Multi-Channel Wireless Networks: Capacity and Protocols

Transcription

1 Multi-Channel Wireless Networks: Capaity and Protools Tehnial Report April 2005 Pradeep Kyasanur Dept. of Computer Siene, and Coordinated Siene Laboratory, University of Illinois at Urbana-Champaign Nitin H. Vaidya Dept. of Eletrial and Computer Engineering, and Coordinated Siene Laboratory, University of Illinois at Urbana-Champaign Abstrat Wireless tehnologies, suh as IEEE 802.a, provide for multiple non-overlapping hannels. Typial multi-hop wireless network onfigurations have only used a single hannel for the network. The available network apaity an be inreased by using multiple hannels, and nodes an be equipped with multiple interfaes to utilize the available hannels. However, the number of interfaes per node is expeted to be smaller than the number of hannels. We establish the apaity of multi-hannel networks under this senario. We develop novel link layer and routing protools that are designed speifially for multi-hannel operation. Simulation results demonstrate the effetiveness of the proposed approah in signifiantly inreasing network apaity, by utilizing all the available hannels, even when the number of interfaes is smaller than the number of hannels. I. INTRODUCTION Wireless tehnologies, suh as IEEE 802. [], provide for multiple non-overlapping hannels. Multiple hannels have been utilized in infrastruture-based networks by assigning different hannels to adjaent aess points, thereby minimizing interferene between aess points. However, typial multi-hop wireless network onfigurations have used a single hannel to ensure all nodes in the network are onneted. For meeting the ever-inreasing throughput demands of appliations, it is neessary to utilize all of the available spetrum, and this requires the development of new protools speifially designed for multi-hannel operation. This researh was supported in part by NSF grant ANI and a Vodafone Graduate Fellowship. Wireless hosts have typially been equipped with one wireless interfae. However, a reent trend of reduing hardware osts [2] has made it feasible to equip nodes with multiple interfaes. Nevertheless, it is still expensive to equip a node with one dediated interfae for eah hannel, as the number of hannels may be large. Even if eah hannel does not have a dediated interfae, urrently available ommodity wireless interfaes (suh as IEEE 802. wireless interfae ards) an be swithed from one hannel to another, albeit at the ost of a swithing lateny, thereby allowing all hannels to be potentially utilized. Thus, it is of pratial interest to develop an arhiteture and protools for the senario wherein the number of interfaes per node is smaller than the number of hannels. Past researh on wireless network apaity [3], [4] has typially onsidered wireless networks with a single hannel, although the results are appliable to a wireless network with multiple hannels as well, provided that at eah node there is a dediated interfae per hannel. When nodes are not equipped with a dediated interfae per hannel, then apaity degradation may our, ompared to using a dediated interfae per hannel. In this paper, we haraterize the impat of number of hannels and interfaes per node on the network apaity, and show that in ertain senarios, even with only a single interfae per node, there is no apaity degradation. This implies that it may be possible to build apaity-optimal multi-hannel networks with as few as one interfae per node. When interfae swithing lateny is aounted for, apaity-optimal performane an be ahieved by using a few interfaes per node instead of just a single interfae.

