RT-Link: A Time-Synchronized Link Protocol for Energy Constrained Multi-hop Wireless Networks

RT-Link: A Time-Synchronized Link Protocol for Energy Constrined Multi-hop Wireless Networks Anthony Rowe, Rhul Mnghrm, Rj Rjkumr Electricl nd Computer Engineering Deprtment, Crnegie Mellon University, Pittsurgh Technicl Report CMU-ECE-TR5-8 Astrct Multi-hop wireless networks of emedded nodes fcilitte pplictions in industril control, surveillnce nd inventory trcking. Our focus is on low-cost lrge-scle deployments where nodes need to e ttery-powered with predictle network lifetimes nd pplictions require ounded end-toend dely. An effective pproch to such energy-constrined networks is to operte t low duty cycles nd mximize the shutdown intervl etween pcket exchnges. The primry chllenge is in coordinting trnsmissions so they re collision-free while minimizing the durtion the nodes re ctive. RT-Link is time-synchronized link protocol for fixed nd moile emedded rdios. We identify three key oservtions in the design nd deployment of RT-Link: () RT-Link offers predictle network lifetime with ounded end-to-end dely. () Achieving glol time synchroniztion is oth economicl nd convenient for indoor nd outdoor deployments. (c) Due to interference etween nodes, our experiments confirm tht nodes with the sme schedule must e spced y minimum of 3 hops. Furthermore, to minimize end-to-end dely, it is more importnt to order time slots thn to minimize the numer of time slots. RT-Link hs een deployed on networks with more thn 3 custom emedded nodes nd uses the IEEE 8.4.5 physicl lyer. It outperforms energyefficient protocols such s B-MAC nd S-MAC in throughput, energy consumption nd end-to-end dely.. Introduction Networks of emedded wireless nodes provide verstile pltform for pplictions in industril control, surveillnce nd inventory trcking. The purpose of such networks is to gther dt nd deliver it cross one or more hops to t lest one gtewy. The principl requirements re low-cost ttery powered rdios, miniml configurtion on set up with simple nd sclle energy-efficient protocols for predictle network lifetime nd ounded end-to-end messge dely. The following deployment scenrios motivte the need for networks of emedded nodes with such requirements: Industril Control Networks: In chemicl nd utomoile plnts, remote control of mchinery nd ccess to performnce dt requires relile rel-time communiction. In such environments, it is necessry tht nodes do not require infrstructure for dt nd power s such provisioning my e oth imprcticl nd expensive. Surveillnce nd Monitoring Networks: Networks of emedded cmers for monitoring motion nd intrusion require ounded end-to-end dely to the gtewy nd deterministic pek throughput for intermittent trnsfer of cptured imges. Inventory Trcking nd Reporting: Networks to clssify nd locte ssets need to e sclle nd must operte in vriety of multi-hop wireless topologies. An effective pproch to energy-efficient service for pplictions with periodic nd periodic flows is to operte ll nodes t low duty cycles so s to mximize the shutdown intervls etween pcket exchnges. The two fundmentl chllenges in delivering dely-ounded service in such networks re () coordinting trnsmissions so tht ll ctive nodes communicte in tightly synchronized mnner nd () ensuring ll trnsmissions re collision-free. Time synchroniztion is importnt ecuse it tightly pcks the ctivity of ll nodes so tht they my mximize common sleep intervl etween ctivities. Furthermore, it provides gurntees on timeliness, throughput nd network lifetime for end-to-end communiction. Such ssurnces re only possile when the link is relile nd collision-free. It is therefore the responsiility of the link lyer protocol to provide exclusive nd interferencefree ccess to the shred wireless chnnel nd mechnism to coordinte sleep intervls of ll nodes. The focus of this pper is on RT-Link, time-synchronized link lyer protocol for collision-free nd energy-efficient rel-time service over multi-hop wireless networks. RT-Link fcilittes dynmic dmission of oth fixed nd moile nodes into tightly synchronized regime. It schedules nodes in time slots such tht concurrent trnsmitters do not interfere with ech other nd the ctivity of ll nodes re coordinted to mximize the sleep durtion. Finlly, RT-Link mintins contention-free opertion y employing n online nd utomtic link conflict detection nd resolution scheme. Such scheme is useful

when topology or environment chnges cuse interference to nodes etween concurrent trnsmitters. RT-Link hs een implemented s link lyer protocol in low-cost nd lowpower emedded nodes developed y us. Ech node includes short-rnge.4hz IEEE 8.5.4 [] physicl lyer for rdio communiction. Through the design nd deployment of RT-Link, we identify the following four oservtions:. RT-Link offers predictle network lifetime with ounded end-to-end dely for pckets etween the gtewy nd every node.. Provision of glol time synchroniztion for emedded multi-hop wireless networks is oth economicl nd convenient. We chieve this y employing n Amplitude Modultion (AM) sed crrier-current method indoors nd with tomic clock receivers for the outdoors. 3. Our experiments show tht due to interference cross the shred chnnel, nodes with the sme schedule (i.e. concurrent trnsmitters) must e spced y minimum 3-hop distnce. 4. In high throughput networks, the scheduling ojective is to mximize the numer of concurrent trnsmitters []. In contrst, in energy-efficient sensor networks, the ordering of the time slots is more importnt thn the numer of time slots. The pper is orgnized s follows: we ddress relted work in the next section followed y description of the RT-Link protocol in Section 3. We study the timeliness, roustness nd efficiency of the protocol in Section 4. Section 5 provides n overview of our implementtion pltform nd deployment experiences. This is followed y comprtive evlution of RT-Link in Section 6 nd our concluding remrks.. Relted Work Severl MAC protocols hve een proposed for lowpower nd distriuted opertion for single nd multi-hop wireless mesh networks. Such protocols my e ctegorized y their use of time synchroniztion s synchronous, loosely synchronous nd fully synchronized protocols. In generl, with greter degree of synchroniztion etween nodes, pcket delivery is more energy-efficient