next century challenges - scalable coordination in sensor networks

8/3/2019 Next Century Challenges - Scalable Coordination in Sensor Networks

1/8

Next Century Challenges: Scalable Coordination in Sensor NetworksDeborah EstrinRamesh GovindanJohn HeidemannSatish Kumar

USC/Information Sciences Institute4676 Admiralty WayMarina de1 Rey, CA 90292, USA{estrin,govinda.n,joh,kkumar}@isi.edu

AbstractNetworked sensors-those that coordinate amongst them-selves to achieve a larger sensing task-will revolutionizeinformation gathering and processing both in urban envi-ronments and in inhospitable terrain. The sheer numbers o fthese sensors and the expected dynamics in these environ-ments present unique challenges in the design of unattendedautonomous sensor networks. These challenges lead us tohypothesize that sensor network coordination applicationsmay need to be structured differently from traditional net-work applications. In particular, we believe that localizedalgorithms (in which simple local node behavior achieves adesired global objective) may be necessary for sensor net-work coordination. In this paper, we describe localized al-gorithms, and then discuss directed diffusion , a simple com-munication model for describing localized algorithms.1 IntroductionIntegrated low-power sensing devices will permit remote ob-ject monitoring and tracking in many different contexts: inthe field (vehicles, equipment, personnel), the office building(projectors, furniture, books, people), the hospital ward (sy-ringes, bandages, IVs) and the factory floor (motors , smallrobotic devices). Networking these sensors-mpoweringthem with the ability to coordinate amongst themselves on alarger sensing task-will revolutionize information gatheringand processing in many situations. Large scale, dynamicallychanging, and robust sensor colonies can be deployed in in-hospitable physical env ironments such as remote geographicregions or toxic urban loca tions. They will also enable lowmaintenance sensing in more benign, but less accessible, en-vironments: large industrial plants, aircraft interiors etc.To motivate the challenges in designing these sensor net-works, consider the following scenario. Several thousandsensors are rapidly deployed (eig., thrown from an aircraft)in remote terrain. The sensors coordinate to establish acommunica tion network, divide the task of mapping andmonitoring the terrain amongst themselves in an energy-

permission to make digital or hard copies of all or part ofthjs work fo,Personal or classroom use is granted w ithout fee provided that copiesare not made or distributed for profit or commercial advantage and thatcopies bear this notice and the full citation on the first page . l-o copyothclwisc, to republish, to post on servers or to redistribute to lists.requires prior specific permission andior a fee.Mobicom 99 Seattle Washington USACopyright ACM 1999 I-581 13-142-9/99/08...$5.00 263

efficient manner, adapt their overall sensing accuracy to theremaining total resources, and re-organize upon sensor fail-ure. When additional sensors are added or old sensors fail,the sensors re-organize themselves to take advantage of theadded system resources.Several aspects of this scenario present systems designchallenges different from those posed by existing computernetworks (Section 2). The sheer numbers of these de-vices, and their unattended deployment, will preclude re-liance on broadcast communica tion or the configuration cur-rently needed to deploy and operate ne tworked devices. De-vices may be battery constrained or subject to hostile en-vironments, so individual device failure will be a regular orcommon event. In addition, the configuration devices willfrequently change in terms of position, reachability, poweravailability, and even task details. Finally, because thesedevices interact with the physical environment, they, andthe network as a whole, will experience a significant rangeof task dynamics.The WINS project [l] has considered device-level com-munication primitives needed to satisfy these requirements.However, these requirements potentially affect many otheraspects of network design: routing and addressing mech-anisms, naming and binding services, application architec-tures, security mechanisms, and so forth. This paper focuseson the principles underlying the design of services and appli-cations in sensor networks. In particular, since the sensingis inherently distributed, we argue that sensor network ap-plications will themselves be distributed.Many of the lessons learned f rom Internet and mobilenetwork design will be applicable to designing sensor net-work applications. However, this paper hypothesizes thatsensor networks have different enough requirements to atleast warrant re-considering the overall structure of appli-cations and services. Specifically, we believe there are sig-nificant robustness and scalability advantages to designingapplications using localized algorithms-where sensors onlyinteract with other senso rs in a restricted vicinity, but nev-ertheless collectively achieve a desired global objective (Sec-tion 3). We also describe directed diffusion, a promisingmodel for describing localized algorithms (Section 4).Our research project is starting to investigate the designof localized algorithms using the directed diffusion model.These ideas were developed in the context of a DARPAISAT s tudy, chaired by one of the authors (Estrin). The