2 2 We also develop link layer and routing protools for multi-hannel networks. Our solution requires at least two interfaes per node to simplify protool design. We propose a novel interfae assignment strategy that keeps one interfae fixed on a speifi hannel, while other interfaes an be swithed, as neessary, among the remaining hannels. The use of a fixed interfae simplifies oordination, while the swithable interfaes enable the utilization of all the available hannels. Traditional routing protools do not aount for hannel diversity. For example, when shortest-path routing is used, a k hop route that traverses all the hops on a single hannel has the same ost as an alternate k hop route that uses different hannels for eah hop. However, the throughput of a route that uses a single hannel on all hops an be substantially smaller than a route that uses multiple hannels, on aount of self interferene along the route. Furthermore, when the swithing lateny is non-negligible, the routing protool has to aount for the ost of interfae swithing when seleting routes. Our proposed routing protool is designed to hoose hanneldiverse routes, while aounting for the ost of swithing lateny. Evaluations show that our proposal is effetive in utilizing multiple hannels. For example, our results show that even with two interfaes, a five hannel network an offer more than five-fold improvement over a single hannel network. The proposed link layer and routing protools are similar to the algorithms used in the onstrutive proof that ahieves the upper bound for multi-hannel apaity. Hene, the proposals are validated to be an effetive hoie by theory as well. The rest of the paper is organized as follows. We desribe related work in Setion II. We present our theoretial results in Setion III. Certain design issues are desribed in Setion IV and assumptions made are disussed in Setion V. Setions VI and VII desribe the details of the proposed link and routing protools. We evaluate our protools in Setion VIII. We disuss possible extensions to our protools in Setion IX, and onlude in Setion X. II. RELATED WORK A. Capaity of wireless networks The detailed apaity results and proofs are inluded in another tehnial report [5]. In their seminal work, Gupta and Kumar [3] derived the apaity of ad ho wireless networks. The results are appliable to single hannel wireless networks, or multi-hannel wireless networks where every node has a dediated interfae per hannel. We extend the results of Gupta and Kumar to those multi-hannel wireless networks where nodes may not have a dediated interfae per hannel, and we also onsider the impat of interfae swithing delay on network apaity. Grossglauser and Tse [4] showed that mobility an improve network apaity, though at the ost of inreased end-to-end delay. Subsequently, other researh [6], [7] has analyzed the trade-off between delay and apaity in mobile networks. Gamal et al. [8] haraterize the optimal throughput-delay trade-off for both stati and mobile networks. In this paper, we adapt some of the proof tehniques presented by Gamal et al. [8] to the multi-hannel apaity problem. Reent results have shown that the apaity of wireless networks an be enhaned by introduing infrastruture support [9] []. Other approahes for improving network apaity inlude the use of diretional antennas [2], and the use of unlimited bandwidth resoures (UWB) albeit with power onstraints [3], [4]. Li et al. [5] have used simulations to evaluate the apaity of multi-hannel networks based on IEEE Other researh on apaity is based on onsiderations of alternate ommuniation models [6] [8]. B. Multi-hannel MAC and link layer protools Several researhers have proposed MAC protools based on IEEE 802. for utilizing multiple hannels. Nasipuri et al. [9], [20], and Jain et al. [2] propose a lass of protools where all nodes have an interfae on eah hannel. The protools differ in the metri used to hoose a hannel for ommuniation between a pair of nodes. The metri may simply be to use an idle hannel [9], or the signal power observed at the sender [20], or the reeived signal power at the reeiver [2]. Wu et al. [22] propose a MAC layer solution that requires two interfaes. One interfae is assigned to a ommon hannel for ontrol purposes, and the seond interfae is swithed between the remaining hannels and used for data exhange. Hung et al. [23] propose a similar two-interfae solution that uses a hannel load-aware algorithm to hoose the appropriate data hannel to be used for data exhange. So et al. [24] propose a MAC solution for multiple hannels that uses a single interfae. All the multi-hannel MAC proposals desribed above require hanges to IEEE 802., and therefore annot be deployed by using ommodity hardware. In ontrast, our proposal an be implemented with standard 802. interfaes.

