A Distributed Three-hop Routing Protocol to Increase the Capacity of Hybrid Wireless Networks

Transcription

1 Ths artcle has been accepted for publcaton n a future ssue of ths journal, but has not been fully edted. Content may change pror to fnal publcaton. Ctaton nformaton: DOI.9/TMC , IEEE Transactons on Moble Computng A Dstrbuted Three-hop Routng Protocol to Increase the Capacty of Hybrd Wreless Networks Hayng Shen*, Senor Member, IEEE, Ze L Chenx Qu Abstract Hybrd wreless networks combnng the advantages of both moble ad-hoc networks nfrastructure wreless networks have been recevng ncreased attenton due to ther ultra-hgh performance. An effcent data routng protocol s mportant n such networks for hgh network capacty scalablty. However, most routng protocols for these networks smply combne the ad-hoc transmsson mode wth the cellular transmsson mode, whch nherts the drawbacks of ad-hoc transmsson. Ths paper presents a Dstrbuted Three-hop Routng protocol () for hybrd wreless networks. To take full advantage of the wdespread base statons, dvdes a message data stream nto segments transmts the segments n a dstrbuted manner. It makes full spatal reuse of a system va ts hgh speed ad-hoc nterface allevates moble gateway congeston va ts cellular nterface. Furthermore, sendng segments to a number of base statons smultaneously ncreases throughput makes full use of wdespread base statons. In addton, sgnfcantly reduces overhead due to short path lengths the elmnaton of route dscovery mantenance. also has a congeston control algorthm to avod overloadng base statons. Theoretcal analyss smulaton results show the superorty of n comparson wth other routng protocols n terms of throughput capacty, scalablty moblty reslence. The results also show the effectveness of the congeston control algorthm n balancng the load between base statons. Index Terms Hybrd wreless networks, Routng algorthm, Load balancng, Congeston control INTRODUCTION Over the past few years, wreless networks ncludng nfrastructure wreless networks moble ad-hoc networks (MANETs) have attracted sgnfcant research nterest. The growng desre to ncrease wreless network capacty for hgh performance applcatons has stmulated the development of hybrd wreless networks [ 6]. A hybrd wreless network conssts of both an nfrastructure wreless network a moble ad-hoc network. Wreless devces such as smart-phones, tablets laptops, have both an nfrastructure nterface an adhoc nterface. As the number of such devces has been ncreasng sharply n recent years, a hybrd transmsson structure wll be wdely used n the near future. Such a structure synergstcally combnes the nherent advantages overcome the dsadvantages of the nfrastructure wreless networks moble ad-hoc networks. In a moble ad-hoc network, wth the absence of a central control nfrastructure, data s routed to ts destnaton through the ntermedate nodes n a mult-hop manner. The mult-hop routng needs on-dem route dscovery or route mantenance [7 ]. Snce the messages are transmtted n wreless channels through dynamc routng paths, moble ad-hoc networks are not as relable as nfrastructure wreless networks. Furthermore, because of the mult-hop transmsson feature, moble ad-hoc networks are only sutable for local area data transmsson. The nfrastructure wreless network (e.g. cellular network) s the major means of wreless communcaton n our daly lves. It excels at nter-cell communcaton (.e., communcaton between nodes n dfferent cells) Internet access. It makes possble the support of unversal network connectvty ubqutous computng * Correspondng Author. Emal: shenh@clemson.edu; Phone: (864) ; Fax: (864) The authors are wth the Department of Electrcal Computer Engneerng, Clemson Unversty, Clemson, SC, E-mal: {shenh, zel, chenxq}@clemson.edu by ntegratng all knds of wreless devces nto the network. In an nfrastructure network, nodes communcate wth each other through base statons (BSes). Because of the long dstance one-hop transmsson between BSes moble nodes, the nfrastructure wreless networks can provde hgher message transmsson relablty channel access effcency, but suffer from hgher power consumpton on moble nodes the sngle pont of falure problem []. A hybrd wreless network synergstcally combnes an nfrastructure wreless network a moble adhoc network to leverage ther advantages overcome ther shortcomngs, fnally ncreases the throughput capacty of a wde-area wreless network. A routng protocol s a crtcal component that affects the throughput capacty of a wreless network n data transmsson. Most current routng protocols n hybrd wreless networks [, 5, 6, 2 8] smply combne the cellular transmsson mode (.e. BS transmsson mode) n nfrastructure wreless networks the ad-hoc transmsson mode n moble ad-hoc networks [8, 9, 7]. That s, as shown n Fgure (a), the protocols use the mult-hop routng to forward a message to the moble gateway nodes that are closest to the BSes or have the hghest bwdth to the BSes. The bwdth of a channel s the maxmum throughput (.e., transmsson rate n bts/s) that can be acheved. The moble gateway nodes then forward the messages to the BSes, functonng as brdges to connect the ad-hoc network the nfrastructure network. However, drect combnaton of the two transmsson modes nherts the followng problems that are rooted n the ad-hoc transmsson mode. Hgh overhead. Route dscovery mantenance ncur hgh overhead. The wreless rom access medum access control (MAC) requred n moble ad-hoc networks, whch utlzes control hshakng a backoff mechansm, further ncreases overhead. Hot spots. The moble gateway nodes can easly become hot spots. The RTS-CTS rom access, n whch most traffc goes through the same gateway, the floodng employed n moble ad-hoc routng to dscover routes (c) 25 IEEE. Translatons content mnng are permtted for academc research only. Personal use s also permtted, but republcaton/redstrbuton requres IEEE permsson. See for more nformaton.

2 Ths artcle has been accepted for publcaton n a future ssue of ths journal, but has not been fully edted. Content may change pror to fnal publcaton. Ctaton nformaton: DOI.9/TMC , IEEE Transactons on Moble Computng 2 Fg. : Tradtonal proposed routng algorthms on the uplnk drecton. may exacerbate the hot spot problem. In addton, moble nodes only use the channel resources n ther route drecton, whch may generate hot spots whle leave resources n other drectons under-utlzed. Hot spots lead to low transmsson rates, severe network congeston, hgh data droppng rates. Low relablty. Dynamc long routng paths lead to unrelable routng. Nose nterference neghbor nterference durng the mult-hop transmsson process cause a hgh data drop rate. Long routng paths ncrease the probablty of the occurrence of path breakdown due to the hghly dynamc nature of wreless ad-hoc networks. These problems become an obstacle n achevng hgh throughput capacty scalablty n hybrd wreless networks. Consderng the wdespread BSes, the moble nodes have a hgh probablty of encounterng a BS whle movng. Takng advantage of ths feature, we propose a Dstrbuted Three-hop Data Routng protocol (). In, as shown n Fgure (b), a source node dvdes a message stream nto a number of segments. Each segment s sent to a neghbor moble node. Based on the QoS requrement, these moble relay nodes choose between drect transmsson or relay transmsson to the BS. In relay transmsson, a segment s forwarded to another moble node wth hgher capacty to a BS than the current node. In drect transmsson, a segment s drectly forwarded to a BS. In the nfrastructure, the segments are rearranged n ther orgnal order sent to the destnaton. The number of routng hops n s confned to three, ncludng at most two hops n the ad-hoc transmsson mode one hop n the cellular transmsson mode. To overcome the aforementoned shortcomngs, tres to lmt the number of hops. The frst hop forwardng dstrbutes the segments of a message n dfferent drectons to fully utlze the resources, the possble second hop forwardng ensures the hgh capacty of the forwarder. also has a congeston control algorthm to balance the traffc load between the nearby BSes n order to avod traffc congeston at BSes. Usng self-adaptve dstrbuted routng wth hghspeed short-path ad-hoc transmsson, sgnfcantly ncreases the throughput capacty scalablty of hybrd wreless networks by overcomng the three shortcomngs of the prevous routng algorthms. It has the followng features: Low overhead. It elmnates overhead caused by route dscovery mantenance n the ad-hoc transmsson mode, especally n a dynamc envronment. Hot spot reducton. It allevates traffc congeston at moble gateway nodes whle makes full use of channel resources through a dstrbuted mult-path relay. Hgh relablty. Because of ts small hop path length wth a short physcal dstance n each step, t allevates nose neghbor nterference avods the adverse effect of route breakdown durng data transmsson. Thus, t reduces the packet drop rate makes full use of spacal reuse, n whch several source destnaton nodes can communcate smultaneously wthout nterference. The rest of ths paper s organzed as follows. Secton 2 presents a revew of representatve hybrd wreless networks mult-hop routng protocols. Secton 3 detals the protocol, wth an emphass on ts routng methods, segment structure, BS congeston control. Secton 4 theoretcally analyzes the performance of the protocol. Secton 5 shows the performance of the protocol n comparson to other routng protocols. Fnally, Secton 6 concludes the paper. 2 RELATED WORK In order to ncrease the capacty of hybrd wreless networks, varous routng methods wth dfferent features have been proposed. One group of routng methods ntegrate the ad-hoc transmsson mode the cellular transmsson mode [, 5, 6, 4, 6 8]. Dousse et al. [6] bult a Posson Boolean model to study how a BS ncreases the capacty of a MANET. Ln et al. [5] proposed a Multhop Cellular Network derved ts throughput. Hseh et al. [4] nvestgated a hybrd IEEE 82. network archtecture wth both a dstrbuted coordnaton functon a pont coordnaton functon. Luo et al. [] proposed a unfed cellular ad-hoc network archtecture for wreless communcaton. Cho et al. [6] studed the mpact of concurrent transmsson n a downlnk drecton (.e. from BSes to moble nodes) on the system capacty of a hybrd wreless network. In [7, 8], a node ntally communcates wth other nodes usng an ad-hoc transmsson mode, swtches to a cellular transmsson mode when ts performance s better than the ad-hoc transmsson. The above methods are only used to assst ntra-cell ad-hoc transmsson rather than nter-cell transmsson. In nter-cell transmsson [, 5, 6], a message s forwarded va the ad-hoc nterface to the gateway moble node that s closest to or has the hghest uplnk transmsson bwdth to a BS. The gateway moble node then forwards the message to the BS usng the cellular nterface. However, most of these routng protocols smply combne routng schemes n ad-hoc networks nfrastructure networks, hence nhert the drawbacks of the ad-hoc transmsson mode as explaned prevously. s smlar to the transmsson protocol [9] n terms of the elmnaton of route mantenance the lmted number of hops n routng. In, when a node s bwdth to a BS s larger than that of each neghbor, t drectly sends a message to the BS. Otherwse, t chooses a neghbor wth a hgher channel sends a message to t, whch further forwards the message to the BS. s dfferent from n three aspects. Frst, only consders the node transmsson wthn a sngle cell, whle can also deal wth nter-cell transmsson, whch s more challengng more common than ntra-cell communcaton n the real world. Second, uses dstrbuted transmsson nvolvng multple cells, whch makes full use of system resources dynamcally balances the traffc load between neghborng cells. In contrast, employs sngle-path transmsson (c) 25 IEEE. Translatons content mnng are permtted for academc research only. Personal use s also permtted, but republcaton/redstrbuton requres IEEE permsson. See for more nformaton.