2/8

idea of applying directed diffusion to this problem domainis due to Van Jacobson, based on experiences with reliablemulticast [2) and adaptive Web caching design.2 Sensor Network ChallengesBy early next century, sensor integration, coupled with un-ceasing electronic miniaturization, will make it possible toproduce extremely inexpensive sensing devices. These de-vices will be able to monitor a wide variety of ambientconditions: temperature, pressure, humidity, soil makeup,vehicular movement, noise levels, lighting conditions, thepresence or absence of certain kinds of objects, mechanicalstress levels on attached objects, and so on. These deviceswill also be equipped with significant (i.e., comparable totodays high-end portable computers) processing, memory,and wireless communica tion capabilities.Emerging low-level and low-power wireless communica -tion protocols will enable us to network these sensors. Thiscapability will add a new dimension to the capabilities ofsensors: Sensors will be able coordinate amongst themselveson a higher-level sensing task (e.g., reporting, with greateraccuracy than possible with a single sensor, the exact speed,direction, size, and other characteristics of an approachingvehicle).Networking inexpensive sensors can revolutionize infor-mation gathering in a variety of situations. Consider the fol-lowing scenarios, arranged in increasing order of complexity:

l Each item of inventory in a factory warehouse or officecomplex has, attached to it, a tag. Stick-on sensors,discreetly attached to walls, or embedded in floors andceilings, track the location history and use of items.The sensor network can automatically locate items,report on those needing servicing, analyze long-termcorrelations between workflow and wear, report unex-pected large-scale movements of items or significantchanges in inventory levels. Some systems today (forexample, those based on bar-codes) provide inventorytracking; full sensor-net based systems will eliminatemanual scanning and provide more data than simplylocation.l Thousands of disposable sensors are densely scatteredover a disaster area. Some of them fall into regions af-fected by the disaster, say a fire-these sensors are de-stroyed. The remaining sensors collectively map theseaffected regions, direct the nearest emergency responseteams to affected sites, or find safe evacuation paths.Disaster recovery today is by comparison very humanintensive.l Every vehicle in a large metropolis has one or more at-tached sensors. These sensors are capable of detectingtheir location; vehicle sizes, speeds and densities; roadconditions and so on. As vehicles pass each other, theyexchange information summaries. These summarieseventually diffuse across sections of the metropolis.Drivers can plan alternate routes, estimate trip times,and be warned of dangerous driving conditions. Wn-like the centralized systems sometimes seen today, one

based on local communication would scale as the num-ber of vehicles grows and provide much greater localdetail.These futuristic scenarios bring out the two key require-ments of sensor networks: support for very large numbers

of unattended autonomous nodes and adaptivity to environ-ment and task dynamics.Many large-scale networks exist today; the Internet is aprime example. Sensor networks present a fundamentallymore difficult problem, though, because the ratio of com-municating nodes to users is much greater. Each personalcomputer on the Internet has a user who can resolve or atleast report all manner of minor errors and problems. Thiihuman element allows the Internet to function with muchless robust software. Sensor networks, by comparison willexist with the ratio of thousands of nodes per user (or more).At such ratios, it is impossible to pay special attention toany individual node. Furthermore, even if it were possible toconsider each node, sensors may be inaccessible, either be-cause they are embedded in physical structures, or throwninto inhospitable terrain. Thus, for such a system to be ef-fective, it must provide ezception-free, unattended operation(the term exception-free is due to Mark Weiser).

It is not completely true that there are no large scaleunattended systems today. Automated factories, for ex-ample, may contain hundreds o f largely unsupervised com-puters. This example illustrates the second requirement ofsensor networks: they operate and must respond to very dy-namic environments. Automated factories are deployed withvery careful planning and react to very few external events.Sensor networks instead will be deployed in a very ad hocmanner (possibly thrown down at random). They will suffersubstantial changes as nodes fail due to battery exhaustionor accident, new nodes are added, nodes move or are carried.User and environmental demands also contribute to dynam-ics as what is being sensed moves and what is consideredinteresting changes. Thus sensor networks must automati-cally adapt to changes in environment and requirements.One hypothesis for the overall design of a sensor net-work is that it is sufficient to design sensor network ap-plications using Internet technologies coupled with ad-hocrouting mechanisms. In such a design, each sensor node isan Internet-capable device (has one or more IP addresses)and can run applications and services. W hen deployed, sen-sor nodes establish an ad-hoc network amongst themselves;thereafter, application instances running on each node cancommunicate with each other. Applications, aided by direc-tory and resource discovery services, are structured muchthe same way as traditional Internet applications.We believe, however, that sensor network requirementsare different enough from those of traditional wired andwireless networks to warrant considering a different design.This design has the following features:Data-Centric Unlike traditional networks, a sensor nodemay not need an identity (e.g., an address) . Thatis, sensor network applications are unlikely to ask thequestion: What is the temperature at sensor #27?Bather, applications focus on the data generated by

In som e situations, for example, for querying a specific faultysensor, the ability to address an individual sensor is clearly necessary.

