usense: A Unified Asymmetric Sensing Coverage Architecture for Wireless Sensor Networks

Transcription

1 usense: A Unified Asymmetic Sensing Covege Achitectue fo Wieless Senso Netwoks Yu Gu, Joengmin Hwng, Tin He, Dvid Hung-Chng Du {yugu,jhwng,tinhe,du}@cs.umn.edu Deptment of Compute Science nd Engineeing, Univesity of Minnesot Abstct As key ppoch to chieve enegy efficiency in senso netwoks, sensing covege hs been studied extensively. Reseches hve designed mny covege potocols to povide vious kinds of sevice guntees on the netwok lifetime, covege tio nd detection dely. While these potocols e effective, they e not flexible enough to meet multiple design gols simultneously. In this ppe, we popose Unified Sensing Covege Achitectue, clled usense, which fetues thee novel ides: Asymmetic Achitectue, Geneic Switching nd Globl Scheduling. We popose symmetic chitectue bsed on the conceptul seption of switching fom scheduling. Switching is efficiently suppoted in senso nodes, while scheduling is done in septed computtionl entity, whee multiple scheduling lgoithms e suppoted. As n instnce, we popose two-level globl covege lgoithm, clled uscn. At the fist level, covege is scheduled to ctivte diffeent potions of n e. We popose n optiml scheduling lgoithm to minimize e bech. At the second level, sets of nodes e selected to cove ctive potions. Impotntly, we show the fesibility to obtin optiml set-cove esults in line time if the lyout of es stisfies cetin conditions. We evlute ou chitectue with netwok of 3 MicZ motes, n extensive simultion with, nodes, s well s theoeticl nlysis. The esults indicte tht usense is pomising chitectue to suppot flexible nd efficient covege in senso netwoks. I. INTRODUCTION Wieless Senso Netwoks (WSNs), consisting of thousnds of low-cost senso nodes, hve been used in mny ppliction domins such s mility suveillnce [], hbitt monitoing [] nd scientific explotion. Limited powe supplies nd difficulties in hvesting mbient enegy mke enegy consevtion citicl issue to ddess. Enegy-efficient sensing covege extends system lifetime by leveging on the edundnt deployment of senso nodes. Within couple of yes, sensing covege hs become well studied subject which povides eithe full covege in both time nd spce [3], [4], [5], [6], [7], [8], covege with gunteed dely nd connectivity [9], [], [], [], o gunteed tget detection within cetin stelth distnce [3], [4]. These lgoithms e designed to be well distibuted nd loclized, poviding solid pefomnce gins with cetin guntee on sevice (e.g., bounded dely in detection o gunteed lifttime). While the stte-of-the-t is encouging, we believe thee e some spects tht need futhe investigtion. Fist, cuently diffeent sensing covege lgoithms focus on diffeent sevice guntees (e.g., covege vs. detection dely). Any single design is not genel enough to meet wide nge of sensing equiements unde diffeent opeting scenios. Second, in most lgoithms, extending system lifetime is chieved essentilly though coodintion mong neighboing nodes. The locl node density, theefoe, imposes theoeticl uppe bound on the system lifetime, if continuous sensing covege o ptil covege is equied. Such bound cn be supssed though globl scheduling. Howeve, the ovehed of globl scheduling would incese significntly if the coodintion mong the nodes goes beyond the neighbohood. To ddess these two issues simultneously, in this ppe, we intoduce new sensing chitectue, clled usense, which fetues thee novel ides: Asymmetic Achitectue, Geneic Switching nd Globl Scheduling. The key concept of usense is the decoupling of sensing covege into two septed functions: scheduling nd switching. The fome clcultes the pmetes of woking schedule fo individul nodes, while the ltte tuns on/off the sensos ccoding to the scheduling pmetes. We employ n symmetic chitectue to impove the flexibility nd extensibility of the design. Switching is lightweight geneic lgoithm, tking two pmetes s inputs. With these pmetes, it cn fithfully execute the sleep/wke schedule of individul nodes decided by ny existing covege lgoithm. Switching must be suppoted in the senso nodes, since node hs to be on to povide covege nd hs to be off to sve enegy. On the othe hnd, it is not bsolutely necessy to implement the scheduling within constined senso nodes. We opt to suppot scheduling on poweful computtionl entity, which cn be implemented t the second-tie nodes (e.g., Intel Stgte used in TENET [5] nd ExScle []), o on n outside seve, o on cluste of seves to void single point of filue. Asymmetic design is now consideed s pomising guiding pinciple fo senso netwoks. By decoupling the scheduling function nd implementing it outside the netwok coe, we cn chieve efficiency nd pefomnce simultneously. This is becuse fistly, with fewe functions shing the limited esouces on node, we cn build the switching functions, the ones tht must be embedded into individul senso node, in less stingent design spce. Secondly, we cn design nd implement the scheduling functions outside of senso netwoks, fee of esouce constins inheent in the senso nodes, theefoe, will be moe sophisticted nd poweful, leding to significntly impoved ovell pefomnce. Befoe descibing the usense design in detil, we identify the objectives nd intellectul contibutions of this wok s follows: Asymmetic Sensing Achitectue: The symmetic chitectue enbles us to design sophisticted covege lgoithms in n unconstined design spce nd epesent such intelligence with lightweight lgoithm implemented in senso nodes. Geneic Switching Algoithm: To the best of ou knowledge, we popose hee the fist geneic nd lightweight switching lgoithm tht is suitble fo senso nodes with

