or atarget behind enemy lines. Instead, myriads of tiny computa-tionally equipped wireless sensor devices may be dropped toform an ad hoc network that operates autonomously to monitorits surroundings, react to distributed events, or alert appropriateauthorities when specific activities are observed.

�The work reported in this paper was supported in part by the National Sci-

ence Foundation under grants CCR-02-05327 and CCR-00-92945, by DARPAunder grant F33615-01-C-1905, and by MURI under grant N00014-01-1-0576.

Sensor networks offer new challenges both from the per-spective of building communication protocols and from theperspective of developing appropriate programming models.These challenges arise due to their large scale, autonomousoperation, massively parallel interactions with a spatially dis-tributed physical environment, and a more stringent set of re-source constraints.

Communication protocols for sensor networks must providereal-time assurances. While ensuring proper timing behaviorof systems has been a topic of real-time research for decades,sensor network applications offer physical space in addition totime, as a new dimension for interaction with the environment.Hence, while traditional real-time computing research has beenconcerned with meeting time constraints, a new branch of the-ory is needed to analyze systems that interact with the theirsurroundings both in real time and in the real dimensions ofphysical space. For example, in a network that tracks vehiclesthrough the sensor field, the application must collect sensorymeasurements in real-time from the actual changing locale inwhich the vehicle is detected. Message communication musttherefore be sensitive to both time and distance constraints,which may depend on external factors such as the physicalspeed of the monitored vehicle. In this paper, we describe aprotocol suite in which both time and distance constraints areaddressed.

A new programming paradigm is needed to facilitate thetask of sensor network application development. Due to thelarge scale of sensor networks, programmers should not haveto concern themselves with low-level abstractions and func-tions such as creating and destroying individual connectionsbetween pairs of nodes. Instead, the programming environ-ment must offer a conceptual view in which global tasks canbe defined in an abstract manner, leaving it for the underlyingsystem to translate them into computational and communica-tion activities on individual sensor nodes. This paper reportson the design of a programming system developed on top ofour communication protocol suite, which provides the requiredhigh-level abstractions. The language allows external events inthe environment to be represented as objects in the computing

system facilitating the monitoring of such events by the appli-cation. The reported architecture is a part of an ongoing re-search effort on developing a sensor network virtual machinefor future distributed deeply embedded applications.

The rest of this paper is organized as follows. In Section 2,we describe a protocol suite that takes into account time andspace constraints, and exports a useful transport-layer abstrac-tion in which logical communication end-points can be asso-ciated with tracked objects in the external environment. Sec-tion 3 describes a new programming model for sensor networkswhich builds upon the aforementioned transport protocol to el-evate environmental objects into first class programming ab-stractions. Related work is summarized in Section 4. The paperconcludes with Section 5 which describes some of the remain-ing challenges and directions for future research.

2 A Protocol Suite for Sensor NetworksCommunication protocols in sensor networks are the funda-mental cornerstone that glues distributed applications together.The deeply embedded nature of sensor networks presents someof the most interesting challenges in the design of their com-munication protocols. New research topics span all protocolstack layers, primarily motivated by a tighter interaction be-tween the network and its physical environment. At the MAClayer, new protocols are needed that enforce message priori-ties consistently with time and distance constraints that arisefrom environmental interactions [22]. Awareness of the phys-ical environment must also be incorporated into the networklayer. For example, location should be an essential attributeof addressable networked objects [15]. Location-assisted rout-ing protocols such as LAR [19] and DREAM [4], as well aslocation services [21] have been described for ad hoc wirelessnetworks. More generally, routing algorithms are needed inwhich destinations are described implicitly by their environ-mental attributes. For example, directed diffusion [18, 14] andthe intentional naming system [3] provide addressing and rout-ing based on data interests. A fundamental rethinking of basicprotocols is required at the transport layer as well. Individ-ual socket-style connections between nodes are too low-levelto be a useful abstraction for the programmer. They must be re-placed with higher-level alternatives that are more suitable forthe main purpose of sensor networks, namely monitoring theexternal surroundings in which they are embedded.

This section describes our answer to the challenge of in-corporating environmental awareness into the design of sensornetwork communication protocols. Our protocol stack featurestwo important contributions. First, it implements new real-timemessage scheduling algorithms in which both time and physi-cal distance requirements are observed. Second, it exports atransport-layer address space that associates unique networkaddresses with external environmental objects. The new ad-dresses serve as connection end-points, thereby raising the levelof connection abstraction to entities of direct interest to the ap-

plication. The layers of our protocol stack are described in thefollowing subsections.

2.1 Real-Time Distance-Aware Scheduling

Message communication in sensor networks must occur inbounded time, for example to prevent delivery of stale data onthe status of detected events or intruders. In general, a sensornetwork may simultaneously carry multiple messages of dif-ferent urgency, communicated among destinations that are dif-ferent distances apart. The network has the responsibility ofordering these messages on the communication medium in away that respects both time and distance constraints.

A protocol that achieves this goal in our architecture iscalled RAP [22]. It supports a notion of packet velocity andimplements velocity monotonic scheduling (VMS) as the de-fault packet scheduling policy on the wireless medium. Ob-serve that for a desired end-to-end latency bound to be met,an in-transit packet must approach its destination at an aver-age velocity given by the ratio of the total distance to be tra-versed to the requested end-to-end latency bound. RAP pri-oritizes messages by their required velocity such that highervelocities imply higher priorities. Two flavors of this algorithmare implemented. The first, called static velocity-monotonicscheduling, computes packet priority at the source and keepsit fixed thereafter regardless of the actual rate of progress ofthe packet towards the destination. The second, called dynamicvelocity-monotonic scheduling, adjusts packet priority en routebased on the remaining time and the remaining distance to des-tination. Hence, a packet’s priority will increase if it suffershigher delays on its path and decrease if it is ahead of schedule.

To achieve consistent prioritization in the wireless network,not only do we need priority queues at nodes, but also a MAClayer that resolves contention on the wireless medium in a man-ner consistent with message priorities. We adopt a schemesimilar to [1] to prioritize access to the wireless medium.The scheme is based on modifying two 802.11 parameters,namely the DIFS counter and the backoff window, such theyare priority-aware. The DIFS counter determines the maximumtime a node waits, after the communication channel becomesidle, prior to transmitting an RTS packet. The actual waitingtime is randomly chosen between 0 and DIFS. An approximateprioritization effect is achieved by letting the DIFS value de-pend on the priority of the outgoing packet at the head of thetransmission queue. A larger value is given to packets of lowerpriority. Hence, more urgent packets tend to contend on themedium more aggressively. The back-off window of 802.11increases the maximum waiting time when collisions occur.To give preferential treatment to higher priority packets, wemake this increase dependent on the priority of the head of thequeue. A higher increase is incurred for packets of lower pri-ority. Hence, collisions tend to be resolved in favor of higher-priority packets.

A detailed performance evaluation of this scheme can be

A Communication Architecture and Programming Abstractions for Real-TimeEmbedded Sensor Networks�

T. Abdelzaher, J. Stankovic, S. Son, B. Blum, T. He, A. WoodDepartment of Computer Science

University of VirginiaCharlottesville, VA 22904

Chenyang LuDepartment of Computer Science

University of Washington, St. LouisSt. Louis, MO 63130

Abstract

Data distribution in embedded real-time sensor networks re-quires new protocols and programming environments thatachieve time-sensitive message delivery and provide useful ab-stractions to the application programmer. Attainment of thesegoals requires changes to multiple layers of the communica-tion protocol stack. In this paper, we review a protocol suitedeveloped by the authors for data communication in embed-ded sensor networks. It takes into account time constraints andexports attribute-based connections that are tightly integratedwith properties of the monitored environment. A programminglanguage is described that allows external physical objects tobe represented as first class abstractions in the computing sys-tem. The language facilitates writing monitoring applications.The system was implemented on a prototypical sensor networkbased on MICA motes.

Keywords: sensor networks, programming paradigms, track-ing, QoS, distributed systems.

1 Introduction

Ad hoc wireless sensor networks, made possible by advances incommunication technology and hardware miniaturization [11],raise the need for a new suite of communication protocols andnew programming abstractions for distributed deeply embed-ded computing. Such sensor networks are especially usefulwhen an inhospitable, poorly accessible, or delicate environ-ment prevents the installation of needed computing infrastruc-ture. An example could be the site of a natural disaster or atarget behind enemy lines. Instead, myriads of tiny computa-tionally equipped wireless sensor devices may be dropped toform an ad hoc network that operates autonomously to monitorits surroundings, react to distributed events, or alert appropriateauthorities when specific activities are observed.

�The work reported in this paper was supported in part by the National Sci-ence Foundation under grants CCR-02-05327 and CCR-00-92945, by DARPAunder grant F33615-01-C-1905, and by MURI under grant N00014-01-1-0576.

Sensor networks offer new challenges both from the per-spective of building communication protocols and from theperspective of developing appropriate programming models.These challenges arise due to their large scale, autonomousoperation, massively parallel interactions with a spatially dis-tributed physical environment, and a more stringent set of re-source constraints.

Communication protocols for sensor networks must providereal-time assurances. While ensuring proper timing behaviorof systems has been a topic of real-time research for decades,sensor network applications offer physicalspacein addition totime, as a new dimension for interaction with the environment.Hence, while traditional real-time computing research has beenconcerned with meeting time constraints, a new branch of the-ory is needed to analyze systems that interact with the theirsurroundings both in real time and in the real dimensions ofphysical space. For example, in a network that tracks vehiclesthrough the sensor field, the application must collect sensorymeasurements in real-time from the actual changing locale inwhich the vehicle is detected. Message communication musttherefore be sensitive to both time and distance constraints,which may depend on external factors such as the physicalspeed of the monitored vehicle. In this paper, we describe aprotocol suite in which both time and distance constraints areaddressed.

A new programming paradigm is needed to facilitate thetask of sensor network application development. Due to thelarge scale of sensor networks, programmers should not haveto concern themselves with low-level abstractions and func-tions such as creating and destroying individual connectionsbetween pairs of nodes. Instead, the programming environ-ment must offer a conceptual view in which global tasks canbe defined in an abstract manner, leaving it for the underlyingsystem to translate them into computational and communica-tion activities on individual sensor nodes. This paper reportson the design of a programming system developed on top ofour communication protocol suite, which provides the requiredhigh-level abstractions. The language allows external events inthe environment to be represented as objects in the computing

system facilitating the monitoring of such events by the appli-cation. The reported architecture is a part of an ongoing re-search effort on developing a sensor network virtual machinefor future distributed deeply embedded applications.

The rest of this paper is organized as follows. In Section 2,we describe a protocol suite that takes into account time andspace constraints, and exports a useful transport-layer abstrac-tion in which logical communication end-points can be asso-ciated with tracked objects in the external environment. Sec-tion 3 describes a new programming model for sensor networkswhich builds upon the aforementioned transport protocol to el-evate environmental objects into first class programming ab-stractions. Related work is summarized in Section 4. The paperconcludes with Section 5 which describes some of the remain-ing challenges and directions for future research.

2 A Protocol Suite for Sensor Networks

Communication protocols in sensor networks are the funda-mental cornerstone that glues distributed applications together.The deeply embedded nature of sensor networks presents someof the most interesting challenges in the design of their com-munication protocols. New research topics span all protocolstack layers, primarily motivated by a tighter interaction be-tween the network and its physical environment. At the MAClayer, new protocols are needed that enforce message priori-ties consistently with time and distance constraints that arisefrom environmental interactions [22]. Awareness of the phys-ical environment must also be incorporated into the networklayer. For example, location should be an essential attributeof addressable networked objects [15]. Location-assisted rout-ing protocols such as LAR [19] and DREAM [4], as well aslocation services [21] have been described for ad hoc wirelessnetworks. More generally, routing algorithms are needed inwhich destinations are described implicitly by their environ-mental attributes. For example, directed diffusion [18, 14] andthe intentional naming system [3] provide addressing and rout-ing based on data interests. A fundamental rethinking of basicprotocols is required at the transport layer as well. Individ-ual socket-style connections between nodes are too low-levelto be a useful abstraction for the programmer. They must be re-placed with higher-level alternatives that are more suitable forthe main purpose of sensor networks, namely monitoring theexternal surroundings in which they are embedded.

This section describes our answer to the challenge of in-corporating environmental awareness into the design of sensornetwork communication protocols. Our protocol stack featurestwo important contributions. First, it implements new real-timemessage scheduling algorithms in which both time and physi-cal distance requirements are observed. Second, it exports atransport-layer address space that associates unique networkaddresses with external environmental objects. The new ad-dresses serve as connection end-points, thereby raising the levelof connection abstraction to entities of direct interest to the ap-

plication. The layers of our protocol stack are described in thefollowing subsections.

2.1 Real-Time Distance-Aware Scheduling

Message communication in sensor networks must occur inbounded time, for example to prevent delivery of stale data onthe status of detected events or intruders. In general, a sensornetwork may simultaneously carry multiple messages of dif-ferent urgency, communicated among destinations that are dif-ferent distances apart. The network has the responsibility ofordering these messages on the communication medium in away that respects both time and distance constraints.

A protocol that achieves this goal in our architecture iscalled RAP [22]. It supports a notion of packet velocity andimplements velocity monotonic scheduling (VMS) as the de-fault packet scheduling policy on the wireless medium. Ob-serve that for a desired end-to-end latency bound to be met,an in-transit packet must approach its destination at an aver-age velocity given by the ratio of the total distance to be tra-versed to the requested end-to-end latency bound. RAP pri-oritizes messages by their required velocity such that highervelocities imply higher priorities. Two flavors of this algorithmare implemented. The first, calledstatic velocity-monotonicscheduling, computes packet priority at the source and keepsit fixed thereafter regardless of the actual rate of progress ofthe packet towards the destination. The second, calleddynamicvelocity-monotonic scheduling, adjusts packet priorityen routebased on the remaining time and the remaining distance to des-tination. Hence, a packet’s priority will increase if it suffershigher delays on its path and decrease if it is ahead of schedule.

To achieve consistent prioritization in the wireless network,not only do we need priority queues at nodes, but also a MAClayer that resolves contention on the wireless medium in a man-ner consistent with message priorities. We adopt a schemesimilar to [1] to prioritize access to the wireless medium.The scheme is based on modifying two 802.11 parameters,namely the DIFS counter and the backoff window, such theyare priority-aware. The DIFS counter determines the maximumtime a node waits, after the communication channel becomesidle, prior to transmitting an RTS packet. The actual waitingtime is randomly chosen between 0 and DIFS. An approximateprioritization effect is achieved by letting the DIFS value de-pend on the priority of the outgoing packet at the head of thetransmission queue. A larger value is given to packets of lowerpriority. Hence, more urgent packets tend to contend on themedium more aggressively. The back-off window of 802.11increases the maximum waiting time when collisions occur.To give preferential treatment to higher priority packets, wemake this increase dependent on the priority of the head of thequeue. A higher increase is incurred for packets of lower pri-ority. Hence, collisions tend to be resolved in favor of higher-priority packets.

A detailed performance evaluation of this scheme can be

Demo Abstract: Electronic Tripwires for Power-Efficient Surveillance and Target Classification

Tian He, Qiuhua Cao, Liqian Luo, Ting Yan, Lin Gu, John Stankovic, Tarek Abdelzaher

Department of Computer Science, University of Virginia {tianhe,qc9b,ll4p,ty4k,lg6e,,stankovic,zaher}@cs.virginia.edu

Categories and Subject Descriptors

C.2.1 [Computer Communication Networks]: Network Architecture and Design

General Terms Design, Performance, Experimentation, Measurement

Keywords Sensor networks, Energy conservation, Tracking, Classification

1. DEMO DESCRIPTION1 This demonstration presents a mini-scale surveillance system developed at the University of Virginia that is currently undergoing a technology transition. The ultimate goal of this project is to develop a clandestinely deployable, environmentally rugged, robust, self-organizing, wireless sensor network for long-term, persistent surveillance, involving detection, tracking, classification, and identification of vehicle and personnel targets over various types of terrain.

One key research challenge for this type of system is to reconcile the need for network longevity (which requires most sensors to remain inactive) with the need for fast and accurate target detection and classification (which requires as many sensors to be awake as possible). In order to reconcile the two conflicting design goals, we developed a suite of services that includes (i) a tripwire-based power management with sentry service, (ii) a radio-based wakeup service, (iii) an entity-based tracking service, and (iv) a three-tier target classification scheme. These services work in conjunction with a localization and time synchronization protocol needed to support their operation.

The conceptual operation of the system is as follows: 1) The sentry service allows most motes to go to sleep, while a small fraction of them (the sentries) remain awake to watch for potential targets. It is ensured that sentries form a communication backbone to deliver important information quickly to long-distance relays in the network when needed. Sentries rotate periodically to balance power consumption.

1 This work was supported in part by NSF grant CCR-0098269, the MURI award N00014-01-1-0576 from ONR, and the DAPRPA IXO office under the NEST project (grant number F336615-01-C-1905).

Copyright is held by the author/owner(s). SenSys’04, November 3–5, 2004, Baltimore, Maryland, USA. ACM 1-58113-879-2/04/0011.

2) When a target enters the field, the wakeup service allows a portion of the network near the target to be quickly awakened. Once awake, motes perform target tracking and classification. With the help of localization and time synchronization, the tracking component computes target location and velocity. Classification is done through a three-tier architecture. In essence, at the first tier, sensors differentiate real targets from false alarms and provide different sensing modalities. At the second tier, sensor groups perform data fusion among adjacent nodes that detect the target concurrently. At the third tier, the base node takes current group-level reports as well as history into account to make final classification decisions. This information is sent to a command center through the nearest relay (a device capable of long-distance communication).

3) To scale the system and allow further power savings, a tripwire service can be activated which partitions the field into regions (called tripwires). Individual regions can be de-activated in which case they go to sleep (including their sentries). Active tripwire regions operate as described above, using sentries to look for targets and waking up the remaining nodes only when targets are detected. A target detected by an active tripwire is tracked and classified. Tracking continues across tripwire boundaries waking up nodes on the target's trajectory.

The complete system is designed to scale to at least 1000 XSM motes and cover at least 100,1000 square meters. In this demonstration, a field of 20 - 40 XSM motes is used to form a mini-size sensor network. Two laptops are used as relays. The network is configured into two tripwire regions, each reporting to a different laptop. Active tripwires are shown to perform sentry selection. Sentry nodes identify themselves by showing different colors on the laptop screen (perhaps using a projection device). The remaining nodes go to sleep. Tripwire configuration commands can also be sent if needed to turn off a tripwire for further energy savings. One or a set of (small) targets then enters the field. As the target crosses an active tripwire, the wake-up service is automatically invoked and non-sentry nodes are awakened in the vicinity of the target. These nodes flash their LEDs to indicate active tracking state. Information about the location, speed, and type of the target is promptly sent to the relay(s) and displayed on the screen. The experiment may be repeated with different targets, different configurations of tripwire regions, and different target paths through the network. When the target departs the network and no further activity is detected, nodes go back to sleep leaving only sentries (in active tripwires) to keep watch. After a period of continued inactivity, sentries are shown to rotate to balance power. Re-introduction of targets at any time repeats the wakeup and tracking scenario described above.