3 3 Adya et al. [25] propose a link-layer solution for striping data over multiple interfaes. The proposal does not use interfae swithing, and for full utilization of available hannels, an interfae is neessary for eah hannel. Bahl et al. [26] propose SSCH, a link-layer solution that uses a single interfae, and an run over unmodified IEEE 802. MAC. In this paper, we propose a new interfae assignment strategy designed for multiple interfaes that an be implemented at the link-layer. Multi-hannel solutions implemented at the MAC or the link layer are not suffiient for effetively utilizing multiple hannels, as the routing protool may selet routes wherein suessive hops interfere with eah other. In this paper, we propose a multi-hannel aware routing protool that omplements our proposed interfae assignment strategy. In the ontext of wired loal area networks, Marsan et al. [27] have studied the performane of multihannel CSMA/CD MAC protools, and shown that signifiant redution in delay average and variane is possible even when the number of interfaes is less than the number of hannels. This paper is motivated by the need to answer a similar question with multi-hannel CSMA/CA based wireless networks. Our evaluations show that in the ase of multi-hannel wireless networks as well, signifiant performane improvement is possible, even if the number of interfaes is less than the number of hannels. C. Multi-hannel routing protools Shaham et al. [28] have proposed a arhiteture for multi-hannel wireless networks that uses a single interfae. Eah node has a default hannel for reeiving data. A node with a paket to transmit has to swith to the hannel of the reeiver before transmitting data. However, the proposal does not onsider the impat of swithing lateny. Furthermore, the routes used in the arhiteture may not utilize all the available hannels. So et al. [29] have proposed a routing protool for multi-hannel networks that uses a single interfae at eah node. We propose to use multiple interfaes, whih an offer better performane than a single interfae solution. Furthermore, the routing protool requires omplex oordination among ommuniating nodes, when setting up routes. Existing routing protools for multi-hop networks suh as DSR [30] and AODV [3] support multiple interfaes at eah node. However, those protools typially selet shortest-path routes, whih may not be suitable for multi-hannel networks [32]. Furthermore, the protools annot exploit all the available hannels, if the number of interfaes is smaller than the number of hannels. We are aware of two routing protools speifially designed for multi-hannel, multi-interfae wireless networks. Draves et al. [32] have proposed LQSR, a soure routing protool for multi-hannel, multi-interfae networks. LQSR uses WCETT, a new metri designed for multi-hannel networks, and ensures high-quality routes are seleted. Our proposal differs from LQSR in the following key aspets: LQSR assumes the number of interfaes is equal to the number of hannels used by the network. In ontrast, our proposal is designed to handle the senario where the number of available interfaes may be smaller than the number of available hannels, and therefore uses interfae swithing. LQSR is designed for stati, multi-hop wireless networks, suh as mesh networks, and does not aount for the impat of node mobility on the routing heuristi. Our proposal is designed for general mobile ad ho wireless networks, and an be used in mesh networks as well. Raniwala et al. [33], [34] propose routing and interfae assignment algorithms for stati networks. Their goal is similar to our work in addressing the senario where the number of available interfaes is less than the number of available hannels. However their approah is different in the following key aspets: The protools are designed for use in stati networks where traffi is direted toward speifi gateway nodes. The ommuniation pattern that arises in suh networks is a tree that is rooted at eah gateway node. In ontrast, our proposal is designed for a more general ommuniation pattern, where any node may ommuniate with any other node. Raniwala s protool assumes nodes are stationary and traffi load between all nodes are known. Using the load information, interfae assignment and route omputation is intelligently done. In ontrast, we assume no suh load information is available, as in an ad ho network, nodes may frequently move, resulting in hanging load onditions over time. Thus, we onsider the multi-hannel, multi-interfae routing problem in more general mobile ad ho networks. We have presented some of the ideas disussed in this report in an earlier paper [35]. This report also extends an earlier tehnial report [36].

4 4 III. CAPACITY ANALYSIS OF MULTI-CHANNEL NETWORKS In this setion, we present the results of our apaity analysis. The detailed proofs are inluded in another tehnial report [5]. A. Modeling multi-hannel multi-interfae networks We onsider a stati wireless network ontaining n nodes. We use the term hannel to refer to a part of the frequeny spetrum with some speified bandwidth. There are hannels, and we assume that every node is equipped with m interfaes, m. We assume that an interfae is apable of transmitting or reeiving data on any one hannel at a given time. We use the notation (m, )-network to refer to a network with m interfaes per node, and hannels. We define two hannel models to represent the data rate supported by eah hannel: Channel Model : In model, we assume that the total data rate possible by using all hannels is W. The total data rate is divided equally among the hannels, and therefore the data rate supported by any one of the hannels is W/. This was the hannel model used by Gupta and Kumar [3], and we primarily use this model in our analysis. In this model, as the number of hannels inreases, eah hannel supports a smaller data rate. This model is appliable to the senario where the total available bandwidth is fixed, and new hannels are reated by partitioning existing hannels. Channel Model 2: In model 2, we assume that eah hannel an support a fixed data rate of W, independent of the number of hannels. Therefore, the aggregate data rate possible by using all hannels is W. This model is appliable