3 Ths artcle has been accepted for publcaton n a future ssue of ths journal, but has not been fully edted. Content may change pror to fnal publcaton. Ctaton nformaton: DOI.9/TMC , IEEE Transactons on Moble Computng 3 Fg. 2: Data transmsson n the protocol. There are other methods proposed to mprove routng performance n hybrd wreless networks. Wu et al. [3] proposed usng ad-hoc relay statons to dynamcally relay traffc from one cell to another n order to avod traffc congeston n BSes. L et al. [2] surveyed a number of mult-hop cellular network (MCN) archtectures n lterature, compared dscussed methods to reduce the cost of deployment for MCNs. The work n [2] nvestgates how to allocate the bwdth to users to mprove the performance of hybrd wreless networks. Thulasraman et al. [22] further consdered the wreless nterference n optmzng the resource allocaton n hybrd wreless networks. The work n [23] proposes a coaltonal game theory based cooperatve packet delvery scheme n hybrd wreless networks. There are also some works [24 26] that study rado frequency allocaton for drecton transmsson relay transmsson n hybrd wreless networks. These works are orthogonal to our study n ths paper can be ncorporated nto to further enhance ts performance. The throughput capacty of the hybrd wreless network under dfferent settngs has also been an actve research topc n the hybrd wreless network. The works n [7, 27] have studed the throughput of hybrd network wth n nodes m statons. Lu et al. [28] theoretcally studed the capacty of hybrd wreless networks under an one-dmensonal network topology a twodmensonal strp topology. Wang et al. [29] studed the multcast throughput of hybrd wreless networks desgned an optmal multcast strategy based on deduced throughput. 3 DISTRIBUTED THREE-HOP ROUTING PRO- TOCOL 3. Assumpton Overvew Snce BSes are connected wth a wred backbone, we assume that there are no bwdth power constrants on transmssons between BSes. We use ntermedate n- odes to denote relay nodes that functon as gateways connectng an nfrastructure wreless network a moble ad-hoc network. We assume every moble node s dual-mode; that s, t has ad-hoc network nterface such as a WLAN rado nterface nfrastructure network nterface such as a 3G cellular nterface. ams to shft the routng burden from the adhoc network to the nfrastructure network by takng advantage of wdespread base statons n a hybrd wreless network. Rather than usng one mult-hop path to forward a message to one BS, uses at most two hops to relay the segments of a message to dfferent BSes n a dstrbuted manner, reles on BSes to combne the segments. Fgure 2 demonstrates the process of n a hybrd wreless network. We smplfy the routngs n the nfrastructure network for clarty. As shown n the fgure, when a source node wants to transmt a message stream to a destnaton node, t dvdes the message stream nto a number of partal streams called segments transmts each segment to a neghbor node. Upon recevng a segment from the source node, a neghbor node locally decdes between drect transmsson relay transmsson based on the QoS requrement of the applcaton. The neghbor nodes forward these segments n a dstrbuted manner to nearby BSes. Relyng on the nfrastructure network routng, the BSes further transmt the segments to the BS where the destnaton node resdes. The fnal BS rearranges the segments nto the orgnal order forwards the segments to the destnaton. It uses the cellular IP transmsson method [3] to send segments to the destnaton f the destnaton moves to another BS durng segment transmsson. Our algorthm avods the shortcomngs of adhoc transmsson n the prevous routng algorthms that drectly combne an ad-hoc transmsson mode a cellular transmsson mode. Rather than usng the multhop ad-hoc transmsson, uses two hop forwardng by relyng on node movement wdespread base statons. All other aspects reman the same as those n the prevous routng algorthms (ncludng the nteracton wth the TCP layer). works on the Internet layer. It receves packets from the TCP layer routes t to the destnaton node, where forwards the packet to the TCP layer. The data routng process n can be dvded nto two steps: uplnk from a source node to the frst BS downlnk from the fnal BS to the data s destnaton. Crtcal problems that need to be solved nclude how a source node or relay node chooses nodes for effcent segment forwardng, how to ensure that the fnal BS sends segments n the rght order so that a destnaton node receves the correct data. Also, snce traffc s not evenly dstrbuted n the network, how to avod overloadng BSes s another problem. Below, Secton 3.2 wll present the detals for forwardng node selecton n uplnk transmsson Secton 3.3 wll present the segment structure that helps ensure the correct fnal order of segments n a message, s strategy for downlnk transmsson. Secton 3.4 wll present the congeston control algorthm for balancng a load between BSes. 3.2 Uplnk Data Routng A long routng path wll lead to hgh overhead, hot spots low relablty. Thus, tres to lmt the path length. It uses one hop to forward the segments of a message n a dstrbuted manner uses another hop to fnd hgh-capacty forwarder for hgh performance routng. As a result, lmts the path length of uplnk routng to two hops n order to avod the problems of long-path mult-hop routng n the ad-hoc networks. Specfcally, n the uplnk routng, a source node ntally dvdes ts message stream nto a number of segments, then transmts the segments to ts neghbor nodes. The neghbor nodes forward segments to BSes, whch wll forward the segments to the BS where the destnaton resdes.below, we frst explan how to defne capacty, then ntroduce the way for a node to collect the capacty nformaton from ts neghbors, fnally present the detals of the routng algorthm. Dfferent applcatons may have dfferent QoS requrements, such as effcency, throughput, routng speed. For example, delay-tolerant applcatons (e.g. voce mal, (c) 25 IEEE. Translatons content mnng are permtted for academc research only. Personal use s also permtted, but republcaton/redstrbuton requres IEEE permsson. See for more nformaton.

4 Ths artcle has been accepted for publcaton n a future ssue of ths journal, but has not been fully edted. Content may change pror to fnal publcaton. Ctaton nformaton: DOI.9/TMC , IEEE Transactons on Moble Computng 4 e-mal text messagng) do not necessarly need fast real-tme transmsson may make throughput the hghest consderaton to ensure successful data transmsson. Some applcatons may take hgh moblty as ther prorty to avod hot spots blank spots. Hot spots are areas where BS channels are congested, whle blank spots are areas wthout sgnals or wth very weak sgnals. Hgh-moblty nodes can quckly move out of a hot spot or blank spot enter a cell wth hgh bwdth to a BS, thus provdng effcent data transmsson. Throughput can be measured by bwdth, moblty can be measured by the speed of node movement, routng speed can be measured by the speed of data forwardng. Bwdth can be estmated usng the nonntrusve technque proposed n [3]. In ths work, we take throughput routng speed as examples for the QoS requrement. We use a bwdth/queue metrc to reflect node capacty n throughput fast data forwardng. The metrc s the rato of a node s channel bwdth to ts message queue sze. A larger bwdth/queue value means hgher throughput message forwardng speed, vce versa. When choosng neghbors for data forwardng, a node needs the capacty nformaton (.e., queue sze bwdth) of ts neghbors. Also, a selected neghbor should have enough storage space for a segment. To keep track of the capacty storage space of ts neghbors, each node perodcally exchanges ts current capacty storage nformaton wth ts neghbors. In the adhoc network component, every node needs to perodcally send hello messages to dentfy ts neghbors. Takng advantage of ths polcy, nodes pggyback the capacty storage nformaton onto the hello messages n order to reduce the overhead caused by the nformaton exchanges. If a node s capacty storage space are changed after ts last hello message sendng when t receves a segment, t sends ts current capacty storage nformaton to the segment forwarder. Then, the segment forwarder wll choose the hghest capacty nodes n ts neghbors based on the most updated nformaton. When a source node sends out message segments, t chooses the neghbors that have enough space for storng a segment, then chooses neghbors that have the hghest capacty. In order to fnd hgher capacty forwarders n a larger neghborhood around the source, each segment recever further forwards ts receved segment to ts neghbor wth the hghest capacty. That s, after a neghbor node m receves a segment from the source, t uses ether drect transmsson or relay transmsson. If the capacty of each of ts neghbors s no greater than tself, relay node m uses drect transmsson. Otherwse, t uses relay transmsson. In drect transmsson, the relay node sends the segment to a BS f t s n a BS s regon. Otherwse, t stores the segment whle movng untl t enters a BS s regon. In relay transmsson, relay node m chooses ts hghest-capacty neghbor as the second relay node based on the QoS requrement. The second relay node wll use drect transmsson to forward the segment drectly to a BS. As a result, the number of transmsson hops n the ad-hoc network component s confned to no more than two. The small number of hops help to ncrease the capacty of the network reduce channel contenton n ad-hoc transmsson. Algorthm shows the pseudo-code for neghbor node selecton message forwardng n. The purpose of the second hop selecton s to fnd a hgher capacty node as the message forwarder n order to mprove the performance of the QoS requre- Base staton A 5 Source (a) Source node has hgher capacty neghbors Moble node B (b) Source node has no hgher capacty neghbors. Fg. 3: Neghbor selecton n. ment. As the neghborhood scope of a node for hghcapacty node searchng grows, the probablty of fndng hgher capacty nodes ncreases. Thus, a source node s neghbors are more lkely to fnd neghbors wth hgher capactes than the source node. Therefore, transmttng data segments to neghbors enablng them to choose the second relays help to fnd hgher capacty nodes to forward data. If a source node has the hghest capacty n ts regon, the segments wll be forwarded back to the source node accordng to the protocol. The source node then forwards the segments to the BSes drectly due to the three-hop lmt. Though sendng data back forth leads to latency bwdth wastage, ths case occurs only when the source nodes s the hghest capacty node wthn ts two-hop neghborhood. Also, ths step s necessary for fndng the hghest capacty nodes wthn the source s two-hop neghborhood, ensures that the hghest capacty nodes are always s- elected as the message forwarders. If the source node does not dstrbute segments to ts neghbors, the hghercapacty node searchng cannot be conducted. Note that the data transmsson rate of the ad-hoc nterface (e.g. IEEE 82.) s more than tmes faster than the cellular nterface (e.g. GSM, 3G). Thus, the transmsson delay for sendng the data back forth n the ad-hoc transmsson s neglgble n the total routng latency. By dstrbutng a message s segments to dfferent n- odes to be forwarded n dfferent drectons, our algorthm reduces the congeston n the prevous routng algorthms n the hybrd wreless networks. When a node selects a relay to forward a segment, t checks the capacty of the node. Only when a node, say node m, has enough capacty, the node wll forward a segment to node m. Therefore, even though the paths are not node-dsjont, there wll be no congeston n the common sub-paths. Fgure 3 shows examples of neghbor selecton n, n whch the source node s n the transmsson range of a BS. In the fgures, the value n the node represents ts capacty. In scenaro (a), there exst nodes that have hgher capacty than the source node wthn the source s two-hop neghborhood. If a routng algorthm drectly let a source node transmt a message to ts BS, the hgh routng performance cannot be guaranteed snce the source node may have very low capacty. In, the source node sends segments to ts neghbors, whch further forward the segments to nodes wth hgher capactes. In scenaro (b), the source node has the hghest capacty among the nodes n ts two-hop neghborhood. After recevng segments from the source node, some neghbors forward the segments back to the source node, whch sends the message to ts BS. Thus, always arranges data to be forwarded by nodes wth hgh (c) 25 IEEE. Translatons content mnng are permtted for academc research only. Personal use s also permtted, but republcaton/redstrbuton requres IEEE permsson. See for more nformaton.