Feedback Control-based Dynamic Resource

Management in Distributed Real-Time Systems

Tian He ∗∗, John A. Stankovic ∗, Michael Marley ∗,

Chenyang Lu ∗ ∗ ∗, Yin Lu ∗ ∗ ∗ ∗ ∗, Tarek Abdelzaher ∗ ∗ ∗∗,

Sang Son ∗, Gang Tao ∗

Department of Computer Science, University of Virginia

151 Engineer’s Way, Charlottesville, Virginia 22904-4740

Abstract

The resource management in distributed real-time systems becomes increasingly unpre-

dictable with the proliferation of data-driven applications. Therefore, it is inefficient to

allocate the resources statically to handle a set of highly dynamic tasks whose resource re-

quirements (e.g., execution time) are unknown a prior. In this paper, we build a distributed

real-time system based on the control theory, focusing on the computational resource man-

agement. Specifically, this work makes three important contributions. First, it allows the

designer to specify the desired temporal behavior of system adaptation, such as the speed

of convergence. This is in contrast to previous literature, specifying only steady-state met-

rics, e.g. the deadline miss ratio. Second, unlike QoS optimization approaches, our solution

meets performance guarantees with no accurate knowledge of task execution parameters –

a key advantage in a poorly modeled environment. Last, in contrast to ad hoc algorithms

based on intuition and testing, we rigorously prove that our approach not only has excellent

steady state behavior, but also meets stability, overshoot, and settling time requirements.

Key words: Real-time, Feedback Control, Quality of Service, Scheduling

Preprint submitted to Elsevier Science 7 July 2006

1 Introduction

Distributed real-time systems are widely used in highly dynamic environments where

the resource requirements are open, fluctuating and not amenable to the traditional

worst-case real-time analysis. For example, a web farm can be used to distribute

time-sensitive contents such as movies and video clips. They need to handle a chang-

ing number of requests with significantly different resource requirements that are

unknown beforehand. In a stock market, a system needs to actively push real-time

stock updates at various interval to a group of users. The number of users served by

a server can change quickly over time. Although these systems differ significantly in

term of applications, they all operate in open environments where both workloads and

available resources are difficult to predict. Monitoring and feedback control are needed

to meet performance constraints. Several difficulties are observed in dynamic resource

management in these systems. One main difficulty lies in their data-dependent re-

source requirements, which cannot be predicted without interpreting input data. For

example, the execution time of an information server (a web or database server) heav-

ily depends on the content of requests, such as the particular web page requested. A

second major challenge is that these systems have highly uncertain arrival workloads;

it is not clear how many users will request some resource in the web. A third chal-

lenge involves the complex interactions among many distributed sites, often across

an environment with poor or unpredictable timing behavior. Consequently, develop-

ing certain types of future real-time systems will involve techniques for modeling the

unpredictability of the environment, handling imprecise or incomplete knowledge, re-

acting to overload and unexpected failures (i.e., those not expressed by design-time

failure hypotheses), and achieving the required performance levels and temporal be-

havior. We envision a trend in real-time computing to provide performance guaran-

∗ Dept of Computer Science, University of Virginia∗∗ Dept of Computer Science & Engineering, University of Minnesota∗ ∗ ∗ Dept of Computer Science & Engineering, Washington Uni. in St. Louis∗ ∗ ∗∗ Dept of Computer Science, University of Illinois, Urban-Champaign∗ ∗ ∗ ∗ ∗Dept of Computer Science & Engineering, University of Nebraska, Lincoln

2

�������������� ��������������������������� �� ���������� �����!"����#� �$�� ������%� �$���&� �� '(���)�� ��*+���-,+� �.$�� �� �

/�0�/1#� �$�� ������32 )546��*���7�8 9:8�;<��=5>�?�?�@�>�?����A ����%! �����B��)C'-� ! ! ����,+� �$�� �� �1D�D�E�?�>�F%>�@�>�?���&���G

��������H�IKJMLONQPKRTSBUVP�SBIKJMPWUVXY�IWJ[Z\NV]W]K^_`VNQX._IWPaSbSBcW`VdeN�SfLO_^_XNVJ[RSBcWJ[gWUQ_.^^NVPW`VU�hK_.X�_%S<PKUQ`VU�S3SBNVJ[R+XI�]KJ[IWgW_iKUkjlcK^^�SBUVP�SB_PKmn`QIKgoUVJ[NQmWUpXI�NSBUQ`VcKJ[_XRrqsSBUVP�SB_X_.goUpNQJ[UVNpY�dK_^UpNVX�XdKUASBNQLOU�X_LOU�Lt_PK_Lt_.uQ_PKm�UQPKUVJ[moR`VIWPaSBcWLO]WX_.IWPvNVPWivUVwoX.UQPKiW_.PWmxSBRCSBXUVLy^_.j�UVX_LOU{zWR|^UQgWUQJ[NVmo_.PWm}XdKUJ[UViWcKPWiKNVPWX~iKUQ]K^IKRrLtUVPKX~IWj�SBUQPaSBIWJ�PKIWiKU�S����[X~_%S�NV^%SBIA]KJ[UVj�UVJ[NVzW^U�jlIWJXdKU�SBUVP�SBIKJ�PKUVXY�IKJ[Z�XI�]KJ[IWgW_iWUniK_jlj�UVJ[UVPWX_.NQXUVi�S.cKJ[goUV_^^.NQPK`QU�SBUQJ[gW_`VUjlIWJ6goNVJ[_IKc�S�X.NQJ[mWUQX}NQJ[UVN�S�Y�_.Xd�iW_.j�jlUVJ[UQPKX}iWUVmoJ[UVU�S}IKj�SBUQ`VcKJ[_XRJ[UV�WcK_J[UVLtUVPWX%S�� �[P�XdK_%S�]WNV]WUVJsh:Y�Uf]WJ[IK]WIaSBU�NfiW_jljlUVJ[UQPKX_NVXUViSBcWJ[gWUQ_.^^NVPW`VU�SBUVJ[go_`VUAjlIKJ5SBUVP�SBIKJpPWUVXY�IWJ[Z�S�zKN�SBUVi�IKP�NQP:NQiKNQ]KXNVzW^.UUVPWUVJ[moRrq[UVjljl_`Q_.UQPKX�SBUQPaSB_PWm+`QIKgoUVJ[NQmWUO]KJ[IKXIW`VIK^%���[P�XdKUk]WJ[IKXIK`QIK^%hWUQNV`VdPKIWiKUv_%S1NVzW^.UvXI�iWRoPWNVLt_`VNV^^R�iWUV`Q_.iWUvNxSB`QdKUViWcK^U�jlIWJ�_.X%SBUQ^j�X.ImocKNVJ[NQPKXUVUvNv`QUVJ[XNV_P�iWUVmoJ[UVU�IWj�`VIKgoUVJ[NQmWU}�[�����~��Y�_Xd�NVgoUVJ[NQmWUUVPWUVJ[moR1`VIWPaSBcWLO]WX_.IWP�_.PWgWUQJsSBUV^R1]WJ[IK]WIKJ[X_IKPKNQ^�XI�X.dWU�PKIKiWU5iKUQPaSB_XRC�� UVgoUVJ[NV^+IK]WX_.Lt_uVNQX._IWPaSnNVPWi�UVwoXUVP�SB_IKPaSnNQJ[U{]KJ[IW]KI�SBUVi|X.I|]KJ[IWgW_iKUUVgoUVP�zKUQXX.UQJ�]KUQJ[jlIKJ[LtNVPW`VU�� � _LOcW^.NQX_.IWP�S.dKIKYpS+XdKNQXpIKcWJn]WJ[IKXIK`QIK^NV`Q`VIWLO]W^._%SBdWU�SbiK_jljlUQJ[UVPWX._NQX.UQi�SBcKJ[goUV_^^NVPW`VU�Y�_Xd�^IKY UQPKUVJ[moR`VIWPaSBcWLO]WX_.IWPa���[XaIKcWX.]WUVJ[jlIWJ[L�S�IKXdKUQJ�SBXNVXUQq[IKjlq[XdKUQq[NVJ[X�SB`QdKUQLOU�S�zKR N�SLtcK`Vd�N�Sp¡£¢K¤�J[UViKcW`VX_IKP¥_.P¥UVPWUVJ[moRA`QIKP�SBcKLt]KX_IKP¥NVPKi¥N�S�LtcK`Qd�N�S¦s§ ¢K¤¨_PK`QJ[UVN�S.U�_PtX.dWU�dKNV^jlq[^_jlU�IKj+XdKU�PKUVXY�IKJ[Z©�

�� � � � � � � � � � �� � � � � � � � � � � � � � ��t� ª��K«�� � � � � � � �� � � � � � � � � � � � � � � � � � ¬Q­V® UQXY�IKJ[Z�¯CJ[IKXIK`QIK^%S

� � � � � � � � � �° ^moIKJ[_XdKLnS�hs¯CUVJ[jlIWJ[LONQPK`QU�hs��U�SB_moP

� � � � � � �� UVPaS.IKJ ® UVXY�IWJ[Z�S3h � UVP�SB_PKm±�~IWgWUQJ[NVmoU�h{²�PKUQJ[mWR��~IKP�SBUVJ[goNVX_IKP�h��_jljlUQJ[UVPWX._NQX.UQi � UVJ[go_`VU

! " � ��# � $ ��" # � �³�_J[UV^U�S3S � UVP�SBIKJ ® UQX.Y�IKJ[Z©S�dKNQgWU�UQLOUQJ[mWUQi1N�S�NnPKUVY�_.PWjlIKJ[LtNVX_IKPWqmoNVXdKUQJ[_.PWm�]KNQJ[NViK_moL´zKN�SBUVi�IKP�XdKU�`QIK^^NVzKIWJ[NVX_gWU�UQjljlIKJ[X�IKjtN5^.NQJ[mWUPKcWLOzWUVJ-IKj�SBUVP�SB_PKm�PKIWiKU�S��k�[P¥SBcK`Qd+PWUVXY�IWJ[Z�S3hrPWIKiKU�S�iWUV]W^.IWRrUVi _P+N

J[UQLOIWX.UµUQPKgo_.J[IWPKLtUVPWX�LOc�SBX�SBUQ^jlq[`VIWPKjl_mocKJ[UµY�_XdKIWcKX�NQPKR}¶�·�¸r¹ ºK¸r¹_PKj�IKJ[LtNVX_IKPANVzKIWcKX�XdKU�PKUVXY�IKJ[ZAX.IW]KIW^.IWmWR�IWJ�mW^IKzWNV^�go_.UQY�� ® IKiKU�SNQ`VX�_.P�J[U�SB]WIKPaS.UnXI�UQPKgo_.J[IWPKLtUVPWX.NQ^�UVgoUVPKX%S�NQPKi¥J[UV^NVR�`VIW^^.UQ`VXUVi¥NVPWi]WIaS�SB_zW^.R+NVmomoJ[UVmoNVXUVi5_PKj�IKJ[LtNVX_IKP5XdKJ[IWcKmod5X.dWU�iKRrPKNQLO_`QNV^^RnjlIWJ[LOUQiLtcK^X_q[dKIK]µY�_.J[UQ^U�S�S�PKUQX.Y�IKJ[Z�_PµNV`Q`VIKJ[iWNVPW`VU�Y�_Xd�iKU�SB_J[UVi�SBRCSBXUVLjlcWPK`QX._IWPKNV^_XRC�+»adKU�SBUkPWUVXY�IWJ[Z�S�`VNVP�jlIKJ[L�XdKUOzKN�S._3S�jlIKJ~LtNVPWR+XRr]KU�SIWj�S.LONQJ[X�UVPWgW_J[IWPKLtUVPKX%S¨SBcW`Vd´N�S¨SBLtNVJ[X:dKI�SB]K_XNV^%S�hnzWNVXX^UVjl_UV^i�S�hUQNVJ[XdK�WcKNVZoU�J[U�SB]WIKPaS.UpSBRCSBXUQL�S�h£NVPWiO^UVNQJ[PK_PKmtUVPWgW_J[IWPKLtUVPKX%S�� ° SBUQX�IWjNQ]K]K^_`VNQX_.IWPaS�h¼SBcK`Qd½N�S�zW_.IWLOUQiK_`V_PWU�h�dWNVuQNVJ[iKIWcaS´UVPWgW_J[IKPWLOUQPKXUQwW]W^.IWJ[NVX_IKP�h�UQPKgo_.J[IWPKLtUVPWX.NQ^{LtIKPK_XIKJ[_PWm�h5LO_^_XNVJ[R¾XJ[NV`QZW_PWm¾NVPWiJ[UQ`VIKPWPKNQ_3S3SBNVPW`VU�SBcWJ[gWUQ_.^^NVPW`VU�NQJ[U�X.dWU�ZWUQR|LtIKX_goNVX_IKPaSnj�IKJ�LONQPKRJ[UQ`VUVPWX�J[U�SBUQNVJ[`QdOUQjljlIKJ[X%S(_.PtXdK_%S�NQJ[UVN��¿aIWY�q[`VI�SBX�iKUQ]K^IKRrLtUVPKX�_%SÀIWPKU�NQ`V`Q^.NQ_LOUQiÁNViWgWNQPKXNVmoU�IKjvS.UVPaS.IKJPWUVXY�IWJ[Z�S�h�Y�dW_`Vd�_LO]W^.RnX.dWNVX~XdKU+J[U�S.IKcKJ[`QU�S�NVgoNV_^NVzK^U+XI¥_PKiW_gW_iKcWNV^PWIKiKU�S+NQJ[U}S.UVgoUVJ[UV^R{^_Lt_.XUQia�Á¿a_LO_XUViµ]WJ[IK`VU�S�SBIWJnzKNQPKiWY�_iKXdµNVPWiSBLtNV^^�LtUVLtIKJ[R�NVJ[U�XY�I�NVJ[mocKNQzK^U�`VIKP�SBXJ[NV_PKX%S�_P}S.UVPaS.IKJtPKUVXY�IKJ[Z©S�hY�dK_`VdyY�_^^�iW_%SBNV]W]KUVNQJ�Y�_Xd�X.dWUfiWUVgoUV^IK]WLOUQPKXvIWj�jlNQzKJ[_`VNQX._IWPXUV`QdKPW_.�WcKU�S���Â�IWY�UQgWUQJsh�XdKU:UVPWUVJ[moR�`VIKP�SBXJ[NV_PKX�_%S cKPK^_ZoUV^R�XI|zWUJ[UQ^._UQgWUQi��KcK_`QZW^R�iWcKUvXI�SB^IKYb]WJ[IKmoJ[U�S�S�_P�iWUVgoUV^IK]W_PKm�zWNVXXUVJ[R`QNV]KNQ`V_XRC��ÃAIKJ[UQIKgoUVJsh<XdKU�cKPKXUQPKiKUQi{PKNQXcKJ[U1IWjASBUVP�SBIKJ PKIWiKU�SONVPWidWNVuVNQJ[iKIWcaS<S.UVPaS._.PWm�UVPKgo_J[IKPWLOUQPKX%S�]WJ[UV`V^cWiKU�zWNVXXUVJ[R�J[UV]W^NV`VUQLOUQPKX(N�S�NjlUQN�SB_zW^.U�SBIW^.cWX_.IWPa����P�XdWUnIKXdWUVJ�dKNQPKi�hCXdKU�S.cKJ[goUV_^^.NQPK`QUnPKNQXcKJ[UnIWjSBUQPaSBIWJ�PKUVXY�IKJ[ZnNQ]K]W^._`QNVX_IKPaS<J[UQ�KcW_.J[U�S<Np^IKPWmn^_jlUVX_LOU�Ä(XdWUVJ[UVj�IKJ[U�hK_X_%S�N�goUVJ[RA_LO]WIKJ[XNVPWX�J[U�SBUVNQJ[`Vd1_%S�SBcWU�XI1]KJ[IWgW_iWU�N�jlIWJ[LÁIKj�UQPKUQJ[mWRrqUQjljl_`V_UVPWX�SBcKJ[goUV_^^NVPW`VUkSBUVJ[go_`VU�jlIKJ�N�moUVIKmoJ[NV]WdK_`�NVJ[UVN��¯CJ[UVgo_.IWcaS�J[U�SBUVNQJ[`Vd�jlIW`Vc�SBU�S�IWP�dKIWY±XI�]KJ[IWgW_iKU|jlcK^^�IKJ�]WNVJ[X_NV^SBUQPaSB_PWmn`VIWgWUQJ[NVmoUp_P�XdKU�`QIKPKXUQwWX-IKj�UVPWUVJ[moRn`VIWPaSBUQJ[gWNQX_.IWPa���[P:SBcK`QdNQPANQ]K]WJ[IKNV`Qdah�PKIWiKU�S~NVJ[U ]WcKX�_PKXI�N iKIWJ[LONQPKX�SBXNQX.U N�S�^IKPWm5N�S�X.dWUV_JPWUV_mWdWzKIWJsS}`VNVPÅ]KJ[IKgo_iKUMS.UVPaS._.PWmf`QIKgoUVJ[NQmWU¼jlIKJvXdKUQL���»adKU�S.USBIW^cKX_IKPaS�J[UQmWNQJ[i�X.dWU+SBUVP�SB_PKmp`QIKgoUVJ[NQmWU�XI�N�`VUQJ[XNV_P�mWUQIKmoJ[NV]WdK_`�NVJ[UQNN�S�zK_PWNVJ[RChoUQ_.XdWUVJ�_Xa]KJ[IWgW_iWU�S�`QIKgoUVJ[NQmWUtIKJ�PWIKX%��Â�IKY�UVgoUVJshoY�UkNQJ[mWcWUXdKNQX%ht_P6LtIaS.X�SB`VUQPKNQJ[_.I�SÀS.cK`Vd¨N�S�zWNVXX^.UQjl_UV^iaS3h�XdKUQJ[UµNVJ[Uµ`VUQJ[X.NQ_PmoUVIWmWJ[NQ]KdK_`�SBUV`QX._IWPaS�SBcW`Vd�N�S<X.dWUpmoUVPKUQJ[NV^�`VIWLOLtNVPWi�`VUQPKXUVJ�XdKNQX-NQJ[ULtcK`QdnLtIKJ[U5S.UV`VcWJ[_X.RrqsSBUQPaSB_X_gWUOXdKNVP�IKXdKUQJsS���Æ~N�SBUVi�IKPnXdWUkjlNQ`VX�X.dWNVX_PKiW_gW_iKcWNV^�SBUVP�SBIKJ-PKIKiWU�SaNVJ[U�PKIWX�J[UV^_NVzW^UtNQPKi¥SBcWzsÇ%UV`QX�XI+j�NV_^cKJ[U�NVPWiSB_PWmW^U|SBUVP�SB_PKm�J[UQNViK_PWm�S�`VNQP�zWU�UVN�S._.^R{iW_%SBXIKJ[XUVi�zWR�zKNQ`VZomWJ[IWcKPWiPWIK_%SBU�NVPKit`VNQcaSBU�jlNQ^3S.U�NQ^.NQJ[L�S3hs_X�_%S�SB_LO]W^RtPWIKX-SBcWjljl_`V_UVPWX(X.IOJ[UV^R�IKPONSB_PWmW^U�S.UVPaS.IKJaXI5SBNQjlUVmocKNQJ[ikN�`QJ[_X._`QNV^�NVJ[UVN����[PpXdK_%S©`VN�SBU�hV_X�_3S©iWU�SB_J[UQiXI1]WJ[IKgo_.iWU�dK_modKUQJtiWUVmoJ[UVU�IWj�`VIWgWUQJ[NVmoU�_P1Y�dW_.`Qd�LtcK^X_]K^U�SBUQPaSBIWJsSLtIKPW_.XIWJaX.dWU�SBNQLOU�^IK`VNQX_.IWPkNQX�XdWU�SBNQLOU�X_LtU�_PpIKJ[iWUVJ�XIpIKzWX.NQ_PpdK_mod`QIKPKj�_.iWUVPW`VU�_PÀiKUVXUQ`VX_IKPa����P�XdKUAIKXdWUVJpdWNVPKi�h�_X�_%StIWgWUQJ[ZW_^^�NVPWiUQPKUVJ[moRA`VIKP�SBcKLt_PKm¥X.I�SBcK]W]KIKJ[X�XdKU�S.NVLtU�dK_mod�iKUQmWJ[UQUnIWj�`VIWgWUQJ[NVmoUjlIWJ�SBIWLOU�PKIWPKq[`VJ[_X_`VNQ^�NVJ[UVN���Æ~N�SBUViOIKP�SBcK`QdkIKz�SBUVJ[goNVX_IKP�S�hQXdK_%SC]WNV]WUVJ^UVgoUVJ[NQmWU�S}]KJ[UQgW_IKc�S�SBIK^cWX._IWPaS}NVPWiÅNQiKiKJ[U�S�SBU�S|XdKUÁ]WJ[IKzW^.UQLÈIWj]WJ[IKgo_.iW_PKmÉNÊiK_jlj�UVJ[UVPWX_.NQXUVi�S.cKJ[goUV_^^.NQPK`QU�SBUQJ[gW_`VUËjlIKJÌS.UVPaS.IKJPWUVXY�IWJ[Z�S��ÍÆ~R}iK_jljlUQJ[UVPKX_NVXUQiÁS.cKJ[goUV_^^.NQPK`QU�h�Y�U�LtUVNQPv]KJ[IKgo_iK_PKmiW_.j�jlUVJ[UQPKX:iKUVmoJ[UVU�S�IKj�SBUQPaSB_PWm¼`QIKgoUVJ[NVmoU�j�IKJ�N´S.UVPaS.IKJ�PKUQXY�IKJ[Z