2 limited esouces. We demonstte ou lightweight design t the senso side is vey effective. Globl Scheduling Algoithms: We design two new globl scheduling lgoithms, using the concept of set-cove. Diffeent fom ll pevious wok, we demonstte the fesibility to identify minimum cove set in line time when the covege e is continuous cuve. System Implementtion nd Extensive Evlution: We hve invested significnt mount of effot to evlute ou design. We hve implemented nd evluted the design on the TinyOS/Mote pltfom, using 3 MicZ motes. We lso povide,-node lge scle simultion to compe with stte-of-t solutions, lowe-bonds nd uppe-bounds. The est of this ppe is ognized s follows. Section II intoduces the oveching chitectue of usense. Section III descibes the design of uscn, two-level globl scheduling lgoithm. Section IV povides n nlytic study of the poposed scheduling lgoithms. Section V descibes ou system implementtion nd povides evlution on the TinyOS/Mote pltfom. The esults of the,-node simultion e pesented in Section VI to compe the pefomnce of usense nd uscn with stte-of-the-t solutions. Section VII discusses the elted wok. Section VIII concludes the ppe. II. USENSE ARCHITECTURE One of key new ides of this wok is the conceptul seption of switching fom scheduling. In this section, we povide n oveview of ou symmetic sensing chitectue. A. Motivtion Ou wok is iming t flexibility nd efficiency in sensing covege. It is often the cse tht senso netwok needs to suppot multiple opeting scenios. Fo exmple, mility suveillnce netwok could be equied to povide full covege duing ed let ( sptil covege equiement), howeve, it my llow cetin detection dely ( tempol covege equiement) on othe occsions to ggessively conseve enegy. Two lgoithms ([7] nd [9]) hve been successfully designed to meet these two design gols. Howeve, neithe of them, unfotuntely, is flexible enough to meet both equiements. Evidently, with these two lgoithms, we cn chieve both functionlities by downloding two septe pogm imges nd switching between them (s suppoted by Deluge [6]). Clely, septe imges intoduce excessive ovehed in tems of communiction bndwidth, enegy nd stoge, putting flexibility nd efficiency t odds with ech othe. We obseve tht the wkeup/sleep schedules of individul senso node cn be descibed by two pmetes, which e independent of the methods to obtin them. Theefoe, the conceptul seption of switching fom scheduling becomes ntul ppoch to esolve the conflict between flexibility nd efficiency. In the usense chitectue, chnging scheduling lgoithms only ffects the vlues of the pmetes, not the switching logic implemented t the nodes. Consequently, flexibility cn be chieved by disseminting only few pmetes. Switching Side Senso Netwok Senso i Senso j Senso k Geneic Switching Algoithm Fig.. Pmetes Connectivity Computtionl Entity g tin s l n T te e m P Oveview of usense Achitectue Scheduling Side Algoithm Algoithm.. Algoithm n B. usense Design The usense chitectue decouples switching fom scheduling s shown in Figue. The switching lgoithm is implemented in the senso nodes, which tkes two scheduling pmetes s input. The scheduling lgoithm, which is implemented septely, is esponsible fo geneting these scheduling pmetes. It tkes the schedules decided by vious sensing covege lgoithms nd tnsltes them into the two pmetes to be used by the switching lgoithm. The usense chitectue equies bi-diectionl communiction, becuse tht (i) most scheduling lgoithms need the loction infomtion of the senso nodes in ode to cete wkeup/sleep schedule fo individul node, nd (ii)usense needs to disseminte the scheduling esults to the nodes within the netwok, whee the ctul switching hppens. Obviously, in sttic senso netwok, these costs e vey smll, comped with the mount of sensing dt tnsmitted by the senso netwok. Futhemoe, vious existing potocols such s SPIN [7] nd PSFQ [8] ledy cn effectively pefom enegy efficient infomtion dissemintion nd gtheing in low powe senso netwoks. In the next two sections, we focus on the sensing covege, descibing the switching nd scheduling lgoithms, espectively. C. Pt I: Geneic Switching Algoithm The switching lgoithm should be lightweight to un on esouce constined nodes nd should be geneic to ccommodte vious types of schedules. In ou switching lgoithm design, two pmetes e used fo ech node, nmely the schedule bits S nd switching te R. Schedule bits S is n infinite biny sting in which denotes the ctive stte nd denotes the inctive stte. The duty cycle of node is the pecentge of s in S. Switching te R defines the te of toggling between sttes. Fo exmple, switching te of.5hz equies node to ed one bit fom the schedule S evey seconds. When tnsient enegy consumption is negligible, idelly high switch te R leds to smll detection dely. Howeve, R cnnot be bitily smll becuse: (i) senso hs wm-up dely befoe it cn pefom elible smplings, (ii) in most detection lgoithms, it is not obust enough to tke just one smple to mke decision, so smpling dely popotionl to the numbe of smples is often intoduced. These delys impose n uppe-bound on the switching te R.

3 Although ou switching lgoithm is simple nd lightweight, it is poweful enough to suppot mny types of covege lgoithms. Theoeticlly, when the switching te R ppoches infinity, schedule bits, s n infinite sting, cn pecisely chcteize on/off behvio geneted by ny covege lgoithm. As mentioned, the switching te is finite in elity, theefoe in the wost cse, node might need to extend wke-up peiod by R seconds to guntee the covege. ) Regul Expession: A switching lgoithm tkes schedule bits S nd switching te R s inputs. Schedule bits S is fomlly defined s n infinite biny sting. Obviously, it is pcticlly impossible to disseminte n infinite biny sting. Fotuntely, the sensing covege schedule is usully peiodic, follows cetin ptten. Theefoe, we cn expess S with egul expession. Fo exmple, () cn be used to denote epeted off-off-ctive-off schedule. Fo vey dense netwok, fte scheduling, the duty cycle of n individul node is usully low. In othe wods, the numbe of s in schedule is comptively smll. In this cse, we cn use index vlues