5 Ths artcle has been accepted for publcaton n a future ssue of ths journal, but has not been fully edted. Content may change pror to fnal publcaton. Ctaton nformaton: DOI.9/TMC , IEEE Transactons on Moble Computng 5 capacty to ther BSes. acheves hgher throughput faster data forwardng speed by takng nto account node capacty n data forwardng. Algorthm Pseudo-code for neghbor node selecton message forwardng. : ChooseRelay( ) { 2: //choose neghbors wth suffcent caches bwdth/queue (b/q) rates 3: Query storage sze QoS requrement nfo. from neghbors 4: for each neghbor n do 5: f n.cache.sze>segment.length && n.b/q>ths.b/q then 6: Add n to R = {r,... r m,...} n a descendng order of b/q 7: end f 8: end for 9: Return R : } : Transmsson( ) { 2: f t s a source node then 3: //routng conducted by a source node 4: //choose relay nodes based on QoS requrement 5: R=ChooseRelay( ); 6: Send segments to {r,... r m} n R 7: else 8: //routng conducted by a neghbor node 9: f ths.b/q b/q of all neghbors then 2: //drect transmsson 2: f wthn the range of a BS then 22: Transmt the segment drectly to the BS 23: end f 24: else 25: //relay transmsson 26: node =gethghestcapablty(chooserelay( )) 27: Send a segment to node 28: end f 29: end f 3: } Algorthm 2 Pseudo-code for a BS to reorder forward segments to destnaton nodes. : //a cache pool s bult for each data stream 2: //there are n cache pools currently 3: f receves a segment (S,D,m,s,q) then 4: f there s no cache pool wth msg sequence num equals m then 5: Create a cache pool n + for the stream m 6: else 7: //the last delvered segment of stream m has sequence num 8: f s == then 9: Send out segment (S,D,m,s,q) to D : + +; : else 2: Add segment (S,D,m,s) nto cache pool m 3: end f 4: end f 5: end f 3.3 Downlnk Data Routng Data Reconstructon As mentoned above, the message stream of a source n- ode s dvded nto several segments. After a BS receves a segment, t needs to forward the segment to the BS, where the destnaton node resdes (.e., the destnaton BS). We use the moble IP protocol [32] to enable BSes to know the destnaton BS. In ths protocol, each moble node s assocated wth a home BS, whch s the BS n the node s home network, regardless of ts current locaton n the network. The home network of a node contans ts regstraton nformaton dentfed by ts home address, whch s a statc IP address assgned by an ISP. In a hybrd wreless network, each BS perodcally emts beacon sgnals to locate the moble nodes n ts range. When a moble node m moves away from ts home BS, the BS where m currently resdes detects m sends ts IP address to the home BS of m. When a BS wants to contact m, t contacts the home BS of m to fnd the destnaton BS where m currently resdes at. However, the destnaton BS recorded n the home BS may not be the most up-to-date destnaton BS snce destnaton moble nodes swtch between the coverage regons of dfferent BSes durng data transmsson to them. For nstance, data s transmtted to BS B that has the data s destnaton, but the destnaton has moved to the range of BS B j before the data arrves at BS B. To deal wth ths problem, we adopt the Cellular IP protocol [3] for trackng node locatons. Wth ths protocol, a BS has a home agent a foregn agent. The foregn agent keeps track of moble nodes movng nto the ranges of other BSes. The home agent ntercepts n-comng segments, reconstructs the orgnal data, re-routes t to the foregn agent, whch then forwards the data to the destnaton moble node. After the destnaton BS receves the segments of a message, t rearranges the segments nto the orgnal message then sends t to the destnaton moble node. A vtal ssue s guaranteeng that the segments are combned n the correct order. For ths purpose, specfes the segment structure format. Each segment contans eght felds, ncludng: () source node IP address (denoted by S); (2) destnaton node IP address (denoted by D); (3) message sequence number (denoted by m); (4) segment sequence number (denoted by s); (5) QoS ndcaton number (denoted by q); (6) data; (7) length of the data; (8) checksum. Felds ()-(5) are n the segment head. The role of the source IP address feld s to nform the destnaton node where the message comes from. The destnaton IP address feld ndcates the destnaton node, s used to locate the fnal BS. After sendng out a message stream to a destnaton, a source node may send out another message stream to the same destnaton node. The message sequence number dfferentates the dfferent message streams ntated by the same source node. The segment sequence number s used to fnd the correct transmsson sequence of the segments for transmsson to a destnaton node. The data s the actual nformaton that a source node wants to transmt to a destnaton node. The length feld specfes the length of the segment ncludng the header n bytes. The checksum s used by the recever node to check whether the receved data has errors. The QoS ndcaton number s used to ndcate the QoS requrement of the applcaton. Thus, each segment s head ncludes the nformaton represented by (S, D, m, s, q)(m, s =, 2, 3,...). When a segment wth head (S, D, m, s, q) arrves at a BS, the BS contacts D s home BS to fnd the destnaton BS where D stays va the moble IP protocol. It then transmts the segment to the destnaton BS through the nfrastructure network component. After arrvng at the BS, the segment wats n the cache for ts turn to be transmtted to ts destnaton node based on ts message segment sequence numbers. At ths tme, f another segment comes wth a head labelled (S, D, (m + ), s, q), whch means that t s from the same source node but belongs to another data stream, the BS wll put t to another stream. If the segment s labeled as (S, D, m, (s+), q), t means that ths segment belongs to the same data stream of the same source node as segment (S, D, m, s, q). The combnaton of the source node s sequence number segment sequence number helps to locate the stream the poston of a segment n the steam. In order to ntegrate the segments nto ther correct order to retreve the orgnal data, the segments n the BS are transmtted (c) 25 IEEE. Translatons content mnng are permtted for academc research only. Personal use s also permtted, but republcaton/redstrbuton requres IEEE permsson. See for more nformaton.

6 Ths artcle has been accepted for publcaton n a future ssue of ths journal, but has not been fully edted. Content may change pror to fnal publcaton. Ctaton nformaton: DOI.9/TMC , IEEE Transactons on Moble Computng 6 to the destnaton node n the order of the segments sequence n the orgnal message. If a segment has not arrved at the fnal BS, ts subsequent segments wll wat n the fnal BS untl ts arrval. Algorthm 2 shows the pseudo-code for a BS to reorder forward segments to ther destnatons. Note that n the cache, we can set the tmer based on the packet rate storage lmt. In other words, the tmer should be set as large as possble to fully utlze the storage on BSes to ensure that a message has a hgh probablty to be recovered. 3.4 Congeston Control n Base Statons Compared to the prevous routng algorthms n hybrd wreless networks, can dstrbute traffc load among moble nodes more evenly. Though the dstrbuted routng n can dstrbute traffc load among nearby B- Ses, f the traffc load s not dstrbuted evenly n the network, some BSes may become overloaded whle other B- Ses reman lghtly loaded. We propose a congeston control algorthm to avod overloadng BSes n uplnk transmsson (e.g., B, B 2 B 3 n Fgure (b)) downlnk transmsson (e.g., B 4 n Fgure (b)), respectvely. In the hybrd wreless network, BSes send beacon messages to dentfy nearby moble nodes. Takng advantage of ths beacon strategy, once the workload of a BS, say B, exceeds a pre-defned threshold, B adds an extra bt n ts beacon message to broadcast to all the nodes n ts transmsson range. Then, nodes near B know that B s overloaded wll not forward segments to B. When a node near B, say m, needs to forward a segment to a BS, t wll send the segment to B based on the algorthm. In our congeston control algorthm, because B s overloaded, rather than targetng B, m wll forward the segment to a lghtly loaded neghborng BS of B. To ths end, node m frst queres a mult-hop path to a lghtly loaded neghborng BS of B. Node m broadcasts a query message nto the system. We set the TTL for the path query forwardng step to a constant (e.g., 3). The query message s forwarded along other nodes untl a node (say m j ) near a lghtly loaded BS (say B j ) s reached. Due to broadcastng, a node may receve multple copes of the same queres. Each node only remembers m the node that forwards the frst query (.e., ts precedng node), gnores all other the same queres. In ths way, a mult-hop path between the source node the lghtly loaded base staton can be formed. Node m j responds to the path query by addng a reply bt the address of m nto ts beacon message to ts precedng node n the path. Ths beacon recever also adds a reply bt the address of m nto ts beacon message to ts precedng node n the path. Ths process repeats untl m receves the beacon. Thus, each node knows ts precedng node succeedng node n the path from m m j based on the address of m. Then, m s message can be forwarded along the observed path along the nodes. The observed path can always be used by m for any subsequent messages to B j as long as t s not broken. The neghborng BSes of an overloaded BS may also be overloaded. As the moble nodes near an overloaded BS know that the BS s overloaded, when they receve a query message to fnd a path to an underloaded BS, they do not forward the message towards ther overloaded BSes. Node m may receve responses from a few nodes near BSes. It can choose b (b ) neghborng BSes of the destnaton to forward the segment. The redundant transmssons enhance the data transmsson relablty whle also ncrease the routng overhead. Thus, the value of b should be carefully determned based on the avalable resources for routng the relablty dem. If b s set to a large value, the routng relablty s hgh at the cost of hgh overhead. If b s set to a small value, the routng relablty s low whle the overhead s reduced. After the neghborng BSes receve the segments, they further forward the segments to the destnaton BS, whch forwards the segments to the destnaton node. In ths way, the heavy traffc from moble nodes to a BS can be dstrbuted among neghborng BSes quckly. Next, we dscuss how to hle the case when the destnaton BS s congested. If a BS has not receved confrmaton from the destnaton BS durng a certan tme perod after t sends out a segment, t assumes that the destnaton BS s overloaded. Then, t sends the segment to b (b ) lghtly loaded neghborng BSes of the destnaton BS from ts routng table. If an attempted neghborng BS does not respond durng a certan tme perod, t s also consdered as overloaded. Then, the BS keeps tryng other neghborng BSes untl fndng lghtly loaded BSes. Redundant neghborng BSes are selected n order to ncrease routng relablty. The constant b should be set to an approprate value consderng factors such as the network sze the amount of traffc n order to acheve an optmal trade-off between overhead relablty. After recevng the message, each lghtly loaded neghborng BS of the destnaton BS fnds a mult-hop path to the destnaton moble node. It broadcasts a path query message, whch ncludes the IDs of the destnaton BS the destnaton node, to the moble nodes n ts regon. The path queryng process s smlar to the prevous path queryng for a lghtly loaded BS. The nodes further forward the path query to ther neghbors untl the query reaches the destnaton node. Here, we do not pggyback the query to beacon messages because ths queryng s for a specfc moble node rather than any moble node near a lghtly loaded BS. Includng the moble node s ID nto beacon messages generates very hgh overhead. B 4 s B B 3 Moble node Source node Destnaton node B 2 D B 5 B 6 In order to reduce the broadcastng overhead, a moble node resdng n the regon of a BS not close to the destnaton BS drops the query. The nodes can determne ther approxmate relatve postons to BSes by sensng the sgnal strengths from dffer- Fg. 4: Congeston control on BSes. ent BSes. Each node adds the strength of ts receved sgnal nto ts beacon message that s perodcally exchanged between neghbor nodes so that the nodes can dentfy ther relatve postons to each other. Only those moble nodes that stay farther than the query forwarder from the forwarder s BS forward the queres n the drecton of the destnaton BS. In ths way, the query can be forwarded to the destnaton BS faster. After the mult-hop path s dscovered, the neghborng BS sends the segment to the destnaton node along the path. Snce the destnaton node s n the neghborng BS s regon, the overhead to dentfy a path to the destnaton node s small. Note that our methods for congeston control n base statons nvolve query broadcastng. However, t s used only when some base statons are overloaded rather than n the normal routng (c) 25 IEEE. Translatons content mnng are permtted for academc research only. Personal use s also permtted, but republcaton/redstrbuton requres IEEE permsson. See for more nformaton.