Î�Ï.ÐBÑkÒ ÓÔÓÔÒ Õ×Ö�ØÔÕ�ÑpÙ�Ú(Ï~ÛVÒ ÜVÒ Ø[Ù�Ý�Õ×ÐCÞ(Ù�Ð�Û�ß�ÕÔà(Ò Ï�Ó�ÕVáCÙ�Ý Ý�ÕÔÐ�à(Ù�Ð.Ø<ÕQáWØ Þ(Ò Ó�â�Õ×Ð.Úná.Õ×Ðà(ÏBÐ3ÓÔÕ×Ö(Ù�ÝãÕÔÐ�ß�Ý Ù�ÓÔÓ Ð�ÕQÕ×Ñ�ä(ÓÔÏ�Ò Ó�ÜÔÐ�Ù�Ö"Ø[Ï�Û�â�Ò Ø Þ�Õ×ä"Ø©áBÏ3Ï(à"Ð3Õ×å�Ò Û"Ï�ÛaØ Þ(Ù�Ø©ß�Õ×à�Ò Ï�Ó�Ù�Ð3ÏÖ(Õ×Ø�ÑkÙ�Û"Ï Õ×РÛVÒ Ó Ø Ð�Ò æ"ä"ØÔÏ�Û�á.Õ×Ð�àVÐ�ÕQáBÒ ØOÕÔÐ�ß�Õ×Ñ�ÑkÏBÐ�ß�Ò Ù�Ý�Ù�Û%å�Ù�ÖVØÔÙ�ÜVÏ�Ù�Ö(ÛkØ Þ(Ù�Øß�Õ×à(Ò Ï�Ó�æ(Ï�Ù�Ð�Ø Þ(Ò Ó�Ö(ÕÔØÔÒ ß�Ï�Ù�Ö(Û�Ø Þ(Ï�á ä�Ý Ý�ß�Ò Ø[Ù�ØÔÒ Õ×ÖAÕÔÖ�Ø Þ(Ï�á.Ò Ð3ÓçØaà�Ù�ÜVÏ�èCé�Õpß�Õ×à"êÕ×Ø Þ�ÏBÐBâ�Ò ÓÔÏ�ë~ÕÔÐ�Ð�ÏBàVä"æ(Ý Ò ÓçÞ�ëCØÔÕkà(ÕQÓçØOÕÔÖ�ÓÔÏBÐ.å�Ï.Ð�ÓOÕ×Ð�ØÔÕkÐ3Ï�Û"Ò ÓçØ Ð�Ò æVä"ØÔÏ�ØÔÕ5Ý Ò ÓçØ[Ó×ëÐ�Ï3ì%ä(Ò Ð3Ï�ÓoàVÐ�Ò ÕÔÐ�Ó à(Ï�ß�Ò á.Ò ß�à(ÏBÐ.ÑpÒ Ó×ÓÔÒ ÕÔÖOÙ�Ö(Û%íÔÕÔÐ�Ù-á.Ï�Ï�èî"ï�ð"î%ñ�ò�ó ô"õ ë%ö�Õ×å�ÏBÑ�æ�ÏBÐ�÷%ø(ùVë%ú(û"û"üVë%ý�ÕVÓrþ-Ö�ÜVÏ�Ý Ï�ÓÔë"ÿaÙ�Ý Ò á.Õ×Ð.Ö(Ò Ù�ë�����þ�èÿaÕ×à"êsÐ3Ò Ü×Þ"Ørú(ûVû"ü�þ�ÿ�����.÷��"ü��Bù"ûVù� �×íÔû"ü%íÔûVû�����"÷"è û"û

NV`Q`VIWJ[iK_PKm´XI iK_jljlUQJ[UVPWX{J[UV�WcK_J[UVLtUVPWX3S3� »aY�I´LtN×Ç%IKJ�moIKNV^%S{NVJ[UNV`QdK_UVgoUVi _P XdK_%S�]WNV]KUQJs�oH�_JsSBX%hoIKcWJ�PKUQY�S.UVPaS._.PWm `VIKgoUVJ[NQmWUtNV^moIKJ[_XdKLIKcWX]KUVJ[j�IKJ[LnS�IKXdKUQJ�SBXNVXUQq[IKjlq[XdKUQq[NVJ[XASBIW^cKX_IKPaS _P|XUVJ[LnS+IWjnUQPKUVJ[moR`VIWPaSBUQJ[gWNQX_.IWPahWUQPKUQJ[mWR+zKNQ^.NQPK`QUpNVPWi�`VIWLOLtcKPW_.`QNVX_IKP�IWgWUQJ[dKUQNVi�Y�dKUVP_XaJ[cKP�S�_P�PWIKPWq[iK_jljlUQJ[UVPKX_NVXUQi�LtIKiWU�� � UQ`VIKPWiahrX.dW_%S�]KNQ]KUQJ�_%S�XdKUtjl_JsSBXXI{]KJ[IW]KI�SBU1iK_jlj�UVJ[UVPWX_.NQXUVi¨SBcWJ[gWUQ_^.^NQPK`VU�jlIKJ¥SBUQPaSBIWJ+PWUVXY�IWJ[Z�SkY�_X.dLt_.PW_LONQ^�IKgoUVJ[dKUQNVi��»adKU¥J[UVLtNV_PKiWUVJ�IWjpXdKU�]WNV]WUVJp_%StIWJ[mWNQPK_uVUQiÀN�StjlIW^^.IWY�S ­ � UV`QX_.IWPÀªiK_%SB`QcaS�S.U�S�]KJ[UVgo_IKc�S�J[U�SBUVNQJ[`Vd´J[UQ^.NQXUVi´XI´XdKU�SBUQPaSB_PKmÁ`VIKgoUVJ[NQmWU]KJ[IWzK^UVL�jlIWcKPKiO_P5SBUQPaSBIWJaPKUVXY�IKJ[Z©S�� � UV`VX_IKP § iWU�SB`QJ[_.zWU�SCXdWU�iKU�SB_mWPIKj~IWcKJkSBUVP�SB_PKm�`QIKgoUVJ[NQmWU�]WJ[IKXIK`QIK^<Y�_X.dWIKcWX<iK_jlj�UVJ[UVPWX_.NQX_.IWPa� � UV`VX_IKP� UVwoX.UQPKi�S�XdKUxiKU�SB_mWP XI ]KJ[IWgW_iKUxiW_.j�jlUVJ[UQPKX_NVXUViySBcWJ[gWUQ_.^^NVPW`VU��� UV`VX_IKP�¡ÀmW_goU�SnN{zWJ[_UVj5iKU�SB`VJ[_]WX._IWP�IKj�zWN�SBUQ^_.PWU�S�XI|Y�dK_`Qd}Y�U`VIWLO]WNVJ[U1IWcKJ Y�IWJ[Z�h<_PW`V^cKiW_.PWm�XY�IÀIKXdKUQJASBXNVXUVq[IWjlq[XdKUQq[NVJ[X�SBUVP�SB_PKm`VIWgWUQJ[NVmoU�]KJ[IKXIW`VIK^%S�� � UV`QX_.IWP��:]WJ[IKgo_iKU�S�N5iWUVXNV_^UVi�]WUVJ[jlIWJ[LONQPK`VUUVgoNV^cWNVX_IKPtNVPKit`VIWLO]WNVJ[_%SBIWPa�~³�U�`QIKPK`Q^cKiKU�XdKU�]KNQ]KUVJ©_.P � UV`VX_IKP��a�

% ! �& ' ��& � � ( # �� �° S.UVX(IKj�J[U�SBUVNQJ[`VdO_3S3SBcKU�SCPKUQUViaS�X.IOzKU�NViKiWJ[U�S�S.UVikzWUVjlIWJ[U�SBcWJ[gWUQ_^.^NQPK`VUQqzKN�SBUViÀNV]W]K^_`VNVX_IKP�S�SBcW`Vd�N�SOLt_.^_XNVJ[R�XJ[NV`QZW_PKm:NVPWiÀUQPKgo_J[IKPKLtUVPWXNV^LtIKPK_XIKJ[_PWm�zKUV`QIKLtU X.UQ`VdWPK_`VNQ^.^Rej�UVN�S._.zW^UÉNQPKi�UQ`VIKPWIKLt_`VNV^^R]KJ[NQ`VX_`VNQ^3���~UQ`VUQPKX^R©haS.UVgoUVJ[NV^�SB`QdKUVLtU�S<NQJ[Up]KJ[IW]KI�SBUVi�XI�NQiKiWJ[U�S�S<XdKUSBUQPaSB_PKm�`VIWgWUQJ[NVmoU}]WJ[IKzK^UQLË_PÁSBUVP�SBIKJ1PWUVXY�IWJ[Z�S3� �[PÁ« ¦�� ¬ h�jlcK^^SBcWJ[gWUQ_.^^NVPW`VU�`VIWgWUQJ[NVmoU�_3SASBcW]K]KIWJ[XpzKR�N:PWIKiWUVqsSB`QdKUViWcK^_PKm¨S.`VdKUQLOUzKN�SBUVi�IKP{IWjljlq[iKcWXRÀUQ^._mo_zK_^_.XR:J[cK^U�S3h-Y�dW_.`Qd�NQ^.^IWY�SOPWIKiKU�SOXI�XcKJ[PXdKUQL�SBUQ^gWU�StIKj�jkN�S�^IKPWm�N�S�X.dWU�PKUV_modKzWIKJ[_PKm�PWIKiKU�S�`QNVP�`VIWgWUQJkXdKUNVJ[UQN5jlIKJtXdKUQL��µ»adK_%S�J[cK^U5mWcWNVJ[NQPKXUVU�S ¦ ¢K¢W¤ SBUVP�SB_PKm1`QIKgoUVJ[NQmWU5N�S^IKPWmON�SCPKIOgWIW_.iOUVwo_%SBX%S���Â�IWY�UVgoUVJshsX.dW_%SCJ[cK^U�cWPKiKUQJ[U�SBX_LtNVXU�S©XdKU�NVJ[UVNXdKNQX<XdKU�PWUV_mWdWzKIWJ�PWIKiKU�S�`VNVP5`QIKgoUVJsh�Y�dW_`Vd5^UVNQiaS�XI5UVwo`VU�S3S�UQPKUVJ[moR`VIWPaSBcWLO]WX_.IWPa�x�[P�«çªWª ¬ h�SBcKJ[goUV_^^NVPW`VU�`VIWgWUQJ[NVmoU1_%SkNV`QdK_UVgoUVi{zWRÀN]KJ[IWzK_PKmkLOUQ`VdKNQPK_%SBLn�t�[P�XdK_%S�SBIW^.cWX_.IWPah"NVj�X.UQJ-N+S.^.UQUV]W_.PWm�PKIWiKU�Y�NQZWU�ScK]�h�_X�zWJ[IKNQiK`VN�SBX%S�N�]WJ[IKzK_PWm1LOU�S�SBNQmWU5Y�_.XdW_.P�N5`QUVJ[XNV_P�J[NQPKmoU5NVPKiY�NQ_X3S�jlIWJ�N�J[UQ]K^RC�A�[j�PKI5J[UQ^Rn_%S�J[UV`VUQ_gWUQi5Y�_XdK_P5N�X_.LtUVIWcKX%h�_X<Y�_^^XNVZoUkXdKUkJ[U�SB]WIKPaS._.zW_^._XR+IKjtSBcWJ[gWUQ_.^^NVPW`VUpcWPKX_^-_X�iKUQ]K^UVXU�S<_X%S<UVPWUVJ[moRC��[P�XdW_3S�SBIK^cKX_IKP�hOXdKU|]KJ[IKzW_PKm�J[NVPKmoU�NVPKi�Y�NQZWUQcK]�J[NVXU�`VNVP�zKUNVi£Ç%caSBXUVitXIONQjljlUQ`VX�XdKU�iKUQmWJ[UQU�IWj©`VIWgWUQJ[NVmoU�_PKiW_J[UV`VX^RC��Â�IWY�UVgoUVJshsX.dW_%S]KJ[IWzK_PKmoq[zKN�SBUVi�NQ]K]WJ[IKNV`Qd�dKN�S�PWI�mWcWNVJ[NQPKXUVU�IWP¨SBUVP�SB_PKmÀ`VIKgoUVJ[NQmWUNVPWi�zK^_PKin]KIK_PWX3S�`QNVP5IK`Q`VcWJs�A��cWJpSBIW^cKX_IKP5_%S�iK_jljlUQJ[UVPKX-jlJ[IKL�X.dWU�SBUSBIW^.cWX_.IWPaS�_P|X.dWUvSBUQPaSBUÀXdKNVX _X�`VNQP|PWIKX�IWPK^R��W¶K¸�¶ ��� ����SBUVP�SB_PKm`VIWgWUQJ[NVmoU�XI1N�`QUVJ[XNV_P1moUVIWmWJ[NQ]KdK_`�NQJ[UVN�h©zKcKX�_X�`QNVP1NQ^%SBI1NViWNV]WX�XdKUiKUQmWJ[UQU IKj�`VIKgoUVJ[NQmWU XI�XdKNVX<NVJ[UVN�h(cW]�X.I�XdKU ^_LO_XNVX_IKP�_Lt]KIaS.UViAzKRXdKU�PKcWLOzWUVJ�IWj~SBUQPaSBIWJ�PKIWiKU�S(]KJ[U�SBUVPWX3�²�PKUQJ[mWR¨zKNQ^NVPK`Q_PKm�_3SANVPKIWXdKUVJ�J[U�SBUQNVJ[`Vd�_%S�SBcWU|NQiKiKJ[U�S�SBUQi6_P�X.dW_%S]KNQ]KUQJs�p« ¦�� ¬ NQPKi5« ªWª ¬ IKPW^.Rt`VIWPaSB_iWUVJ�XdKU�LtUVXJ[_`�_PpXUVJ[LnS©IKjaXdKU�X.IWXNV^NVLtIKcWPKX-IKj�UVPWUVJ[moRn`VIWPaSBcWLOUQi5J[UVmoNVJ[iW^.U�S�S~IKj�X.dWU iK_%SBXJ[_zKcKX_IKP�IWj�XdKUUVPWUVJ[moR�NVLtIKPWmvXdKUµPKIWiKU�S3�y³�U�NVJ[mocKU�XdKNQXncWPKzKNQ^NVPK`QUVivUQPKUVJ[moRiK_%S�S._.]WNVX_IKP�`VNVc�SBU�S�SBIKLtUkPWIKiKU�S-XIniW_.UkLOcW`Vd�jlN�SBXUVJ�XdKNQP�IKXdKUQJsS<iKI�hXdKUQJ[UVjlIWJ[U�h�XdWU�dWNV^jlq[^_jlU�IWj�XdKU+PKUQXY�IKJ[Z�_%S�iWJ[NVLtNVX_`VNQ^.^R5J[UQiKcK`QUVi¥_.PXdKU1cWPKq[zWNV^NVPW`VUVi�NV]K]WJ[IKNQ`Vda���~U�SBUQNVJ[`VdÀdKN�SONViWiKJ[U�S3SBUViÀX.dWU�UQPKUVJ[moRzKNQ^NVPK`QUk_%S�SBcWUkjlJ[IWL6iW_jljlUVJ[UQPKX�N�S.]KUV`QX%S-IKj�S.UVPaS.IKJ~PKUQXY�IKJ[Z©S�� � ¯©²�²��« ¦ ¢ ¬ zKNV^NQPK`VU�S�XdKU+XJ[NVjlj�_.`+zKR5PWIKPWq[iKUVXUQJ[LO_PK_%SBX_`VNQ^^.R5j�IKJ[Y�NQJ[iK_PKm�XdKU]KNQ`VZoUVX�XdKJ[IWcKmod¼LtcK^X_]K^U�J[IKcKXU�S���� ° Hy« ªK¢ ¬ ]KUQJ[jlIKJ[LnS�^UQNViKUQJJ[IKXNQX._IWPpNVLtIKPWmkXdKU�PWIKiKU�S�_.P�SB_iKU�N�go_J[X.cWNV^CmWJ[_iah"_P�IKJ[iWUVJ�XI�zKNV^NQPK`VUUVPWUVJ[moR�`VIWPaSBcWLO]WX._IWPa��~U�S.UVNVJ[`QdOIWPOPWUVXY�IWJ[Z�X.IW]KIW^.IWmWR�`QIKPWX.J[IW^-SBcK`QdkN�S � ¯ ° ® « � ¬ hQ¿a² ° �~« ¦ ¢ ¬ NVPWi�� ° H « ªK¢ ¬ NViWiKJ[U�S�SBU�S�XdKUM]KJ[IWzK^UVL IWj¼]WJ[IKgo_iK_PKm`VIWLOLtcKPW_`VNVX_IKPA`VIWgWUQJ[NVmoU�Y�_X.dW_P¥NQP�UVPWUVJ[moR�`VIKP�SBUVJ[goNVX_IKP¥`VIWPKXUVwoX%�»adKU�SBU�dKNQgWU�N�SB_LO_^NVJ-jl^NVgoIKJ-N�SaX.dWU�S.cKJ[goUV_^^.NQPK`QUO`QIKgoUVJ[NVmoUt]KJ[IKzW^UVLn�H�IKJ~UVwoNVLt]K^U�hW¿a² ° �~Âx« ¦ ¢ ¬ ]KNVJ[X_X_IKPaS<NkPKUVXY�IKJ[Zn_PWX.In`V^caSBXUQJsS<NVPKi