to epesent the positions of ctive bits. ) Timed Finite Automt: Fo given schedule S descibed s egul expession, node builds finite utomt (FA). A stightfowd method is to use time t the te of R to dive the stte tnsitions within the FA. Howeve, this equies node to wke-up the pocesso peiodiclly, intoducing excessive enegy consumption. To ddess this issue, we theefoe build Timed Finite Automt (TFA). In TFA, stte tnsition is tiggeed by o segments in the schedule S, nd the delys of tnsitions e the gps between o segments. D. Pt II: Scheduling Algoithms As shown in Figue, senso scheduling lgoithm is implemented septely (e.g., t the second tie). Fee of esouce constints inheent in the senso nodes, we cn suppot lge numbe of covege lgoithms. In this section we focus on the geneic scheduling fmewok, befoe poposing concete scheduling lgoithms in the next section. To ccommodte diffeent sensing covege lgoithms, we need to convet the output of covege lgoithm into two pmetes undestndble by the geneic switching lgoithm. To illustte the ide, we use the lgoithm pesented in [8] s cse study. Round ( Dution T ) Round Round i T font Ref T end T font Ref T end... T font Fig.. Schedule Tnsltion In [8], the sensing phse of nodes is divided into ounds with equl dution T. The schedule fo node is detemined by tuple with fou pmetes: (T, Ref, T font, T end ). As shown in Figue, Ref is ndom time instnce chosen by node within [, T). T font is the dution of time pio to the efeence point Ref, ndt end is the dution of time fte efeence point Ref. To Ref T end build schedule bits S, we check whethe time instnce k R,k [,T R] is locted within the ctive peiods. Fo exmple, the schedule shown in Figue cn be expessed s (). Computtionl Entity g tin s l n T te e m P uscn Algoithm.. Algoithm n Fig. 3. uscn Tile-Level Scheduling Line Scnning Systolic Scnning Node-Level Scheduling The Design of uscn III. GLOBAL SCHEDULING ALGORITHMS Conceptully, usense cn suppot mny existing covege lgoithms. Due to its symmetic chitectue, it is especilly fiendly to the globl scheduling lgoithms. Since globl scheduling llows mny moe nodes to ctivte in tun the thn the loclized ones tht only schedule the nodes within neighbohood, it leds to significnt enegy svings. In this section, we popose globl scheduling lgoithm clled uscn. Figue 3 shows the eltion between uscn nd usense. Essentilly, uscn is one of sensing covege lgoithms tht e suppoted by the usense chitectue. The outputs of uscn e the schedule bits S nd switching te R fo individul nodes. uscn is two-level schedule lgoithm, which woks s follows: Suppose we povide sensing covege to given e using uscn s shown in Figue 4. Fist, uscn divides the e into smll egions, nd decides the woking schedules fo these egions. This level of scheduling is conceptully independent of the deployment of the nodes. At the secondlevel, we ssign nodes to cove the ctive egions t diffeent time intevls, using set-cove technique. By combining the fist-level schedule nd the set-cove ssignment, we cn decide the schedule bits S fo individul nodes. The biggest dvntge of this two-level schedule lgoithm is the seption of sensing ptten fom the undelying node scheduling. The ppliction only needs to specify the desied sensing behvio on the field in the fist level of scheduling, nd the second level of uscn cn tke vious specified sensing pttens s input nd poduce the finl woking schedule of ech individul senso device. The two-level schedule of the uscn povides flexibility, eusbility nd efficiency to the sensing component of senso pplictions by feeing vious sensonet pplictions fom designing thei own scheduling potocols unde diffeent ppliction equiements. A. Assumptions Fo the clity of the potocol desciption in the est of the ppe, we ssume tht nodes e time-synchonized nd thei loctions e pecise. We efe the sensing e of node s cicle with nominl dius centeed t the loction of the node. These e common ssumptions fo mny senso netwok pplictions [], []. 3

4 A Tile Active Sleep A node with schedule bits... Fig. 4. Regul Tesselltions Fig. 5. Hoizontl Scn Fig. 6. Systolic Scn B. Level I: Tile Scheduling In uscn, we ptition n e unde suveillnce into some smll egions of the sme shpe, pocess clled tesselltion. These smll egions e clled tiles, which cn be egul tingles, ectngles o egul hexgons in -D spce. One simple exmple of tesselltion is the ectngle-bsed ptition, s shown in Figue 4. The size of tiles is set to be smlle thn the minimum tget size, so tht tget is detected s long s potion of tile is coveed. As eminde, nodes within senso netwok only suppot geneic switching lgoithm, which hs neithe the concept of tiles no the ptition infomtion of the tiles. All the complex logic esides outside of senso nodes. In this section, we descibe two simple, yet effective methods fo the tile-level scheduling. They diffe in the enegy consumption te nd the detection dely. ) Line Scn: We stt with simple tile-level scheduling s shown in Figue 5. Insted of tying to cove ll tiles, we only cove column/ow of tiles in cetin intevl of time duing one ound of scn. The coveed columns/ows e incesing o decesing consecutively. Becuse only smll pecentge of tiles e sensed t specific point of time, line scn leds to significnt eduction in enegy consumption, comped with full covege [8]. Specificlly, in the line-bsed globl scnning, we intoduce the concept of scnning speed v, which epesents the speed of scn fom one end to the othe, hoizontlly o veticlly. This scnning speed detemines the mximum detection dely netwok expeiences. The scnning speed v cn be tnsfomed into the switch te R. Suppose we hve ectngle-bsed tesselltion, the length nd width of tile e L l nd L w, espectively. Fo given scn speed v, if we wnt to scn hoizontlly, the switch te R is set to be v L l. Similly, the v switch te is L w, when we scn veticlly. Stting fom, we index the tiles in ow-mjo