264


3/8

sensors. Data is named by attributes and applicationsrequest data matching certain attribute values. So,the communication primitive in this system is a re-quest: Where are nodes whose temperatures recentlyexceeded 30 degrees? This approach decouples datafrom the sensor that produced it. This allows for morerobust application design: even if sensor #27 dies,the data it generates can be cached in other (possi-bly neighboring) sensors for later retrieval.Application-Specific Traditional networks are designedto accommodate a wide variety of applications. Webelieve it is reasonable to assume that sensor networkscan be tailored to the sensing task at hand. In partic-ular, this means that intermediate nodes can performapplication-specific data aggregation and caching, orinformed forwarding of requests for data. This is incontrast to routers that facilitate node-to-node packetswitching in traditional networks.

If we admit this architecture, how might we design appli-cations on top of a sensor network that provided this kind ofcommunication? Recall that sensor network applications ofinterest to us are those in which sensor nodes coordinate toperform a higher-level sensing task (e.g., Is it time to ordermore inventory? At what speed and in what direction wasthat elephant traveling?). Clearly, this kind of coordinationcan be structured in a centralized manner. Individual sen-sors report their data to a central node, which then performsthe computation required fo r the application. This central-ized structure is a bad choice for several reasons: it providesa single point of failure, it can be energy inefficient, and itdoesnt scale to large networks.We hypothesize that sensor network coordination appli-cations are better realized using localized algorithms. Weuse this term to mean a distributed computation in whichsensor nodes only communicate with sensors within someneighborhood, yet the overall computation achieves a de-sired global objective. What is the rationale for using local-ized algorithms in sensor networks? Since the sensors them-selves are physically distributed, it is not unnatural to designsensor networks using distributed algorithms. Furthermore,localized algorithms have two attractive properties. First,because each node communicates only with other nodes insome neighborhood, the communication overhead scales wellwith increase in network size. Second , for a similar reasonthese algorithms are robust to network partitions and nodefailures. We are just beginning the work of validating thishypothesis through design and experimentation.In the next section, we describe the challenges posed bythe design of localized algorithms in data-centric, application-specific sensor networks.3 Localized Algorithms for CoordinationClustering allows sensors to efficiently coordinate their localinteractions in order to achieve global goals. In particular,localized c lustering can contribute to more scalable behavioras number of nodes increase, improved robustness, and moreefficient resource utilization for many distributed sensor co-ordination tasks. One such sensor coordination task is theelection of extremal sensors to form the widest baseline fo r

locating external objects. Especially when this triangula-tion is performed frequently, it may be more energy efficientfor cluster heads alone, rather than all the sensors in the net-work, to participate in this election. Data aggregation is asecond example of the use of clustering. Consider an officeenvironment where sensors monitor the location of varioustagged objects such as projectors and books. Cluster-headscould summarize the objects located in their clusters to pro-vide a less detailed view to distant nodes. The disseminatedsummary information can then be used to locate objectssuch as the nearest projector or a missing book.We first present a localized clustering algorithm and laterdiscuss an application that makes use of the clustered sen-sors to efficiently pinpoint the location of objects. We as-sume that a link level procedure is run on each sensor thatadjusts the transmission power and thus the communica-tion range to a minimum value that ma intains full networkconnectivity. The clustering algorithm then elects cluster-head sensors such that each sensor in the multi-hop networkis associated with a cluster-head sensor as its parent. Theparent-child relationships are established only between sen-sors that are able to communicate with each other thus pre-venting inconsistencies due to asymmetric communication .The clusters adapt to network dynamics and changing en-ergy levels of nodes. For simplicity, we describe a two-levelcluster formation algorithm in this paper. The algorithmcan be recursively applied to build a cluster hierarchy.In our algorithm, we associate sensors at a particularlevel with a radius. The radius specifies the number of phys-ical hops that a sensors advertisements will travel. Sensorsat a higher level are associated with larger radii than thoseat lower levels. All sensors start off at the lowest level of 0.Each sensor then sends out periodic advertisements to sen-sors within radius hops. The sensor advertisements carrythe sensors hierarchical level, parent ID (if any) and re-maining energy of the sensor. Sensors then wait for a cer-tain wait time that is proportional to their radius in orderto allow advertisements from various sensors to reach eachother. At the end of the above wait period, a level 0 sen-sor starts a promotion timer if it does not have a parent.The promotion timer is set to be inversely proportional tothe sensors remaining energy and the number of other sen-sors from whom level 0 advertisements were received. Thiswould cause sensors located in relatively dense regions andwith higher remaining energy to have smaller timeout val-UeS. When the promotion timer expires, a sensor promotes it-self to level 1 and starts sending periodic advertisements atthe level 1 radius. In these advertisements, the newly pro-moted sensor lists its potential child sensors that consistsof the level 0 sensors whose advertisements it previously re-ceived. Only the level 0 sensors that appear in this potentialchildren list can choose the level 1 sensor to be their parent.This ensures that parent-child relationships are establishedonly between sensors that can see each others advertise-ments and thus are able to communicate with each other. Alevel 0 sensor picks the closest potential parent that it seesto be its parent. Once a level 0 sensor picks a parent, it can-cels its promotion timer if running and thus drops out of theelection process. After promotion, the level 1 sensors starta wait timer proportional to their new larger radius. At the