J[NQPKiKIWLO^R¥J[IKXNVXU�S�XdWU�`V^caS.X.UQJ�^.UQNViWUVJ�_P1IKJ[iWUVJ�XI1UVgoUVPW^.RAiK_%SBXJ[_zKcWXUXdKU�UQPKUQJ[mWRO`VIKP�SBcKLt]KX_IKP�NQLOIWPKmpXdWU+SBUVP�SBIKJsS�� � ¯ ° ® « � ¬ _%S�NQPKIKXdWUVJJ[NQPKiKIWLO_uQUVixNV^moIKJ[_XdKL Y�dWUVJ[U¨PKIWiKU�S�LtNVZoU¨^.IW`VNQ^1iKUQ`V_%SB_IKP�S�IWPY�dKUVXdWUVJ�X.IAS.^.UQUV]�IWJ-XI�Ç%IK_P�N�zKNQ`VZozKIWPKU�PKUQXY�IKJ[Zk_.P�IWJ[iKUQJ-XI�J[UViWcK`QUUQPKUVJ[moR�`VIWPaSBcWLO]WX_.IWPa�~»adKU�LONÔÇ%IKJ�iW_.j�jlUVJ[UQPK`VU�zKUQXY�UVUQPOXdWIaSBU�NVPKiOXdK_%SY�IKJ[Z�_%S�XdWNVX�`QIKLtLOcWPK_`VNQX._IWP+`QIKgoUVJ[NQmWU�`VIWPaSB_iWUVJsS�IWPK^R�`VIWPKPWUV`VX_go_.XRzWUVXY�UQUVPAX.dWUnPWIKiKU�S��À�[P¥`VIKPWXJ[N�SBX%h�SBcWJ[gWUQ_^.^NQPK`VU��sSBUVP�SB_PKmo��`VIWgWUQJ[NVmoUNQiKiKJ[U�S�SBU�SkX.dWU1`VIWgWUQJ[NVmoU1]KJ[IWzK^UVLfXI{UVgoUVJ[R:]KdWR©S._.`QNV^t]KIW_.PWXO_P{XdWUXUVJ[J[NQ_Pa� ° S�NAJ[U�SBcK^X%haPKUVYfSB`QdKUQiKcK^_PKm�_%S�J[UV�WcK_J[UVi�XI�jlIKJ[`QU�S.IKLtUPWIKiKU�S�SBXNVR NQY�NVZoUkj�IKJtSBcWJ[gWUQ_.^^NVPW`VUp]WcKJ[]WIaSBU�S<UVgoUVP�XdKIWcKmod�XdKUQR�NQJ[UPWIKX�]KNQJ[X_.`Q_]KNVX_PKm�_PtiKNVXN�jlIKJ[Y�NVJ[iW_.PWm��Æ~U�SB_iKU�S�XdKU�NQjlIKJ[UQLOUQPKX_IKPWUVi�Y�IWJ[Z5_.P�`QIKLtLOcWPK_`VNQX._IWPANQPKiÀSBUQPaSB_PKm`QIKgoUVJ[NVmoU�XdKNQX<`VIWPaSBUQJ[gWU�S�UVPKUQJ[mWRCh�Y�IWJ[Z�_P5UVPWUVJ[moR�`VIWPaSBUQJ[gWNQX._IWPAj�IKJmoUVPWUVJ[NV^�SBUQPaSBIWJ�PWUVXY�IWJ[Z�S<dWN�S<zWUVUVPn`VIKP�SB_iKUQJ[UVi�NQX-goNVJ[_IKc�S�^.UQgWUQ^%S~IWjXdKU�`QIKLtLOcWPK_`VNQX._IWP}S.X.NQ`VZ©�}HCJ[IWL´XdKU�zKIWX.XIWL´X.I�XdKU�XIK]�h�SB]KUQ`V_NV^dWNVJ[iKY�NVJ[Uf« ¦ ¡ ¬ _%S�iWU�SB_moPKUQi´Y�_Xd´LOcW^.X_]K^U�UQPKUVJ[moRÁiW_%S�SB_]KNQX._IWPSBUQXX._PWm�S��Mà ° �f^NQRoUQJ¥]WJ[IKXIK`QIK^%SniKUQgWUQ^IK]KUQi�jlIWJAUQPKUQJ[mWR�SBNVgo_PKm©SLtIaSBX^R�XNQZWU�NQiKgoNVPKXNQmWU�IWj+IWgWUQJ[dKUQNVJ[_PKmÀNVPKivSB`VdWUViKcW^_.PWm�XI{NQ^.^IWYPWIKiKU�S�XIfSB^UVUQ]xY�dK_^U¨XdKUVR6NVJ[UvPWIKX�XJ[NQPaSBLt_X.X_PKm�IKJ:J[UV`VUQ_gW_PKmLtU�S�S.NVmoU�S�« � ¬ « ¦K¦ ¬ � ° X�XdKUnPKUVXY�IKJ[Z�NQPKi1J[IWcKX_PKmA^NVRrUVJsS�h<SBIK^cKX_IKP�SNQJ[U:iW_gWUQJsSB_jl_UVia����NVXN�]K^NQ`VUVLtUVPWX5SB`QdKUQLOU�SA« § ¬ Lt_.PW_LO_uVU�UQPKUQJ[mWRNQ^.IWPKm:XdKU�XJ[NQPaSBLt_%S�SB_IKP:]KNQX.d:XdKJ[IWcKmodÀiWNVXN¥`QNV`QdK_PKm©�¨�[P}« ¦ � ¬ h��t��~NQLONQPKNVXdWNVP¾UQX3�nNQ^%��NVi£Ç3c�SBX�`VIWLOLtcKPW_.`QNVX_IKP¼J[NVPKmoU�iWRrPKNVLt_`VNQ^.^RzWN�SBUQi�IKP�XdWU�PKIWiKU�iWUVPaS._.XR¨XI�`QIKPaS.UVJ[goU}UVPWUVJ[moR¨`VIKP�SBcKLtUVi6_PXJ[NVP�SBLt_3S3SB_IKPa�~ÃAH��µ« ¦ � ¬ zKR�»aNQZWNQmW_(UVX%�QNQ^%�QcaSBU�SCN�Lt_PK_LtNV^(dKIW]k]KNQXdXIµJ[UViWcK`VU�XdKU�XIKXNQ^pPKcWLOzWUVJ+IKj XJ[NVPaS.LO_%S�SB_IWPaS��f«çª ¦ ¬ SBUVX%SpJ[IKcWX.U�SNQ`V`VIWJ[iK_PKm5XI�XdKU UVPWUVJ[moR�J[UVLtNV_PW_.PWm5NVX�PWIKiWU�S~NV^IKPWm5X.dWNVX�]WNVXdah(NVPWi« ¦ ¢ ¬ c�SBU�S�LtUV`VdWNVPW_3S.L�S�XI:SBNVgoU�UVPWUVJ[moRnXdKJ[IWcKmod5XdKU iK_%SBXJ[_zKcWX._IWPAIWjLtU�S�S.NVmoU�S�NVLtIKPWmpgWNQJ[_IKcaS�]WNVXdaS�jlJ[IWLxSBIKcWJ[`VU�XI�iKU�SBX_PKNVX_IKP�� ° X�XdWUNQ]K]K^_`VNQX_.IWP1^NVRrUVJsh©XdWUn]KJ[IWXIK`VIW^%S�_PK`VIWJ[]KIWJ[NVXUnJ[IWcKX_PKm�SBUVLtNVPWX._`�S�XIjlIWJ[L�moJ[IKcW]aS<NVPWi5J[IKXNVXU�^UVNQiKUQJsSBdK_]5J[U�S.]KIKP�SB_zK_^_X._U�S�h�NV^^IKY�_PWm�PKIWPKq^UVNQiKUQJ-PWIKiKU�S�X.IAS.^.UQUV]�NQPKi `VIWPaSBUQJ[gWU�XdKUQ_.J-UVPWUVJ[moR�« � ¬ �kHC_PWNV^^R©hãiKNQXNNQmWmoJ[UVmoNVX_IKP�XUQ`VdKPW_�KcKU�S�_.P�SB_iKU�XdKU�PWUVXY�IWJ[Z�NVPWi5NV]W]K^_`VNVX_IKP5^NQRoUQJsSNQ^3S.I�]WJ[IKgo_iKUµUVPWUVJ[moR�`VIWPaSBUQJ[gWNQX._IWP�j�UVNVXcWJ[U�S5_.PvzWIKXdvNV]W]K^_`VNQX._IWP_PKiWUV]WUVPKiWUVPWX�« � ¬ NVPWipNV]W]K^_`VNQX._IWPpiKUQ]KUVPWiKUQPKX�jlN�SBdW_IKPaS~« ¦ ª ¬ « ¦s§ ¬ � ° ^^IWj�XdKU�S.U�]KJ[IWXIK`VIW^%SCXNV`VZo^U�XdKU�UVPWUVJ[moRt_%S�S.cKU�j�J[IKL}iW_jljlUVJ[UQPKX�N�SB]WUV`QX3S©IWjSBUQPaSBIWJ�PKUQX.Y�IKJ[Z©S�h-XdKc�SOY�U¥`VIWPaSB_iKUQJ�XdKUVL `QIKLt]K^UVLtUVPWX.NQJ[R�XIÀIWcKJY�IKJ[Z©��~UQ`VUVPWX^.RCh�J[U�S.UVNVJ[`Qd�dKN�S+S.X.NQJ[XUVi�XI�NViWiKJ[U�S3S���I � _%S�S.cKU�S�XIµS.cK]K]WIKJ[XiW_.j�jlUVJ[UQPKX_NVXUViÁSBUVJ[go_`VU�SAjlIWJµSBUVP�SBIKJ�PKUVXY�IKJ[Z©S�� � ³ ° ® « ¦ ¬ caSBU�SjlUQUViWzKNV`QZ _.PWjlIKJ[LtNVX_IKP�jlJ[IWL XdWU¼Ã ° � ^NQRoUQJ�XIfJ[UQmWcW^NVXU¼XdWUXJ[NVP�SBLt_3S3SB_IKP�J[NQX.UkIKj~PKIWPKq[J[UQNV^q[X_.LtUpXJ[NVj�jl_`p_P�IKJ[iWUVJ�XI:SBcaSBXNQ_.P5J[UQNV^qX_LOU�X.J[NQjljl_`���� ° ¯x« ¦�� ¬ c�SBU�SOgoUV^IW`V_XR:LtIKPWIKXIKPW_`�SB`QdKUViWcK^_PKm:XI]WJ[_.IWJ[_X._uQU�J[UVNQ^q[X._LtU�XJ[NVjlj�_.`�NVPKikUVPKj�IKJ[`VU�S~SBcK`Qdp]KJ[_IWJ[_.X_uVNQX_.IWPpXdKJ[IWcKmodN�iW_.j�jlUVJ[UQPKX_NVXUVi�à ° �:¿aNVRrUVJs���[PA« ª ¬ h"X.dWU�iKUQ^._goUVJ[ROJ[NVX_IaS�IKj�]KNQ`VZoUVX%SNQX�iK_jljlUQJ[UVPWX�]KJ[_IKJ[_XR ^UVgoUV^%SvNVJ[U NVj�jlUV`QX.UQiMzKR XdKU´j�IKJ[Y�NQJ[iK_PKm]WJ[IKzKNQzK_^_X.R:NVX�_PKXUQJ[LOUQiK_NVXU�PWIKiWU�S��6»aIÀiKNQX.U�h-XI�XdKU¥zKU�SBX�IKjpIWcKJZoPKIWY�^UViWmWU�hAPKI�NV^moIKJ[_XdKL dKN�S�zWUVUVPÍSB]WUV`V_jl_`QNV^^R¾iKU�SB_mWPWUVi XINQiKiKJ[U�S�S¥dWIKYÍXI6]KJ[IKgo_iKU|iK_jljlUQJ[UVPKX_NVXUQi´SBcWJ[gWUQ_.^^NVPW`VU|jlIKJµS.UVPaS.IKJPWUVXY�IWJ[Z�S��k��cKUtX.I+XdKUnSBcKJ[goUV_^^NVPW`VUtPKNVXcWJ[UOIWj<LtIaSBX�SBUQPaSBIWJ�PKUQXY�IKJ[ZNQ]K]K^_`VNQX_.IWPaS�h£Y�U�iKUQUVLµXdKNVX�XdW_3S-SBUVJ[go_`VU�_%S�U�S3SBUVPWX._NQ^3�

) ! ���" �� * �# �# �# ' � � & �" � � ��[P�XdW_3S�S.UV`VX_IKP}NVPKi�SBUV`QX._IWP � h�Y�U�_PKXJ[IWiKcK`QU�XdKUÀiKU�SB_mWP|IKj�NQPNQiKNV]WX_.goU�S.UVPaS._.PWmk`QIKgoUVJ[NVmoU SB`QdKUQLOU�j�IKJ�SBUVP�SBIKJ�PWUVXY�IWJ[Z�S���»adWU�zKN�SB_`iWU�SB_moPpY�_XdKIWcKX�iK_jljlUQJ[UVPKX_NVX_IKPk_3S©_PWX.J[IWiKcW`VUVikjl_JsSBXC_P�S.UV`VX_IKP § �"»adK_%S_%SCXdKUQPkXdKU�zKN�SB_%SCjlIKJaXdWU�UVwoXUVP�SB_IKPkj�IKJaiK_jlj�UVJ[UVPWX_.NQXUVi5S.cKJ[goUV_^^.NQPK`QU�XISBcW]K]WIKJ[X�SB`VUQPKNVJ[_I�S|_PÅY�dW_`VdÅN¼]KNQJ[X._NQ^�`QIKgoUVJ[NQmWU¼��� ¦ ¢K¢K¤+�v_%SSBcWjljl_`V_UQPKX1IWJÀNvdW_.mod�iKUQmWJ[UQUvIKj�`VIWgWUQJ[NVmoUv��� ¦ ¢K¢W¤��:_%S�iWU�SB_J[UQi�sSBUQ`VX_IKP � �s�

Achieving Repeatability of Asynchronous Events inWireless Sensor Networks with EnviroLog

Liqian Luo†, Tian He‡, Gang Zhou, Lin Gu, Tarek F. Abdelzaher†, John A. StankovicComputer Science Department, University of Virginia, Charlottesville, Virginia 22904

{ll4p, tianhe, gz5d, lg6e, zaher, stankovic}@cs.virginia.edu

Abstract— Sensing events from dynamic environments arenormally asynchronous and non-repeatable. This lack of repeata-bility makes it particularly difficult to statistically evaluate theperformance of sensor network applications. Hence, it is essentialto have the capability to capture and replay sensing events,providing a basis not only for system evaluation, but also forrealistic protocol comparison and parameter tuning. To achievethat, we design and implement EnviroLog, a distributed servicethat improves repeatability of experimental testing of sensornetworks via asynchronous event recording and replay. To useEnviroLog, an application programmer needs only to specify twotypes of simple annotations to the source code. Automatically,the preprocessor embeds EnviroLog into any desired level of anevent-driven architecture. It records all events generated by lowerlayers and can replay them later to upper layers on demand.We validate the accuracy and performance of recording andreplay through a set of microbenchmarks, using the latest XSMplatforms. We further demonstrate the strength of EnviroLog insystem tuning and performance evaluation for sensor networkapplications in an outdoor environment with 37 XSMs.

I. INTRODUCTION

With the increase in maturity of sensor networks researchand with recent solutions to several practical systems and de-ployment problems, sensor network applications are enteringthe real world. Representative experiences of such evolutionare documented in recent literature such as military surveil-lance [7] [19], habitat monitoring [13] [12], and environmentalmonitoring [23] [2], just to name a few. Quite different fromcontrolled lab settings, physical environments introduce a highdegree of uncertainty that makes it hard to conduct repro-ducible experiments. Consequently it is hard for researchersto obtain statistically consistent empirical results. With thegrowing number of applications developed and deployed, thereis an increasing need for tools and services to assist withsystem evaluation and debugging, as well as with performancetuning of applications in outdoor environments.

To address this issue, we propose EnviroLog, a tool to im-prove repeatability of experimental testing of distributed event-driven sensor network applications. Unlike time-driven appli-cations, such as periodic sampling of environmental condi-tions, the state of event-driven systems can change dependingon the particular sequence of events received and their timing.

The work reported in this paper was supported in part by the NationalScience Foundation grants EHS-0208769, CSR-0509233, and NETS-0435060,and by MURI grant N00014-01-1-0576.†Liqian Luo and Tarek F. Abdelzaher are now with University of Illinois

at Urbana-Champaign.‡Tian He is now with University of Minnesota.

Matters such as tuning protocol parameters for particular eventscenarios or comparing performance of different protocols typ-ically require the same distributed event traces to be replayed(e.g., to ensure a common basis for comparison). To addressthis requirement, we provide an event recording and replayservice that can capture and reproduce distributed events ondemand. The service provides the abstraction of a completelyrepeatable environment as observed by the sensory subsystemfor the sake of experimental testing. Communication propertiesremain stochastic. Hence, a separation is achieved between theeffects of communication non-determinism and the effects ofenvironmental non-repeatability in the study of sensor networkprotocols. The service is especially valuable in the study ofrare, unsafe, or hard-to-reproduce events such as the motionof tracked animals through a sensor field.

Our service is geared for the final stages of testing, typicallyperformed in-field, where the effects of environmental realitiescan be studied. Early testing can use simulators that may begood for initial debugging since they allow fully controlled andrepeatable experiments to be conducted. However, simulatorsare notoriously inaccurate when it comes to sensor networkapplications, since: (i) certain practical issues (e.g., the dis-tribution of radio irregularity) are not adequately capturedin most available simulators, resulting in large discrepanciesbetween simulation results and empirical measurements, and(ii) simulators do not faithfully mimic environmental eventsignatures which affects the performance of sensor networkapplications. The problem is especially severe in event-drivenarchitectures, where application behavior is more sensitive tothe sequence, timing, and parameters of events received.

In addition to improving the repeatability of field testing,our service significantly reduces experimentation cost. In theabsence of a recording and replay service, the investigatorswould have to either physically reproduce or passively waitfor environmental events of interest, which entails additionalcosts. For example, the authors of [7] have developed asurveillance system that tracks persons and vehicles in thefield. The need for walking or driving through the fieldhundreds of times while tuning an array of protocol parametershas proved to be a major practical impediment imposing asignificant limitation on the rate at which experiments couldbe conducted in practice. EnviroLog, the asynchronous eventrecording and replay service described in this paper, providesa comprehensive solution to this problem.

Most sensor platforms employ flash memory for persistent

storage. For example, Mica, Mica2, Mica2Dot [15] and XSM[17] hardware platforms incorporate 128 KB internal flashfor code storage and 512 KB external flash for other usage.EnviroLog logs environmental events into such persistentstorage devices. Later on, in replay mode, EnviroLog replacesenvironmental inputs with retrieved logs and re-issues thelogged events in their original time sequence as asynchronousinputs. In addition to environmental events, EnviroLog canalso log system runtime status for future analysis by recordingselected variable values at runtime as specified by the program-mer using simple annotations.

EnviroLog has two unique features. First, EnviroLog canoperate at any layer of an application. In other words, eventsrecorded and replayed by EnviroLog are not limited to directreflection of environmental events such as raw sensory read-ings. They can be any system-level events. This characteristicof EnviroLog enables the debugging or tuning of any specificlayer using controlled and repeatable inputs from lower layersin an event-driven system. Second, EnviroLog provided avery friendly user interface. Users only need to add simpleannotations before events or variables to be logged. Applica-tions with EnviroLog annotations can be compiled either intoproduction code that ignores all EnviroLog annotations or intodevelopment code that allows on-demand recording and replay.

The potential uses of EnviroLog include (i) in-field de-bugging and performance tuning of specific parameters ofan application, (ii) collecting statistical results from a largenumber of repeated experiments, and (iii) generating tracesfor mixed simulation environments that accept experimentalmeasurements as inputs.

The remainder of the paper is organized as follows. SectionII reviews related work. Section III describes the design goalsand system architecture for EnviroLog. Section IV describesthe implementation details. Section V evaluates EnviroLog,using XSM platforms, through a series of in-field experimentsbased on several sample applications provided by TinyOS [8]and a surveillance system [7] built upon TinyOS. Section VIconcludes the paper.

II. RELATED WORK

In recent years, sensor network researchers have proposedseveral tools and middleware that aid the debugging andevaluation of sensor network applications. Generally, theycan be divided into four categories: simulators, emulators,test-beds and services. The section compares EnviroLog withrelated work in each of these categories.