ode. Theefoe tile with coodintes (ow, col) is ssigned the index of ow col mx +col (col mx is the mximl column index). To cove tile t(i) with coodintes (ow, col) in scnning ound, schedule bits S fo this tile is s follows: S h (i) = (.. }{{} col S v (i) = (.. } {{ } ow.. }{{} ) (Hscn) col mx col ) (Vscn).. }{{} ow mx ow Moeove, we cn pefom hoizontl nd veticl scns simultneously. Both diections she the sme scnning speed v. () We cn obtin the schedule bits of two-wy scn by pplying the bitwise OR opetion on S h nd S v obtined in Eqution : S(i) =S h (i) S v (i) () ) Systolic Scn: Systolic Scn emultes the cdic cycles of beting het. Figue 6 shows the design of systolic scn. The tiles e scnned fom the inne lye to the oute lye, s denoted by diffeent gy-levels in Figue 6. Without loss of genelity, we descibe the method with simple cse whee col mx = ow mx = N. Clely, the length of schedule bits in scnning ound is N/ Fo the fist time intevl (epesented by the fist digit in the schedule bits), the tiles t the cente of the e set thei fist digit of schedule bits to, nd the schedule bits fo these tiles e ( }.. {{} ). N/ Similly, fo the n th time intevl, the tiles whose coodintes meet one of follow fou conditions: ow == n & col > n & col N n ow == N n & col > n & col N n col == n & ow > n & ow N n col == N n & ow > n & ow N n set thei schedule bits s follows: S(i) = ( }.. {{} }.. {{} N/ n n i = ow N + col whee n =,,..., N/ nd i is the index of tiles tht stisfies the equiements. Both line scn nd systolic scn specify only the set of tiles need to be ctivted (coveed) t given point of time. The tsk of coveing ech tile set is ccomplished by the second-level node scheduling, which will be descibed in the next section. C. Level II: Node Scheduling Tile-level scheduling detemines the set of ctive tiles TS i t the time intevl i. Fo exmple, in hoizontl line scn, the i th column is ctivted t time intevl i. In this section, we descibe how we cn tnslte known tile schedule into coesponding node schedule bits S, which cn be intepeted diectly by geneic switching lgoithm. ) (3) (4) 4

5 N T T T 3 N T T N N N N N N v v v 3 N N N T 4 T 3 N 3 N 4 T 5 N 3 N 4 N 3 v 4 T 4 N 4 N 5 N 4 v 5 N 3 T 5 N 5 v 6 N 5 Fig. 7. Physicl Covege Fig. 8. Biptite Gph Fig. 9. MSC using DAG ) Min Ide: Befoe we discuss the complete lgoithm, we fist illustte ou ppoch with simple exmple. Figue 7 shows one column of tiles TS = {T,T,T 3,T 4,T 5 } tht is coveed by set of nodes NS = {N,N,N 3,N 4,N 5 }.Since we set the tile size smlle thn the miniml tget size, tile is sid to be coveed s long s potion of this tile is coveed. Figue 7 cn be mpped to the Covege Biptite Gph shown in Figue 8 ccoding to the covege eltionship. Node scheduling consists of two steps. Fist, we keep identifying one-cove set with miniml numbe of nodes, until the size of one-cove set is bove cetin theshold. Fo exmple, s shown in Figue 7, we identify thee one-cove sets fo TS: CS = {N,N 5 }, CS = {N,N 3 } nd CS 3 = {N,N 4 } to ensue tht ll nodes e used. Thee sets CS, CS nd CS 3 cn povide covege to the tile set TS in ound-obin fshion. To do this, we cete node schedule tht hs thee segments, ech of which hs length of the tile schedule. If node belongs to the CS k set, the k th segment hs the sme vlue s the tile schedule. Othewise, the k th segment hs n ll-zeo vlue. Fo exmple, if the tile schedule of TS is, the finl schedules fo the nodes shown in Figue 7 e: S = ( ) S = ( 3 3 S 3 = ( ) S 4 = ( ) (5) 3 3 S 5 = ( ) 3 ) Identifying Minimum Set-Cove within Line Time: To sve enegy t ech time intevl, we need to identify minimum set of nodes to cove n ctive tile set. This is typicl set-cove poblem, which cn be fomlly defined s: Definition Given collection C of subsets of finite set T,find set cove C (C C) fot, such tht evey element in T belongs to t lest one membe of C. The geneic Minimum Set Cove (MSC) poblem hs been poven NP-Hd nd ny polynomil lgoithm cn only find esults of +ln T optimum [9]. Fotuntely, line scn covege is specil cse of the geneic set cove poblem, becuse node cn cove only continuous segment of tiles. The min ide of ou polynomil lgoithm is to mp Covege Biptite Gph ) (figue 8) into Diected Acyclic Gph (DAG) (figue 9). The one-to-one mpping ules e s follows: ) We mp N tiles in TS i into N vetices V = {v,..., v N } nd dd one ext vetex v N+. ) If node coves set of tiles {T i,..., T i+n }, we cete n diectionl edges (v i,v j ) whee v j = v i+,..., v i+n+. Ech edge hs unit cost. Though this mpping, the tile set cove poblem cn be educed to the poblem of finding out the shotest pths fom v to v N+. The mpping pocess tkes O( V )+O( E ) time. Fo n bity gph, the shotest pth lgoithm finishes within O( V ) using the Dijkst lgoithm. Since the gph we cete is DAG, we cn find the shotest pth in O( E ) time, using fst eching lgoithm []. Theefoe the whole lgoithm finishes in O( V )+O( E ). To illustte the ide, Figue 9 shows DAG which is mpped fom Figue 8. To identify the miniml set cove, we need to find out the shotest pth fom v to v 6. In this simple cse, we cn cove ll the tiles using one of following node sets: {N,N 3 }, {N,N 4 }, {N,N 5 }, {N,N 3 } o {N,N 4 },which e five coesponding shotest pths fom v to v 6. We note tht the poposed polynomil lgoithm does not pply to geneic tile scheduling. When tile set does not fom continuous cuve o node cn cove multiple segments of tile set simultneously, the polynomil lgoithm cn not guntee the complete covege of ctive tiles. In these cses, we dopt geedy set-cove method by choosing the node tht coves the most numbe of tiles fist. 3) Selecting Cove Sets fo Multiple TS: Up to now, node scheduling hs been descibed using simple exmple tht ssumes node only needs to cove one tile set. Obviously, to suppot line scn o systolic scn in -D spce, we need to identify cove sets fo the whole e (not just fo single column). Thus node my need to cove multiple tile sets TS i. The detiled pocess to cove the e is s follows: ) Ech node mintins counte SC to ecod how mny times it hs been selected into finl Cove Sets (fo the pupose of enegy blnce). ) Fo tile set TS i, the lgoithm clcultes the minimum cove set MCS i mong the nodes with minimum SC vlues. If the nodes with minimum SC vlues cn not fom complete cove set, nodes with highe SC vlues 5