7 Ths artcle has been accepted for publcaton n a future ssue of ths journal, but has not been fully edted. Content may change pror to fnal publcaton. Ctaton nformaton: DOI.9/TMC , IEEE Transactons on Moble Computng 7 algorthm n order to avod load congeston n BSes. Fgure 4 shows an example of the congeston control on BSes when b = 2. As shown n fgure, BS B s congested. Then, the relay nodes of the source node s message broadcast locally by beacon pggybackng to fnd mult-hop paths whch lead to B 3 B 4. The relay nodes then send segments along the paths. In ths way, the traffc orgnally targetng overloaded B can be spread out to the neghborng BSes B 3 B 4. B 3 B 4 further forward the segments to the destnaton BS B 6 f B 6 s not congested. If B 6 s also congested, B 3 B 4 send the segments to the neghborng BSes of B 6. Specfcally, B 4 sends the segment to B 3. B 3 does not forward the segment to another BS snce t already s close to B 6. B 3 then fnds a mult-hop path to the destnaton node uses ad-hoc transmsson to forward the segments to the destnaton node. Smlarly, when B 2 wants to send a segment to the destnaton node, t also uses a mult-hop path for the segment transmsson. 4 PERFORMANCE ANALYSIS OF THE PROTOCOL σ Node densty M Number of BSes l Segment s length s h Area sze of a cell n(s) Number of nodes n area S R Transmsson range W Bwdth of a node v m Moble node P (σ, M)Throughput n(σ, M) Number of nodes TABLE : Parameter table. In ths secton, we analyze the effectveness of the protocol at enhancng the capacty scalablty of hybrd wreless networks. In our analyss, we use the same scenaro n [7] for hybrd wreless networks, use the same scenaro n [33] for the ad-hoc network component. We present the scenaros some concepts below. We consder a large number of moble nodes unformly romly deployed over a 2-D feld. The movng drectons of the nodes are ndependent dentcally dstrbuted (..d.). The dstrbuton of moble nodes can be modeled as a homogeneous Posson process wth node densty σ [34]. That s, gven an area of sze S n the feld, the number of nodes n the area, denoted by n(s), follows the Posson dstrbuton wth the parameter σs, Pr (n(s) = k) = (σs)k e σs, k =,, 2,... () k! Besdes moble nodes, there are M BSes regularly deployed n the feld. The BSes dvde the area nto a hexagon tessellaton, n whch each hexagon has sde length h. The BSes are assumed to be connected together by a wred network. We assume that the lnk bwdths n the wred network are large enough so that there are no bwdth constrants between BSes. In snglepath transmsson, a message s sequentally transmtted n one routng path. In mult-path transmsson, a message s dvded nto a number of segments that are forwarded along multple paths n a dstrbuted manner. We assume each segment has the same length l. Table lsts the notatons used n our analyss. We assume that the transmsson range of all moble nodes all BSes s R (R > h). In ths paper, we use protocol model [7, 33] to descrbe the nterference among nodes; that s, a transmsson from a node (here node can be ether moble node or BS) v to another node v j s successful f the followng two condtons are satsfed: ) v j s wthn the transmsson range of v,.e., v v j R (2) where v v j represents the Eucldean dstance between v v j n the plane. 2) For any other node v k that s smultaneously transmttng over the same channel, v k v j ( + ) v v j. (3) Formula (3) guarantees a guard zone around the recevng node to prevent a neghborng node from transmttng on the same channel at the same tme. The radus of the guard zone s ( + ) tmes the dstance between the sender the recever. The parameter defnes the sze of the guard zone we requre that >. We frst adopt a concept called aggregate throughput capacty ntroduced n [7, 33] to measure the throughput of the network. Defnton (Aggregate Throughput Capacty of Hybrd Networks) The aggregate throughput capacty of a hybrd wreless network s of order Θ(f(σ, M)) f there are determnstc constants α >, α < + such that lm Pr (P (σ, M) = αf(σ, M) s feasble) = (4) lm nf Pr ( P (σ, M) = α f(σ, M) s feasble ) <. (5) Snce the workng frequency of nfrastructure networks s around 7MHz whle that of ad-hoc networks s 2.4 GHz, the communcatons n nfrastructure mode (between moble nodes BSes through cellular nterface) would not generate nterference to ad-hoc mode. We dvde the channel for nfrastructure mode transmssons nto uplnk downlnk parts, accordng to the transmsson drecton relatve to the BSes. Accordngly, n the protocol, the traffc of each S-D par s composed of at most two ntra-cell traffcs, one uplnk traffc one download traffc. Snce uplnk traffc downlnk traffc use dfferent sub-channels, there s also no nterference between these two types of traffcs. For each node v, we denote the bwdth assgned to ntracell, uplnk, downlnk sub-channels by W nt, W down, respectvely. We let = W down because there are the same amount of uplnk traffc downlnk traffc. The transmsson rates should sum to W,.e., W nt + + W down = W. Though no nterference exsts between ntra-cell, uplnk, downlnk traffcs, nterference exsts between the same type of traffc n a cell between dfferent cells. Fortunately, there s an effcent spatal transmsson schedule that can prevent such nterferences [7]. Frst, to avod the nterference n a cell, any two nodes wthn the cell are not allowed to transmt wth the same traffc mode at the same tme. Second, to avod the nterference between dfferent cells, the cells are spatally dvded nto a number of groups transmssons n the cells of the same group do not nterfere wth each other. If the groups are scheduled to transmt n a round robn fashon, each cell wll be able to transmt once every fxed amount of tme wthout nterferng wth each other. Below, we show how many groups we need to dvde the cells to prevent nterference. We adopt the noton of nterferng neghbors ntroduced n [7], gve the number of cells that can be affected by a transmsson n one cell. Two cells are defned to be nterferng neghbors f there s a pont n one cell whch s wthn a dstance (2+ )R of a pont n the other cell. Accordngly, f two cells are not nterferng neghbors, transmssons n one cell do not nterfere wth transmssons n the other cell. [7] has proved that () each cell has no more than c nterferng neghbors (Lemma n [7]), where c s a constant c = 4 ( ) 2 3l + 2R + R, (6) 3 3l (c) 25 IEEE. Translatons content mnng are permtted for academc research only. Personal use s also permtted, but republcaton/redstrbuton requres IEEE permsson. See for more nformaton.

8 Ths artcle has been accepted for publcaton n a future ssue of ths journal, but has not been fully edted. Content may change pror to fnal publcaton. Ctaton nformaton: DOI.9/TMC , IEEE Transactons on Moble Computng 8 (2) all cells should be dvded nto c + groups the whole channel should be dvded nto c + subchannels, where each subchannel s allocated to the cells n one group. Thus, the number of group we need to dvde the cells to prevent nterference s c +. Before calculatng the aggregate throughput capacty of, we frst ntroduce Lemma 4.. Lemma 4.: The number of cells that have moble n- odes s Θ(M). Proof: Denote the number of cells havng moble nodes by M. To prove M = Θ(M), we need to prove that there exsts determnstc constants α > α < + such that lm Pr (M = αm) =, (7) lm nf Pr ( M = α M ) <. (8) For Formula (8), let α = 2. Because the number of cells havng moble nodes s upper bounded by M, then lm nf Pr (M = 2M s feasble) =. (9) Now, we prove that Formula (7) can also be satsfed for some constant α. Because the number of nodes n a cell follows a Posson dstrbuton the sze of each cell (hexagon) s s h = 3 3h 2, then we can derve the probablty that no moble node s n a cell equals Pr (n(s h ) = ) = σ e s h = e s h. ()! Consder an arbtrary cell k, let X, X 2,..., X k,..., X M be..d. rom varables, where X k represents whether cell k has{ moble nodes. Then, X k s defned as follows: cell k has moble nodes X k = cell k does not have moble nodes () E(X k ) = e s h. For smplcty, let c 2 = e s h. Then, M = M k= X k. By the Strong Law of Large Number (SLLN) [34], ( ) M k= Pr lm X k = c 2 =, (2) M whch mples that lm Pr (M = c 2 M) =, whch ndcates that when α = c 2, Formula (7) can also be satsfed. Lemma 4.2: Let n(σ, M) denote the number of moble nodes n the whole network. Then, lm Pr (n(σ, M) = s hm) =. (3) Proof: Let Z, Z 2,..., Z M be..d. rom varables representng the number of nodes n cell, 2,..., M, respectvely. Then, n(σ, M) = M k= Z k. Because each Z k follows a Posson dstrbuton wth parameter s h, E(Z k ) = s h, ( k M. Accordng) to SLLN, M k= Pr lm Z k M = s h =, (4) ( M ) whch mples that lm Pr k= Z k = s h M =, hence lm Pr (n(σ, M) = s h M) =. Theorem 4.: For a hybrd network of M BSes σ moble node densty, where each node has the ntra-cell, uplnk downlnk sub-channel bwdth satsfyng W down = = = W/4, W nt = W nt = W/2 (5) the aggregate throughput capacty of s P (σ, M) = Θ(MW ). (6) Proof: To prove P (σ, M) = Θ(M W ), we need to prove that there exsts determnstc constants α > α < such that lm Pr{P (σ, M) = αmw s feasble} = (7) lm nf Pr{P (σ, M) = α MW s feasble} <. (8) Recall that any two nodes wthn a cell cannot transmt smultaneously n the same traffc mode, the throughput P s upper bounded by MW/4, whch can be acheved only f each cell has one node to send the message. Hence, Formula (8) can be satsfed by settng α to /2. Then, we wll show how Formula (7) can be satsfed. Snce the same message has to go through an uplnk a downlnk t s counted only once n the throughput capacty, calculatng the throughput of the whole network s equvalent to calculatng the throughput of uplnk traffc P up or the throughput of downlnk traffc P down. Notce calculatng ntra-cell traffc throughput s not accurate because a message may transmt twce wth ntra-cell mode. In ths proof, we calculate P up. Frst, we consder the throughput of the uplnk traffc of an arbtrary cell k, denoted by Pup. k Snce the schedule allocates /(c + ) tme slots to ths cell, then Pup k = c +. (9) Then, we consder the throughput of the whole network. Let P up = M = P upx represent the throughput of uplnk traffc, then ( we have lm Pr P up = c2mw ) 3(c + ) ( M ) = lm Pr P up c2mw upx = c + = ( M ) = lm Pr X = c 2M = (By Lemma 4.) = Accordngly, Formula (7) can be satsfed when α s set c to 2 3(c. +) Corollary 4.: Wth the restrcton n Theorem 4., can acheve Θ(W ) throughput per S-D par. Proof: Denote the throughput of per S-D par by P, whch equals P (σ, M) P =. (2) n Obvously, P s upper bounded by W 4 because each node has at most W 4 for uplnk traffc (or downlnk traffc), whch equals ts S-D par throughput. By Lemma 4.2 Theorem 4., ( we can derve ) that lm Pr c 2W P = 3(c + )s h ( ) P (σ, M) = lm Pr n(σ, M) = c 2W 3(c + )s h ( lm Pr P (σ, M) = c2w M ) Pr (n(σ, M) = s h M) 3(c + ) ( = lm Pr P (σ, M) = c2w M ) = 3(c + ( ) ) whch mples that lm Pr P = c2w 3(c +)s h =. Corollary 4. shows that produces a constant throughput for each par of nodes regardless of the number of nodes n each cell due to ts spacal reuse of the system. Theorem 4. Corollary 4. show that the aggregate throughput capacty the throughput per S-D par of are Θ(M W ) Θ(W ), respectvely. The work n [7] proves that DHybrd acheves Θ(M W ) nfrastructure aggregate throughput, the work n [33] proves that the pure ad-hoc transmsson W acheves Θ( n logn ) throughput per S-D par. The results demonstrate that the throughput rates of DHybrd are hgher than that of the pure ad-hoc transmsson. Ths s because the pure ad-hoc transmsson s not effcent n a large scale network [35]. A large network sze reduces the path utlzaton effcency ncreases node nterference. Facltated by the nfrastructure network, (c) 25 IEEE. Translatons content mnng are permtted for academc research only. Personal use s also permtted, but republcaton/redstrbuton requres IEEE permsson. See for more nformaton.