Simulators are popular tools in debugging and evaluation ofsensor network applications since they don’t usually requirethe deployment of sensor hardware. NS-2 [16], GloMoSim[24] and TOSSIM [11] are good examples. NS-2 is a discreteevent simulator supporting various networking protocols overwired and wireless networks. GloMoSim focuses more onmobile, wireless networks. It allows comparison of multipleprotocols at a given layer. TOSSIM is a simulator especiallydesigned for TinyOS applications, which provides scalablesimulations of sensor network software. Current simulators,

however, do not adequately capture the real behavior of sensornetworks. This is due to the difficulty in modeling practicalimperfections such as radio irregularity as well as due to thelack of good models of environmental inputs. The ability ofEnviroLog to record environmental events can presumably beutilized to improve these tools by importing recorded eventdata to simulate environmental inputs.

Another category of debugging and performance evaluationtools in sensor networks is emulators that mimic sensor deviceseither in software or hardware. AVR JTAG ICE [1], a realtime in-circuit emulator, is a good representative of hardwareemulators. It uses the JTAG interface to enable a user to doreal-time emulation of the microcontroller of sensor devices.A drawback of such in-circuit emulators is that they have tobe physically connected to emulated devices, which causeslogistical difficulties in conducting experiments especially forlarge-scale applications covering a wide field. Atemu [18] isa software emulator for AVR-processor-based systems thatemulates AVR processors as well as other peripheral deviceson the MICA2 platform. Like TOSSIM, Atemu also sim-ulates wireless communication. Such software emulators donot introduce the logistical difficulties exhibited in hardwareemulators, but they are usually less realistic in reproducingnetwork behavior.

The final stages of debugging and performance tuningtypically use actual testbeds to evaluate sensor network appli-cations. For example, Motelab [22] is a public testbed usingMICA2 platforms, which allows users to upload executablesand receive execution results via the Internet. Kansei [20]is another testbed. It employs XSM, MICA2, and Stargateplatforms. EmStar [5] is a combination of emulators andtestbeds for Linux-based sensor network applications, whichruns applications using either a modeled radio channel or thechannel of real nodes. EmTOS [6] extends EmStar to runTinyOS applications by compiling them into EmStar binaries.These testbeds ease the development and evaluation a lotwithout requiring full-scale deployment. However, they do notfocus on repeatability of environmental inputs like EnviroLogdoes.

We categorize all other software facilitating field tests ofsensor network applications as services. EnviroLog belongs tothis category. Monitoring tools such as Message Center [21]aid field tests by capturing messages in the air, filtering anddisplaying them to users. Closest to EnviroLog is TOSHILT[10], a middleware for hardware-in-the-loop testing. TOSHILTdefines emulated stimuli to replace the real environmentalevents, so that applications can be evaluated repeatedly beforethe final deployment. Since TOSHILT uses synthetic andparametric event profiles, the detail of accuracy is less thanwhat can be captured by EnviroLog. In addition, TOSHILTdoesn’t provide abstractions similar to EnviroLog annotationsto ease the integration of the middleware into user applications.All these difference make EnviroLog unique. In the following,EnviroLog is described in more detail.

EnviroSuite: An Environmentally Immersive

Programming Framework for Sensor Networks

LIQIAN LUO and TAREK F. ABDELZAHER

University of Illinois at Urbana-Champaign

TIAN HE

University of Minnesota

and

JOHN A. STANKOVIC

University of Virginia

Sensor networks open a new frontier for embedded distributed computing. Paradigms for sensornetwork programming in the large have been identified as a significant challenge towards devel-oping large-scale applications. Classical programming languages are too low-level. This paperpresents the design, implementation, and evaluation of EnviroSuite, a programming framework

that introduces a new paradigm, called environmentally immersive programming, to abstract dis-tributed interactions with the environment. Environmentally immersive programming refers toan object-based programming model in which individual objects represent physical elements inthe external environment. It allows the programmer to think directly in terms of environmentalabstractions. EnviroSuite provides language primitives for environmentally immersive program-ming that map transparently into a support library of distributed algorithms for tracking andenvironmental monitoring. We show how nesC code of realistic applications is significantly sim-plified using EnviroSuite, and demonstrate the resulting system performance on Mica2 and XSMplatforms.

Categories and Subject Descriptors: C.2.4 [Computer-communication Networks]: DistributedSystems—Distributed Applications; D.2.11 [Software Engineering]: Software Architectures—Domain-specific Architectures; D.3.2 [Software Engineering]: Language Classifications—Spe-cialized Application Languages

General Terms: Design, Experimentation, Languages, Performance

Additional Key Words and Phrases: Abstractions, embedded systems, middleware, programmingmodels, tracking, sensor networks

1. INTRODUCTION

This paper presents EnviroSuite, the first sensor network programming frameworkfor environmentally immersive programming. The need to facilitate software devel-opment for sensor networks motivates new high-level abstractions for programming-in-the-large. These abstractions must hide the details of distributed monitoring and

This work is supported in part by Microsoft and NSF grants EHS-0208769, ITR-0205327, andEHS-0509233.Permission to make digital/hard copy of all or part of this material without fee for personalor classroom use provided that the copies are not made or distributed for profit or commercialadvantage, the ACM copyright/server notice, the title of the publication, and its date appear, andnotice is given that copying is by permission of the ACM, Inc. To copy otherwise, to republish,to post on servers, or to redistribute to lists requires prior specific permission and/or a fee.c© 2005 ACM 1529-3785/2005/0700-0001 $5.00

ACM Transactions on Computational Logic, Vol. V, No. N, October 2005, Pages 1–31.

2 · Liqian Luo et al.

tracking algorithms, capture the unique properties of these networks such as theirdistributed interactions with a physical environment, and address the issue of scale.

Traditional programming languages such as Java and C, as well as their sensornetwork adaptations, such as nesC [Gay et al. 2003] and its extension galsC [Cheonget al. 2003; Cheong and Liu 2005], are too low-level. Their basic computation,communication and actuation unit is typically the sensor node. Programmers mustthink in terms of single node activities and explicitly encode interactions amongnodes. For example, programmers are exposed to reading sensing data from appro-priate sensor devices, aggregating data pertaining to the same external stimulus,deciding where to send it, and communicating with actuators if needed. If themonitored activity moves in the environment, programmers are responsible for spa-tial and temporal correlation of measurements obtained about the activity acrossa changing set of sensor nodes, and associating that data with event progress.

A more desirable approach would be for the programmers to encode only overallnetwork behavior, leaving it to the underlying system to decompose such behaviorinto node-level algorithms. Examples of higher-level abstractions that address thisconcern include logical-node-based primitives [Gummadi et al. 2005], token-basedlanguages [Newton et al. 2005], database-centric abstractions [Madden et al. 2003;Yao and Gehrke 2002; Madden et al. 2005], event-based systems [Li et al. 2004],group-based primitives [Blum et al. 2003; Whitehouse et al. 2004; Welsh and Main-land 2004; Liu et al. 2003], state-centric approaches [Liu et al. 2003] and virtualmachines [Levis and Culler 2002; Boulis et al. 2003]. These paradigms offer logicalnodes, tokens, queries, events, sensor node groups, and logical state, respectively,as the underlying abstractions with which the programmer operates.

EnviroSuite is an object-based programming system. Its abstractions revolvedirectly around elements of the environment as opposed to sensor network con-structs such as regions, neighborhoods, or sensor groups. The existence of thesensor network is thus made more transparent. EnviroSuite is different from otherobject-based systems in that its objects are representations of elements in the ex-ternal environment. Dynamic object instances are created automatically by therun-time system when the corresponding external elements are detected and aredestroyed when these elements leave the network. A unique mapping betweenobject instances and the corresponding environmental elements is maintained bythe system. Object instances float across the network following (geographically)the elements they represent. The execution of object code at the location of thecorresponding physical element is ideal for sensing and actuation tasks. Objectsencapsulate the aggregate state of the elements they represent, making such stateavailable to their methods. These objects (as opposed to the individual nodes) aretherefore the units that encapsulate program data, computation, communication,sensing and actuation. Classical objects (that do not represent any environmentalelements) are also supported. We call the above model, environmentally immersive

programming (EIP).

This paper presents the first comprehensive design and implementation of anenvironmentally immersive programming framework. EnviroSuite abstractions aresupported by an underlying library called EIPLib, which is implemented in nesCon TinyOS [Hill et al. 2000], an operating system designed specifically for sen-

ACM Transactions on Computational Logic, Vol. V, No. N, October 2005.

EnviroTrack: Towards an Environmental Computing Paradigm for DistributedSensor Networks �

Abdelzaher T., Blum B., Cao Q., Chen Y., Evans D., George J., George S., Gu L., He T., Krishnamurthy S.,Luo L., Son S., Stankovic J., Stoleru R., Wood A.

Department of Computer Science, University of Virginia, Charlottesville, VA 22904

AbstractDistributed sensor networks are quickly gaining recognition asviable embedded computing platforms. Current techniques forprogramming sensor networks are cumbersome, inflexible, andlow-level. This paper introduces EnviroTrack, an object-baseddistributed middleware system that raises the level of program-ming abstraction by providing a convenient and powerful in-terface to the application developer geared towards trackingthe physical environment. EnviroTrack is novel in its seamlessintegration of objects that live in physical time and space intothe computational environment of the application. The perfor-mance of an initial implementation of the system is evaluatedon an actual sensor network based on MICA motes. Resultsdemonstrate the ability of the middleware to track realistic tar-gets.

Keywords: sensor networks, programming paradigms, track-ing, QoS, distributed systems

1 Introduction

The work reported in this paper is prompted by the increas-ing importance of large-scale wireless sensor networks [15] asa future platform for a growing number of applications suchas habitat monitoring [7, 21], intrusion detection [29], defense,and scientific exploration. Advances in hardware miniaturiza-tion [10] have made it economically viable to develop embed-ded systems of massively distributed disposable sensor nodes,characterized by coordination of a very large number of tinywireless computing elements. A great impediment to rapid de-ployment of such systems lies in the lack of distributed soft-ware and programming support for sensor network applica-tions. A new distributed computing paradigm is needed thatexports appropriate abstractions and implements efficient infor-mation management protocols in large-scale sensor networks.EnviroTrack is an attempt to develop such a paradigm.

EnviroTrack is a middleware layer that exports a new ad-dress space in the sensor network. In this space, physical eventsin the external environment are the addressable entities. This

�The work reported in this paper was supported in part by the NationalScience Foundation grants EHS-0208769, CCR-0205327, and CCR-00-92945,DARPA grant F33615-01-C-1905, and MURI grant N00014-01-1-0576.

type of addressing is convenient for applications that need tomonitor environmental events. For example, a surveillance ap-plication that monitors vehicle movement behind enemy linesmay assign unique labels to individual vehicles. Their statecan then be addressed by reference to these labels. Moreover,computing or actuation objects can be attached to individualaddresses in much the same way computation is assigned toIP hosts in an Internet-like environment. Such attached com-putation or actuation is then performed in the physical neigh-borhood of the named entity. Hence, for example, a micro-phone could be turned on at some network address (e.g., onethat names a vehicle in the external environment) to listen-inon the corresponding environmental object. As the named vehi-cle moves, the middleware will turn on the appropriate nearbynode microphones such that a non-interrupted audio stream isdelivered to the receiver despite the mobile nature of the source.Communication can also occur between two mobile endpoints.For example, a walking soldier with a PDA may track the po-sition of a suspect vehicle detected elsewhere in the network.In short, we (i) export a novel logical address space in whichexternal environmental objects are the labeled entities, and (ii)allow arbitrary data, computation, or actuation to be attached tosuch logical network addresses. These data, computation, andactuation are encapsulated in an abstraction we call trackingobjects.

The EnviroTrack middleware library implements a set ofprotocols that off-load from an application developer the de-tails of inter-object communication, object mobility, as well asthe maintenance of tracking objects and their state. It abstractsaway the fact that computation associated with the object maybe distributed and performed by all sensor nodes in the vicin-ity of the tracked physical entity. As the tracked entity moves,the identity and location of the sensor nodes in its neighbor-hood change, but the tracking object representing it remains thesame. The programmer thus interacts with a changing group ofsensor nodes through a simple, uniquely addressable, object in-terface.

EnviroTrack has been implemented and tested on a pop-ular sensor network platform based on MICA motes [16].Our initial implementation of this infrastructure uses compiledNesC [13] programs on TinyOS [15], an operating system for

sensor networks. Recent advances in programming supportfor sensor networks, such as the development of a virtual ma-chine [19], will significantly simplify the code developmentand dissemination effort in the future. We present evaluation re-sults, which illustrate how typical sensor-network applicationsthat use EnviroTrack will perform on the current hardware plat-form.

The rest of this paper is organized as follows. Section 2defines the tracking problem in more detail, describes our pro-gramming system architecture, and elaborates on the main ab-stractions provided by EnviroTrack. Section 3 illustrates howa sample tracking application can be written in EnviroTrack.Section 4 provides implementation details. Section 5 presentsa performance evaluation. An overview of related work is pre-sented in Section 6. The paper concludes with Section 7.

2 Programming Model

The programmer’s view of an application written in Enviro-Track is depicted in Figure 1. Sensors which detect certainuser-defined entities in the physical environment form groups,one around each entity. A network abstraction layer associatesa context label with each such group to represent the corre-sponding tracked entity in the computing system. Context la-bels can be thought of as logical addresses of virtual hosts (con-texts) which follow the external tracked entity around in thephysical environment. In the following, we use contexts andcontext labels interchangeably. Objects can be attached to con-text labels to perform context-specific computation. These at-tached objects are called tracking objects. They are executedon the sensor group of the context label. Since the actual loca-tion of the tracking object is the nodes in the physical vicinityof the target, the object can perform local sensing and actuationto interact directly with the target’s locale. For completeness,EnviroTrack also supports conventional static objects that arenot attached to context labels.

Context labels have types depending on the entity tracked.For example, a context label of type CAR is created wherevera car is observed. To declare a context label of some type e

(named after the tracked event type), the programmer must sup-ply three pieces of information. First, the programmer suppliesa function sensee�� that describes the sensory signature iden-tifying the tracked environmental target. For example, if thecontext type is to identify moving vehicles, sensee�� might bea function of magnetometer and motion sensor readings. Themiddleware watches for the specified sensory pattern in the en-vironment and creates a sensor group around the detected targetwhen the pattern occurs. This function is also used to maintainthe membership of the sensor group around the tracked targetwhen the target moves. Group membership, in this case, is re-stricted to those nodes that sense the given target (i.e., for whichsensee�� is true).

Second, the programmer declares what constitutes the envi-ronmental state to be encapsulated in the context label. This

Sensor Network Abstraction Layer

Aggreg.State

Aggreg.State

Sensing CAR Sensing FIREGroupGroup

Context type: CARContext Label: Car02 Context Label: Fire01

VehicleTrackingObjects

Fire

ObjectsTracking

Context type: FIRE

Figure 1. Programming Model

state is shared by all tracking objects attached to this label.State is declared by defining an aggregation function statee��that acts on the readings of all sensors for which sensee�� istrue or was true within a recent past defined by a freshnessconstraint. The aggregation is carried out locally by a sen-sor node that acts as the group leader of all sensors sensingthe named target. The aggregation function can also includea critical mass constraint that specifies the minimum numberof sensors that must be involved in the aggregation for the re-sult to be statistically meaningful. EnviroTrack provides a li-brary of the most common distributed aggregation functionsto choose from, such as addition, averaging, and median com-putation. These functions can also be location-aware, for ex-ample, to compute the center of gravity of the measurements.The underlying infrastructure includes a data collection proto-col executed by the leader to collect, timestamp, and log sensordata (i.e., the arguments for the statee�� function) from sen-sor group members satisfying sensee��. The statee�� functionis then applied on the collected data in a way that satisfies thefreshness and critical mass conditions. Finally, the program-mer specifies which objects are to be attached to the contextlabel. Attached object code can reference the aggregate statemaintained by the leader in this context.

In the following, we describe in more detail the network ab-straction layer, tracking objects, and aggregate state.

2.1 Network Abstraction Layer

Context labels abstract sensor groups for the programmer. Theprogrammer is aware that a distributed computation, associ-ated with the context label, is executed on multiple sensors inthe vicinity of a tracked entity. The programmer, however, isnot involved in managing the membership, leader election, and

Energy-Efficient Surveillance System Using WirelessSensor Networks∗

Tian He, Sudha Krishnamurthy, John A. Stankovic, Tarek Abdelzaher,Liqian Luo, Radu Stoleru, Ting Yan, Lin Gu

Department of Computer ScienceUniversity of Virginia, Charlottesville, VA 22903

{tianhe, skrish, stankvoic, zaher, ll4p, rs6bd, ty4k, lg6e}@cs.virginia.edu

Jonathan Hui, Bruce Krogh

Department of Electrical and Computer EngineeringCarnegie-Mellon University, Pittsburgh, PA 15213

jwhui@cs.berkeley.edu, krogh@ece.cmu.edu

ABSTRACTThe focus of surveillance missions is to acquire and verify in-formation about enemy capabilities and positions of hostiletargets. Such missions often involve a high element of riskfor human personnel and require a high degree of stealth-iness. Hence, the ability to deploy unmanned surveillancemissions, by using wireless sensor networks, is of great prac-tical importance for the military. Because of the energyconstraints of sensor devices, such systems necessitate anenergy-aware design to ensure the longevity of surveillancemissions. Solutions proposed recently for this type of systemshow promising results through simulations. However, thesimplified assumptions they make about the system in thesimulator often do not hold well in practice and energy con-sumption is narrowly accounted for within a single protocol.In this paper, we describe the design and implementation ofa running system for energy-efficient surveillance. The sys-tem allows a group of cooperating sensor devices to detectand track the positions of moving vehicles in an energy-efficient and stealthy manner. We can trade off energy-awareness and surveillance performance by adaptively ad-justing the sensitivity of the system. We evaluate the per-formance on a network of 70 MICA2 motes equipped withdual-axis magnetometers. Our results show that our surveil-lance strategy is adaptable and achieves a significant exten-sion of network lifetime. Finally, we share lessons learned inbuilding such a complete running system.

∗This work was supported by the DAPRPA IXO offices un-der the NEST project (grant number F336615-01-C-1905).

Permission to make digital or hard copies of all or part of this work forpersonal or classroom use is granted without fee provided that copies arenot made or distributed for profit or commercial advantage and that copiesbear this notice and the full citation on the first page. To copy otherwise, torepublish, to post on servers or to redistribute to lists, requires prior specificpermission and/or a fee.MobiSYS’04, June 6–9, 2004, Boston, Massachusetts, USA.Copyright 2004 ACM 1-58113-793-1/04/0006 ...$5.00.