due to the minimiztion of idle listening when there is no communiction, etter collision voidnce nd elimintion of overhering of neighor converstions. We riefly review key lowpower link protocols sed on their support for low-power listen, multi-hop opertion with hidden terminl voidnce, sclility with node degree nd offered lod... Asynchronous Link Protocols The Berkeley MAC (B-MAC) [3] protocol performs the est in terms of energy conservtion nd simplicity in design. B-MAC supports Low Power Listening (LPL) where ech node periodiclly wkes up fter smple intervl nd checks the chnnel for ctivity for short durtion of.5ms. If the chnnel is found to e ctive, the node stys wke to receive the pylod following n extended premle. Using this scheme, nodes my efficiently check for neighor ctivity. The mjor drwck of B-MAC is tht the trnsmitter must remin ctive for the durtion of the smpling intervl for ech sent pcket. For exmple, if receiver nodes periodiclly wke up every 8ms, then trnsmitter would need to continuously trnsmit for 8ms to e detected y neighor. This coupling of the receiver s smpling intervl nd the durtion of the trnsmitter s premle severely restricts the sclility of B-MAC when operting in dense networks nd cross multiple hops. B-MAC does not inherently support collision voidnce due to the hidden terminl prolem nd the use of RTS-CTS hndshking is expensive nd inefficient. In multi-hop network, it is necessry to use topology-wre pcket scheduling for collision voidnce. Furthermore, upon wke up, B-MAC employs Crrier Sense Multiple Access (CSMA) nd is prone to wsting energy nd dding non-deterministic ltency due to pcket collisions... Loosely Synchronous Link Protocols Protocols such s T-MAC [4], WiseMAC [5], nd S-MAC [6] employ locl sleep-wke schedules etween node pirs to coordinte pcket exchnges while reducing idle opertion. All three schemes exchnge synchronizing pckets to inform their neighors of the intervl until their next ctivity nd use CSMA prior to trnsmissions. Both T-MAC nd WiseMAC use LPL to minimize energy consumption during chnnel smpling. WiseMAC, however, is designed for point-to-multipoint communiction nd does not cter to multi-hop networks. S-MAC is similr to ut simpler thn T-MAC, ut does not implement LPL. Both schemes do not scle well ecuse ll the neighors of node cnnot her ech other nd this forces the node to set multiple wke up schedules for different groups of neighors. Furthermore, the use of CSMA degrdes performnce severely with incresing node degree nd trffic..3. Fully Synchronized Link Protocols With the provision of glol time synchroniztion, TDMA protocols such s TRAMA [7] re le to communicte etween node pirs in dedicted time slots. TRAMA supports oth scheduled slots nd CSMA-sed contention slots for node dmission nd network mngement. RT- Link hs similr support for contention slots ut employs Slotted-ALOHA [8] rther thn CSMA s it is more energy efficient with LPL. While TRAMA outlines schedule exchnge protocol, it does not explicitly specify node scheduling scheme. The uthors do not ddress n energyefficient nd prcticl time synchroniztion scheme. RT- Link hs een inspired y systems such s [9, ] which used dul-rdio solutions for low power wke-up. However neither system hs een used for time synchronized opertion. RT-Link employs tight glol time synchroniztion to

estlish common wke-up instnce nd durtion for ll nodes. The procedure fter wke up is time synchronized into scheduled trnsmission slots nd therey elimintes ny collisions in the network. To the est of our knowledge, RT-Link presents the first deployment of glolly synchronized low-power sensor networks with energy-efficient nd economicl time synchroniztion 3. RT-Link Protocol Design RT-Link is Time Division Multiple Access (TDMA) sed link lyer protocol designed for networks tht require predictility in throughput, ltency nd energy consumption. All pcket exchnges occur in well-defined time slots. lol time synchroniztion is provided to ll fixed nodes y roust nd low-cost out-of-nd chnnel. We descrie in detil the RT-Link protocol nd its opertion modes. 3.. Protocol Overview Ech fixed (i.e. sttionry) node hs two rdios which operte on seprte chnnels. A low-power receiver (e.g. AM, FM, tomic clock) to detect periodic synchroniztion pulse nd n RF trnsceiver for dt communiction fter the sync pulse is received. As shown in Figure, periodic synchroniztion pulse is detected first nd is followed y finely slotted dt communiction period. This period is defined s frme nd the intervl etween time sync pulses is defined s cycle. The sync pulse serves s n indictor of the eginning of the first frme. After frme is complete, ech node schedules timer to wke up just efore the expected time of the next sync pulse nd promptly switches to sleep mode. One or more frmes my e scheduled within cycle. Ech frme is divided into two regimes of time synchronized opertion: () series of Scheduled Slots (SS) within which nodes re ssigned specific trnsmit nd receive time slots nd () series of Unscheduled or Contention Slots (CS) where nodes, which re not ssigned slots in the SS, select trnsmit slot t rndom. Nodes operting in SS re provided timeliness gurntees s they re grnted exclusive ccess of the shred chnnel nd hence enjoy the privilege of interference-free nd hence collision-free communiction. Fixed nodes tht hve not yet een ssigned time slot in SS continue to operte in CS nd re suject to finite proility of pcket collision. We ssume ll nodes re wre of the fixed numer of slots within the SS nd CS. The methods for ssigning collision-free time slots re descried in Section 4. 