9 Ths artcle has been accepted for publcaton n a future ssue of ths journal, but has not been fully edted. Content may change pror to fnal publcaton. Ctaton nformaton: DOI.9/TMC , IEEE Transactons on Moble Computng 9 R: Transmsson range of a base staton a moble node B: Base staton m : moble node D: Dstance between B m h D B m R (a) DHybrd h m B D R (b) Fg. 5: The traffc load n DHybrd. DHybrd avod long dstance transmssons, leadng to a hgher transmsson throughput. Proposton 4.: Suppose a moble node needs to allocate totally U segments wth the same length to L neghborng moble nodes m,..., m L, whch has uplnk bwdth up,..., WL, respectvely. Let U denote the number of segments to be allocate to m ( =, 2,..., L). To mnmze the average latency of these segments, the optmal allocaton should satsfy =... = U L U Ul 2 L = L. The mnmzed average latency equals. Proof: Recall that each segment has length l. Then, l for each moble node m t requres tme to transmt a segment. Therefore, the j th segment that m needs to transmt has to wat (j )l slots. Hence, the total latency of the segments that m needs to transmt to ts BS equals U (j )l ( (U ))l = U 2 l. (2) j= 2 Hence, the average latency of transmttng all the messages should be K U 2 l = /U. Accordng to Cauchy- 2 Schwarz nequalty [34], the average latency s lower bounded L U 2 l l L U 2 = 2U U 2 L L = When =... = U L U 2 W up W up L = =... = l 2U L = Ul 2 L = U L 2 L W up L = ( L = = U 2 ) 2. (22), or equvalently,, the average segment latency L can acheve the mnmum value Ul 2 L =. = U U 2 l 2 Proposton 4. ndcates that forwardng segments to the nearby nodes wth the hghest capacty can mnmze the average latency of messages n the cell. It also balances the transmsson load of the moble nodes wthn a cell. Proposton 4.2: A source node n can fnd relay nodes for message forwardng wth probablty k= k c k r e cr k k!, where c r = πr 2. Proof: Let m denote the number of nodes wthn m s transmsson area defne the ndcator varable Q by { m s the hghest capacty node Q = m s not the hghest capacty node (23) then, = Pr{m can fnd relays for message forwardng} k c k Pr (Q = m = k) Pr (m = k) = re cr k k! k= k= = /U Proposton 4.2 ndcates that n a hgh-densty network, a source node n can fnd relay nodes for message forwardng wth a hgh probablty. For example, assume the average number of neghbor nodes of a source node s. Wth the daly ncreasng number of moble devces, such an assumpton s realstc. Then, the probablty of not beng able to fnd any node n the range of a node s k= k k e k k!.2, whch s very small. Therefore, n a hgh-densty network, a source node can fnd neghbors for message forwardng wth a hgh probablty. We use DHybrd to denote the group of routng protocols n hybrd wreless networks that drectly combne the ad-hoc transmsson mode the nfrastructure transmsson mode [, 5, 6, 2 8]. Proposton 4.3: In a hybrd wreless network, the D- Hybrd routng protocol leads to load mbalance among the moble nodes n a cell. Proof: Fgure 5 (a) shows a cell wth a BS a romly pcked moble node m n the range of the BS. The shaded regon represents all possble postons of the source nodes that choose m as the relay node n DHybrd. The total traffc passng through node m s the sum of the traffc generated by the nodes n the shaded regon. The area of shaded regon s S = s h πd 2 ( < D < h), (24) where D s the dstance between the BS relay node m s h s the area sze of a cell. Therefore, the expected value of traffc passng through node m s W σ (s h πd 2 ) ( < D < h), (25) where W s the data transmsson rate of a source node, σ s the densty of the nodes n a regon. Equaton (25) shows that the traffc passng through node m decreases as D ncreases. That s, the nodes closer to the BS have a hgher load than the nodes stayng at the brm of the cell. Proposton 4.4: In a hybrd wreless network, acheves more balanced load dstrbuton among the moble nodes n each cell. Proof: The shaded regon n Fgure 5 (b) represents all possble postons of the source relay nodes that choose node m as relay node. Suppose m neghbor nodes are chosen as relay nodes, then the expected traffc passng through node m s W m σ πr2 whch shows that the traffc gong through node m s ndependent of ts locaton relatve to ts BS. Snce every node n the cell has an equal probablty of generatng traffc, the traffc load s balanced among the nodes n the cell. 5 PERFORMANCE EVALUATION Ths secton demonstrates the propertes of through smulatons on NS-2 [36] n comparson to DHybrd [7], [9] AODV [8]. In DHybrd, a node frst uses broadcastng to observe a mult-hop path to ts own BS then forwards a message n the ad-hoc transmsson mode along the path. Durng the routng process, f the transmsson rate (.e., bwdth) of the next hop to the BS s lower than a threshold, rather than forwardng the message to the neghbor, the node forwards the message drectly to ts BS. The source node wll be notfed f an establshed path s broken durng data transmsson. If a source sends a message to the same destnaton next tme, t uses the prevously establshed path f t s not broken. In the protocol, a source node selects the better transmsson mode between drect transmsson relay transmsson. If the source node can fnd a neghbor that has hgher bwdth to the (c) 25 IEEE. Translatons content mnng are permtted for academc research only. Personal use s also permtted, but republcaton/redstrbuton requres IEEE permsson. See for more nformaton.

10 Ths artcle has been accepted for publcaton n a future ssue of ths journal, but has not been fully edted. Content may change pror to fnal publcaton. Ctaton nformaton: DOI.9/TMC , IEEE Transactons on Moble Computng Througput per S-D par (kbps) DHybrd AODV Network sze Throughput per S-D par (kbps) DHybrd AODV Number of base statons Network sze Fg. 6: Throughput vs. networkfg. 7: Throughput vs. number Fg. 8: Delay vs. network sze. sze (smulaton). of BSes. BS than tself, t transmts the message to the neghbor. Otherwse, t drectly transmts the message to the BS. Unless otherwse specfed, the smulated network conssts of 5 moble nodes 4 BSes. In the ad-hoc component of the hybrd wreless network, moble nodes are romly deployed around the BSes n a feld of square meters. We used the Dstrbuted Coordnaton Functon (DCF) of the IEEE 82. as the MAC layer protocol. The transmsson range of the cellular nterface was set to 25 meters, the raw physcal lnk bwdth was set to 2Mbts/s. The transmsson power of the ad-hoc nterface was set to the mnmum value requred to keep the network connected for most tmes, even when nodes are n moton n the network. Then, the nfluence of the transmsson range on dfferent methods performance s controlled. Specfcally, we set the transmsson range through the ad-hoc nterface to.5 tmes of the average dstance between neghborng nodes, whch can be obtaned by measurng the smulated network. We used the two-ray propagaton model for the physcal layer model. Constant bt rate (CBR) was selected as the traffc mode n the experment wth a rate of 64kbps. In the experment, we romly chose 4 source nodes to contnuously send messages to romly chosen destnaton nodes. The number of channels for each BS was set to. We set the number of redundant routng paths b n Secton 3.4 to. We assumed that there was no capacty degradaton durng transmsson between BSes. Ths assumpton s realstc consderng the advanced technologes hardware presently used n wred nfrastructure networks. There was no message retransmsson for faled transmssons n the experments. We employed the rom way-pont moblty model [37] to generate the movng drecton, speed, pause duraton of each node. In ths model, each node moves to a rom poston wth a speed romly chosen from ( 2)m/s. The pause tme of each node was set to. We set the number of segments of a message to the connecton degree of the source node. The smulaton warmup tme was set to s the smulaton tme was set to s. We conducted the experments 5 tmes used the average value as the fnal expermental result. To make the methods comparable, we dd not use the congeston control algorthm n unless otherwse ndcated. 5. Scalablty Fgure 6 shows the average throughput measured n kbps per S-D par of dfferent routng protocols versus the number of moble nodes n the system. The fgure shows the throughput of remans almost the same wth dfferent network szes. Ths result conforms to Corollary 4.. uses dstrbuted mult-path routng to fully take advantage of the spatal reuse avod transmsson congeston n a sngle path. Unlke the mult-hop Delay per S-D par (s) DHybrd AODV Delay per S-D par (s) DHybrd AODV Number of base statons Fg. 9: Delay vs. number of B- Ses. routng n moble ad-hoc networks, does not need path query mantenance. Also, t lmts the path length to three to avod problems n long-path transmsson. The throughput of DHybrd AODV decreases as the number of nodes n the network ncreases. Ths s manly because when the network sze ncreases, more beacon messages are generated n the network. Also, the long transmsson path also leads to hgh transmsson nterference. Then, nodes n these methods suffer from ntense nterference, leadng to more transmsson falure degraded overall throughput. Also, the moble node ncrease n the system leads to hgh network dynamsm, resultng n frequent route re-establshments. The short routng paths n reduce congeston sgnal nterference, thus enablng better spatal reuse as n. Meanwhle, enables nodes to adaptvely swtch between drect transmsson relay transmsson. Hence, part of the transmsson load s transferred to relay nodes, whch carry the messages untl meetng the BSes. As a result, gateway nodes connectng moble nodes BSes are not easly overloaded. Therefore, the throughput of s hgher than DHybrd. However, snce the number of message routng hops s confned to one, may not fnd the node wth the best transmsson rate to the BSes because of the short transmsson range of the ad-hoc nterface. Therefore, the throughput of s lower than, especally n a network wth hgh node densty. The reason that AODV has the lowest throughput per S-D par s ts long transmsson paths. Fgure 7 shows the throughput per S-D par versus the number of BSes n dfferent routng protocols. The number of BSes was vared from 3 to 6. The BSes are unformly dstrbuted n the network. We can see from the fgure that as the number of BSes ncreases, the throughputs of,, DHybrd ncrease whle the throughput of AODV stays nearly constant. In,, DHybrd, as the number of BSes ncreases, the total number of nodes close to the BSes ncreases. Then, more nodes have hgh transmsson rates to the BSes, leadng to a throughput ncrease. In AODV, snce the traffc between S-D pars does not travel though BSes, the throughput between an S-D par s not affected by the ncreased number of BSes n the network. The fgure also shows that the throughput of s constantly larger than the throughput of s constantly larger than DHybrd. AODV constantly has the lowest transmsson delay due to the same reasons as n Fgure Transmsson Delay Fgure 8 shows the average transmsson delay of S- D pars for successfully delvered messages n dfferent routng protocols versus network sze. The network sze was vared from 2 to wth 2 ncrease n each step (c) 25 IEEE. Translatons content mnng are permtted for academc research only. Personal use s also permtted, but republcaton/redstrbuton requres IEEE permsson. See for more nformaton.