Categories and Subject DescriptorsC.2.1 [Computer Communication Networks]: NetworkArchitecture and Design

General TermsDesign, Performance, Experimentation, Measurement

KeywordsSensor networks, Energy conservation, Tracking, Wireless

1. MOTIVATIONOne of the key advantages of wireless sensor networks

(WSN) is their ability to bridge the gap between the phys-ical and logical worlds, by gathering certain useful infor-mation from the physical world and communicating thatinformation to more powerful logical devices that can pro-cess it. If the ability of the WSN is suitably harnessed, itis envisioned that WSNs can reduce or eliminate the needfor human involvement in information gathering in certaincivilian and military applications. In the near future, sensordevices will be produced in large quantities at a very lowcost and densely deployed to improve robustness and reli-ability. They can be miniaturized into a cubic millimeterpackage (e.g., smart dust [16]) in order to be stealthy in ahostile environment. Cost and size considerations imply thatthe resources available to individual nodes are severely lim-ited. We believe, however ,that limited processor bandwidthand memory are temporary constraints in sensor networks.They will disappear with fast developing fabrication tech-niques. The energy constraints on the other hand are morefundamental. According to R.A. Powers [20], battery capac-ity only doubles in 35 years. Energy constraints are unlikelyto be solved in the near future with the slow progress inbattery capacity and energy scavenging. Moreover, the un-tended nature of sensor nodes and the hazardous sensing en-vironment preclude manual battery replacement. For thesereasons, energy awareness becomes the key research chal-lenge for sensor network protocol design. Several researchershave addressed energy conservation recently. Most of them

Figure 1: Sensor Network Deployment

focus on particular protocols and investigate whether theirenergy conservation goal can be achieved. To the best of ourknowledge, none of them investigate energy-conservation fora running system as whole. Normally they evaluate their ap-proach through simulations. Simulation approaches tend tomake simplified assumptions that often do not hold well inpractice and they are subject to incompleteness. For ex-ample, in [22][23][21], several sensing coverage schemes areproposed for energy conservation. None of them considerenergy consumption in activities other than sensing.

In this paper, we describe our effort that involves systemdesign and implementation on a MICA2 platform with 70MICA2 motes. The primary goal of the system is to sup-port the ability to track the position of moving targets inan energy-efficient and stealthy manner. Our experimentalresults show that the probability of false alarms observedreaches zero when aggregation is achieved among more than3 member motes. The experimental results we obtained alsoshow that with 5% of deployed motes serving as sentries andthe non-sentries operating at a 4% duty cycle, our algorithmextends the lifetime of a sensor network by up to 900%.

The main contributions of this paper are: 1) the de-sign and implementation of a running system with energy-awareness as the main design principle across multiple com-ponents, 2) Mechanisms for dynamic control, which allowtradeoffs between energy-efficiency and system performanceby adjusting the sensitivity of the system, and 3) a physicalimplementation and field evaluation that reveal the practical

issues that are hard to capture in simulation.The remainder of this paper is organized as follows. Sec-

tion 2 describes the requirements of a typical ground surveil-lance application. In Section 3, we describe the system setupand hardware components. In Section 4, we provide anoverview of our system design. In Section 5, we elaborate onhow the individual components of the system contribute toenergy-efficient tracking. In Section 6, we discuss implemen-tation issues concerning our system. We present experimen-tal results in Section 7, and summarize the lessons learnedfrom our experience in Section 8. Finally we conclude inSection 9 and discuss some future work in Section 10.

2. APPLICATION REQUIREMENTSOur system design is motivated by the requirements of a

typical ground surveillance application. The general objec-tive of such an application is to alert the military commandand control unit in advance to the occurrence of events of in-terest in hostile regions. The event of interest for our workis the presence of moving vehicles in the deployed region.The deployed sensor devices must have the ability to de-tect and track vehicles in the region of interest. Successfuldetection and tracking requires that the application obtainthe current position of a vehicle with acceptable precisionand confidence. When the information is obtained, it has tobe reported to a remote base station within an acceptablelatency. Several application requirements must be satisfiedto make this system useful in practice:

• Longevity: The mission of a surveillance applicationtypically lasts from a few days to several months. Dueto the confidential nature of the mission and the inac-cessibility of the hostile territory, it may not be pos-sible to manually replenish the energy of the power-constrained sensor devices during the course of themission. Hence, the application requires energy-awareschemes that can extend the lifetime of the sensor de-vices, so that they remain available for the duration ofthe mission.

• Adjustable Sensitivity: The system should have anadjustable sensitivity to accommodate different kindsof environments and security requirements. In criticalmissions, a high degree of sensitivity is desired to cap-ture all potential targets even at expense of possiblefalse alarms. In other case, we want to decrease thesensitivity of the system, maintaining a low probabilityof false alarms in order to avoid inappropriate actionsand unnecessary power dissipation.

• Stealthiness: It is crucial for military surveillancesystems to have a very low possibility of being detectedand intercepted. Miniaturization makes sensor deviceshard to detect physically; however, RF signals can beeasily intercepted if sensor devices actively communi-cate during the surveillance stage. A zero communi-cation exposure is desired in the absence of significantevents.

• Effectiveness: The precision in the location estimate,and the latency in reporting an event are the metricsthat determine the effectiveness of a surveillance sys-tem. Accuracy and latency are normally consideredimportant metrics of tracking performance. However,

Achieving Real-Time Target Tracking Using Wireless Sensor Networks

Tian He§, Pascal Vicaire†, Ting Yan†, Liqian Luo∗, Lin Gu†, Gang Zhou†,Radu Stoleru†, Qing Cao∗, John A. Stankovic† and Tarek Abdelzaher∗

†Department of Computer Science, University of Virginia§Department of Computer Science and Engineering, University of Minnesota

Abstract

Target tracking systems, consisting of thousands of low-costsensor nodes, have been used in many application domains suchas battlefield surveillance, wildlife monitoring and border se-curity. These applications need to meet certain real-time con-straints in response to transient events, such as fast-moving tar-gets. While the real-time performance is a major concern inthese applications, it should be compatible with other impor-tant system properties such as energy consumption and accu-racy. Hence, it is desirable to have the ability to exploit thetradeoffs among them. This work presents the real-time designand analysis of VigilNet, a large-scale sensor network systemwhich tracks, detects and classifies targets in a timely and en-ergy efficient manner. Based on a deadline partition methodand theoretical derivations of each sub-deadline, we are able tomake guided engineering decisions to meet the end-to-end track-ing deadline. To confirm our design and obtain an empiricalunderstanding of these tradeoffs, we invest significant efforts toperform large-scale simulations with 10,000 nodes as well asa field test with 200 XSM motes, running VigilNet. The resultsfrom both analysis and evaluation can serve as general designguidelines to build similar real-time systems.

1 Introduction

Recent developments in sensor techniques make wirelesssensor networks (WSNs) available to many application do-mains [6, 12, 17, 26, 32]. Most of these applications, such as bat-tlefield surveillance, disaster and emergency response, deal withvarious kinds of real-time constraints in response to the physicalworld. For example, surveillance may require a sensor node todetect and classify a fast moving target within 1 second before itmoves out of the sensing range. Compared with the traditionaldistributed systems, the real-time guarantee for sensor networksis more challenging due to the following reasons. First, sen-

∗Liqian Luo, Qing Cao and Tarek Abdelzaher are now with University ofIllinois, Urbana-Champaign

sor networks directly interact with the real world, in which thephysical events may exhibit unpredictable spatiotemporal prop-erties. These properties are hard to characterize with the tradi-tional methods. Second, although the real-time performance isa key concern, it should be performance compatible with manyother critical issues such as energy efficiency and system robust-ness. For example, it is not efficient to activate the sensors all thetime only for the benefit of a fast response. This naive approachseverely reduces the system lifetime [12]. Third, the resourceconstraints restrict the design space we could trade off. For ex-ample, the limited computation power in sensor nodes makes theFast Fourier Transformation not quite suitable for real-time de-tection. All these issues challenge us with two questions. Howto make the design of a large-scale real-time sensor network sys-tem manageable? And how to trade off among system metricswhile maintaining the real-time guarantee? Our answer to thesequestions, presented in this paper, is a case study of the VigilNetsystem, a real-time outdoor tracking system using a large-scalewireless sensor network.

Our contribution lies in the following aspects: 1) This workaddresses a real-world application with a running real-time sys-tem, designed and implemented over last few years. 2) We inves-tigate multi-dimensional tradeoffs between the real-time perfor-mance and other system properties. Such investigation providesthe guidance for the future development of similar systems. 3)The real-time design and tradeoffs are validated by a large-scalefield evaluation with 200 XSM motes and an extensive simu-lation with 10,000 nodes. These evaluations reveal quite a fewpractical design suggestions that can be applied to other real-timesensor systems.

The remainder of the paper is organized as follows: Section 2introduces the tracking process in VigilNet. Section 3 identi-fies the real-time requirements. Section 4 provides a real-timeanalysis of VigilNet’s tracking performance and its tradeoffs. InSection 5, we evaluate the real-time performance of VigilNet inan outdoor field test. In Section 6, we conduct a large-scalesimulation to further validate and analyze the real-time issuesin VigilNet. Section 7 discusses the related work. Section 8 con-cludes the paper.

1

����� � � ���� ��� ������ ���

��� ��� ���� � � �� ��� ��� � ��� �����

� ��� � �"! ! � �#! ��� � ���

$%��&�' � � '�$(�)&* �+� �����

,-�#. �/���� �� .0.0� � !

12��� !�� �$(��� �3��� ��

4(� ���+��* ��� �����

5+607�8 9;: 5+607�8 9 < 5+607�8 9�=

5�607�8 9;> 5�607�8 9 ? 5�60738 9�@

Figure 1: The Delay Breakdown in Tracking Operation

2 Overview of VigilNet Tracking Operations

VigilNet is an energy-efficient surveillance and tracking sys-tem, designed for spontaneous military operations in remote ar-eas. In these areas, the events of interest happen at a relativelylow rate, i.e. the duration of significant events (e.g., intrud-ers) is very short, compared with the overall system lifetime(e.g., 5-minute tracking per day). According to our empiricalresults [13], nearly 99% of energy is consumed during the idle-waiting period for potential targets. Therefore to conserve en-ergy, the most effective approach is to selectively turn a subset ofnodes off, and wake them up on demand in the presence of sig-nificant events. This power management technique fundamen-tally shapes the VigilNet tracking process. It introduces a set ofnew delays that traditional tracking systems do not experience.

In this section, we give a brief overview of the VigilNet track-ing operation, serving as a background for the real-time designand analysis in the following sections. As shown in Figure 1,after a target enters the area, it activates the first sensor node thatcan confirm the detection, then other nodes nearby are wakenup to form a group to deliver the aggregated reports to the base.More specifically, the VigilNet tracking operation has six phases:

A. Initial Activation: VigilNet stays in the power manage-ment state when there are no targets. The power manage-ment protocol puts nodes into either one of two states: sen-try and non-sentry. In brief, a node becomes a sentry nodeif it is a part of the routing infrastructure or it needs to pro-vide the sensing coverage. Otherwise, it becomes an inac-tive non-sentry node. The details of sentry selection can befound in [12]. If the sentry nodes are active 100% of time(i.e. the deployed area is always covered), any incomingtarget is covered by at least one sentry node immediately.On the other hand, if the sentry nodes have a certain dutycycle (i.e. they go to sleep and wake up periodically to saveenergy), there will be an initial activation delay, denoted asTinitial, before the first sentry node starts to sense the in-coming target.

B. Initial Target Detection: After the initial activation, ittakes a certain delay, defined as Tdetection, for the first sen-try node to confirm the detection. This delay consists of thehardware response delay, the discrete sampling delay andthe delay to accumulate a sufficient number of samples be-fore a detection algorithm recognizes the target.

C. Wake-up: Normally, the detection from a single sentrynode does not provide sufficient confidence in detection andclassification, therefore a group-based tracking is designedin VigilNet. In order to form a group with a reasonable size,non-sentry nodes need to be waken up after the initial targetdetection by a sentry node in Phase B. We define the wake-up delay Twakeup as the time required for a sentry node towake up other sleeping non-sentry nodes. This delay is de-termined by the time to broadcast the wake-up messages.

D. Group Aggregation: Once awaken, all nodes that detectthe same target join the same logic group to establish aunique one-to-one mapping between this logic group andthe detected target. Each group is represented by a leaderto maintain the identity of the group as the target movesthrough the area. Group members (who by definition cansense the target) periodically report to the group leader. Aleader starts to report detection to the base after the numberof member reports exceeds a certain threshold, defined asthe degree of aggregation (DOA). We use Taggregation todenote the group aggregation delay, which is the time re-quired to collect and process the detection reports from themember nodes.

E. End-to-End Report: After the group aggregation, theleader node reports the event to the nearest base. Multiplebases are used to partition a network into several sections,in order to bound the end-to-end delivery delay Te2e.

F. Base Processing (Tbase): A base is in charge of processingthe reports from the leaders of different logic groups. Sincethe reports from the same logic group are spatiotemporalcorrelated, a string of consecutive reports can be further an-alyzed and summarized for end users. For example, takingthe time stamps and the locations of targets as the inputs, abase uses the least-square estimation to obtain the velocityof each target.

3 Real-Time Requirement in VigilNet

To ensure the effectiveness of the target tracking, VigilNetmust meet a certain real-time constraint. Specifically, VigilNetshould detect, classify and analyze the incoming targets within acertain end-to-end deadline (e.g.,5 seconds from Phase A to F).As shown in Section 2, this deadline involves complex analysisof the whole tracking process. It is not scalable for us to identifya system-wide feasible region within such a high-dimensionaldesign space. Therefore, we adopt the deadline partition methodto divide the end-to-end deadline into multiple sub-deadlines.The sub-deadline partition varies with the system configurations.As a concrete example, supposing a target enters the field with aspeed up to 20 mph, to guarantee that this target can be detectedby the first sentry node with a probability higher than 90% , weneed to design a detection algorithm with a sub-deadline lessthan 1 second, assuming the detection range is 10 meters.

2

Achieving Real-Time Target Tracking Using

Wireless Sensor Networks

Tian He§, Pascal Vicaire†, Ting Yan†, Liqian Luo‡, Lin Gu†, Gang Zhou†,

Radu Stoleru†, Qing Cao‡, John A. Stankovic† and Tarek Abdelzaher‡

†Department of Computer Science, University of Virginia§Department of Computer Science and Engineering, University of Minnesota‡Department of Computer Science, University of Illinois, Urbana-Champaign

Target tracking systems need to meet certain real-time constraints in response to transientevents, such as fast-moving targets. While the real-time performance is a major concern in theseapplications, it should be compatible with other important system properties such as energyconsumption and accuracy. This work presents the real-time design and analysis of VigilNet,

a large-scale sensor network system which tracks, detects and classifies targets in a timely andenergy efficient manner. Based on a deadline partition method and theoretical derivations to

guarantee each sub-deadline, we are able to make guided engineering decisions to meet the end-to-end tracking deadline. The results from 10,000-node simulation and 200 mote field test revealthe effectiveness of our design.

Categories and Subject Descriptors: C.3. [Special-Purpose and Application-Based Sys-

tems]: Real-Time and Embedded Systems; C.2. [Computer Communication Networks]:

Distributed Systems

General Terms: Design, Performance, Experimentation

Additional Key Words and Phrases: Sensor Networks, Real-Time, Tracking, Energy Conservation

1. INTRODUCTION

Recent developments in sensor techniques make wireless sensor networks (WSNs)available to many application domains [Dutta et al. 2005; He et al. 2004; Juanget al. 2002; Simon et al. 2004; Xu et al. 2004; Arora and et al. 2005]. Most of theseapplications, such as battlefield surveillance, disaster and emergency response, dealwith various kinds of real-time constraints in response to the physical world. Forexample, surveillance may require a sensor node to detect and classify a fast movingtarget within 1 second before it moves out of the sensing range. Compared with thetraditional distributed systems, the real-time guarantee for sensor networks is morechallenging due to the following reasons. First, sensor networks directly interactwith the real world, in which the physical events may exhibit unpredictable spa-tiotemporal properties. These properties are hard to characterize with traditionalmethods. Second, although the real-time performance is a key concern, it shouldbe performance compatible with many other critical issues such as energy efficiencyand system robustness. For example, it is not efficient to activate the sensors all thetime only for the benefit of a fast response. This naive approach severely reduces

the system lifetime [He et al. 2004]. Third, the resource constraints restrict thedesign space we could trade off. For example, the limited computation power insensor nodes makes the Fast Fourier Transformation not quite suitable for real-timedetection. All these issues challenge us with two questions. How to make the designof a large-scale real-time sensor network system manageable? And how to trade offamong system metrics while maintaining the real-time guarantee? Our answer tothese questions, presented in this paper, is a case study of the VigilNet system, areal-time outdoor tracking system using a large-scale wireless sensor network.

Our contribution lies in the following aspects: 1) This work addresses a real-world application with a running real-time system, designed and implemented overthe last few years. 2) We demonstrate how to guarantee the end-to-end trackingdeadline in a complex sensor system. For a given sub-deadline partition, we identifythe system configurations that meet the sub-deadlines without compromising otherimportant system properties. 3) The real-time design and tradeoffs are validatedby a large-scale field evaluation with 200 XSM motes and an extensive simulationwith 10,000 nodes. These evaluations reveal quite a few practical design suggestionsthat can be applied to other real-time sensor systems.

The remainder of the paper is organized as follows: Section 2 introduces the track-ing process in VigilNet. Section 3 identifies the real-time requirements. Section 4provides a real-time analysis of VigilNet’s tracking performance and its tradeoffs. InSection 5 we describe the implementation of the VigilNet system. In Section 6, weevaluate the real-time performance of VigilNet in an outdoor field test. In Section 7,we conduct a large-scale simulation to further validate and analyze the real-timeissues in VigilNet. Section 8 discusses the related work. Section 9 concludes thepaper.

2. OVERVIEW OF VIGILNET TRACKING OPERATIONS

VigilNet is an energy-efficient surveillance and tracking system, designed for spon-taneous military operations in remote areas. In these areas, the events of interesthappen at a relatively low rate, i.e. the duration of significant events (e.g., in-truders) is very short, compared with the overall system lifetime (e.g., 5-minutetracking per day). According to our empirical results [He et al. 2006], nearly 99%of energy is consumed during the idle-waiting period for potential targets. There-fore to conserve energy, the most effective approach is to selectively turn a subset ofnodes off, and wake them up on demand in the presence of significant events. Thispower management technique fundamentally shapes the VigilNet tracking process.It introduces a set of new delays that traditional tracking systems do not experience.

In this section, we give a brief overview of the VigilNet tracking operation, servingas a background for the real-time design and analysis in the following sections. Asshown in Figure 1, after a target enters the area, it activates the first sensor nodethat can confirm the detection, then other nodes nearby are awakened to form agroup to deliver the aggregated reports to the base. More specifically, the VigilNettracking operation has six phases:

A. Initial Activation: VigilNet stays in the power management state when thereare no targets. The power management protocol puts nodes into either one oftwo states: sentry and non-sentry. In brief, a node becomes a sentry node

An Overview of the VigilNet Architecture

Tian He, Liqian Luo, Ting Yan, Lin Gu, Qing Cao, Gang Zhou, Radu Stoleru

Pascal Vicaire, Qiuhua Cao, John A. Stankovic, Sang H. Son and Tarek F. Abdelzaher

Department of Computer Science

University of Virginia, Charlottesville, VA 22903

Abstract

Battlefield surveillance often involves a high element of risk

for military operators. Hence, it is very important for the mili-

tary to execute unmanned surveillance by using large-scale wire-

less sensor systems. This invited paper summarizes the archi-

tecture of the VigilNet system – a long-term real-time networked

sensor system for military surveillance. Specifically, we review

the design of several major subsystems within VigilNet includ-

ing sensing and classification, localization, tracking, networking,

power management, reconfiguration, graphic user interface, and

the debugging subsystem. High-level programming abstractions

are also presented. This is a balanced design to achieve real-

time response, high confidence detection, accurate tracking and

energy efficiency simultaneously.

1. INTRODUCTION

Recently, many efforts have been undertaken to support new

military applications by using large-scale wireless sensor networks.