265


4/8

Figure 1: Object 7Gngulation in Sensor Networks: The high-level task of this sensor network is to accurately pinpoint the location o fthe object (represented by a dark square). Under certain assumptions, there exists a localized algorithm which can do this, which maybe non-optimal under certain circumstances.

end of the wait period, the level 1 sensor may demote itselfif it does not have any child sensors or if its energy level isless than a certain threshold function of its childrens en-ergy (e.g., less than 50% of the maximum energy among itschildren). All level 0 and level 1 sensors periodically enterthe wait state. Thus, any change in network conditions, orin sensor energy levels results in m-clustering with boundeddelay.Many clustering proposals can be found in the literature.However, they do not adequately address the primary con-straints of the wireless sensor networks environment. Someproposals [3, 41 exhibit non-localized behavior where nodesneed to communicate with othe r distant node(s) to electleaders. The Landmark hierarchy [5] a and other localizedclustering proposals [6, 71 suggest significant improvementsbut do not handle two key design constraints: asymmetriccommunica tion in the network and limited energy of sen-sors. Asymmetric communications may cause a sensor inthe network to have inconsistent information about its owncluster. For example, a chiid sensor might choose a partic-ular cluster-head to be its parent (thereby joining its clus-ter) even though the cluster-head cannot see advertisementsfrom the child sensor (due to the childs advertising radiusbeing smaller than the longer reverse path to the parent).Energy-insensitive design may lead to quick depletion of en-ergy levels of sensors thus reducing the life of the network.We next illustrate an application of the clustering algo-rithm in pinpointing object locations. Consider the scenarioshown in figure. This figure depicts a sensor network orga-nized into c lusters where no single level 1 cluster encom-passes all sensors. The only active sensors in the networkare the cluster-head sensors (the shaded ones) that detectan object (the dark square). Each sensor can determine thegeneral direction of the object. The task of this sensor net-work is to pinpoint, in an energy-efficient manner, the exactlocation of the object. To accurately determine the loca-tion of the object, we need the widest possible measurementbaseline. To achieve energy efficiency, we need the fewestnumber of sensors participating in this triangulation. Thatis, for the network shown in Figure l(a), we would like todesign a localized algorithm that results in cluster-heads Aand B participating in the triangulation.Assume that each sensor can determine its position ina-space, and that each sensor can specify the approximatedirection of the object relative to its own location. Then,there exists a simple rule whereby each cluster-head sensor

Recursive applicatio n of our algorithm leads to the constructio nof a Landmark hierarchy

can locally determine (based on information from neighbourcluster-heads alone) whether it should participate in the tri-angulation computation: If all the neighboring cluster-headsof a cluster-head sensor lie on the same side of a line d rawnbetween the sensor and the object, then that cluster-headsensor elects itself as a participant in the computation. Bythii rule, for example, the cluster-head X in Figure l(a)does not elect itself to be a participant. This rule will electthe cluster-head sensor at each extremity (or more than oneif those sensors are aligned with respect to the object) Onceelected, these extremal sensors report their readings to anexternal observer.To implement this rule, a single message exchange be-tween neighboring cluster-head nodes suffices. The abovecluster-based approach for base-line estimaton has severalnice properties. First, because these sensor algorithms useonly local information, they should have generally lower en-ergy consumption than those that entail global communi-cation. Intuitively, these algorithms have the potential todemonstrate scaling complexity such that the overhead ofthe algorithm run at each sensor node is a sublinear func-tion of the total number of sensor nodes, and is proportionalto the local population density. Second, the algorithm is ro-bust to link or node failures and network partitions. As weshow below, however, it could be slightly inefficient whenthese pathologies occur. Third, because all commun icationis inherently localized, mechanisms for self-configuration canbe simpler than for other networks. This enables rapid de-ployment and robust unattended operation. Finally, localcommunica tion and per-hop data filtering can avoid trans-mitting large amounts of data over long distances, therebypreserving node energy resources. Node energy resources arealso better utilized since the cluster-heads adapt to chang-ing energy levels. Essentially, the sensors in a cluster taketurns at being the cluster-head based on their cu rrent energylevels thus leading to more efficient energy usage.The previously described rules achieve our objective: elect-ing sensors that form the longest-baseline for triangulation.So, what is hard about designing such localized algorithms?This simplified algorithm can be non-optimal under certainterrain conditions. For example, if some cluster-head sensorsare behind obstacles and cannot see the object, while theirneighbor cluster-heads can, the rule can cause several sen-sors to elect themselves (Figure l(b)). One way to alleviatethe impact of such conditions might be to allow a cluster-head to switch on some number of child sensors in its clusterto do object location. A cluster-head can then communicateto its neighbour cluster-heads that it detects an object if any