6 e used. 3) Afte we obtin ll MCS i, the smllest eligible MCS i (SMCS) is selected nd ecoded fo the pupose of node scheduling, nd the SC vlues of nodes within this SMCS set e incemented. 4) Ech TS i hs covege theshold, denoting the mximum numbe of nodes tht cn be used in selected MCS i. These thesholds e clculted bsed on the concept of edundncy. If the fist MCS i chosen fo TS i includes M nodes, the numbe of nodes in the following MCS i fo TS i should not diffe significntly, in ode to educe edundncy in covege. We set the theshold fo TS i s M π 7, ccoding to the edundncy in cicle coveing []. 5) The SMCS selection pocess is epeted until the size of ll MCS i e lge thn thei thesholds. 4) Cete Node Schedule Bits R : Suppose K one-cove sets e selected fo tile set TS i with tile schedule S T (fom Section III-B), we cete node schedule S i fo node N i,which hs K segments. The vlue of ech segment is eithe S T o zeo. If node belongs to the k th one-cove set, the vlue of the k th segments is S T.Othewise,thek th segment hs n ll-zeo vlue. Since in -D spce, node might need to cove diffeent tile sets in single ound. Supposing node needs to cove M diffeent tile sets, the finl node schedule S is: S = OR M (S i ). (6) i= 5) Suppot Diffeentited/Robust Suveillnce: Diffeentited suveillnce [8] cn be suppoted esily by uscn, due to its setcove bsed ppoch. Insted of tuning on one set of nodes to cove column/ow, uscn cn tun on multiple disjoint set of nodes to incese the degee of covege. This leds to highe detection confidence, but t the cost of netwok lifetime. Similly, fult tolence cn be chieved by tuning multiple sets on. It is inteesting to emphsize tht nodes ctully hve no concept of set, which leds to nice popety fo fult tolence: To fix the filue of nodes, we only need to modify the schedule bits S of the nodes in the neighbohood of filed node nd no coodintion between nodes is needed. IV. DESIGN ANALYSIS Diffeent fom full covege lgoithms in [7], [], uscn coves only pt of netwok. On one hnd, this ppoch significntly inceses the netwok lifetime, but on the othe hnd, it intoduces cetin dely in tget detection. In this section, we povide nlytic esults on the pefomnce of uscn. Hee, we focus on the tile-level nlysis insted of the node-level. Let s conside n e with N by N tiles. A. Detection Dely fo Sttic Tgets To evlute the detection dely fo sttic tgets, we ssume tht tget is ndomly locted in n e nd is detected fte neighboing node tuns on fo R seconds. In line nd systolic scn, in ode to guntee detection, tile must be tuned on once pe ound. The miniml detection dely hppens when tget shows up in tile ight befoe this tile is tuned on. In this cse, the detection dely is R. The mximum detection dely N Beched Ae N ( R + ) N ( R + ) Un-beched Ae N ( R + ) Tget Diection N ( R + ) Beched Ae N R Scn Diection N R Un-beched Ae N R ) Systolic Scn (Outwd) b) Line Scn (Left to Right) Fig.. The Beched Ae N R + hppens when tget shows up in tile ight fte this tile is tuned on. In this cse, the detection dely is +N R fo line scn nd + N/ R fo systolic scn. Since the dely is unifomly distibuted in ound, the expected dely is +N R fo line scn nd + N/ R fo systolic scn. Unde the sme configution, the detection dely fo full covege lgoithms is zeo. To educe the detection dely in uscn, we cn divide netwok into sub-netwoks, whee multiple line scns nd systolic scns e executed in pllel. B. Beched Ae fo Mobile Tgets In full covege scenio, the wost-cse bech e is zeo. A mobile tget is detected once it entes into the e. In the scnning ppoch, tget would ech cetin potion of the e befoe it is detected. We define the lgest pecentge of the e tht tget cn ech without being detected s the Wost-Cse Bech (WCB). A smlle WCB indictes bette pefomnce in mobile tget detection. To clculte WCB, we ssume the following mobility model: A tget cn only ente fom outside of the netwok, nd the mximum speed of ny tget is tiles pe second. WCB WCB of Systolic Scn WCB of Line Scn WCB of Line Scn t Double speed Tget Speed (Scn speed R= ) Fig.. Pefomnce Compison The wost cse bech scenio fo systolic scn (outwd diection) is shown in Figue (). In this scenio, the wostcse bech hppens when tget entes t the beginning of the scn ound. The distnce this tget cn bech without detection N is (R+). Theefoe, the WCB s(, R) fo systolic scn unde the tget speed nd switching R is: WCB s (, R) = (R + ) (R + ). (7) 6

7 Fig.. usense System Setup Fig. 3. Injecting Events Fig ,, 5 Lyouts The wost cse bech scenio fo line scn is shown in Figue (b), which hs fou moe sophisticted cses: ) If tget entes fom the left edge (the sme diection s the scn), the distnce this tget cn bech is N R. ) If tget entes fom the ight edge (opposite diection s the scn), the distnce this tget cn bech is N R+. 