3.. Node Types nd Pcket Types RT-Link supports oth fixed nd moile nodes. Only fixed nodes hve sync pulse receiver in ddition to the RF dt trnsceiver nd re le to operte within SS. Fixed nodes mintin time synchroniztion y the glol sync pulse nd my e ssigned specific trnsmit nd receive time slots in the intervl following the sync pulse. On the other hnd, moile nodes re not ssigned specific scheduled time slots s their neighors my chnge frequently nd therefore operte Sync Pulse Frme Scheduled Slots Time-sync Cycle Contention Slots Figure. RT-Link time slot lloction with out-of-nd synchroniztion pulses solely in CS. Moile nodes otin time synchroniztion y listening to neighors operting in SS. Upon detecting one such pcket, the moile node is informed of the current slot numer nd determines the strt of the CS. It then rndomly selects trnsmit slot in the CS. Only nodes operting in the SS re permitted to route dt to nd from the gtewy. Moile nodes, typiclly rodcst dt to the gtewy vi one or more fixed nodes. We now descrie the six pcket types shown in Figure required for sic network opertion.. RT-Link Pcket Heder Every pcket includes the common link lyer heder. The common heder, s shown in Figure (), contins 6-it Medium Access Control (MAC) ddresses of the source, destintion nd current forwrding node. As fixed nodes operte either in the SS or CS, ech pcket is tgged explicitly with the trnsmit slot numer. For exmple, in our network deployment, cycle contins 3 slots including 4 (i.e. slots to 3) slots within the SS nd 8 slots within the CS. As gurnteed service is provided only to nodes ssigned explicit slots within the SS, only these nodes re entitled to receive n cknowledgement. The receiver operting within SS implicitly cknowledges ll pckets received y setting it msk mpping the slots within which the pckets were received. This provides n efficient mechnism to identify nd cknowledge trnsmitters operting within the SS. Moile nodes nd nodes operting within the CS, only trnsmit rodcst pckets destined for the gtewy nd re hence never cknowledged.. HELLO Pcket The HELLO rodcst pcket, s shown in Figure (), dvertises node s list of neighors. This pcket serves s keep-live pcket to inform neighors of one s continued presence nd lso informs the gtewy of the network topology. A node is considered neighor s long s t lest one HELLO messge is received from it within the pst k (e.g. 5) cycles. 3. SCHEDULE Pcket When the gtewy receives multiple HELLO pckets from node nd is stisfied out the stility of 3

) Pcket Heder Mgic ) Hello Pcket c) Schedule Pcket Heder d) Dt Pcket Src MAC DST MAC Lst Hop Slot # 4 Implicit ACK Msk TTL Retry # Pkts Priority Heder HELLO SCHED # Segs Offset 4 TX Msk RX Msk TX MAC Chn... TXn MAC Chn Heder 3- DATA e) Route Pcket f) Error Pcket Heder ROUTE # Segs Trget MAC Heder ERROR # Segs Size Dst MAC Action Src MAC 4 Neighor MAC Appliction Defined Dt Next Hop... Type... set routes. We only provide primitive support for source routing s the focus is on link communiction. The ER- ROR pcket (Figure (f)) is sent to neighors when conflict, due to overlpping time slots mong neighors, is detected. A node experiencing the conflict informs its neighors nd the gtewy of the error nd my choose to relinquish its slots. A node relinquishing slot needs to restrt the ssocition procedure (descried elow) to request one or more scheduled slots from the gtewy. The ERROR pcket informs the gtewy of scheduling conflict. Ech pcket is ssigned priority, which my e sed on the lst-hop ddress or pcket type. When more pckets hve ccumulted t node thn its ville uffer spce, pckets re dropped in the scending order of their priority. Pcket priorities re useful in ggregtion nd the highest priority mong the ggregted pckets is ssigned to the entire forwrded pcket. 3.3. Network Opertion Procedures iven the generl slot structure, we now descrie the rules node follows upon ssocition, scheduled opertion, during slot ssignment conflicts nd disssocition. Figure. Six pcket types required for sic opertion. node s neighorhood, it sends unicst SCHEDULE pcket Figure (c) to the node with its unique slot numers. A node my e ssigned one or more slots depending on higher-level ttriute such s ppliction ndwidth requirements. The gtewy executes node coloring lgorithm to determine the schedule for ll known nodes in the network such tht concurrently ctive nodes do not interfere with ech other. The node scheduling lgorithm is network lyer service nd is out of the scope of this pper. However, in section 4 nd 5 we provide insights nd descrie sic lgorithm for lef to gtewy tree dt collection. A node s schedule is descried y trnsmit nd receive it-msks. A node is ctive in only those slots set in ech msk. The rdio is turned off in ll other slots. This enles node to void listening to neighoring nodes through which no common flow is routed. In our implementtion, we use 3-it msks for trnsmit nd receive schedules. 4. DATA, ROUTE nd ERROR Pckets The DATA pcket (Figure (d)) cn contin mximum of ytes of dt s the mximum size of n IEEE 8.5.4 pcket is 8 ytes. DATA pckets support pcket ggregtion nd forwrd ggregted dt with the sender s ddress. The ROUTE pcket Figure (e) is sent to node vi unicst from the gtewy to explicitly 3.3.. Network Assocition nd Disssocition RT-Link opertes on simple 3-stte stte mchine s shown in Figure 3. In generl, nodes operting in the CS re considered uests, while nodes with scheduled slots re considered Memers of the network. A fixed node tht is currently uest ecomes Memer once it is ssigned one or more slots in the SS. On the other hnd, moile nodes re never ssigned scheduled slot nd re considered uests for the lifetime of their opertion. When fixed node is powered on, it is first initilized s uest nd opertes in the CS. It initilly keeps its sync rdio receiver on until it receives sync pulse. Following this, it wits for set numer of slots (spnning the SS) nd then rndomly selects slot mong the CS to send HELLO messge with its node ID. This messge is then forwrded vi flooding to the gtewy nd the node is eventully scheduled slot in the SS. In ootstrpping mnner, uest nodes closer to the gtewy should e scheduled first in order to llow fster delivery of scheduling pckets further wy from the gtewy. During norml opertion, Memers nd uest wke up efore the expected instnce of the sync pulse nd operte in the SS nd CS. On the other hnd, when moile node needs to trnsmit, it first stys on until it overhers neighor operte in its SS. The moile node chieves synchroniztion y oserving the Memer s slot numer nd computes the time until the strt of the CS. It then rndomly selects slot in the CS nd trnsmits. The moile node remins silent until Memer is identified. During norml opertion, ll nodes trnsmit on their ssigned slots within the SS, nd turn on their receiver during the receive slots from ech neighor. 4