266


5/8

of its selected child sensors detect the object. A cluster-headthat is elected to participate in the triangulation can reportback readings from the extremal child sensor in its clusterthat detects the object.The preceding paragraph shows by example the difficultyof designing localized algorithms. Localized algorithms arehard to design for two main reasons. First, local algorithmsmust provide a desired global behavior with at best indirectglobal knowledge. Thus the process of crafting local algo-rithms from the global behavior is akin to the process of con-verting a centralized algorithm to a completely distributedone. The resulting rules often bear little resemblance tothe original distributed computation. Second, some kindsof localized algorithms are parametrically sensitive; differ-ent choices of algorithm parameters can lead to radicallydifferent kinds of global behavior. An example of this is thereaction-diffusion systems studied by Turing [8]. It is dif-ficult to design localized a lgorithms that both empiricallyadapt to a wide range of environments and converge to thedesired global behavior over that entire range.We believe that the following two-pronged approach canbe used to overcome these difficulties.

l First, develop intuition for localized algorithms by de-signing and prototyping some algorithms. A class ofalgorithms pertinent to sensor networks that we planto explore are adaptive fidelity algorithms. An adap-tive fidelity algorithm is one where the quality (fi-delity) of the answer can be traded against batterylifetime, network bandwidth, or number of active sen-sors. (Of course, the resulting fidelity must still fallwithin acceptable bounds.) One illustration of thisidea is shown in Figure 2. Consider, as before, asensor network that determines the exact location ofsensed objects. Now, instead o f every cluster-head sen-sor participating in the baseline determination, somecluster-head sensors turn themselves off to conservepower (the grayed sensors in Figure 2(a)). This has theeffect of a smaller baseline and, consequently, lower fi-delity triangulation. Subsequently, as some of the cur-rently active sensors die (due to battery failure o raccidents, shown with dotted lines), new cluster-headsare elected and their neighbor cluster-heads take overand continue the baseline determination for other ob-jects (Figure 2(b)). If designed correctly, this can re-sult in only slightly degraded baseline determinationbut a sensor network with nearly double the lifetime.

l Second, develop techniques for characterizing the per-formance of localized algorithms. Localized algorithmsexhibit good robustness and scaling properties. Toachieve these properties, however, these algorithms maysacrifice resource utilization or sensing fidelity, respon-siveness, or immunity to cascading failures. It is desir-able to develop a methodology that can characterizethese tradeoffs.

4 Directed DiffusionLocalized algorithms have many desirable properties in thecontext of sensor networks. However, these algorithms arehard to design and characterize. It would be convenient to