3) If tget entes fom the top o bottom edge. In ode to chieve the mximum bech, tget should ente with n ngle α =ctn( R ). In this cse, the mximum N distnce tht this tget cn bech is. R 4) If tget hs speed of ( )R o gete, it cn ente the e fom the left to ech t lest pecent of the e nd it cn lso ente the e fom the ight to ech t lest pecent of the e. Theefoe, the whole e is beched. By combining these fou cses, we get WCB l (, R) fo line scn, ssuming tget speed nd switching R: WCB = ( R R + R ( ) <( )R R R ( )R (8) Now we e edy to compe two globl scheduling lgoithms. As shown in Section III-B., fo given switching te R, systolic scn consumes twice s much enegy s line scn does. To obtin fi compison, we thus double the switching te of line scn. By comping WCB l nd WCB s, it is esy to pove tht WCB l (, R) WCB s (, R) t ll tget speeds. In othe wods, systolic scn is bette thn line scn in tems of minimizing the beching e. Actully we hve poven tht systolic scn is n optiml scnning lgoithm in tems of peventing e bech when the tget speed is vey fst. Due to the spce constint, we omit the detiled poof hee. On the othe hnd, the line scn lgoithm hs its own dvntges. As we hve shown in Section III-C., we e ble to obtin optiml setcove esults fo line scn within polynomil time. To illustte the diffeence futhe, Figue shows the WCB vlues unde diffeent tget speeds (-), when the switching te R is. Clely, the diffeence is significnt. Fo exmple, when the tget speed is hlf of scnning speed(=5), systolic scn potects bout hlf of the e, while line scn cnnot potect ny potion of the netwok. V. IMPLEMENTATION AND EVALUATION We hve implemented complete vesion of usense (with uscn) s designed in Section II nd III. The geneic switching lgoithm is witten with the NesC lnguge, unning on the TinyOS/Mote pltfom. The compiled imge of full implementtion occupies,4 bytes of code memoy nd 97 bytes of dt memoy. A simple time-diven FA logic is implemented to tun mote on/off ccoding to the schedule bits. We use FTSP [3] fo the pupose of time synchoniztion mong motes nd Deluge [6] fo the pupose of wieless epogmming. The synchoniztion ccucy is t tens of micoseconds, which is sufficient fo most sensing scheduling lgoithms. The scheduling lgoithms e witten in Jv, uns on lptop. Since senso nodes hve no concept of scheduling, the globl scheduling lgoithm uscn nd othe covege lgoithms e witten only in Jv. The schedule bits S nd the switching te R e disseminted fom the bse, using single pcket. We lso implement n evlution engine using Jv, which genetes vitul tgets using the light. To ccutely mesue the dely, we implement n NTPlike two-wy hndshking synchoniztion potocol ove seil cble to synchonize the bse mote nd lptop. This synchoniztion potocol is not pt of usense nd is only used fo the evlution pupose. As shown in Figue to evlute usense nd uscn, we ttch 5 MicZ motes on veltex bod (4 feet by feet) using velco stps. We use DELL 3MP pojecto to genete light spots on the veltex bod. These light spots e used to emulte sttic nd mobile events. Fo exmple, the sttic events e ndomly geneted in the gid to tigge the detection, s shown in Figue 3. Afte the nodes detect the light events, they epot timestmps to the lptop, whee the dely is clculted. As shown in Figue 4, thee gid lyouts e used in the expeiments: 5 by, by nd 5 by 5. In ddition, we evlute uscn with ndom plcement s well. The loctions of nodes in the ndom plcement e obtined though ndom geneto. Ech node is ssigned n enegy budget (the numbe of times node cn be tuned on), which is used to evlute the lifetime of the sensos. Events e epeted hundeds of times to obtin esults with high sttisticl confidence. 7

8 Tget Size = Tget Size = Detection Possibility CDF Rndom Plcement Gid Plcement Time Detection Dely Fig. 5. Detection Ove Time unde Diffeent Node Plcements Line Scn Systolic Scn Fig. 6. Detection Dely of fo Sttic Tgets unde Diffeent Tget Sizes CDF.4 CDF Detection Dely. Lyout 5 by 5,Line Scn-Sme diection Lyout *,Line Scn-sme diection Lyout 5*,Line Scn-Sme diection Detection Dely Fig. 7. Detection Dely of Line nd Systolic Scn fo Sttic Tgets Fig. 8. Detection Dely of Line Scn fo Mobile Tgets Detection Dely Line Scn-Sme diection Line Scn-Pependicul diection Line Scn-Opposite diection Systolic Scn Detection Possibility Line Scn-Sme diection Line Scn-Pependicul diection Line Scn-Opposite diection Systolic Scn Switching Dely Switching Dely Fig. 9. Detection Dely Unde Diffeent Switching Delys Fig.. Detection Pobbility Unde Diffeent Switching Delys A. Testbed Evlution The system evlution focuses on the detection dely fo sttic nd mobile tgets, the detection pobbility fo mobile tgets nd the node lifetime. All the expeiment e epeted times. ) Detection Pobbility Ove Time: In this expeiment, we inject sttic tgets though the DELL pojecto into the netwok to evlute the detection pobbility ove time. A tget is missed only if tile is not coveed (i.e., node uns out of powe). We test uscn unde the ndom plcement nd gid plcement using MicZ motes. The esults e shown in Figue 5. Since nodes in the gid plcement hve well blnced dutycycles, they povide full covege until ll of them un out of enegy simultneously. In the ndom plcement, the spse es become uncoveed fist, nd covege degdes gdully ove time. Fo exmple, ndom plcement still keeps bout 4% covege when covege educes to zeo in the gid plcement. ) Detection Dely fo Sttic Tgets: In this expeiment, we investigte the detection dely fo sttic tgets unde diffeent minimum tget sizes. Sttic tgets e injected t ndom time intevls into the e. Since the tile size is detemined by the minimum tget size, the numbe of columns needed to cove the sme e educes when the minimum tget size inceses, theefoe the detection dely educes s well. Figue 6 shows the Cumultive Density Function (CDF) cuves of the delys fo two tget sizes. Clely, lge tget size leds to smlle delys. Fo exmple, when the tget size is one, the mximum dely on detection is 4834ms, while the mximum dely is 378ms with tget size of two. 3) Compison of Detection Dely: In this expeiment, we use 5 MicZ to fom 5 by 5 gid nd compe the detection dely of sttic tgets fo line nd systolic scn, gin using CDF cuves. Fom Figue 7, we cn see tht unde the sme switching te, the detection dely of systolic scn is bout one-hlf of line scn, which is consistent with the length of schedule bits shown in Section III-B. In the 5 by 5 gid lyout, the vege detection delys fo systolic nd line scn e 38ms nd 76ms. 4) Impct of the Netwok Size nd Scn Diection: In this expeiment, we study the impct of the netwok size using thee netwok lyouts. The tget moves in the sme diection s line scn. As shown in Figue 8, s the netwok size educes, the detection dely deceses ccodingly. Fo exmple, the vege detection delys fo the by 5, by nd 5 by 5 lyouts e 3758ms, 84ms nd 5ms, espectively. This indictes tht to guntee cetin detection dely, we should ptition lge e nd pefom scns within the sub-es. 5) Impct of the Switching Dely nd Scn Diection: Figue 9 studies the detection dely unde diffeent switching delys (the ecipocl of the switching te R) nd diffeent 8