11 Ths artcle has been accepted for publcaton n a future ssue of ths journal, but has not been fully edted. Content may change pror to fnal publcaton. Ctaton nformaton: DOI.9/TMC , IEEE Transactons on Moble Computng Communcaton overhead (kpbs) DHybrd AODV Network sze Fg. : Overhead vs. network sze Total throughput (kbps) DHybrd 5 AODV Node movng speed (m/s) Fg. : Throughput vs. moblty. Total thorughput of base statons (kbps) DHybrd Number of source nodes Fg. 2: Throughput of BSes vs. number of source nodes. Transmsson delay s the amount of tme t takes for a message to be transmtted from ts source node to ts destnaton node. From the fgure, we see that generates the smallest delay. In, each source node frst dvdes ts messages nto smaller segments then forwards them to the nearby nodes wth the hghest capacty, whch leads to more balanced transmsson load dstrbuton among nodes than the prevous methods. Accordng to Proposton 4., average latency can be mnmzed when the transmsson loads of all the nodes are balanced. Hence, has smaller latency than the prevous methods. The delay of DHybrd s 5-6 tmes larger than. DHybrd uses a sngle transmsson path, whle uses multple paths. Recall that we set the number of segments of a message to the connecton degree of the source node n. Thus, the rato of delay tme of DHybrd to that of equals the average connecton degree. As the number of nodes n the system ncreases, the connecton degree of each node ncreases, the ncrease rate of the rato grows. Ths s caused by two reasons. Frst, a hgher node densty leads to longer path lengths n DHybrd, resultng n a longer delay because of a hgher lkelhood of lnk breaks. Second, a hgher node densty enables a node to quckly fnd relay nodes to forward messages n, as ndcated n Proposton 4.2. also produces a shorter transmsson delay than for two reasons. Frst, the mult-path parallel routng of saves much transmsson tme as shown n Proposton 4.. Second, the dstrbuted routng of enables some messages to be forwarded to the destnaton BS s neghborng cells wth hgh transmsson rates rather than watng n the current hot cell for a transmsson channel. We can also observe that produces lower delay than DHybrd. Ths s because the delay of DHybrd ncludes the tme for establshng a path for data transmsson. Also, the mult-hop transmsson component of DHybrd results n a hgher delay due to the queung delay n each hop. Because of the long dstance transmssons wthout support from an nfrastructure network, AODV generates the longest delay. Fgure 9 plots the average communcaton delay per S-D par for successfully delvered messages versus the number of BSes n dfferent routng protocols. The fgure shows that the ncreasng number of BSes n the system leads to a communcaton delay decrease between nodes n,, DHybrd, but does not affect the communcaton delay n AODV. In,, DHybrd, as the number of BSes ncreases, more nodes can stay close to the BSes, leadng to fewer communcaton hops better transmsson lnks between nodes BSes. Thus, the transmsson delay between the nodes s reduced. Snce the communcaton between S-D pars n AODV does not rely on BSes, AODV mantans constant communcaton delay. The fgure also shows that the communcaton delay between S-D pars follows <<DHybrd<AODV for the same reason as n Fgure Communcaton Overhead We use the generaton rate of control messages n the network MAC layers n kbps to represent the communcaton overhead of the routng protocols. Fgure llustrates the communcaton overhead of, Twohop, DHybrd, AODV versus network sze. We can see that the communcaton overheads of are very close. Ths s because both are transmsson protocols of short dstance small hops. has slghtly hgher communcaton overhead than because utlzes three hop transmsson, whch has one more hop than two hop transmsson. However, the margnal overhead ncrease leads to a much hgher transmsson throughput as shown n Fgure 6. DHybrd generates much hgher overhead than because of the hgh overhead of routng path queryng. The pure AODV routng protocol results n much more overhead than the others. Ths s because wthout an nfrastructure network, the messages n AODV travel a long way from the source node to the destnaton node through much longer paths. 5.4 Effect of Moblty In order to see how the node moblty nfluences the performance of the routng protocols, we evaluated the throughput of these four transmsson protocols wth dfferent node mobltes. Fgure plots the throughput of, DHybrd,, AODV versus node movng speed. From the fgure, we can see that the ncreasng moblty of the nodes does not adversely affect the performance of. It s ntrgung to fnd that hgh moblty can even help to ncrease ts throughput that generates constant throughput regardless of the moblty. Ths s because the transmsson modes do not need to query rely on mult-hop paths; thus, they are not affected by the network partton topology changes. Moreover, snce transmts segments of a message n a dstrbuted manner, as the moblty ncreases, a moble node can meet more nodes n a shorter tme perod. Therefore, enables the segments to be quckly sent to hgh-capacty nodes. As node moblty ncreases, the throughput of DHybrd decreases. In DHybrd, the messages are routed n a mult-hop fashon. When the lnks between nodes are broken because of node moblty, the messages are dropped. Therefore, when nodes have smaller moblty, the lnks between the moble nodes last longer more messages can be transmtted. Hence, the throughput of DHybrd s adversely affected by node moblty. However, snce DHybrd can adaptvely adjust (c) 25 IEEE. Translatons content mnng are permtted for academc research only. Personal use s also permtted, but republcaton/redstrbuton requres IEEE permsson. See for more nformaton.

12 Ths artcle has been accepted for publcaton n a future ssue of ths journal, but has not been fully edted. Content may change pror to fnal publcaton. Ctaton nformaton: DOI.9/TMC , IEEE Transactons on Moble Computng 2 Throughput per S-D par (kbps) hop 2 hops 3 hops 4 hops Network sze The dstance from a moble node to ts base staton Fg. 3: Throughput vs. number of Fg. 4: Load dstrbuton n a cell. hops. the routng between the ad-hoc transmsson cellular transmsson, the throughput of DHybrd s much hgher than AODV s. Wth no nfrastructure network, AODV produces much lower throughput than the others. Its throughput also drops as node moblty ncreases for the same reasons as DHybrd. Average load per node (kbps) 5.5 Effect of Workload We measured the total throughput of BSes on the messages receved by BSes. Fgure 2 shows the total throughput of the BSes versus the number of source nodes. We can see that have much hgher throughput ncrease rates than DHybrd. Ths s because n, the number of transmsson hops from a source node to a BS s small. Meanwhle, each node can adaptvely swtch between relay transmsson drect transmsson based on the transmsson rate of ts neghbors. Hence, part of a source node s transmsson load s transferred to a few relay nodes, whch carry the messages untl meetng the BSes. Therefore, the gateway moble nodes are less lkely to be congested. However, nodes n DHybrd cannot adaptvely adjust the next forwardng hop because t s predetermned n the routng path. Messages are always forwarded to the moble gateway nodes that are closer to the BSes or that have hgher transmsson rates. Therefore, these moble gateway nodes can easly become congested as the workload of the system ncreases, leadng to many message drops. Therefore, when the number of the source nodes s larger than 4, the throughput of DHybrd remans nearly constant. Ths s also the reason that the throughput of DHybrd s constantly lower than those of. Addtonally, the fgure shows that the overall throughput of s lower than that of. Ths s because most of the traffc n s confned to a sngle cell. When a BS n a cell s congested, the traffc cannot be transferred to other cells. In contrast, s three-hop dstrbuted forwardng mechansm enables t to dstrbute the traffc among the BSes n a balance. Therefore, the BSes n wll not become congested easly. In addton, as the forwardng mechansm gves nodes more flexblty n choosng relay nodes wth hgher transmsson rates for message forwardng to the BSes, the overall BS throughput n s larger than n. 5.6 Effect of the Number of Routng Hops We conducted experments to show the optmal number of routng hops for the routng n hybrd wreless networks. We tested the throughput per S-D par for x- hop, where x was vared from to 4. In the -hop routng, a node drectly transmts a message to the BS wthout message dvson. In the other routng protocols, the (x ) th hop chooses the best transmsson mode between drect transmsson relay transmsson. Also, n the 4-hop routng, the second relay node romly chooses the thrd relay node DHybrd Throughput of a base staton (kbps) DHybrd Rank of base statons Fg. 5: Load dstrbuton among BSes. Fgure 3 shows the average throughput per S-D par versus network sze n. As the fgure shows, as the network sze ncreases, the node throughput keeps constant regardless of the number of forwardng hops n a routng. The reason s the same as n Fgure 6. We can also see from the fgure that the throughput of the four protocols follows 3-hop>4-hop>2-hop>-hop. In the -hop routng, each node only transmts segments drectly to a BS regardless of ts current transmsson rate. In the 2-hop routng, f the transmsson rate of a node s neghbor s hgher than that of the node, t asks ts neghbor node to forward the segment to a BS. Therefore, the 2-hop routng has hgher throughput than the -hop routng. The 3-hop routng can greatly ncrease the number of node optons for segment routng snce the number of nodes that the source node can encounter ncreases from d to d 2, where d s the average node degree. Thus, a node wth a greater transmsson rate can be chosen as the forwardng node. Meanwhle, the 3-hop routng can greatly facltate nter-cell communcaton because a node has a hgher probablty of reachng a neghborng BS wthn a 3-hop path length than wthn a 2-hop path length. Therefore, the throughput of the 3-hop routng s much hgher than that of the 2-hop routng. The fgure also shows that the 4-hop routng produces lower throughput than the 3-hop routng. The reason s that 3 hops are enough to fnd a hop wth hgh transmsson rate acheve nter-cell communcaton because of wdespread BSes. The 4-hop routng ncreases the forwardng delay due to the greater number of nodes n a route; thus, t cannot ncrease the uploadng transmsson rate of messages. 5.7 Load Dstrbuton Wthn a Cell In ths experment, we tested the load dstrbuton of moble nodes n a romly chosen cell n the hybrd wreless network that employs each of the, DHybrd, protocols. We normalzed the dstance from a moble node to ts base staton accordng to the functon D R b, where D s the actual dstance R b s the radus of ts cell. We dvded the space of the cell nto several concentrc crcles measured the loads of the nodes on each crcle to show the load dstrbuton. Fgure 4 shows the average load of a node correspondng to the normalzed dstance from tself to the BS n the chosen cell. The fgure shows that most of the traffc load of DHybrd s located at nodes near the BS. The nodes far from the BS have very low load. The results conform to Proposton 4.3. In DHybrd, f a source node wants to access the Internet backbone or engage n nter-cell communcaton, t transmts the messages to the BSes n a mult-hop fashon. Therefore, the nodes near the BSes wll have the hghest load. On the other h, snce there s lttle traffc gong through the nodes at the brm of a cell, the load of these nodes s small. As a result, some nodes can easly become hot spots whle the resources of other nodes are not (c) 25 IEEE. Translatons content mnng are permtted for academc research only. Personal use s also permtted, but republcaton/redstrbuton requres IEEE permsson. See for more nformaton.