Moile Unsynchronized Contention Mode x Synchronize off of overherd pckets ot Scheduling Pcket uest Synchronized Contention Mode Memer Synchronized Scheduled Mode Figure 3. RT-Link node stte mchine. ) ) c) Figure 4. Scheduling Conflict Cses. Node is trying to communicte with node, while node x shres s timeslot. The dotted line indictes n interfering link tht cnnot send vlid dt. All nodes with scheduled slots listen on every slot in the CS. As descried in Section 5, ll nodes use low power listen mechnism to quickly nd economiclly detect if there is ctivity in the eginning of slot nd turn the receiver off if there is none. When node chooses to leve the network, it ceses rodcsting HELLO pckets nd is grcefully evicted from the neighor list from ech of its neighors. The gtewy eventully detects the sence of the deprted node from ech of the neighors HELLO updtes nd my reschedule the network if necessry. x Scheduling Conflict x 3.3.. Conflict Detection nd Resolution While ll nodes re expected to operte in tightly synchronized mnner, there re severl cses when the slot ssignment results in conflicts. A synchroniztion conflict occurs when two neighors within communiction or interference rnge re ssigned overlpping or prtilly overlpping time slots. Such cses my occur due to topology chnges, some nodes experiencing extremely lrge clock drift or jitter, nd incorrect slot ssignment y the gtewy, mong other resons. Figure 4 illustrtes three potentil conflict cses. In Figure 4(), we oserve tht oth node A nd X re ssigned the sme slot resulting in collisions t node B. In this cse, X is lso within the communiction rnge of A. In Figure 4(), we see similr cse, except X is hidden from A. Finlly, in Figure 4(c) while B is within A s communiction rnge, it is lso within X s interference rnge. B is unle to decode A s pckets s X s interference rises the noise floor. We present two mechnisms, Active Listener nd Active Trnsmitter, to detect nd correct the ove timing errors. The Active Listener instructs nodes to ctively listen on their trnsmit slot when they do not hve ny dt to send. This wy, node is le to overher potentil interferer nd detect the conflict. For exmple, in the cse of Figure 4(), if node A were quiet in its trnsmission slot, it would overher node X s trnsmission. The Active Trnsmitter pproch enles oth conflict detection nd resolution. During every kth cycle, e.g. k=5, scheduled node rodcsts pcket with its neighor list nd slot ssignment of ech neighor from slot within CS. This informs ll neighors of timing conflicts with potentil hidden terminls. For exmple, in the cse of Figure 4(), if B rodcsts its node list, then it would list oth A nd X s owners of the sme receive slot. Either A or X my e nominted to relinquish their slot nd re-ssocite with new slot numer. In the finl cse, s B is le to only decode A s pckets ut suffers high pcket loss due to interference from X, it requests A to relinquish its slot nd re-ssocite y sending HELLO pcket to the gtewy with its previous slot s foridden slot. 4. RT-Link Protocol Enhncements In this section we riefly discuss enhncements tht compliment the sic RT-Link protocol. These my e executed prior to or during network deployment to improve the overll throughput, end-to-end ltency nd network lifetime. 4.. Topology-sed Predictle Lifetime We hve developed hyrid network simultor to predict network lifetime for given topology nd offered ppliction trffic. The network topology my e generted y the simultor or e cquired from deployed network. In order to cquire the network connectivity grph, we deploy n ctul network nd ggregte the neighor lists from ech node t the gtewy. We then construct connectivity nd interference grphs nd schedule nodes sed on k-hop coloring, such tht two nodes with the sme slot schedule re mutully seprted y t lest k+ hops. The time-triggered simultor then simultes given trffic worklod during ech cycle nd clcultes the worst -cse network lifetime. The network lifetime is defined s the time until the first node completely consumes the ville ttery energy nd disconnects the grph (i.e. ottleneck node). Through this exercise, we re le to nswer questions such s, "Wht is the expected lifetime for the given topology?" nd lterntively, "Wht is the est topology for the given ppliction demnds?" we illustrte this point y clculting the expected lifetime for three different network topologies. In Figure 5() we show the connectivity grph of rndomly generted topology with nodes. The grph ws colored sed on the connectivity to ensure tht it is free of collisions. Links cn then e removed y instructing n djcent node to no longer wkeup to listen on tht prticulr timeslot. Using this principle, we generte spnning tree strting from the gtewy shown in Figure 5() tht gurntees full connectivity. One could either mesure network lifetime s the time efore the first node runs out of power, 5

connect spnning d3 () Network Connectivity rph () Spnning Tree Cover of Network (c) Network with Blnced Degree of 3 Figure 5. Network rph Constructed For Energy Efficiency or the time t which the network connectivity flls elow specific threshold. In either cse, dding higher degree of redundncy without incresing the mximum degree of the network will increse totl lifetime. Figure 5(c) shows the spnning tree topology with links dded ck into the grph such tht whenever possile nodes hve degree equl to the mximium degree of the originl spnning tree. The originl network hs simulted lifetime of. yers, while the prunned networks do not experience node loss for. yers. We oserve tht well-lnced networks with uniform node degree outperform networks with symmetric or highly dense su grphs. Well-lnced grphs efficiently exploit sptil reuse nd evenly spred the network lod. 4.. Interference-free Node Scheduling In order to chieve high throughput in multi-hop wireless network, it is necessry to minimize the numer of collisions long ech trnsmission hop. This prolem hs trditionlly een solved s distnce-k node coloring (slot scheduling). To determine the interference rnge of node, we plced set of nodes long line in n open field nd first mesured the pcket loss etween trnsmitter nd receiver s the trnsmitter s distnce ws vried. Once the stle communiction distnce etween trnsmitter nd receiver ws determined, we evluted the effect of constntly trnsmitting node (i.e. jmmer) on the receiver. Our experimentl results for stle trnsmit power level 8 re shown in Figure 6. We notice