to the senario where new hannels are reated by utilizing additional frequeny spetrum. The results presented in this paper are derived assuming hannel model. However, all the derivations are appliable for hannel model 2 as well, and the results for model 2 an be obtained by replaing W in the results of model by W. B. Definitions We study the apaity of stati multi-hannel wireless networks under the two settings introdued by Gupta and Kumar [3]. Arbitrary Networks: In the arbitrary network setting, the loation of nodes, and traffi patterns an be ontrolled. Sine any suitable traffi pattern and node plaement an be used, the bounds for this senario are appliable to any network. The arbitrary network bounds may be viewed as the best ase bounds on network apaity. The network apaity is measured in terms of bit-meters/se (originally introdued by Gupta and Kumar [3]). The network is said to transport one bitmeter/se when one bit has been transported aross a distane of one meter in one seond. Random Networks: In the random network setting, node loations are randomly hosen, and eah node sets up one flow to a randomly hosen destination. The network apaity is defined to be the aggregate throughput over all the flows in the network, and is measured in terms of bits/se. We use the following notation to represent bounds: ) f(n) = O(g(n)) implies there exists some onstant d and integer N suh that f(n) dg(n) for n > N. f(n) 2) f(n) = o(g(n)) implies that lim n g(n) = 0. 3) f(n) = Ω(g(n)) implies g(n) = O(f(n)). 4) f(n) = ω(g(n)) implies g(n) = o(f(n)). 5) f(n) = Θ(g(n)) implies f(n) = O(g(n)) and g(n) = O(f(n)). 6) MIN O (f(n), g(n)) is equal to f(n), if f(n) = O(g(n)), else, is equal to g(n). The bounds for random networks hold with high probability (whp). In this paper, whp implies with probability when n. C. Main Results Gupta and Kumar [3] have shown that in an arbitrary network, the network apaity sales as Θ (W n) bitmeters/se, and in a random) network, the network apaity sales as Θ ( W n log n bits/se. Under the hannel model, whih was the model used by Gupta and Kumar [3], the apaity of a network with a single hannel and one interfae per node (that is, a (, )-network in our notation) is equal to the apaity of a network with hannels and m = interfaes per node (that is, a (, )-network). Furthermore, under both hannel models, the apaity of a (, )-network is at least as large as the apaity of a (m, )-network, when m (this is trivially true, by not using m interfaes in the (, )-network). In the results presented in this paper, we apture the impat of using fewer than interfaes per node by establishing the loss in apaity, if any, of a (m, )-network in omparison to a (, )-network. The goal of this work is to study the impat of the number of hannels, and the number of interfaes per node m, on the apaity of arbitrary and random networks. Our results show that the apaity is

5 5 Network apaity W n W n log n W W n A Capaity when = m Capaity loss B C Network apaity W n log n W log n log log n W log log n n log n D E Capaity when = m F Capaity loss G log n n n 2 Ratio of hannels to interfaes ( ) m log n n ( ) 2 log log n log n n 2 Ratio of hannels to interfaes ( m ) Fig.. Impat of number of hannels on apaity saling in arbitrary networks (figure is not to sale) Fig. 2. Impat of number of hannels on apaity saling in random networks (figure is not to sale) dependent on the ratio m, and not on the exat values of either or m. We now state our main results under hannel model.. Results for arbitrary network: The network apaity of a (m, )-network has two regions (see Figure ) as follows: ) When ( m ) is O(n), the network apaity is Θ W nm bit-meters/se (segment A-B in Figure ). Compared to a (, )-network, there is a apaity loss by a fator of m. 2) When m is Ω(n), the network apaity is Θ ( W nm ) bit-meters/se (line B-C in Figure ). In this ase, there is a larger apaity degradation than ase, as nm nm when m n. Therefore, there is always a apaity loss in arbitrary networks whenever the number of interfaes per node is fewer than the number of hannels. 2. Results for random network: The network apaity of a (m, )-network has three regions (see Figure 2) as follows: ) When ( m is O(log n), the network apaity is Θ W ) n log n bits/se (segment D-E in Figure 2). In this ase, there is no loss ompared to a (, )- network. Hene, in many pratial senarios where may be onstant or small, a single interfae per node suffies. ( ) ) 2) When m (n is Ω(log n) and also O log log n 2 log n, ( ) the network apaity is Θ W nm bits/se (segment E-F in Figure 2). In this ase, there is some apaity loss. Furthermore, in this region, the apaity of a (m, )-random network is the same as that of a (m, )-arbitrary network (segment E- F in Figure 2 overlaps part of segment A-B in Figure ), implying randomness does not inur a apaity penalty. ( ) ) 3) When m (n is Ω log log n 2 log n, the network ) ( W nm log log n log n apaity is Θ bits/se (line F-G in Figure 2). In this ase, there is a larger apaity degradation than ase 2. Furthermore, in this region, the apaity of a (m, )-random network is smaller than that of a (m, )-arbitrary network, in ontrast to ase Other results: The results presented above are derived under the assumption that there is no delay in swithing an interfae from one hannel to another. However, it an be shown [5] that even if interfae swithing delay is onsidered, the network apaity is not redued, provided a few additional interfaes are provisioned for at eah node. This implies that it is possible to hide the interfae swithing delay by using extra interfaes in onjuntion with arefully designed routing and transmission sheduling protools. D. Impliations of apaity results A ommon senario of operation is when the number of hannels is not too large ( m = O(n)). Under this senario, the apaity ( of a (m,) )-network in the arbitrary setting sales as Θ W nm under hannel model. Similarly, under hannel model 2, the apaity of the network sales as Θ (W nm). Under either model, the apaity of a (m, )-network goes down by a fator of m, when ompared with a (, )-network. Therefore, doubling the number of interfaes at eah