Unmanned real-time surveillance is one of the most promising ap-

plications. By combining the computation, sensing, actuation and

wireless networking together, large-scale sensor networks have

several advantages over many other distributed systems. First,

sensor networks can be quickly deployed in an infrastructure-free

environment, which is highly desired in military operations. Sec-

ond, the redundancy introduced by a large-scale dense deploy-

ment makes a sensor system robust to node failures, which is

critical in a hostile environment. Third, massively distributed in-

network data processing allows exploration of new techniques for

detection and classification, not posible with only a small number

of devices. Along with these advantages, however, several chal-

lenges arise. The constrained resources in wireless sensor nodes,

such as limited memory, power, processing, and communication

bandwidth, impose problems previous research did not need to

address. To realize the vision of wireless surveillance systems,

many research efforts have been published that address challeng-

ing problems concerning networking [14], self-organization [4],

energy-conservation [27, 30] and tracking [16] in this type of sys-

tems. However, most existing efforts address individual protocols

or subproblems in the design space. Few systems actually pro-

vide a complete architecture and a running implementation tested

in outdoor environments. The VigilNet project is, therefore, dis-

tinguished in this aspect. It is a large-scale sensor network sys-

tem which has been successfully designed, built, demonstrated

and delivered to the military for realistic deployment. To accom-

plish different mission objectives, the VigilNet system consists

of 40,000 lines of code, supporting multiple existing mote plat-

forms including MICA2DOT, MICA2, and XSM. To accommo-

date various mission requirements, VigilNet is dynamically re-

configurable and reprogrammable. For example, we can flexibly

explore trade-offs between surveillance quality parameters and

network lifetime by adjusting various system parameters online.

The remainder of this paper is organized as follows: Section 2

discusses the related work. Section 3 presents the overarching

architecture of VigilNet and the details of individual subsystems.

We list several lessons learned from our experience in Section 4

and conclude the paper in Section 5.

2. RELATED WORK

Middleware services form the basis of sensor network sys-

tems. Localization is a key service to identify the locations from

where sensor readings are obtained. Two categories of local-

ization have been proposed: range-based schemes [22, 24] and

range-free schemes [2, 11]. The former category uses absolute

point-to-point distance estimates (range) or angle estimates to lo-

calize nodes; The latter makes no assumptions about the availabil-

ity of such information. Time synchronization is another critical

middleware service. The reference broadcast scheme (RBS) pro-

posed in [7] maintains information about the phase and frequency

of each pair of clocks in the neighborhood of a node. While

RBS achieves a precision of about 1 µs, the message overhead

in maintaining the neighborhood information is high and may not

be energy-efficient in large-scale systems. Maroti [21] synchro-

nizes the network through limited flooding with timestamp values

reassigned at intermediate nodes immediately prior to transmis-

sion. This scheme reduces the synchronization error introduced

by uncertainty due to MAC contention. Our VigilNet system uses

a variation of this approach. In addition, power management is

employed to ensure network longevity. Other important proto-

1

cols developed for modern sensor network hardware [5, 6] in-

clude medium access control [23, 31], sensing coverage [27, 30],

energy aware routing [29], data aggregation [20, 10], topology

management [4] and energy-aware applications [12, 26, 28].

With the assistance of various middleware services, several

outdoor sensor network systems have recently been built. The

Great Duck Island Project [26] explores long-term habitat moni-

toring in remote environments. ZebraNet [15] focuses on wildlife

tracking in Africa. The Extreme Scaling project [6] investigates

scalability issues in surveillance systems. The shooter localiza-

tion system [8] combines precise time synchronization with accu-

rate acoustic sensing to localize positions of snipers. The Wisden

system [28] monitors the health of building structures by con-

tinuously retrieving structural response data. To complement the

aforementioned efforts, VigilNet aims at building a practical mil-

itary surveillance system, which can survive in harsh and hostile

environments for a long period of time, and exhibit a high de-

tection, classification and tracking performance. These require-

ments necessitate the design of unique solutions to various sens-

ing, communication, and tracking problems [9, 17, 18, 19, 25, 32]

and call for seamless integration of the resulting middleware com-

ponents [12].

3. VIGILNET ARCHITECTURE

The VigilNet system has a layered architecture as shown in

Figure 1. This architecture provides an end-to-end solution for

supporting military surveillance applications. As an overview,

this paper focuses on the design of individual components of this

system, but omits the details of system implementation and per-

formance evaluation. Such details can be found in other publica-

tions by the authors [9, 12, 17, 18, 19, 25, 32]

3.1 Sensing Subsystem

Sensing is the basis for any surveillance system. The VigilNet

sensing subsystem implements detection and classification of tar-

gets using continuous online sensor calibration (to a changing en-

vironment) and frequency filters to determine critical target fea-

tures. These filters extract the target signatures from a specific

spectrum band, eliminating the burden of applying a computation-

intensive Fast Fourier Transform. The sensing subsystem con-

tains three detection algorithms for the magnetic sensor, acoustic

sensor and passive infrared sensor (PIR), respectively.

• The magnetic sensor detection algorithm computes two mov-

ing averages of most recent magnetic readings. The slower

moving average, with more weight on previous readings,

establishes a baseline to follow the thermal drift noise caused

by the changing temperature during the day. The faster

moving average, with more weights on the current read-

ing, detects the swift change in magnetic filed caused by

ferrous targets. To make a detection decision, the differ-

ence between the two moving average values is compared

to a dynamic threshold, which is established during the cal-

ibration phase.

• The acoustic sensor detection algorithm uses a lightweight

power-based approach. It first computes a moving aver-

age of multiple recent acoustic readings, then establishes an

auto-adapting acoustic threshold by calculating a moving

standard deviation of readings over a certain time window.

If an acoustic reading is larger than the sum of the moving

average and its corresponding moving standard deviation,

we consider it is a crossover. If the number of crossovers

exceeds a certain threshold during a unit of time, this algo-

rithm signals a detection to the upper layer components.

• The passive infrared sensor is designed to sense changes

in thermal radiation that are indicative of motion. When

there is no movement, the thermal reading is stable and

does not trigger detections. If an object is moving in front

of a PIR sensor, this object causes a thermal disturbance,

triggering the PIR. Most moving objects, such as shaking

leaves, rain drops, and vehicles, can trigger the PIR sen-

sor. However, different thermal signatures generate trig-

ger events with different frequencies. Low frequency de-

tections (< 2Hz) are normally triggered by wind-induced

motion and other slow moving objects. On the other hand,

fast-moving targets such as vehicles generate signals with a

much higher frequency. Therefore, it is sufficient to design

a high pass ARMA filter to filter out the frequency com-

ponents lower than 2Hz. Similar to other detection algo-

rithms, this threshold is dynamically adaptated to accom-

modate the changing environment. For example, thermal

noise is much higher in a hot and humid environment than

that in a dry and cool one. In fact, thermal and humidity

variations over the course of a day can significantly change

threshold values.

Due to the space constraints, we do not detail the specific im-

plementation of each sensing algorithm. More information on this

subsystem can be found in [9].

3.2 Context-Awareness Subsystem

Sensed data is meaningful only when it is interpreted along

with the context in which it is obtained. For example, a temper-

ature reading is useless, if we don’t know where and when such

value is measured. The Context-Awareness subsystem comprises

lower-level context detection components such as time synchro-

nization and localization. These components form the basis for

implementing other subsystems, such as the tracking subsystem.

Localization ensures that each node is aware of its location, so

that we can determine the location of detected targets. Time syn-

chronization is responsible for synchronizing the local clocks of

nodes with the clock of the base station, so that every node in the

network has a consistent global view of time. Combining time

synchronization and localization, we are able to estimate the ve-

locity of targets.

2

Robust and Timely Communication over HighlyDynamic Sensor Networks

Tian He§, Brian M. Blum†, Qing Cao‡,John A. Stankovic†, Sang H. Son† and Tarek F. Abdelzaher‡

†Department of Computer Science, University of Virginia§Department of Computer Science and Engineering, University of Minnesota‡Department of Computer Science, University of Illinois, Urbana-Champaign

Abstract

Highly dynamic sensor networks, such as mobile robotic sensor networks, have been applied in various kinds of applicationscenarios such as real-time planet exploration and deep-ocean discovery. In these types of networks, mobility and energymanagement protocols change the connectivity among the neighboring nodes quickly. Traditional state-based protocols, designedfor static and/or low-mobility networks, suffer excessivedelay in updating their routing or neighborhood tables, leading to severepacket loss and communication delay in the highly dynamic situations. To provide robust and timely communication, we exploitthe concept ofLazy-Binding to deal with the elevated network dynamics. Based on this concept and the knowledge of the nodepositions, we introduce Implicit Geographic Forwarding (IGF), a new protocol for highly dynamic sensor networks that is altogetherstate-free. We compare our work against several typical routing protocols in static, mobile and energy-conserving networks undera wide range of system and workload configurations. In the presence of mobility and other dynamics, IGF achieves as much as10times improvement in the delivery ratio and significant reduction in both the end-to-end delay and control overhead. In additionto extensive simulations, we also implement and evaluate the IGF protocol on the Berkeley mote platform.

I. INTRODUCTION

Highly dynamic sensor networks, such as mobile robotic sensor networks, have been widely used to explore environmentsthat are difficult and dangerous for humans. In the exploration missions of the red planet, scientists employ the roboticsensordevices (Rover) to discover the possibility of water activity. In the deep-ocean exploration, robotic sensors are usedto access therisk of potential earthquakes, volcanoes and tsunamis. In addition, robotic sensors are used to investigate dangeroussites suchas radioactive environments and mine fields. These robotic sensors are normally equipped with various kinds of sensors suchas cameras, spectrometers and magnetometers. They can localize themselves in real-time through either GPS [1] or positiontracking [2]. Distributed data processing, such as the collaborative exploration for map construction [3], requires ateam ofrobotic sensors to communicate with each other constantly in real-time. This type of communication imposes several challenges.First, when robotic sensors move, the constant changes of the connectivity among the neighboring nodes make it difficulttomaintain freshness of the routing states in traditional state-based routing protocols. Second, energy management protocols[4]–[6] transit the robotic sensors into and out of sleep states, and their participation in the network becomes probabilisticat any given point in time. These unique challenges demand a new routing solution. In this paper, we exploit the concept oflazy-binding, which is widely used in other research areas, such as the programming language design and operating systems.Specifically here, we definelazy-binding as deferring mapping the system physics (e.g., the network topologies) into the volatilestates (e.g., the route state), required by a certain operation, to the last possible moment allowed by the operations. Sincelazy binding defers binding the volatile states as late as possible, it enables the system to cope with real-time changesin thenetwork topology. Our first installment based on this concept is a location-based routing protocol, called Implicit GeographicForwarding (IGF). IGF allows a sender to determine a packet’s next-hop online in real-time. By combining lazy-binding andlocation-address semantics, IGF becomes a pure state-freeprotocol, which does not depend on the knowledge of the networktopology or the presence/absence of other nodes. This characteristic of being state-free is valuable to the highly dynamicsensor networks, as it supports fault tolerance and makes protocols robust to real-time topology shifts or node state transitions.Further, a state-free solution eliminates the bandwidth-consuming packets required in the state-based solutions forrouting andneighbor table upkeep.

The remainder of the paper is organized as follows: Section II motivates the need for lazy-binding in highly dynamicnetworks. Section III describes the IGF protocol in detail.Section IV presents our simulation experiments and analysis inmobile and other environments. Section V describes our implementation on the MICA2 platform and its evaluation. Wediscuss the state-of-the-art and future work in Sections VIand VII, respectively. We conclude the paper in Section VIII.

II. THE MOTIVATION FOR LAZY-BINDING

Advances in the protocol design [7]–[11] continuously expand our ability to deal with the high dynamics inside the network.These protocols have been designed with robustness in mind,however, the level of fault tolerance is usually designed toadapt tooccasional node failures and infrequent topology migration. In order to cope with the elevated transition of network topologies,the state-based solutions are required to refresh the routing states in real time to reflect changes, which consequentlyintroducessignificant overhead and network congestion. Eventually, the performance of these algorithms might degrade dramatically, asthe real-time maintenance of the routing states might not keep up with the transition rate of the network topologies. We observethat the delay betweenthe time when a physical network topology maps to the routing states and the time when thesestates are actually used for packet forwarding is the root cause of state invalidation and routing failures. We term this delayas the binding delay. A long binding delay leads to a high probability that recorded states are invalid by the time they areused.This problem increases as the network dynamics increase. Inaddition, since routing states are volatile and become outdatedat a much faster rate in highly dynamic networks, it is inefficient to maintain state proactively and eagerly. According to thebinding time, we categorize the routing protocols into fourcategories as shown in Figure 1.

� � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � � �� � � � � � � � � �� � � ! " # � � � � " $ % � ! � # �& ' ( ) * + , - . / 0 1 2 3 4 5 1 6 , 3 7 8 , 9 : 4 ; / 3 7 / 3 <= > ? @ A B C D E @ F C > G H IJ K L M N O

Fig. 1. Difference in Binding Time

• Fixed routing schemes are rarely used due to their rigid early-binding at the deployment time. The binding delay of thistype of network could be infinite.

• Proactive schemes such as DSDV [10] and GPSR [11] maintain the network states aggressively. The routing states arerefreshed regardless whether there is need of data delivery. This eager and proactive binding property is suitable for thenetworks with a small rate of topology change. The binding delay is the interval between consecutive routing updates.

• On-demand algorithms such as DSR [8], AODV [9] and Directed Diffusion [7] bind the routing state to physical topologieswith a lazier approach - On-demand Route Discovery. The on-demand property allows them to defer the binding of therouting states to the physical network topologies until there is a need for end-to-end delivery. Those schemes have beenproven effective [12] in dealing with moderate mobility andfailures. The binding delay of on-demand algorithm is thetime since the route discovery.

• Different from the on-demand schemes, the IGF protocol proposed in this work goes one-step further. It defers the bindingof the routing states to the physical network topologyuntil the packet forwarding operation actually happens at a sendingnode. This design allows: 1) The elimination of the communicationoverhead to maintain the state proactively, reducing theunnecessary update of the volatile routing states, 2) The real-time detection of the node failure, migration, and transitioninto a sleep state and 3) The real-time utilization of recently awoken or newly arriving nodes.

III. IGF PROTOCOL DESIGN

In this section, we introduce the IGF protocol as an exemplarinstance of the lazy-binding concept applied to routing.

A. System Model and Assumptions

IGF is targeting to the high-end sensor networks (e.g., mobile sensor networks), where each sensor node can obtain itslocation (x, y) through GPS [1] or a position tracking technique [2]. The IGFcommunication supports the location-addresssemantic, in which locations are specified as the routing destinations, instead of using a particular node ID. This location-address semantic are valid in many sensor networks, becausesensor data, such as temperature readings, are normally taggedwith the location-context, and therefore can be addressed directly by the location, eliminating the overhead to translate thetarget destinations into a set of node IDs. Since the packet size in high-end sensor networks is relatively large, our maindesign uses RTS-CTS handshaking to avoid the hidden and exposed terminal problems in wireless communication [13], andan alternative solution for small-packet delivery (e.g., Tinyos Message) is discussed separately in Section III-H.4.For the sakeof simplicity, we describe IGF in Section III-B assuming a sufficient node density. The issues related to the density, radioirregularity and localization error are resolved in the later sections.

Towards Optimal Sleep Scheduling in Sensor Networks forRare-Event Detection

Qing Cao, Tarek Abdelzaher, Tian He, John StankovicDepartment of Computer Science, University of Virginia, Charlottesville, VA 22904

email:�qingcao, zaher, tianhe, stankovic � @cs.virginia.edu

Abstract— Lifetime maximization is one key element in the design ofsensor-network-based surveillance applications. We propose a protocol fornode sleep scheduling that guarantees a bounded-delay sensing coveragewhile maximizing network lifetime. Our sleep scheduling ensures thatcoverage rotates such that each point in the environment is sensed withinsome finite interval of time, called the detection delay. The framework isoptimized for rare event detection and allows favorable compromises tobe achieved between event detection delay and lifetime without sacrificing(eventual) coverage for each point. We compare different sleep schedulingpolicies in terms of average detection delay, and show that ours is closestto the detection delay lower bound for stationary event surveillance. Wealso explain the inherent relationship between detection delay, whichapplies to persistent events, and detection probability, which appliesto temporary events. Finally, a connectivity maintenance protocol isproposed to minimize the delay of multi-hop delivery to a base-station.The resulting sleep schedule achieves the lowest overall target surveillancedelay given constraints on energy consumption.

I. INTRODUCTION

Sensor networks promise surveillance of large areas with possiblyunprecedented accuracy. Currently, energy supply is one fundamentalbottleneck. It is very expensive to replace sensor node batteries oncethey are deployed, both because of the large number of sensing nodesand because of the typically hazardous or unfriendly environment inwhich these nodes are deployed. Hence, prolonging battery life is aprime consideration in network design. Current literature advocatesemploying redundancy to allow some nodes to go to sleep with-out jeopardizing sensory coverage. These approaches imply that aminimum number of nodes must remain awake for the right degreeof coverage to remain satisfied. A trade-off exists between energysavings and coverage. For example, in [1], [2] partial coverageschemes are investigated to increase energy saving gains. In theseefforts, both random and synchronized sleep schedules are proposedand studied for certain scenarios. The former refers to the case whereeach node independently chooses random sleep and wakeup times.The latter refers to the case where all nodes go to sleep and wakeup together in a synchronized fashion.

In contrast, we develop a near-optimal deterministically rotatingsensory coverage. In this scheme, the area is only partially coveredat any point in time. However, any point is eventually sensed withina finite delay bound. The energy/coverage trade-off is thus moremeaningfully expressed as one between energy savings and theaverage detection delay, defined as the average time elapsed betweenevent occurrence at a point and its detection by a nearby sensor. It isdesired to minimize average detection delay subject to a constrainton energy consumption (expressed as a duty-cycle constraint). Thegoal of this paper is to develop a localized distributed protocol for(near-optimally) solving the aforementioned constrained optimizationproblem while ensuring upper bounds on the worst-case detectiondelay.

The paper also addresses sleep scheduling schemes for minimizingpacket delivery latency to a common base-station. Observe that atvery low duty cycles, it is likely that sensor nodes that are awake at

any given time do not form a connected graph unless their wakeuptimes are appropriately synchronized. Such synchronization, however,may deviate from the optimal sleep schedule from the perspectiveof minimizing average detection delay. We develop a heuristic thatprovides partial synchronization to reduce delivery latency withoutsignificantly impacting the average detection delay.

The combination of detection delay and packet delivery latencyis the perceived surveillance delay, which refers to the time elapsedfrom the occurrence of an event in the system to the time the eventis reported to a base-station. Hence, the overall contribution of thispaper is to develop a protocol for minimizing the surveillance delaysubject to energy (namely, duty cycle) constraints.

Our protocol is optimized for detection of rare (but urgent) events.In such applications, network longevity is especially important, sincemission lifetime must be appropriately large. Nodes operate at verylow duty cycles and do not communicate unless an event is detected.Therefore, we consider sensing power as the predominant energydrain over the system lifetime. Once detection occurs, a promptreaction may be needed (e.g., activating a camera or reporting anemergency). Consider, for example, the detection of forest fires. Thereare two natural concerns with this application: first, how long willthe network last once deployed? Second, how responsive will it be inreacting to fire events? Our design translates these two questions intotwo related design parameters; namely, the energy consumption rate(i.e., the duty cycle which determines lifetime) and the surveillancedelay. Our protocol offers a design space in which the designer cantrade-off these parameters in a near-optimal fashion.

The rest of the paper is organized as follows. Section II presentsthe general framework and assumptions underlying our approach. Alocalized distributed optimization algorithm is presented in Section IIIto produce a sleep schedule that approaches the optimal on detectiondelay. This algorithm is subsequently enhanced to reduce deliverylatency as well. Simulation results are presented in Section IV.We survey related work in Section V, and conclude this paper inSection VI with our summary and directions for future work.