tht % or the pckets re received up to trnsmitter-receiver distnce of m. Following this, we plced the trnsmitter t distnce of, 4, 6, 8, nd meters nd for ech trnsmitter position, jmmer ws plced t vrious distnces. At ech point, the trnsmitter sent one pcket every cycle to the receiver for cycles. We mesure the impct of the jmmer y oserving the percentge of successfully received pckets. We oserve two effects of the jmmer: First, the effect of the jmmer is lrgely function of the distnce of the jmmer from the receiver nd not of the trnsmitter from the receiver. Between -8 meters, Pcket Percentge.9.8.7.6.5.4.3.. No Jmmer RX TX m RX TX 4m RX TX 6m RX TX 8m RX TX m 5 5 5 3 Distnce (meters) Figure 6. Pcket Success rte while trnsmitting in collision domin. the impct of the jmmer is similr cross ll trnsmitter distnces. Second, when the trnsmitter nd jmmer re close to the receiver, (i.e. under 9m), the trnsmitter demonstrtes cpture effect nd mintins n pproximtely % pcket reception rte. The ove results show tht the jmmer hs no impct eyond twice the stle reception distnce (i.e. m) nd concurrent trnsmitter my e plced t thrice the stle reception distnce (i.e. 3m). Such prmeters re incorported y the node coloring lgorithm in the gtewy to determine collision-free slot schedules. Results for scheduled multi-hop network re presented in Section 6. 4.3. Coloring nd Ordering In multi-hop wireless networks, the gol for higher throughput hs trditionlly een pproched from the perspective of mximizing the set of concurrent trnsmitters in the network []. This is chieved either y scheduling nodes or links such tht they operte without ny collisions. In the networks considered here, the pplictions generte stedy 6

or low dt rte flows ut require low end-to-end dely. In Figure 7 we see two schedules, one with the miniml numer of timeslots, the other contining extr slots ut provisioned such tht lef nodes deliver dt to the gtewy in single TDMA cycle. The miniml timeslot schedule mximizes concurrent trnsmissions, ut does not minimize the ltency of ll nodes. Insted, the mximl concurrency schedule will cuse quequeing delys tht will hurt overll network performnce given ll nodes eqully contriute trffic. In Figure 8, we illustrte set of nodes communicting to nd from gtewy. By ssigning the time slots lierlly nd in preference to fster uplink nd downlink routes, we show tht for networks with dely-sensitive dt, ordering of slots should tke priority over mximizing sptil reuse. As nodes sleep etween slots not ssigned in their trnsmit or receive it-msks, the energy sving in using fewer slots s compred to ll ville slots in cycle is nominl. The est strtegy is to use s mny slots required to minimize the end-to-end dely long oth the uplink nd downlink. The genertion of schedules is similr to the distnce-two grph coloring prolem tht is known to e NP-complete []. In prctice, mny heuristics cn work well given tht devitions from the optiml color selection re typiclly overcome y excess time slots. Though not the focus of this pper, we will discuss one such heuristic tht provides sic scheduling ility for networks where the trffic consists of smll pckets eing routed up tree to single gtewy. The heuristic consists of four steps. The first step uilds spnning tree over the network rooted t the gtewy. Using Dijkstr s shortest pth lgorithm ny connected grph cn e converted into spnning tree. As cn e seen in Figure 9(), the spnning tree must mintin "hidden" links tht re not used when iterting through the tree, ut cn e used to mintin tht the two hop constrint is still stisfied in the originl grph. Once spnning tree is constructed, redth first serch is performed strting from root of the tree. The heuristic egins with n initilly empty set of colors. As ech node is trversed y the redth first serch, it is ssigned the lowest vlue in the color set tht is unique from ny single or two hop neighors. If there re no free colors, new color must e dded into the current set. The next step in the heuristic tries to eliminte redundnt slots tht lie deeper in the tree y replcing them with lrger vlued slots. As will ecome pprent in the next step, this mnipultion llows dt from the leves of the tree to move s fr s possile towrds the gtewy in single TDMA cycle. Figure 9(c) shows how the previous three nodes re given lrger vlues in order to minimize pcket ltencies. The finl step in the heuristic inverts ll of the slot ssignments such tht lower slot vlues re towrds the edge of the tree llowing informtion to e propgted nd ggregted in cscding mnner towrds the gtewy. c d 3 e f g h g c d e h f 6 5 3 c d 4 e f g h g e c d h f Figure 7. Mximl concurrency schedule (left) compred to dely sensitive schedule (right). Note tht the mximl concurrency schedule needs two frmes to deliver ll dt. Even if the schedule is duplicted, it will still require two extr cycles compred to the dely sensitive schedule. 3 c Left grph shows mximl concurrency tht needs two frmes to deliver ll dt. Even if you duplicte the schedule, you will require extr cycles compred to the left grph. d 3 c d 5 3 4 4 5 3 c d c 6 7 d c d c d c d c d Figure 8. A set of nodes communicting with different schedules. 4 3 5 4 3 5 6 ) ) c) d) 4 3 Figure 9. ) shows mesh network topology. ) shows how this topology cn e turned into spnning tree nd colored using Dijkstr s lgorithm nd greedy redth first serch heuristic. c) Repeted colors tht lie further down the tree re mde lrger so tht dt cn flow in single cycle towrds the gtewy. d) All slots re inverted such tht erlier slots cn cscde up the tree towrds the gtewy. 6 5 7