13 Ths artcle has been accepted for publcaton n a future ssue of ths journal, but has not been fully edted. Content may change pror to fnal publcaton. Ctaton nformaton: DOI.9/TMC , IEEE Transactons on Moble Computng 3 Throughput of a base staton (kbps) Rank 2 Rank 2 Rank 3 Rank Smulaton tme (s) (a) Throughput of a base staton (kpbs) fully utlzed. Ths load mbalance prevents DHybrd from fully utlzng system resources. The traffc load of s almost evenly dstrbuted n the system, whch s n lne wth Proposton 4.4. In, the traffc from a source node s dstrbuted among a number of relay neghbors for further data forwardng. The nodes at the brm of the cell also take responsblty for message forwardng, snce the neghbor nodes of the brm nodes could be located n other cells wth good transmsson channels. In, the source node consders drect transmsson or one-hop relay transmsson based on the channel condton. Snce the node s chosen wthn one hop, the messages wll not gather close to the BS due to the lmted transmsson range. However, because of ts sequental transmsson, cannot acheve load balance among nodes n a cell as well as. 5.8 Load Balance Between Cells In ths experment, we tested the effectveness of the congeston control algorthm n. We also added a congeston control algorthm to DHybrd. In the algorthm, when a node receves beacon messages from ts BS ndcatng that t s overloaded, the node broadcasts a query message to fnd a path to a nearby uncongested BS. We selected two BSes out of the total four BSes. In the range of each of the two selected BSes, we romly selected one moble node as the source node to send messages to a romly selected destnaton node n the network. Once the source node moves out of the range of the selected BS, another moble node n the range was selected as the source node. In order to show the load dstrbuton of the BSes n dfferent protocols, we ranked the BSes based on BS throughput. The BS wth the hghest throughput has a rank of. Fgure 5 shows the throughput of each BS versus the BS rank. We can see from the fgure that n, the throughput of the frst two BSes s extremely hgh whle the throughput of the last two BSes s extremely small. Ths s because the two hop routng path length n s not long enough to forward messages from a congested BS to a lghtly loaded BS. Therefore, the traffc cannot be shfted to the neghborng lghtly loaded BSes, leadng to an unbalanced load dstrbuton. We can also see from the fgure that n, the varance of the throughputs n dfferent BSes s small. The reason s that three forwardng hops are enough for a moble node to reach a neghborng BS hence to balance the load between the BSes. Meanwhle, the congeston control algorthm n can effectvely swtch the traffc from a hghly loaded cell to a lghtly loaded cell. Because the BSes of ranks 2 n are not congested, ther throughput s less than the correspondng BSes n Twohop; also, the throughput of the BSes of ranks 3 4 n s much hgher than that of the correspondng BSes n. DHybrd acheves more balanced load dstrbuton between BSes than snce t employs Rank Rank 2 Rank 3 Rank Smulaton tme (s) (b) Fg. 6: Base staton load vs. smulaton tme. Throughput of a base staton (kbps) Rank Rank 2 Rank 3 Rank Smulaton tme (s) (c) DHybrd a congeston control algorthm. In DHybrd, f a prevously establshed path to a destnaton s not broken, a node stll uses ths path to transmt messages to the same destnaton. Thus, the nodes cannot dynamcally balance load between BSes. Also, when a node fnds that ts current BS s congested, t takes a long tme for t to fnd a path to a non-congested BS by re-ssung a query message to the neghborng non-congested BS, whch greatly reduces the throughput of the system. Fgure 6 further shows the throughput of the BSes versus smulaton tme n the three routng protocols. At the begnnng, the BSes wth ranks 2 are congested those wth ranks 3 4 do not have much traffc. Thus, the three fgures show that the BSes wth ranks 2 have hgh throughput but those wth ranks 3 4 have extremely low throughput at the begnnng n all three protocols. Fgure 6 (a) shows the throughput of the BSes n. As shown n the fgure, snce can adaptvely adjust the traffc among the BSes usng ts congeston control algorthm, the throughput of the two hghly congested BSes s dstrbuted to the neghborng BSes. As the traffc s forwarded from the BSes of ranks 2 to the BSes of ranks 3 4, the throughputs of these BSes are very smlar later n the smulaton. Ths result ndcates the effectveness of the congeston control algorthm n for load balance between cells. Fgure 6 (b) shows the throughput of the BSes n. In, snce the source nodes cannot effectvely move the traffc between BSes, the BSes wth rank rank 2 constantly have the hghest throughput, whle the BSes wth rank 3 rank 4 constantly have low throughput. The low throughput s produced when the mmedate neghbors of the source node are n the range of the neghborng BSes of the source node s BS. However, the probablty of such cases s very small. Fgure 6 (c) shows the throughput of the BSes n DHybrd. As the nodes n DHybrd cannot effectvely balance the load between the BSes, the throughput of the BSes of rank rank 2 s much larger than that of the BSes of rank 3 rank 4. Comparng Fgure 6 (b) Fgure 6 (c), we can fnd that the throughput n DHybrd s lower than that n. Ths s because the mult-hop transmsson n the ad-hoc network n DHybrd greatly reduces the throughput. Meanwhle, the moble gateway nodes n DHybrd easly become congested, leadng to more message drops. 6 CONCLUSIONS Hybrd wreless networks have been recevng ncreasng attenton n recent years. A hybrd wreless network combnng an nfrastructure wreless network a moble ad-hoc network leverages ther advantages to ncrease the throughput capacty of the system. However, current hybrd wreless networks smply combne the routng protocols n the two types of networks for (c) 25 IEEE. Translatons content mnng are permtted for academc research only. Personal use s also permtted, but republcaton/redstrbuton requres IEEE permsson. See for more nformaton.