9 tget diections. This investigtion is intended to evel the impotnce of designing fst hdwe nd detection lgoithms. We use 5 by 5 lyout, genete mobile tgets fom thee diffeent diections, nd mesue the delys befoe the mobile tgets e detected. Figue 9 shows fou cuves, epesenting thee tget moving diections on line scn nd one diection on systolic scn. Systolic scn hs the smllest detection dely t ll switching tes. Line scn in the opposite diection of tget moving diection povides the second smllest detection dely. The longest dely hppens when we scn t the sme diection s the tget moving diection. In ddition, Figue 9 shows tht, genelly, when the switching dely inceses, the detection dely inceses linely. Inteestingly, thee e two dt points tht do not follow this tend. It is becuse tht line scn misses the tgets when the scn speed is below twice the tget moving speed (when both move in the sme diection). In ou setup, this hppens when the switching dely is longe thn 3ms. Unde these slow scnning speeds, uscn my miss tgets nd ecod only the shot detection delys, leding to smll vege detection dely. This is lso confimed by the detection pobbility esults in Figue, which indicte when tgets move oppositely to the diection of line scn, we cn ensue the % detection of mobile tgets. Howeve, if the scnning diection is the sme s the tget moving diection, the detection pobbility dops to 45% t the long switching dely of 6ms. VI. LARGE SCALE SIMULATION: COMPARING WITH STATE-OF-THE-ART Expeiments on the test-bed indicte tht usense cn be efficiently implemented on the esouce constined devices nd evel its nice fetues. In this section, we compe the pefomnce of uscn with sevel stte-of-the-t sensing covege lgoithms integted into usense chitectue, vlidte the benefit of symmetic sensing chitectue nd the two-level scheduling ppoch. In this simultion, up to, senso nodes e ndomly distibuted in 3m 3m sque field. The sensing nge is m. The senso nodes e deployed with ndom distibution into the sque field. To void pefomnce distotion due to the edge effect, we set the covege e s the 9m 9m sque in the cente of the sque field nd do sttistics on the centl 75m 75m field. The following bselines nd bounds e dopted fo puposes of compison: ) Bseline-I: All-Woking: The full covege mode with ll nodes on. ) Bseline-II: DiffSuv: Diffeentited Suveillnce fo senso netwoks poposed in [8] integted into the usense chitectue. 3) Bseline-III: Vitul Ptol: Covege-oiented ptols using wieless senso netwoks poposed in [4] integted into the usense chitectue. 4) Uppe Bound-I: Optiml full covege using / 7 cicles with honeycomb lyout. 5) Uppe Bound-II: Idel uppe bound fo line scn, which ssumes tht nodes e plced optimlly such tht MCS i fo ech time intevl is minimized. All the expeiments e epeted times with diffeent ndom seeds nd node deployments. The 95% confidence intevls e within 5% of the men, which is not plotted fo the ske of legibility. The following metic is used to evluted the pefomnce of uscn. Netwok Hlf-life: We define the hlf-life of senso netwok s the time fom the beginning of the deployment until exctly hlf of the nodes e still live. This metic indictes the enegy efficiency of the netwok s whole. A. Pefomnce unde Full Covege Mode If the tile set TS to be coveed includes ll the tiles in the netwok, uscn essentilly povides full covege. In this expeiment, we compe uscn with othe solutions unde Full Covege Mode to evlute the effectiveness of the set-cove ppoch in uscn. Figue shows the netwok hlf life of fou diffeent solutions: usense, DiffSuv, ll-woking lowe bound nd theoeticl uppe bound (by ssuming honeycomb lyout). Fom Figue, we cn see tht the hlf lives fo ll cses e incesing linely when the node density inceses. The slope of usense is lge thn DiffSuv, which implies tht s the node density inceses, the diffeence in hlf-life between usense nd DiffSuv inceses s well. Fo exmple, t the node density, the hlf life fo DiffSuv nd usense e.9 nd.35, espectively. And when the node density eches 4, the coesponding two hlf lives e 4.9 nd 6.99 nd the hlf-life diffeence inceses fom 3.4% to 4%. System Hlf Life Idel Uppe Bound DiffSuv usense All Woking Fig.. Node Density System Hlf Life vs. Node Densities B. Pefomnce unde Scnning Mode In [4], Gui nd Mohpt popose vitul ptol solution simil to ou line scn scheme. The min diffeence between vitul ptol nd uscn is tht vitul ptol only povides singlelevel of scheduling. At ech time, if node s distnce to the ptolle is within its sensing nge, the node is ctive; othewise it is in sleep stte. Figue shows the hlf life of the line scn fo usense, vitul ptol nd idel uppe bound. Fom the Figue we cn see tht s the node density inceses, the system hlf life of the usense inceses lmost linely. On the conty, the hlf life of the vitul ptol emins stedy nd cnnot tke the dvntge of the incesed node density. When the node density eches, the hlf life of the usense is , while the vitul ptol hs hlf life of 3.99, which is bout 7 times enegy inefficiency thn the usense. The mjo eson fo such lge pefomnce gp is tht vitul ptol ctivtes ll the nodes cn sense the ptolle, while the usense only use minimized subset of the eligible nodes to povide the desied covege. 9