Figure. FireFly nd FireFly Jr ord with AM synchroniztion module 5. RT-Link Implementtion In the following section, we descrie our hrdwre pltform s well s two different hrdwre-ided out-of-nd time synchroniztion solutions. First, we introduce Fire- Fly, custom 8.5.4 wireless sensor node. Following this, we descrie n dd-on ord for receiving the tomic clock rodcst for outdoor synchroniztion nd ord for receiving n AM rodcst synchroniztion pulse for indoors. We then evlute the timing nd energy impct of our synchroniztion hrdwre on the MAC protocol. 5.. Hrdwre Figure shows our custom sensor node, FireFly. The ord uses n Atmel Atmeg3L [] 8-it microcontroller nd Chipcon CC4 [3] IEEE 8.5.4 wireless trnsceiver. The microcontroller opertes t 8Mhz nd hs 3KB of ROM nd KB of RAM. The FireFly ord includes light, temperture, udio, dul-xis ccelertion nd pssive infrred motion sensors. We hve lso developed lower-cost version of the ord clled the FireFly Jr. tht does not include sensors, nd is used to forwrd pckets in the network. The FireFly ord interfces with computer using n externl USB dongle. 5.. Time Synchroniztion In order to chieve the highly ccurte time synchroniztion required for TDMA t pcket level grnulrity, we use two out-of-nd time synchroniztion sources. One uses the WWVB tomic clock rodcst, nd the other relies on crrier-current AM trnsmitter. In generl, the synchroniztion device should e low power, inexpensive, nd consist of simple receiver. The time synchroniztion trnsmitter must e cple of covering lrge re. 5... Implementtion The WWVB tomic rodcst is pulse width modulted signl with it strting ech second. Our system uses n off-the-shelf WWVB receiver (Figure ) to detect these rising edges, nd does not need to decode the entire time string. When ctive, the ord drws.6ma t 3 volts nd requires Figure. Left to Right: WWVB tomic clock receiver, AM receiver nd USB interfce ord. less thn 5uA when powered off. Inside uildings, tomic clock nd PS receivers re typiclly unle to receive ny signl, so we use crrier-current AM rodcst. Crriercurrent uses uilding s power infrstructure s n ntenn to rdite the time synchroniztion pulse. We used n offthe-shelf low-power AM trnsmitter nd power coupler [4] tht dhere to the FCC Prt 5 regultions without requiring license. The trnsmitter provides time synchroniztion to two 5-storey cmpus uildings which operte on AC phses. Figure shows n dd-on AM receiver module cple of decoding our AM time sync pulse. We use commercil AM receiver module nd then designed custom supporting-ord which thresholds the demodulted signl to decode the pulse. The supporting AM ord is cple of controlling the power to the AM receiver. 5... Energy Consumption The energy required to ctivte the AM receiver module nd to receive pulse is equivlent to sending one nd hlf 8.5.4 pckets. The use of more dvnced single chip AM rdio [5] would ring these vlues lower nd llow for more compct design. We estimte tht using single chip AM rdio receiver, the synchroniztion energy cost would e less thn one tenth the energy of sending or receiving single in nd pcket. We lso investigted using sucrrier FM trnsmission from locl rdio sttion to trnsmit the synchroniztion pulse. Commercil FM rdio sttions re typiclly issued two sucrrier chnnels y the FCC for trnsmitting digitl informtion such s song nmes, wether nd trffic informtion. We hve not yet pursued such technology since it would mke control of the timing source more difficult during our erly development phse since the trnsmitter would e physiclly locted t rdio sttion. 5..3. Sclility nd Performnce In order to mintin sclility cross multiple uildings, our AM trnsmitter loclly rerodcsts the tomic clock time signl. The synchroniztion pulse for the AM trnsmitter is line-lnced 5us squre wve generted y modified 8

3..7 node node node node 3 node 4.3 Percentge of Sync Pulses 8.8 7.3 5.9 4.4.9.5 Figure 3. RT-Link opertion nd timing prmeters. 5 5 5 Time (microseconds) Figure. Distriutions of AM crrier current time synchroniztion jitter over 4 hour period. FireFly node cple of tomic clock synchroniztion. In order to evlute the effectiveness of the synchroniztion, we plced five nodes t different points inside five storey uilding. Ech node ws connected to dt collection ord using severl hundred feet of cles. The dt collection ord timed the difference etween when the synchroniztion pulse ws generted nd when ech node cknowledged the pulse. This test ws performed while the MAC protocol ws ctive in order to get n ccurte ide of the possile jitter including MAC relted processing overhed. Figure shows histogrm with the distriution of ech node s synchroniztion time jitter. An AM pulse ws sent once per second for 4 hours during norml opertion of clssroom uilding. The grph shows tht the jitter is ounded to within us. 99.6% of the synchroniztion pulses were correctly detected. We found tht with more refined tuning of the AM rdios, the jitter could e ounded to well within 5us. In order to mintin synchroniztion over n entire TDMA cycle durtion, it is necessry to mesure the drift ssocited with the clock crystl on the processor. We oserved tht the worst of our clocks ws drifting y us/s giving it drift rte of e-5. Our previous experiment illustrtes tht the jitter from AM rdio ws t worst us indicting tht the drift would not ecome prolem for t lest seconds. The drift due to the clock crystl ws lso reltively consistent, nd hence could e ccounted for in softwre y timing the difference etween synchroniztion pulses nd performing clock-rte djustment. In our finl implementtion with line-lnced input to the trnsmitter, we re le to mintin glolly synchroniztion to within us. 5.3. TDMA Slot Mechnics When node is first powered on, it ctivtes the AM receiver nd wits for the first synchroniztion pulse. Figure 3 shows the ctul timing ssocited with our TDMA frmes. Once the node detects pulse, it resets the TDMA frme counter mintined in the microcontroller which then powers down the AM receiver. When the node receives its synchroniztion pulse, it egins the ctive TDMA time cycle. After checking its receive nd trnsmit msks, the node determines which slots it should trnsmit nd receive on. During receive timeslot, the node immeditely turns on the receiver. The receiver will wit for pcket, or if no premle is detected it will time out. The received pcket is red from the CC4 chip into memory ddress tht ws llocted to tht prticulr slot. We employ zero-copy uffer scheme to move pckets from the receive to the trnsmit queue. In the cse of utomtic pcket ggregtion, the pylod informtion from pcket is explicitly copied to the end of the trnsmit uffer. When the node reches trnsmit timeslot, it must wit for gurd time to elpse efore sending dt. Accounting for the possiility tht the receiver hs drifted hed or ehind the trnsmitter, the trnsmitter hs gurd time efore sending nd the receiver premle-check hs gurd time extending eyond the expected pcket. Tle in the next section shows the different timeout vlues tht work well for our hrdwre configurtion. Once the timeslot is complete, there needs to e n dditionl gurd time efore the next slot. We provide this gurd time plus configurle inter-slot processing time tht llows the MAC to do the miniml processing required for inter-slot pcket forwrding. This feture is motivted y memory limittions nd reduction of network queue sizes. After the TDMA cycle hs completed, the interrupt hndler returns leving the flow of execution to continue fter the lst processor sleep cll. At this point the ppliction cn run sensing tsk, schedule pcket for trnsmission, or return to sleep until the next interrupt is clled. Figure 4 shows smple trce of two nodes communicting with ech other. The rpid receiver checks t the end of the cycle show the contention period with low-power listening for the durtion of premle. 