define a set of abstractions that describe the communica-tion patterns underlying many localized algorithms. In thissection, we briefly explore one such set, directed diffus ion.A sensor network based on directed diffusion exhibits thefollowing properties. Each sensor node names data that itgenerates with one or more attributes. Other nodes mayexpress interests, based on these attributes. Network nodespropagate interests. Interests establish gradients that directthe diffusion of data. As it propagates, data may be locallytransformed at each node. We explain these concepts withthe aid of the simple scenario shown in Figure 3.Figure 3 shows a sensor network in which each node candetect motion (and possibly other information) within somevicinity. One or more sink nodes may query the sensor net-work for motion information from a particular section of theterrain (e.g., from the southeast quadrant). One goal of thesensor network is to robustly compute a data disseminationpath from source to sink. The following four paragraphs de-scribe this path finding algorithm using the diffusion model.Attribute-based naming is the first characteristic of di-rected diffusion systems. In our example, each sensor namesdata that it generates using a single attribute motion, whichhas a geographic location (e.g., latitude/ longitude, or rel-ative location with respect to some landmark) as its value.In general, m otion data may be described using several at-tributes: (type=seismic, id=12, timestamp=99.01.22/21:08:15,location=75N/120E, footprint= vehicle/wheeled/over-40-ton).As this example shows, an attributes value may also havea hierarchical structure.A sink (such as the node a in Figure 3(a)) may query formotion information by disseminating an interest. Syntacti-cally, an interest is simply a range of values for one or moreattributes. In our example, the node a specifies south-eastquadrant as the value of the motion attribute in its inter-est. More generally, interests may have complex structure(type=seismic, timestamp=99.01.22/*, location=70-80N/lOO-140E).Each node disseminates interests based on the contentsof the interest. In our example, intermediate nodes send theinterest towards the neighbor in the direction of the south-east quadrant. Conceptually, the path of interest propaga-tion sets up a reverse data path for data that matches theinterest. Then, when nodes x and y in the southeast quad-rant detect motion, the motion signature travels towards aalong data propagation path.In the diffusion model, we say that this data propagationpath has an associated gradient. The notion o f gradient isuseful when, for robustness, each intermediate node propa-gates the interest towards multiple neighbors (Figure 3(b)).We say that the strength of the interest is different to-wards different neighbors, resulting in source-to-sink pathswith different gradients. In its simplest form, a gradient isa scalar quantity. Negative gradients inhibit the distribu-tion of data along a particular path, and positive gradientsencourage the transmission of data along the path. Thevalue of a particular gradient may have application-specificsemantics. In our motion sensing scenario, for instance, ifa node has two outgoing paths, one with a gradient of 0.8and another with a gradient of 0.4, then the node may sendtwice as much detail along the higher gradient path thanalong the lower.

267


6/8

Figure 2: Adaptive F idelity Algorithms: These localized algorithms selectively turn off some sensors to conserve system resources. Asindividual sensors die, others take their place. This technique extends network lifetime while possibly reducing sensing fidelity.

Although not described in our scenario, the diffusionmodel allows intermediate nodes to cache or locally tt-ans-form (e.g., aggregate) data. This aspect of the model lever-ages the application-specificity that is possible in sensor net-works. Caching and aggregation can increase the efficiency,robustness and scalability of coordination. Locally cacheddata may be accessed by other users with lower energy con-sumption than if the data were to be resent end to end.Intermediate node storage increases availability of the data,thereby improving robustness. F inally, intermediate nodescan increase the scalability of coordination by using cachedinformation to carefully direct interest propagation.The diffusion models data naming and local data trans-formation features capture the datai-centricity and application-specificity inherent in sensor networks (Section 2). Themodel allows for neighbor-to-neighbor interest propagationand local data transformation rules; these elements cap-ture the communication patterns expected of localized al-gorithms. Finally, gradients model the network-wide resultsof these local interactions. We believe that the diffusionmodel can be used to describe not only other data dissemina-tion patterns (shortest path multicast trees, energy-efficientspanning tree multicast), but also other coordination algo-rithms (such as the triangulation example of Section 3).To the extent that diffusion primitives help set up com-munication paths between nodes in sensor networks, theyplay the role of the routing system in traditional data net-works. Because we expect most sensor applications to belocalized, we think sensor networks are unlikely to incor-porate a reactive routing system like that found in todaysInternet. Instead, we expect the routing function in a sen-sor network to be tightly integrated with the application.Applications will use a combination of proactive and reac-tive schemes [9, 10, 11, 12, 13,141 to achieve energy-efficientcommunication.5 Related WorkSeveral projects have already demonstrated the feasibilityof low power integrated sensors [l, 15, 16, 171 and MEMS-based microsensors [ 181.Different contexts have been proposed for the applicationof small-scale networks of devices for sensing and actuationtasks.Various p rocess control and automation tasks in facto-ries have traditionally used networks of embedded systemsknown as control network [19] to efficiently perform theirmonitoring tasks. These control networks today consist of

a centralized processor coordinating the communica tion be-tween many tens of sensors and actuators.Simple networks of integrated sensors have been pro-posed for military awareness situations [l]. These networksschedule sensor transmission using TDMA methods, andprovide for automatic discovery of neighbors and cooper-ative detection by sensors.The Piconet [20] project is developing a prototype em-bedded network. Piconet is a low-rate (about 40 Kb/s), low-range (5 meters) ad-hoc radio network. Sensors can use thePiconet module to enable wireless connectivity. The projecthas designed a low level radio protocol for communicationbetween various embedded objects. They have also prototyped some interesting home and office information discov-ery applications [21].Embedded networks have long been used for personallocation, equipment tracking, or information gathering. Ac-tive badges [22], the ParcTab [23], the ORL location sys-tem [24], the Pinpoint Positioning System (LPS) [25 ], andthe Factoid [26] are all examples of this class of networkedsensors.The Ubiquitous Computing project at Xerox PARC [27]explored a generalized version of these applications: seam-less integration of computing devices into the environment.Finally, the MIT smart room project [28] and the Forest ofSensors project [29] analyze data from video images to makeinferences about the presence or absence of various objectsor people.Several research efforts provide insights into the designof coordination algorithms for sensor networks.Biological Systems: The reaction-diffusion models formorphogenesis [8] describe the mechanism by which the ini-tially homogeneous human cells eventually differentiate them-selves to form various tissues and organs. Models of antcolonies (301 are based on the fact that the almost blindants seem to be able to find shortest paths to destinationsusing the pheromone trails deposited by other ants as theonly information. These contain some elements of the diffu-sion model: localized interactions that lead to the formationof network gradients.