10 System Hlf Life Uppe Bound-II Vitul Ptol usense Fig.. Node Density System Hlf Life vs. Node Densities VII. RELATED WORK Physicl sensing covege is the esech focus in the beginning. In [7], uthos suppot full suveillnce covege bsed on n off-duty eligibility ule. DiffSuv [8] povides diffeentited suveillnce to n e with cetin degee of covege. In [], suveillnce covege is chieved though pobing. Sevel othe woks focus moe on the theoeticl esults of sensing covege. Kum et l. [5] identify citicl bound fo k-covege in netwok, ssuming node is ndomly tuned on with cetin pobbility. In [5], Kum et l. investigte the k-bie covege poblem, identifying the citicl condition fo wek k- bie covege. Sevel lgoithms e designed bsed on the concept of set cove. In [4], Cdei et l. popose two heuistic lgoithms to identify mximum numbe of set coves to monito set of sttic tgets t known loctions. In [3], Abms et l. popose thee ppoximtion lgoithms fo elxed vesion of the peviously defined SET K-COVER poblem [6]. To chieve highe enegy efficiency, sevel ecent woks focus on the ptil covege within fixed time dely. In [9], nodes coodinte to guntee the wost-cse detection dely nd to minimize the vege detection dely. Anothe type of tempol guntee is to nlyze the detection dely in the context of tcking. In [3], [4], [4], they ssume the netwok is ptilly coveed nd povide theoeticl nlysis nd simultion on the dely befoe tget is detected. Ou wok is unique in the following spects: (i) usense is unified chitectue insted of n individul solution. We e the fist to popose the concept of geneic switching. (ii) uscn demonsttes novel two-level globl scheduling method tht cn significntly educe enegy consumption. (iii) Ou set-cove is uniquely implemented t the second level, llowing the fistlevel optimiztion. Fo exmple, we demonstte optiml cove sets cn be obtined in line time when the covege e is continuous cuve. In contst, single-level set-cove ppoches [4], [3] could be inefficient unde non-unifom node distibution. VIII. CONCLUSION In this wok, we popose unified sensing chitectue clled usense. It fetues n symmetic design, which suppots vious kinds of covege lgoithms with simple geneic switching lgoithm in senso nodes. It llows us to flexibly chnge covege lgoithms with only two pmetes. Anothe mjo contibution of this wok is two-level globl scheduling lgoithm clled uscn, which is semlessly suppoted by the usense chitectue. In the fist level, we popose n optiml scheduling lgoithm in tems of minimizing e bech. In the second level, we popose line lgoithm to ddess the set-cove poblem when the lyout of tiles stisfies cetin conditions. We hve invested significnt effot to evlute ou design, which includes netwok of 3 MicZ motes, n extensive simultion with, nodes, s well s theoeticl nlysis. With thee novel ides: Asymmetic Achitectue, Geneic Switching nd Globl Scheduling, ou wok hs successfully chieved flexibility nd efficiency fo the senso netwok covege poblem. ACKNOWLEDGEMENT This wok ws suppoted in pt by NSF gnt CNS-6664 nd Genel Dynmics Resech Fellowship. REFERENCES [] A. Ao nd et l., A Wieless Senso Netwok fo Tget Detection, Clssifiction, nd Tcking, Compute Netwoks (Elsevie), 4. [] R. Szewczyk, A. Minwing, J. Andeson, nd D. Culle, An Anlysis of Lge Scle Hbit Monitoing Appliction, in SenSys 4, 4. [3] Z. Abms, A. Goel, nd S. Plotkin, Set K-Cove Algoithms fo Enegy Efficient Monitoing in Wieless Senso Netwoks, in IEEE IPSN, 4. [4] M. Cdei, M. T. Thi, Y. Li, nd W. Wu, Enegy-Efficient Tget Covege in Wieless Senso Netwoks, in IEEE INFOCOM, 5. [5] S. Kum, T. H. Li, nd J. Blogh., On k-covege in Mostly Sleeping Senso Netwok, in Mobicom, 4. [6] S. Slijepcevic nd M. Potkonjk, Powe Efficient Ogniztion of Wieless Senso Netwoks, in IEEE ICC,. [7] D. Tin nd N. Geogns, A Node Scheduling Scheme fo Enegy Consevtion in Lge Wieless Senso Netwoks, Wieless Communictions nd Mobile Computing Jounl, 3. [8] T. Yn, T. He, nd J. A. Stnkovic, Diffeentited Suveillnce Sevice fo Senso Netwoks, in SenSys 3, Novembe 3. [9] Q. Co, T. Abdelzhe, T. He, nd J. Stnkovic, Towds Optiml Sleep Scheduling in Senso Netwoks fo Re Event Detection, in IPSN 5, 5. [] C.-F. Chisseini nd M. Getto, Modeling the pefomnce of wieless senso netwoks, in IEEE INFOCOM, 4. [] C.-F. Hsin nd M. Liu, Netwok Covege Using Low Duty-Cycle Sensos: Rndom & Coodinted Sleep Algoithms, in IPSN 4, 4. [] X. Wng, G. Xing, Y. Zhng, C. Lu, R. Pless, nd C. Gill, Integted Covege nd Connectivity Configution in Wieless Senso Netwoks, in Sensys 3, Novembe 3. [3] C. Gui nd P. Mohpt, Powe Consevtion nd Qulity of Suveillnce in Tget Tcking Senso Netwoks, in MobiCom 4, 4. [4] S. Ren, Q. Li, H. Wng, X. Chen, nd X. Zhng, Anlyzing Object Tcking Qulity unde Pobbilistic Covege in Senso Netwoks, ACM MCR, vol. 9, no., 5. [5] O. Gnwli, B. Geenstein, K.-Y. Jng, A. Joki, J. Pek, M. Viei, D. Estin, R. Govindn, nd E. Kohle, The tenet chitectue fo tieed senso netwoks, in SENSYS, 6. [6] J. Hui nd D. Culle, The Dynmic Behvio of Dt Dissemintion Potocol fo Netwok Pogmming t Scle, in SenSys, Novembe 4. [7] W. R. Heinzelmn, J. Kulik, nd H. Blkishnn, Adptive Potocols fo Infomtion Dissemintion in Wieless Senso Netwoks, in Poc. of MOBICOM, August 999. [8] C. Y. Wn, A. T. Cmpbell, nd L. Kishnmuthy, PSFQ: A Relible Tnspot Potocol fo Wieless Senso Netwoks, in WSNA,. [9] V. T. Pschos, A Suvey of Appoximtely Optiml Solutions to Some Coveing nd Pcking Poblems, in ACM Computing Suveys, June 997. [] Kolmogoov-sminov test., E.W. Weisstein. MthWold, 4, [] R. Willims, Geometicl Foundtion of Ntul Stuctue:A Souce Book of Design. Dove Publictions Inc, New Yok, 979. [] F. Ye, G. Zhong, S. Lu, nd L. Zhng, PEAS: A Robust Enegy Conseving Potocol fo Long-lived Senso Netwoks, in ICDCS, My 3. [3] M. Moti, B. Kusy, G. Simon, nd A. Ledeczi, The Flooding Time Synchoniztion Potocol, in SenSys 4, 4. [4] C. Gui nd P. Mohpt, Vitul ptol: new powe consevtion design fo suveillnce using senso netwoks, in IPSN, 5. [5] S. Kum, T. H. Li, nd A. Ao, Bie Covege With Wieless Sensos, in MobiCom 5, 5.