5.4. Modeling Lifetime To clculte the node duty cycle nd lifetime we sum the node s energy consumptions over TDMA frme. Tle shows the power consumed y ech component ssuming 9

9 x 3 8 7 degree= degree= degree= degree=3 Averge Power (mw) 6 5 4 3 4 6 8 4 6 8 TDMA Frme Size (Seconds) Figure 5. Avg. power consumed y node with 3 time slots with respect to neighor degree nd TDMA frme size. Figure 4. Chnnels nd show trnsmit nd receiver ctivity for one node. Chnnels 3 nd 4 show rdio ctivity for second node tht receives pcket from the first node nd trnsmits response few slots lter. The smll pulses represent RX checks tht timed out. Longer pulses show pckets of dt eing trnsmitted. The group of pulses towrds the right side show the contention slots. opertion t 3 volts. Tle nd Tle 3 show the timing prmeters nd the energy of ech opertion during the TDMA frme. The ctive time of ech TDMA slot, T ctive, is dependent on the totl numer of slots, N slots, the mximum slots trnsmit time T mx_pylod, the AM synchroniztion setup T sync_setup nd cpture T sync s well s inter slot processing time T ISS. T ctive = T sync_setup + T sync + N slots (T mx_pylod + T ISS) () The idle time, T idle, etween slots is the difference etween the ctive time nd the totl frme time, T frme. This is typiclly customized for the specific ppliction since it hs impct on oth ttery life nd ltency. T idle = T frme T ctive () The three customizle prmeters tht define the lifetime of node re the TDMA frme time, the numer of TDMA slots (including the numer of contention slots Ncontention) nd the degree d of the node. As the degree increses, the node must check the strt of dditionl time slots nd my potentilly hve to receive pckets from its neighors. The minimum energy tht the node will require during single TDMA frme E min is the sum of the different possile energy consumers ssuming no pckets re received nd the node does not trnsmit pckets: E min = E sync + (d + N contention ) E RX + E CP U_ctive +E CP U_sleep + E rdio_idle + E rdio_sleep (3) Power Prmeters Symol I(m) Power(mW) Rdio Trnsmitter P rdio_t X 7.4 5. Rdio Receiver P rdio_rx 9.7 59. Rdio Idle P rdio_idle.46.8 Rdio Sleep P rdio_sleep e 3 3e 3 CPU Active P CP U_ctive. 3.3 CPU Sleep P CP U_sleep e 3 3e 3 AM Sync Active P sync 5 5 Tle. Power Consumption of the min components. The mximum energy the node cn consume during single TDMA frme is the miniml energy consumed during tht frme summed with the possile rdio trnsmissions tht cn occur during TDMA frme. E mx = E min+(d+n contention) E RX+N T X_slots E T X (4) The mximum power consumed y node over TDMA frme cn e computed s follows: P vg = E mx/t frme (5) The lifetime of the node cn e computed s follows: Lifetime = (E cpcity/e mx) T F rme (6) Figure 5 shows the verge power of node with respect to TDMA frme size with different neighor degrees using 3 time slots, 8 of which re used during the contention segment of the protocol. As the node degree increses, in order to mintin the sme verge power nd hence lifetime, the TDMA frme size must e incresed. 6. Performnce Evlution In this section, we compre multi-hop performnce of RT- Link with tht of S-MAC nd B-MAC. We first vlidted our implementtion of RT-Link in -node test-ed. Following this, we use simultion to compre ltency nd throughput.

Timing Prmeters Symol Time (ms) Mx Pcket Trnsfer T mx_pylod 4 Sync Pulse Jitter T sync e 3 Sync Pulse Setup T sync_setup + (ρ T frme ) RX Timeout T RX 3e 3 TX urd Time T T X e 3 Inter Slot Spcing T ISS 5e 3 Clock Drift Rte ρ e s/s Tle. Timing Prmeters for min components. Energy Prmeters Symol Energy (mw) Synchroniztion E sync P sync (T sync + T sync_setup) Active CPU E CP U_ctive P CP U_ctive T ctive Sleep CPU E CP U_sleep P CP U_sleep T idle TX Rdio E rdio_tx P rdio_tx (T mx_pylod + T T X) RX Rdio E rdio_rx P rdio_rx T mx_pylod Idle Rdio E rdio_idle P rdio_idle T ctive Sleep Rdio E rdio_sleep P rdio_sleep T idle RX Rdio Check E RX P rdio_rx T RX Tle 3. Energy of components with respect to power nd time. 6.. Multi-hop Network Performnce In order to determine the sptil seprtion etween nodes for interference-free communiction, we plced ten nodes in line nd fixed the power so tht ech node could only relily communicte with its direct neighors. The nodes generted 5 yte pcket of dt ech second for seconds. Ech node trnsmitted dt only to the neighor closer to the gtewy. In the first test, ll nodes were ssigned unique time-slots s shown in Figure 6() so tht there were no concurrent trnsmitters. We repeted this test three times nd oserved tht every node received % of its neighor s pckets. This provided snity check tht the 5us inter-slot processing time provided the necessry temporl seprtion for oth the synchroniztion jitter nd the pcket processing (i.e. pcket reception nd ggregtion) etween slots. We lso oserved t the gtewy tht the end-to-end dely for node s pckets ws consistently under 5ms. This confirmed our expecttion of ggregting nd forwrding pckets on slot-y-slot sis. In the second test, s shown in Figure 6), two time slots (i.e. slots nd ) were lternted cross the nodes. This -hop coloring resulted in n verge pcket loss of 67% due to interference. Following this, three slots (i.e., nd ) were lternted cross the nodes nd ech node received ll trnsmitted pckets. The tests were repeted thrice in two different outdoor loctions ( footll stdium nd n open field in prk) nd the results were consistent cross ll eighteen runs. ) ) c) 9 8 7 6 5 4 3 Figure 6. Multi-hop schedules in RT-Link test-ed.. n Figure 7. Multi-hop network topology with hidden terminl prolem. 