6 6 node (as long as number of interfaes is smaller than the number of hannels) inreases the hannel apaity by a fator of only 2. Furthermore, the ratio between m and deides the apaity, rather than the individual values of m and. Inreasing the number of interfaes may result in a linear inrease in the ost but only a sub-linear (proportional to square-root of number of interfaes) inrease in the apaity. Therefore, the optimal number of interfaes to use may be smaller than the number of hannels depending on the relationship between ost of interfaes and utility obtained by higher apaity. Different network arhitetures have been proposed for utilizing multiple hannels when the number of available interfaes is smaller than the number of available hannels [32], [34], [35]. The onstrution we use in proving lower bound [5] implies that maximal apaity is ahieved when all hannels are utilized. One arhiteture used in the past [32] is to use only m hannels when m interfaes are available, leading to wastage of the remaining m hannels. That arhiteture results in a fator of m loss in apaity whih an be signifiantly higher than the optimal m loss (when m = O(n)). Hene, in general, higher apaity may be ahievable by arhitetures that use all hannels, possibly by dynamially swithing hannels. Our results suggest that the apaity of multi-hannel random networks with total hannel data rate of W is the same as that of a single hannel network with data rate W as long as the ratio m is O(log n). When the number of nodes n in the network inreases, we an also sale the number of hannels (for example, by using additional bandwidth, or by dividing available bandwidth into multiple sub-hannels). Even then, as long as the hannels are saled at a rate not more than log n, there is no loss in apaity even if a single interfae is available at eah node. In partiular, if the number of hannels is a fixed onstant, independent of the node density, then as the node density inreases beyond some threshold density (at whih point log n), there is no loss in apaity even if just a single interfae is available per node. Thus, this result may be used to roughly estimate the number of interfaes eah node has to be equipped with for a given node density and a given number of hannels. In a single hannel random network, i.e., a (, )- network, the apaity bottlenek arises out of the hannel beoming fully utilized, and not beause interfae at any node is fully utilized. On an average, the interfae of a node in a single hannel network is busy only for X fration of the time, where X is the average number of nodes that interfere with a given node. In a (, )- random network with n nodes, eah node on an average has Θ(log n) neighbors to maintain onnetivity [3]. This implies that in a( single) hannel network, eah interfae is busy for only Θ log n time. Intuitively, our onstrution above utilizes this slak time of interfaes to support up to O(log n) hannels without loss in apaity. In general, there is no loss in apaity in a random network as long as the number of hannels is smaller than the average number of nodes in the neighborhood of a node. When the number of hannels is large (speifially, ω(log n)) and eah node has a single interfae, there is a apaity loss when ompared to a single hannel network. This apaity loss arises beause the number of hannels is more than the number of interfaes in a neighborhood. The lower bound onstrution [5] suggests that an optimal strategy for maximizing apaity when number of hannels is large is to suffiiently inrease the transmission power used to ensure neighborhood size is suffiiently large. This ensures that number of interfaes in a neighborhood will then be equal to the number of hannels. However, there is still some apaity loss beause larger transmission power (than that is needed for onnetivity alone) lowers apaity by onsuming more area. Some of the impliations of our apaity results on protool design are disussed in Setion IV-B. IV. DESIGN ISSUES In this setion, we first motivate the benefits of using a multi-interfae solution for exploiting multiple hannels. We then identify the need for speialized routing protools for multi-hannel, multi-interfae networks. A. Benefits of using multiple interfaes We define interfae to be a network interfae ard equipped with a half-duplex radio transeiver, e.g., a ommodity 802. wireless ard. In most multi-hop networks, a single hannel is used, and therefore a single interfae suffies. However, when multiple hannels are available, having more than one interfae is benefiial. Our apaity analysis has shown that when swithing delay is negligible, or when there are no lateny onstraints, a single interfae per node may suffie in ahieving optimal apaity in random networks. However, in pratie, swithing delay is often non-negligible, The neighborhood of a node onsists of all other nodes that may interfere with it.