II. GENERAL FRAMEWORK

We consider an area covered by sensing nodes. Let some event(e.g., a fire) occur at one point in the area. The maximal detectiondelay for an event occurring at this point is defined as the longesttime that may elapse before the event is detected by a nearby node.The average detection delay for this point is defined as the averagetime elapsed until the event is detected. The maximal detection delayfor the entire area is the largest value of all maximal detection delaysat points in the area. Similarly, the average detection delay for thearea, denoted � , is the average value for all detection delays of allpoints. Trivially, when the area is sensing covered, both the maximaldetection delay and the average detection delay for the area are � ,since all events are detected immediately.

Sensors in the area are duty-cycled. Most sensors have a finite“warm-up” time ��� upon startup before reliable readings can bereported. Following the warm-up time, a sensor takes a sample of theenvironment, which itself takes time ��� (possibly including repeatedsensor readings). This may be followed by other necessary processing(such as data logging) which takes time � . Hence, from the instant anode is powered on, a minimum time interval, ��� ����� ��� ��� � �� ,must elapse before the node can go to sleep again. Given a dutycycle constraint � which defines the maximum percentage of timea node can be awake, the node must sleep for at least a duration��� , where � � ����� � � � � ��������� . Hence, any event is detected in atmost ��� � ����� time units. It is desired to minimize the average eventdetection time. We are especially interested in very low duty-cycleoperation where � � �! " ��� .

We propose a two-level sleep scheduling framework. The first levelselects a minimal subset of all deployed nodes, called the primarysubset, such that sensing coverage is maintained using the fewestprimary nodes. We assume that there are enough nodes in the networkfor sensory coverage to be achieved. The remaining nodes are turnedoff. This process is repeated periodically at a fairly large period (e.g.,of the order of tens of hours) to change the set of primary nodesso that their energy is not depleted. Algorithms for such rotationhave been proposed in prior literature and are not considered in thispaper. The second level focuses on the current primary nodes. Itcontributes further energy savings by duty-cycling these nodes at ahigher frequency (e.g., seconds or minutes). That is to say, each nodein the primary subset sleeps for ��� then wakes up for ��� � , where� � ����� � � � � � � �#�$� , the desired duty cycle. Our purpose, in thispaper, is to coordinate the duty cycles of primary nodes such that theaverage detection delay in the area is minimized.

One interesting remark is that although the maximum energysavings by first level scheduling are bounded by the need to maintainsensory coverage, the second level savings can be made arbitrarilylarge by decreasing the duty cycle of primary nodes. In principle,there is no lower bound on energy consumption after the secondlevel scheduling. The only consideration is that lowering the dutycycle increases average detection delay.

If the average number of primary nodes within a sensory radius is% , any point in the environment is sensed by % nodes on average.Since each node sleeps for ��� and wakes up for � � � , at low dutycycles (i.e., when ����� " ��� ), a point is sensed on average nomore than once every � �&� % time units. An event arriving randomlybetween sense instants will thus suffer an average detection delay nolower than ��� ��' % . This value establishes a lower bound on detectiondelay given the sensor wakeup period, � � � , and the chosen duty cycle,�(�)� � ����� � � � � ����� , which uniquely determine the minimum ��� ,and hence the minimum ��� ��' % .

On the other extreme, if all primary nodes sleep and wake up inunison, each point is sensed only once every ��� , and the averagedetection delay for a randomly arriving event is ��� �&' . Our purposeis to design a sleep scheduling protocol that approaches the lowerbound, ��� �&' % , on the average detection delay.

It can be shown that minimizing detection delay leads to mini-mizing the variance in detection delay as well. Intuitively, this isbecause the sum of the squares (or higher powers) of numbers thatadd up to a constant is minimized when these numbers are equal.Hence, equally spacing sensor wakeup times within an interval ���leads to minimizing both the mean and variance of detection delay.The complete proof is omitted for space limitations.

Finally, observe a relationship between detection delay and de-tection probability. An event with a short lifespan can be detected

as long as its lifespan intersects any of the waking periods ofneighboring sensor nodes. It is easy to show that the probability ofsuch intersection is maximized when the wakeup periods are equallyspaced. Thus, the sleep scheduling that optimizes the detection delayalso maximizes the detection probability of short-lived events. Nextwe present a protocol that produces a near-optimal sleep schedule.

III. SLEEP SCHEDULE OPTIMIZATION

In this section, we describe a sleep scheduling protocol thatoutperforms both random and synchronized scheduling in terms ofaverage detection delay. The protocol is distributed, and has thefavorable feature that it guarantees local optimality in that every nodeends up with a wakeup point that cannot be further improved interms of the average detection delay within its sensing range. Wealso present a protocol for optimizing end-to-end delivery latency.The combination of these two protocols is explored to reduce overallsurveillance delay.

A. Detection Delay Optimization

Our overall algorithm for minimizing detection delay is a threestage transition process, shown in Figure 1.

1

Stage II: Each node re-calculates its wakeup time, exactly once in each

iteration Tc, based on the most recently updated neighbor schedules. If one node does not receive any updates within an iteration and has not changed its own

wakeup time, proceed to stage III.

Stage III: The wakeup time is

finalized.3

2

Stage I: Each node chooses a wakeup time regardless of

other nodes.

Receive neighbor beacon, update neighbor wakeup

time

Fig. 1. State Transition of Optimization Algorithm

We assume that neighboring nodes have approximately synchro-nized clocks. Protocols for clock synchronization in sensor networkscan be found in [3]. Each node * starts at Stage 1, where it randomlypicks an initial wakeup time, +-,/. �10 for itself on a common timeline inthe cyclic interval [ �32-��� � ����� ). For the purposes of this analysis, thewakeup time denotes the instant at which the node’s wakeup interval� � � starts. The initial selection of the wakeup times of different nodesis completely uncoordinated. Each node communicates its randomlychosen wakeup time to its neighbors, sets up an iteration timer tofire at a period ��4 , and enters Stage 2. Observe that in this stage allprimary nodes are still awake (i.e., have not yet started their duty-cycling). The period � 4 is called the schedule iteration period, whichis different from the period � � � � ��� of the would-be duty cycles.

In Stage 2, each node undergoes multiple schedule iterations.Within a single iteration, a node makes at most one adjustment to itswakeup time to reduce the average detection delay. Ultimately, a local

Design and Comparison of Lightweight Group

Management Strategies in EnviroSuite⋆

Liqian Luo, Tarek Abdelzaher, Tian He, and John A. Stankovic

Department of Computer Science, University of Virginia,Charlottesville, VA 22904

{ll4p, zaher, th7c, stankovic}@cs.virginia.edu

Abstract. Tracking is one of the major applications of wireless sensornetworks. EnviroSuite, as a programming paradigm, provides a compre-hensive solution for programming tracking applications, wherein movingenvironmental targets are uniquely and identically mapped to logicalobjects to raise the level of programming abstraction. Such mapping isdone through distributed group management algorithms, which organizenodes in the vicinity of targets into groups, and maintain the uniquenessand identity of target representation such that each target is given aconsistent name. Challenged by tracking fast-moving targets, this paperexplores, in a systematic way, various group management optimizationsincluding semi-dynamic leader election, piggy-backed heartbeats, and im-plicit leader election. The resulting tracking protocol, Lightweight Envi-

roSuite, is integrated into a surveillance system. Empirical performanceevaluation on a network of 200 XSM motes shows that, due to these op-timizations, Lightweight EnviroSuite is able to track targets more than 3times faster than the fastest targets trackable by the original EnviroSuiteeven when 20% of nodes fail.

1 Introduction

The increasing popularity of sensor networks in large-scale applications such asenvironmental monitoring and military surveillance motivates new high-levelabstractions for programming-in-the-large. As a result, several programmingmodels that encode the overall network behavior (rather than per-node behav-iors) have been proposed in recent years. Examples include virtual machines[1][2], and database-centric [3][4][5], space-centric [6][7], group-based [8][9], andenvironment-based [10] programming models, which offer virtual machine in-struction sets, queries, sensor node groups and environmental events, respec-tively, as the underlying abstractions with which the programmer operates.These abstractions capture the unique properties of distributed wireless sensornetworks and expedite software development.

⋆ The work reported in this paper is funded in part by NSF under grants CCR-0208769and ITR EIA-0205327.

V. Prasanna et al. (Eds.): DCOSS 2005, LNCS 3560, pp. 155–172, 2005.c© Springer-Verlag Berlin Heidelberg 2005

156 L. Luo et al.

An important category of sensor network applications involves tracking en-vironmental targets. In these applications, some internal representation of theexternal tracked entity is maintained such as a state record, a logical agent, or alogical object representing the physical target. A given target in the environmentshould be represented uniquely and its identity should be preserved consistentlyover time. One way to ensure unique, consistent representation is to relay allsensor readings to a centralized base-station which runs spatial and temporalcorrelation algorithms to infer the presence of targets, assign them unique iden-tities, and maintain such identities consistently. Such a centralized approach,however, is both inefficient and vulnerable. In addition to relying on a singlepoint of failure, it results in excessive power consumption due to communicationwith a centralized bottleneck and may unduly increase latency, especially whentargets move far away from the base-station.

To avoid these limitations, in EnviroSuite, we take the alternative approachof processing target data at or near the location where the target is sensed.Hence, appropriate distributed group management policies are needed to ensurethe uniqueness and identity of target representation such that targets are givenconsistent names and sensors agree on which target they are sensing. This papersystematically investigates different system optimizations in the design of suchgroup management algorithms in a sensor network. The resulting tracking sys-tem, Lightweight EnviroSuite, is used in a sensor network surveillance prototypethat has since been transferred to the Defense Intelligence Agency (DIA). It isevaluated on a sensor network of 200 XSM motes [11]. Results from field testsof the overall system are provided, focusing on tracking performance. (Otherperformance aspects of the system such as efficacy of energy management algo-rithms will be reported elsewhere.) It is seen that realistic targets can indeed betracked correctly despite environmental noise using low-range sensors. Our fieldtest results show that even when 20% of nodes fail, Lightweight EnviroSuite isable to track targets more than 3 times faster than the fastest targets trackableby the original EnviroSuite, due to the optimizations described in this paper.The improved tracking coincides with reduced communication cost.

The remainder of the paper is organized as follows. Section 2 presents thebackground information and enumerates limitations of the original EnviroSuite.Section 3 explores group management strategies and their effects that consti-tute Lightweight EnviroSuite. Section 4 analyzes the performance results of asurveillance system that is constructed from Lightweight EnviroSuite. Finally,Section 5 supplies a summary and concludes the paper.

2 Background

Target tracking has received special attention in recent ad hoc and sensor net-works literature. Many prior approaches (e.g., in the ubiquitous computing andcommunication domains) focused on tracking cooperative targets. Cooperativetargets are those that allow themselves to be tracked typically by exporting aunique identifier to the infrastructure (such as a cell-phone number). Examples

Sensor Localization with Ring Overlapping Based onComparison of Received Signal Strength Indicator

Chong LiuComputer Science Dept.

University of VictoriaBC, Canada V8W 3P6chongliu@cs.uvic.ca

Kui WuComputer Science Dept.

University of VictoriaBC, Canada V8W 3P6

wkui@cs.uvic.ca

Tian HeComputer Science Dept.

University of VirginiaCharlottesville, Virginia 22904-4740

tianhe@cs.virginia.edu

Abstract—Sensor localization has become an essential requirement for re-

alistic applications over Wireless Sensor Networks (WSNs). Ra-dio propagation irregularity and the stringent constraint on hard-ware cost, however, make localization in WSNs very challenging.Range-free localizations are more appealing for range-based ones,since it does not depend on received signal strength to estimatedistance and thus needs simple and cheap hardware only. In thispaper, we propose a ring-overlapping, range-free approach us-ing Ring Overlapping based on Comparison of Received SignalStrength Indicator (ROCRSSI). Simulation results have verifiedthe high estimation accuracy achieved with ROCRSSI.

Key Words—Range-Free Localization, Wireless Sensor Networks

I. INTRODUCTION AND BACKGROUND

Wireless Sensor Networks(WSNs) become the current hotspot of networking area and have been used for various appli-cations, such as habitant monitoring, environment monitoring,and target tracking. Location information plays a crucial rolein understanding the application context in WSNs [5] [6], andmany localization algorithms for WSNs have been proposed toprovide per-node location information. They can be dividedinto two categories: ranged-based methods [1] [3] and ranged-free methods [2] [7] [4]. Range-based localization depends onthe assumption that the absolute distance between a sender anda receiver can be estimated by received signal strength or by thetime-of-flight of communication signal from the sender to thereceiver. The accuracy of such estimation, however, is subjectto the transmission medium and surrounding environment andusually relies on complex hardware [8]. In contrast, range-freelocalization never tries to estimate the absolute point-to-pointdistance based on received signal strength. As such, the designof hardware can be greatly simplified, making rang-free local-ization very appealing for WSNs.

In this paper, we propose another range-free localizationapproach, Ring Overlapping based on Comparison of Re-ceived Signal Strength Indicator (ROCRSSI), that uses ring-overlapping to estimate nodes’ location. This approach hasthe following prominent features. First, it does not requiresensor nodes to send out control messages, and the commu-nication cost is light and on anchor nodes only. Second, ring-overlapping, compared to triangle-overlapping in APIT [4] thatis demonstrated to perform best for randomly deployed WSNsamong existing range-free localization approaches, generates

small intersection area and results in more accurate locationestimation. Finally, the proposed ring-overlapping method isrobust under irregular radio propagation patterns.

II. RING OVERLAPPING BASED ON COMPARISON OFRECEIVED SIGNAL STRENGTH INDICATOR (ROCRSSI)

A. Introduction of ROCRSSI

The motivation of ROCRSSI is to get accurate estimation andreduce communication overhead with small number of anchors.Anchors, which is generally required to deploy with sensornodes by most of range-free localization methods [2] [7] [4], arethose nodes who know their locations and usually have largertransmission power than normal sensor nodes. The general ideaof ROCRSSI is that each sensor node uses a series of overlap-ping rings to narrow down the possible area in which it resides.As the example shown in Fig. 1(b), if S can determine that itsdistance to A is larger than the distance between A and B, butless than the distance between A and C, it can conclude that itfalls within the ring center at A with the inner radius equal to thedistance between A and B and the outer radius equal to the dis-tance between A and C. Similarly, S can figure out another ringcentered at anchor B, and a circle centered at anchor C. Thenit calculates the intersection area of these rings (or circles) andtakes the gravity of this area as its estimated location.

The rings can be generated by comparison of the signalstrength a sensor node receives from a specific anchor and thesignal strength other anchors receive from the same anchor. Forexample, in Fig. 1(a), assume that

�, � , and � are three anchor

nodes and � is a sensor node. Assume that anchor�

sends outbeacon messages and the signal strength received by anchor � ,anchor � , and sensor � is � � � � � , � � � � � , and � � � � �respectively. If � � � � � � � � � � � � � � � � � , then � islikely to fall within the shadowed ring area if anchor B, anchorC, and node S are roughly in the same direction from anchor A.Since ROCRSSI does not try to map the received signal strengthto absolute point-to-point distance, it belongs to range-free lo-calization approaches. Notably, ROCRSSI only compares therelative strength of RSSI and does not depend on absolute RSSIvalues.

The correctness of ROCRSSI is based on the assumption thatin a certain range of direction, with the increase of distance be-tween a sender and a receiver, the signal strength at the receiverdecreases monotonically. This assumption is usually true for

5160-7803-8815-1/04/$20.00 ©2004 IEEE

2

Fig. 1. Examples of ROCRSSI

realistic sensor networks [4]. If omni-directional antennas areused in a homogeneous environment, we can assume isotropicradio attenuation model and ROCRSSI can get the best esti-mation accuracy. Note that although ROCRSSI depends on thehomogeneity of radio transmission in a large range of direction,the estimated location errors by ROCRSSI are generally smalleven under circumstances with very irregular radio propagation.Briefly, this is because we introduce extra constraints on usableRSSI information. The detailed reasons will be illustrated inSection II-D.

ROCRSSI works in a purely distributed fashion. In RO-CRSSI, the received signal strength from a certain anchor, sayanchor

�, can be measured by neighbouring nodes (anchors

and sensors) that fall within the radio transmission range of an-chor

�. All neighbouring anchors will broadcast the measured

signal strength from anchor�

. In this way, each sensor nodewill be able to collect enough information to generate a seriesof overlapping rings that it believes itself falls within. Note thatonly anchor nodes are required to send out control messages.

The ROCRSSI algorithm can be broken into two phases:RSSI propagation and location estimation, which are describedin detail in the sequel.

B. RSSI PropagationAt initial period, each anchor broadcasts a specific number

of beacon messages. During this period, each of its neighbour-ing anchors and sensors will constantly sample received signalstrength. At the end of this period (for example, the numberof broadcast beacon messages in each anchor have reached apredefined number), all neighbouring anchors and sensors willcalculate the mean of the measured signal strength. For a sen-sor node, it will store the mean value for later use. For an an-chor node, it will broadcast a RSSI message, including the meanvalue of the measured signal strength, its own ID, the ID of thereference anchor (the one who has been measured), and its ownlocation information. Any sensor node that receives the RSSImessage will record the related information.

C. Location EstimationWhen sensor node S obtains enough information after the ini-

tial RSSI propagation stage, it can make use of the informationto calculate its own location. The pseudo-code of the algorithmis shown in Fig. 2. The basic idea is to generate a series ofrings, each of which S believes itself falls within. S calculatesthe intersection area of these rings, and takes the gravity of thefinal intersection area as its estimated location.

// Assume that sensor node S wants to calculate its location// Input:// Sn denotes the set of neighbouring anchors of sensor S.// SA denotes the set of neighbouring anchors of anchor A,// where A is one of neighbouring anchors of sensor S.// RRSIAB denotes the signal strength recorded from node A to node B.Process LocationEstimation( ){

RingSet R = {};While(Sn has more elements){

step 1: Anchor A= Sn.nextElements();step 2: Split SA into two parts: SA1 and SA2, such that

each element I in SA1 has larger RRSIAI than RRSIAS andeach element J in SA2 has smaller RRSIAJ than RRSIAS.

step 3: if ((SA == null) or (SA2==null)) goto step 1.J = the element with the largest value in SA2;d2 = distance between J and A;if (SA1 ==null) { d1 = d2;}

else {I = the element with the smallest value in SA1;

d1 = distance between I and A;}

Generate a ring r centered at A with inner radius d1 and outer radius d2;R = R + {r}; // insert r into R.

}step 4: Calculate the intersection area of all rings in R;step 5: return {the gravity of this intersection area};

}

Fig. 2. Algorithm for Location Estimation

Fig. 3. Grid scan algorithm alleviates the influence of wrong rings

Note that in the algorithm, if the set � � or the set � � � isempty, then no ring will be generated. If the set � � is emptybut the set � � � is not, only one circle is generated. If both sets

� � and � � � are not empty, � will generate a ring.In step 4, a grid-scan algorithm [4] is used to calculate the

gravity of the intersection area. In this algorithm, the whole ter-rain is divided into small pieces of grids. Each grid maintains acounter which is initialized to 0. Every time a ring is generatedby comparison of signal strength, the counter values associatedwith the grids within the ring are increased. Once all possiblerings are calculated, we scan the whole grid array to find thearea with the maximum counter value, then take this area as thefinal intersection area and calculate its gravity. It is easy to seethat the size of the grids determines the possible smallest granu-larity of location error. Small grids are thus preferred but smallgrids need more calculation time.

D. Handling Radio Irregularity

According to the measurement results over real sensor de-vices, radio propagation is usually not homogenous in all direc-tions [9], that is, different directions have different radio atten-uation rates. So a good localization algorithm should accom-modate radio irregularity by not assuming isotropic path losses.ROCRSSI does not exclude generating wrong rings and thusmakes incorrect estimation because of the irregularity of radio

517