6.. End-to-end Ltency In order to investigte the performnce of RT-Link, we simulted its opertion to compre the end-to-end ltency nd throughput with synchronous nd loosely synchronized protocols cross vrious topologies. To study the ltency in multi-hop scenrio we focused on the impct of the hidden terminl prolem on the performnce of B-MAC nd S- MAC. All the tests in [3] were designed to void the hidden terminl prolem nd essentilly focused on extremely lowlod nd one-hop scenrios. We simulted network topology of two "ckone" nodes connected to gtewy. One or more lef nodes were connected to the lower ckone node s shown in Figure 7. Only the lef nodes generted trffic to the gtewy. The totl trffic issued y ll nodes ws fixed to -yte pckets. At ech hop, if node received multiple pckets efore its next trnsmission, it ws le to ggregte them up to -yte frgments. The tested topology is the se cse for the hidden terminl prolem s the trnsmission opportunity of the ckone nodes is directly ffected y the degree of the lower ckone node. We compre the performnce of RT-Link with ms nd 3ms cycle durtion with RTS-CTS enled B-MAC operting with 5ms nd ms check times. The RTS-CTS cpility ws implemented s outlined in [3]. When node wkes up nd detects the chnnel to e cler, RTS nd CTS with long premles re exchnged followed y dt pcket with short premle. We ssume B-MAC is cple of perfect cler chnnel ssessment, zero pcket loss trnsmissions nd zero cost cknowledgement of pcket reception. We oserve tht s the node degree increses (Figure 8), B-MAC suffers liner increse in collisions, leding to n exponentil increse in ltency. With check time of ms, B-MAC sturtes t degree of 4. Incresing the check time to 5ms, pushes the sturtion point out to degree of 8. Using the schedule generte y the heuristic in Section 4, RT-Link demonstrtes flt end-to-end ltency. The cler drwck to sic B-MAC with RTS-CTS is tht upon hidden terminl collisions, the nodes immeditely retry fter smll rndom ckoff. To llevite prolem, we

Ltency (ms) 7 6 5 4 3 BMAC Adptive ms BMAC RTS/CTS ms BMAC RTS/CTS 5ms BMAC Adptive 5ms RT Link 3ms RT Link ms Throughput in Percentge of Chnnel Cpcity.9.8.7.6.5.4.3 S MAC B MAC RT-Link. 3 4 5 6 7 8 9 Degree Figure 8. Impct of Ltency with node degree. 4 6 8 4 6 8 Numer of Nodes Figure. Effect of node degree on throughput for single hop with no hidden terminls. BMAC Adptive ms Pcket Collisions 8 6 4 BMAC RTS/CTS ms BMAC RTS/CTS 5ms BMAC Adptive 5ms RT Link 3 4 5 6 7 8 9 Degree Figure 9. Effect of node degree on pcket collisions for B-MAC provided nodes with topology informtion such tht node s contention window size is proportion to the product of the degree nd the time to trnsmit pcket. As cn e seen in Figure 9, this llows for reltively constnt numer of collisions since ech node shres the chnnel more efficiently. This extr ckoff, in turn increses ltency linerly with the node degree. We see tht RT-Link suffers zero collisions nd mintins constnt ltency. 6.3. Throughput Figure shows the effect of node degree on throughput. In this exmple, ll nodes re within single hop of the receiver nd do not exhiit the hidden terminl prolem. As in [3, 6], ech node constntly trnsmits dt to the one-hop wy gtewy. We oserve tht s the numer of nodes communicting with the gtewy increses, RT-Link is le to ssign dditionl time slots nd mintin fixed throughput. In such scenrio, RT-Link cn support pproximtely,5 unique time slots in one second while re-synchronizing every seconds. RT-Link mintins stedy 8% throughput. The % loss in throughput (i.e. pproximtely 8us for every 4ms pcket) is due to inter-slot spcing used for pcket processing nd ggregtion. On the other hnd, the throughput offered y B-MAC decreses s the chnnel contention round the gtewy increses. The throughput offered y S-MAC is limited due to the fixed 5ms sleep durtion enforced on ech node. 7. Conclusion In this pper we explore the design, implementtion nd performnce of link lyer protocol for energy-constrined multi-hop wireless networks with end-to-end dely constrints. We introduced RT-Link, time-synchronized link protocol for fixed nd moile emedded rdios. We identify three key oservtions in the design nd deployment of RT- Link: () RT-Link offers predictle network lifetime with ounded end-to-end dely. () Achieving glol time synchroniztion is oth economicl nd convenient for indoor nd outdoor deployments. (c) Due to interference etween nodes, nodes with the sme schedule must e spced y minimum of 3 hops. RT-Link hs een implemented in Fire- Fly, our sensor network pltform, nd hs een deployed on networks with over 3 IEEE 8.4.5 nodes. It outperforms energy-efficient protocols such s B-MAC nd S-MAC in throughput, energy consumption nd end-to-end dely. References [] IEEE Std 8.5.4, 3. [] Rndolph D. Nelson nd Leonrd Kleinrock. Mximum proility of successful trnsmission in rndom plnr pcket rdio network. INFOCOM, pges 365 37, 983. [3] J. Polstre, J. Hill, nd D. Culler. Verstile low power medi ccess for wireless sensor networks. SenSys, Novemer 5. [4] T. Dm nd K. Lngendoen. An dptive energy-efficient mc protocol for wireless sensor networks. SenSys, Novemer 3. [5] A. El-Hoiydi nd J. Decotignie. Wisemc: An ultr low power mc protocol for the downlink of infrstructure wireless sensor networks. ISCC, 4. [6] W. Ye, J. Heidemnn, nd D. Estrin. An energy-efficient mc protocol for wireless sensor networks. INFOCOM, June. [7] V. Rjendrn, K. Orczk, nd J. J. rci-lun-aceves. Energy-efficient, collision-free medium ccess control for wireless sensor networks. Sensys, 3. [8] L.. Roerts. Aloh pcket system with nd without slots nd cpture. SICOMM, 5():8 4, 975.

[9] C. Schurgers, V. Tsitsis, S. neriwl, nd M. Srivstv. Topology mngement for sensor networks: Exploiting ltency nd density. MoiHoc,. [] C. uo, L. C. Zhong, nd J. Rey. Low power distriuted mc for d hoc sensor rdio networks. loecom,. [] Hri Blkrishnn et l. The distnce- mtching prolem nd its reltionship to the mc-lyer cpcity of d hoc wireless networks. IEEE Journl on Selected Ares in Comm., (6):69 79, August 4. [] Atmel corportion, tmeg3 dt sheet, Mrch 5. [3] Chipcon inc., chipcon cc4 dt sheet, Octoer 3. [4] Rdio systems 3w tr-6 m trnsmitter dt sheet, Mrch. [5] Te555t -chip m rdio philips semiconductors, Octoer 99. 3