Distributed robotics: Of relevance here are coordinationprotocols being designed for troops of low-cost robots to ex-plore and acquire maps of unknown environments (311. Inthese schemes, the robots move about in a partially ran-dom fashion and cooperate with each other by transferringthe collected information when they meet. After the robotscomplete their exploratory run, they deliver their partialmaps to a host computer that then derives a complete mapof the area.

268


7/8

(4 (b) (cl (4Figure 3: Diffusion for path finding: This illustrates the basic diffusion cons tructs: data naming, interests that set up gradients, datadiffusing along gradients.

Amorphous Computing: Coore et al. [6] present dis-tributed algorithms to organize unstructured processing el-ements into a hierarchy of cooperative groups, called an AChierarchy. However adaptation to network failures has notbeen discussed.Internet Multicast and Web Caching: The design of In-ternet multicast tools [32] has led to significantly improvedunderstanding of robust application design. Specifically,techniques such as lightweight sessions and soft state man-agement are also applicable in the sensor network context.In adaptive Web caching schemes [33], cache servers self-organize themselves into overlapping multicast groups. Themesh of overlapping groups form a scalable, implicit hierar-chy that is used to diffuse popular web content towards thedemand. These techniques are relevant because frequentlyqueried info rmation in sensor networks also needs to be ef-ficiently diffused to the interested users.Perhaps mos t directly relevant to sensor networks is on-going work on ad-hoc networks. Ad-hoc networks refer toself-organizing networks of mobile wireless nodes that donot depend on any fixed infrastructure. A central focusof the work on ad-hoc ne tworks has been the design ofproactive [Q, 10, 111 and reactive [12] routing protocols, andcombinations thereof [13, 141. Proactive routing protocolscontinuously compute routes to all nodes so that a routeis already available when a packet needs to be sent to aparticular node. Such continuous route computation mayenergy-inefficient. Reactive routing protocols on the otherhand start a route computation process only when a packetneeds to be sent to some other node. However, they mayredundantly flood requests throughout the network. A com-bination proactive and reactive scheme may overcome thesedisadvantages, but may still not perform as effectively asschemes that use application knowledge to route data andqueries.AcknowledgementsMany of the ideas presented in this paper were the outcomeof a DAHPA ISAT study, chaired by one of the authors (Es-trin). Two of the other authors (Govindan, Heidemann)participated in the study. The other study participantswere: Hod Brooks, Mani Chandy, Dave Clark, Steve Deer-ing, Rich Ivanetich, Van Jacobson, Butler Lampson, BillMark, Ve rn Paxson, Greg Pottie, and Mark Weiser. Otherswho contributed to the ideas described in this paper include:

Cengiz Alaettinoglu, Nirupama Bulusu, Padma Haldar, andChalermek Intanagonwiwat. The authors are also gratefulto the anonymous reviewers for suggestions which led to theimproved readability of the paper.ReferencesPIPI