14 Ths artcle has been accepted for publcaton n a future ssue of ths journal, but has not been fully edted. Content may change pror to fnal publcaton. Ctaton nformaton: DOI.9/TMC , IEEE Transactons on Moble Computng [7] Y. Yao, X. Tang, E. Lm, In-network processng of nearest [34] In how many ways can m balls be dstrbuted nto n boxes? 4 neghbor queres for wreless sensor networks, n Proc. of DAShttp:// otekman/dsc/usef.pdf. FAA6, 26. [35] G. Wang, G. Cao, T. L. Porta, W. Zhang, Sensor relocaton n moble sensor networks, n Proc. of IEEE INFOCOM, 25. [8] R. Szewczyk, J. Polastre, A. Manwarng, D. Culler, Lessons [36] H. L. Hu D. Evans, moble sensor networks, from a Sensor Network Proc. from of EWSN, 24. data transmsson, whchexpedton, prevents n them achevng [8] Y. Hseh R. Localzaton Svakumar. for A hybrd network model for n Proc. of packet MobCom, 24. [9] C. Intanagonwwat, R. Govndan, D. Estrn, Drected Dffuwreless data networks. In Proc. of GLOBECOM, 22. hgher system capacty. In ths paper, we propose a [37] Y. F. Lu, X. Cheng, D. Hua, D. Chen, Locaton dscovery son: A Scalable Robust Communcaton Paradgm for Sensor [9] We D. Gtln. -relay archtecture for for nextdstrbuted Three-hop () data routng prosensor networks wth short range beacons, IJAHUC, 29. CommuNetworks, n Proc. ofroutng Mobcom, 2. generaton WWAN/WLAN ntegraton. IEEE Wreless tocol that ntegrates the dual features of hybrd wreless [38] R. Fonseca, S. Ratnasamy, J. Zhao, C. T. Ee, Beacon vector [] W. Zhang, G. Cao, T. L. Porta, Data Dssemnaton wth ncaton, 24. pont-to-pont routng n wrelesscellular sensornets, n X. J. L, scalable B. C. Seet, P. H. J. Chong. Multhop networks: networks n theindex data process.ninproc., a [2] routng: Rng-Based for transmsson Wreless Sensor Networks, of ICNP, Technology economcs. Computer Networks, 28. Proc. of NSDI, 25. pp.dvdes a message stream nto segments source23, node Zhang, X. Du. resource allocaton [39] B. N.Bengfort, Bulusu, J.W.Hedemann, D. Effcent Estrn, Gps-Less Low-Cost n [] S. Madden, M. J.ts Frankln, neghbors, J. M. H. W. Hong, TAG: a Tny [2] transmts them to moble whch further hybrd wreless networks. InSmall Proc. Devces, of WCNC,IEEE 2. Outdoor Localzaton For Very Personal ComAGgregaton Servce for Ad-Hoc Sensor Networks, n Proc. of forward the segments to ther destnaton through an [22] muncatons P. Thulasraman vol. X. Shen. Magazne, 7, no. Interference 5, pp , aware 2. resource allocaosdi, 22. ton hybrd herarchcal wreless networks. Computer Networks, nfrastructure network. lmts thea Robust routng Thefor one smulator. [2] F. Ye G. Zhong, GRAdent Broadcast: Datapath Delv- [4] 54(3): , lengthery to Protocol three, for always arranges for hgh-capacty [4] H. Shen, T. L, L.2. Zhao, Z. L, SDS: Dstrbuted SpatalLarge Scale Sensor Networks, WINET, 25. [23] K. Akkarajtsakul, E. Hossan, D. Nyato. Cooperatve packet Temporal Smlarty Data Storage n Wreless Networks, nodes forward data. S.Unlke most exstng routng [3] to H. Luo, F. Ye, J. Cheng, Lu, L. Zhang, Ttdd: Two-ter data delvery n hybrd wreless moble networks:sensor A coaltonal game n Proc. of ICCCN, 29. Mob. Comput., 2(5):84 854, 23. dssemnaton n large-scale sensor networks, Wreless Networks, protocols, produces sgnfcantly lower overhead approach. IEEE Trans. vol., pp. 6 75, by elmnatng route 22. dscovery mantenance. In [24] T. Lu, M. Rong, P. L, D. Yu, Y. Xue, E. Schulz. Rado resource allocaton n two-hop cellular relayng network. In Proc. of VTC, [4] S. Ratnasamy, B. Karp, S. Shenker, D. Estrn, R.of Grovndan, addton, ts dstngushng characterstcs short 26. L. Yn, F. Yu, Data-centrc storage n sensornet wth ght: A path length, short-dstance transmsson, balanced [25] T. Lu, M. Rong, H. Sh, D. Yu, Y. Xue, E. Schulz. Reuse geographc hash table, n Proc. of MONET, 23. parttonng n fxed two-hop cellular relayng network. In Proc. load[5]dstrbuton S. Ratnasamy, provde B. Karp, S. hgh Shenker,routng D. Estrn,relablty L. Yn, Dataof WCNC, 26. effcency. also has a congeston centrc storage n sensornets wth GHT, a control geographcalgorthm hash table, [26] L. Guan, J. Zhang, J. L, G. Lu,receved P. Zhang. Spectral Hayng Shen the BS degree n effcent ComMONET, vol. 8, pp , 23. to avod load congeston n BSes n the case of frequency allocaton scheme n multhop cellular network. puter Scence Engneerng from Tongj Un- In [6] X. L, Y. J. Km,dstrbutons W. Hong, Mult-dmensonal range queres unbalanced traffc n networks. Theoretcal Proc. of VTC, 27. versty, Chna n 2, the MS Ph.D. n sensor n Proc. of SenSys, 23. that can [27] D. M. Shla, Y. Cheng, T. Anjal. Throughput delay analyss networks, smulaton results show n networks Computer wth Engneerng from Wayne In [7] D. Ganesan, DIMENSIONS: Why do we need a new data analyss of hybrddegrees wreless mult-hop uplnks. dramatcally mprove the throughput capacty State Unversty n 24 26, respectvely. hlng archtecture for sensor networks, n Proc. of the ACM Proc. of INFOCOM, 2. scalablty of 22, hybrd wreless networks due to ts hgh [28] B. Lu, P. Thran, She Assstant thehoc HotNets, pp s D.currently Towsley.anCapacty of Professor a wrelessnad Department In of Proc. Electrcal Computer scalablty, effcency, relablty W. lowye,overhead. [8] D. Ganesan, A. Cerpa, Y. Yu, D. Estrn, J. Zhao, network wth nfrastructure. of Mobhoc, 27. Engthe Drector of Multcast the Pervasve Com[29] C. Wang, X. L, C.neerng, Jang, S. Tang, Y. Lu. throughput Networkng ssues n wreless sensor networks, JPDC, 24. muncatons of Clemson Unversty. for hybrd wreless networkslaboratory under gaussan channel model. [9] B. Greensten, D. Estrn, R. Govndan, S. Ratnasamy, ACKNOWLEDGEMENTS Her 2. research nterests nclude dstrbuted comtmc, (6): , S. Shenker, Dfs: A dstrbuted ndex for features n sensor net[3] A. G. Valko. Cellular p:systems A new approach to nternet hostwth moblty. puter computer networks, an Ths research works, nwas Proc.supported of SNPA, 23.n part by U.S. NSF grants ACM Computer Commncaton, 999. IIS-35423, CNS-2546, [2] J. L, J. Jannott, D. S. J. De, CNS-24963, C. Davd, R. Karger,CNS-49947, R. Morrs, A emphass on peer-to-peer content delvery networks, moble comc. wreless Sarr, C. sensor Chaudet, G. Chelus, computng. I. G. Lassous. A Nodeputng, networks, grd Her research CNS-9756 CNS-25652, Mcrosoft Research Fac-of [3] scalable locaton servce for geographc ad hoc routng, n Proc. Based Avalable Bwdth Evaluaton In IEEE 82. Ad Hoc work has been publshed n top journals conferences n these MobCom, 2. ulty Fellowshp 8375, the Unted States Depart- areas.networks. IJPEDS, (): 2, 25. She was the Program Co-Char for a number of nternatonal [2]of J. Newsome D. Song,We GEM: Graphlke EMbeddng for routng X. P. Costa, M. T. Moreno, H. Hartensten. A smulaton study ment Defense would to thank Mr. [32] conferences member of Program Commttees of many leadng on the performance ofthe herarchcal moble pv6. In Proc. of ITC, data-centrc storage n sensor networks wthout geographc Kang Chen for hs help n addressng revew comments. conferences. 23. She s a member of the IEEE ACM. She s Mcrosoft nformaton, n Proc. of SenSys, 23. Research Faculty Fellow of 2. [22] A. Caruso, S. Chessa, S. De, R. Urp, GPS free coordnate [33] P. Gupta P. R. Kumar. The capacty of wreless networks. IEEE TIT, 2. assgnment routng n wreless sensor networks, n Proc. of R EFERENCES [34] L. B. Koralov Y. G. Sna. Theory of probablty rom IEEE R. INFOCOM, pp.l.5 6. [] H Luo, Ramjee, P.25, Snha, L, S. Lu. Ucan: A unfed cell processes. Berln New York Sprnger, 27. [23] Desnoyers, D. archtecture. Ganesan, Shenoy, Tsar: A 23. two ter [35] H. Y. Hseh R. Svakumar. Performance comparson of P.ad-hoc network In P.Proc. of MOBICOM, sensor storageh.archtecture usng nterval skp graphs, n Proc. [2] P. K. McKnley, Xu, A. H. Esfahanan, L. M. N. Uncastcellular mult-hop wreless networks: A quanttatve study. of SenSys5. Press, 25, n pp.wormhole-routed based multcast ACM communcaton drect netin Proc. of SIGMETRIC, 2. [24] C. T.TPDS, Ee 992. S. Ratnasamy, Practcal data-centrc storage, n Proc. [36] The network smulator - ns-2. works. Lanyu Zhao receved the BS MS degrees [3] H. of Wu, C. Qao, NSDI, 26. S. De, O. Tonguz. Integrated cell ad hoc [37] M. Grossglauser n D. Tse. Scence Moblty ncreases the capacty Computer from Jln Unversty, Chna. of relayng systems: CAR. J-SAC, 2. ad hoc wreless networks. In Proc. of TON, 22.n the Depart[25] M. Aly, K. Pruhs, P. K. Chrysanths, KDDCS: A loadhe s currently a Ph.D. student [4] Y. H. Tam, H. S. Hassanen, S. G. Akl, R. Benkocz. Optmal balanced n-network data-centrc storage scheme n sensor nethayng Shen receved the BS degree n Comment of Electrcal Computer Engneerng mult-hop work, ncellular Proc. ofarchtecture CIKM, 26,for pp.wreless communcatons. In puter Scence Engneerng fromnterests Tongj Uof Clemson Unversty. Hs research Proc. of LCN, 26. [26] F. Ban, X. L, R. Govndan, S. Schenker, Usng herarchcal nversty, Chna sensor n 2, theroutng MS Ph.D. nclude wreless network, proto[5] Y. D. Ln Y. C. Hsu. Mult-hop cellular: A new archtecture degrees n Computer Engneerng from for scalable routng rendezvous n wreless cols, applcatons securty ssues n Wayne P2P forlocaton wrelessnames communcatons. In Proc. of INFOCOM, 2. State Unversty n 24 26, respectvely. sensor networks, Proc. SenSys,Connectvty 24, pp networks. [6] P. T. Olver, Dousse, n M. ofhasler. n ad hoc She s currently an Assocate Professor n the [27] J. Xu, X. Tang,In W.ofchen Lee, A22. new storage scheme for hybrd networks. Proc. INFOCOM, Department of Electrcal Computer Englocaton queres n object trackng sensordestnaton networks, [7] E. approxmate P. Charles P. Bhagwat. Hghly dynamc neerng at Clemson Unversty. Her research IEEE TPDS, vol. 9, pp , 28. for moble computers. sequenced dstance vector routng (DSDV) nterests nclude dstrbuted computer systems [28] M. L Y. Lu, 994. Rendered path: range-free localzaton n In Proc. of SIGCOMM, computer networks, wth an emphass on networks wth n Proc. MobCom, [8] C. ansotropc Perkns, E.sensor Beldng-Royer, holes, S. Das. RFCof 356: Ad 27. hoc content delvery networks, moble computng, on dem dstance vector (AODV) TechncalSearchng report, wreless sensor networks, cloud. She s [29] H. Shen, T. L, T. Schweger, An routng. Effcent Smlarty a Mcrosoft Faculty Fellow of 2, a senor Internet Engneerng Task Force, 23. Hashng, n Proc. of ICDT, Scheme Based on Localty Senstve member of the IEEE a member of the ACM. [9] D. 28. B. Johnson D. A. Maltz. Dynamc source routng n ad hoc D. wreless IEEE Computng, 996. ZeLLreceved receved BS degree n Electroncs [3] Karger,networks. E. Lehman, T. Moble Leghton, M. Levne, D. Lewn, Ze the the BS degree n Electroncs [] V. R. D. Pangrahy, Park Consstent M. Scott Corson. hghly adaptve Informaton Engneerng from Huazhong HashngA Rom Trees:dstrbuted Dstrbuted Informaton Engneerng from Huazhong Unver- Uroutng algorthm for moble wreless networks. In Proc. of nversty of Scence Technology, Chna, n Cachng Protocols for Relevng Hot Spots on the World Wde sty of Scence Technology, Chna, n 27. of INFOCOM, the Ph.D. degree n the Department Web, n Proc. STOC, 997, He s currently Ph.D. student n the Depart[] R. S. Chang, W. Y.ofChen, Y. pp. F. Wen. Hybrd wreless network Electrcal a Computer Engneerng of Clem[3] W. Nejdl, W.Transacton Sbersk, M. C. Schmntz, ment of Electrcal Computer Engneerng protocols. IEEE on Wolpers, Vehcular Technology, 23. Routng son Unversty, n 22. Hs research nterests clusterng n R. schema-based super peer networks, of Clemson Unversty. Hs research nclude dstrbuted networks, wth annterests emphass [2] G. N. Aggelou Tafazoll. On the relayng capactynofproc. next-of IPTPS, 23. on content delverynetworks, networks, wreless mult-hop nclude dstrbuted wth an emphass generaton gsm cellular networks. IEEE Personal Communcatons cellular networks, theory data mnng. [32] P. A. Bernsten, F. Gunchgla, A. Kementsetsds, J. Mylopoulos, Magazne, 2. on peer-to-peer game content delvery networks, Serafn, I. Zahrayeu, Data management for peer-to-peer [3] T. L. Rouse, I. B, S. McLaughln. Capacty power wreless mult-hop cellular networks, game thenvestgaton opportunty multple computng: of A vson, n Proc.drven of WebDB, 22. access (ODMA) ory data mnng. He s a student member of networks n TDD-CDMA systems. Proc. of ICC, 22. [33] A. Y. Halevy, Z. G. Ives,based P. Mork, I.In Tatarnov, Pazza: Data IEEE. Chenx Qu receved the BS degree n Telecom[4] H. management Y. Hseh nfrastructure R. Svakumar. On Usng the Ad-hoc Network for semantc web applcatons, n Proc. muncaton Engneerng from Xdan Unversty, Model n Wreless Packet Data Networks. In Proc. of MOBIHOC, Chna, n 29. He currently s a Ph.D. student of WWW, n the Department of Electrcal Computer [5] L. M. Feeney, B. Cetn, D. Hollos, M. Kubsch, S. Mengesha, Engneerng at Clemson Unversty, SC, Unted H. Karl. Mult-rate relayng for performance mprovement n eee States. Hs research nterests nclude sensor 82. wlans. In Proc. of WWIC, 27. networks wreless networks. [6] J. Cho Z. J. Haas. On the throughput enhancement of the downstream channel n cellular rado networks through multhop relayng. IEEE JSAC, 24. [7] B. Lu, Z. Lu, D. Towsley. On the capacty of hybrd wreless networks. In Proc. of INFOCOM, (c) 25 IEEE. Translatons content mnng are permtted for academc research only. Personal use s also permtted, but republcaton/redstrbuton requres IEEE permsson. See for more nformaton.