131

[41

[51

PI

171

PI

PIWI

WIWI

P31

G. Pottie, W. Kaiser, L. Glare and H. Marcy. Wireless Inte-grated Network Sensors. Submitted for publication, 1998.S. Floyd, V. Jacobson, C. Liu, S. Mccanne and L. Zhang.A Reliable Multicast Framework for Light-weight Seasionsand Application Level Framing. IEEE/ACM lhnsoctionson Networking, November 1996.David G. Thaler and Chinya V. Ravishankar. Distributedtopdown hierarchy construction. In Pmt. of the IEEE IN-FOCOM, 1998.P. Sharma, D. Estrin, S. Floyd and L. Zhang. Scalable Ses-sion Messages in SRM. Technical Report 98-670, Universityof Southern California, 1998.Paul F. Tsuchiya. The Landmark hierarchy: A new hierarchyfor routing in very large networks. In Pmt. of the ACMSIGCOMM, 1988.D. Coore, R. Nagpal and R. Weiss. Paradigms for Struc-ture in an Amorphous Computer. Technical Report 1614,Massachusetts Institute of Technology Artificial IntelligenceLaboratory, October 1997.A. Rosenstein, 3. Li and S.Y. Tong. MASH: The MulticastingArchie Server Hierarchy. ACM Computer CommunicationReview, 27(3), July 1997.A. M. Turing. The chemical basis of morphogenesis.Phil. Ifans. Roy. Sot. of London, Series B: BiologicalSciences(237):37-72, 1952.A reaction-difision model for development.J.J. Garcia-Luna-Aceves. Loop-free Routing using DiffusingComputations. IEEE/ACM lhnsactions on Networking,February 1993.C. Perkins and P. Bhagwat. Highly Dynamic Destination-Sequenced Distance-Vector Routing (DSDV) for MobileComputers. In Pmt. of the ACM SIGCOMM, October 1994.P. Jacquet, P. Muhlethaler and A. Qayyum. Optimized LinkState Routing Protocol. Internet Dmft, dmft-ietf-manet-oIs&lg.tzt, November 1998 . Work in progress.J. Broth, D. A. Maltz, D. B. Johnson, Y. C. Hu, and J.Jetcheva. A Performa&e Comparison of Multi-H&p Wire-less Ad Hoc Network Routing Protocols. In Pmt. of theACM/IEEE MobiCom, October 1998.Z. Hass and M. Pearlman. The Performance of Query Con-trol Schemes for the Zone Routing Protocol. In Pruc. of theACM SIGCOMM, Vonwuver, Canada, September 1998.

269


8/8

(141 M. Jiang, J. Li and Y. Tay. Cluster Based Routing Protocol(CBRP) Functional Specification. Internet Dmft, dmft-ietf-monet-cbrp-apec-OO.t, August 1998. Work in progress.[15] IhIWINS project. http://wz.janet.ucla.edu/lpe.[16] The Ultra Low Power Wireless Sensors project. http://ww-mtl.mit.edu/jimg/project-top.html.(171 Wearable Computing. http://vuu.cs.cmu.edu/uearabla/.(181 MEMS Technology Applications Center. http: //mema .mcnc.org.[19] Echelon, The LonWorka Company. LonWorks Solutions.http://ww.echslon.com/Solutions/.[ZO] F. Bennett, D. Clarke, J. Evans, A. Hopper, A. Jones and

PllWIP31

P41

P51

PI1301

1311

1321

I331

D. Leask. Piconet: Embedded Mobile Networking. IEEEPersonal Commun ications, 4(5), October 1997.F. Stajano and A. Jones. The Thinnest of Clients: Con-trolling it all via cellphone. ACM Mobile Computing ondCommunications Review, October 1998.R. Want et al. The Active Badge Location System. ACMDunsoctions on Information Systems, January 1992.N. Adams. An Infrared Network for Mobile Computers.In Proc. USENIX Mobile and Location-Independent Comp.Spmp., Cambridge, MA, August 1993.A. Ward, A. Jones and A. Hopper. A New Location Tech-nique for the Active Office. IEEE Personal Commun ications,4(5), October 1997.J. Werb and C. Lanzl. Designing a positioning system foriinding things and people indoors. IEEE Spectrum, Septem-ber 1998.The Factoid project. http://vuv.research.digital.com/wrl/projectz/Factoid/index.html.M. We&r. The Computer for the 21st Century. ScientificAmerican, September 1991.A. Pentland. Machine Understanding of Human Action. InProceedings of 7th International Forum Fron tier of Telecom-munication Tech., November 1995.W. Grimson, C. Stauffer, R. Roman0 and L. Lee. Usingadaptive tracking to classify and monitor activities in a site.CVPR, 1998.M. Dorigo, V. Maniezzo and A. Colomi. The Ant Sys-tem: Optimization by a colony of cooperating agents. IEEED-znractions on Systems, Mon and Cybernetics - Port B,26(l), 1996.M. L6pez-S&nchez, R. L6pez de Mantams, and C. Sierra.Incremental map generation by low const robots based onpossibility/necessity grids. In Pmt. of the Conf. on Uncer-tointy in Artif. Intell., pages 351-357, Providence, R I, 1997.S. McCanne. Scalable Multimedia Communica tion withInternet Multicast, Lightweight Sessions, and the Mbone.Technical Report Technical Report CSD 98-1002, Universityof California, Berkeley, March 1998.L. Zhang, S. Michel, S. Floyd, V. Jacobson, K. Nguyenand A. Rosenstein. Adaptive Web Caching: Towards aNew Global Caching Architecture. In ThirdIntemotionolCaching Workshop, June 1998. Available from http: //irl.cr.ucla.edu/AWC/.

270

next century challenges - scalable coordination in sensor networks

Documents