7 7 and appliations typially expet end-to-end lateny to be reasonably small. Under these onstraints, our analysis shows that multiple interfaes per node may be required for ahieving optimal apaity. Apart from the theoretial need for multiple interfaes to ahieve asymptotially optimal apaity, many pratial onerns, desribed below, motivate the use of multiple interfaes as well. When using a single interfae per node, if the interfaes of two nodes are on different hannels, then they annot ommuniate. For reduing synhronization requirements and overheads, eah interfae has to stay on a hannel for many paket transmission durations (00ms in [24] and 0ms in [26]). As a result, when pakets are traversing multi-hop paths, pakets may be delayed at eah hop, unless the next hop is on the same hannel as well. Thus, when a single interfae is used, there is an inrease in the end-to-end lateny if different hops traversed are on different hannels. Otherwise, if most hops are on the same hannel, transmissions on onseutive hops interfere, reduing the maximum apaity. In either ase, TCP throughput is signifiantly affeted. When at least two interfaes are available, we propose keeping one interfae permanently assigned to a hannel to greatly simplify oordination, while swithing the seond interfae (based on traffi requirements) to avoid delaying a paket at eah hop. We defer disussion of the proposed approahes till later in the paper, but multiple interfaes are required to derive both simpliity in oordination and minimal delays. A seond benefit is the ability to reeive and transmit data in parallel. Half-duplex wireless interfaes annot simultaneously transmit and reeive data. However, when multiple (say two) interfaes and multiple hannels are available, while one interfae is reeiving data on one hannel, the seond interfae an simultaneously transmit data on a different hannel. In many ases, this an double the maximum throughput ahievable on a multihop route. Our proposed arhiteture exploits this benefit of using multiple interfaes as well. B. Insights obtained from apaity analysis The onstrutions used in establishing apaity results [5] offer insights into optimal link layer and routing strategies that need to be used for multi-hannel networks. In the transmission sheduling sheme used in our lower bound onstrution proof (presented in [5]), it suffies for a node to always transmit on a speifi hannel without requiring to swith hannels for different pakets. However, a node may have to swith hannels for reeiving data. An alternate onstrution is to use a sheduling sheme whih ensures that a node reeives all data on a speifi hannel, but may have to swith hannels when sending data. It an be shown that the alternate onstrution is equivalent to the lower bound onstrution. This intuition an be used in developing a pratial sheme that uses two interfaes per node. One interfae an be used for reeiving data and is always fixed to a single hannel (for long time intervals). The seond interfae an be used for sending data and is swithed between hannels, as neessary. Existing multihannel protools have often required tight synhronization among nodes. The use of two interfaes, with a dediated interfae on a fixed hannel obviates the need for tight synhronization as a node reeives data on a well-known hannel. Furthermore, using a fixed hannel for reeption does not degrade apaity sine it is based on the (optimal) alternate onstrution. The lower bound onstrution also suggests that load balaning (i.e., distributing routes) among nodes in a given neighborhood is essential for full utilization of multiple hannels. Existing routing protools for multihop networks suh as DSR [30] and AODV [3] typially selet shortest-hop routes, and do not inorporate load balaning. In addition, route seletion does not onsider the interfae swithing ost, and the hosen routes may require frequent hannel swithing, degrading network performane. In a single hannel network, load balaning is sometimes used to balane energy onsumption aross nodes, or to improve resiliene of the network. However, load balaning in the same neighborhood is not always required in single hannel networks for maximizing apaity. Thus, there is a need for a ustomized routing protool for multi-interfae, multi-hannel networks. Figure 3 illustrates a senario that highlights the need for speialized routing protools for multi-hannel networks. In the figure, node A is ommuniating with node D using route A-C-D. Node E wishes to ommuniate with node F, and either of B or C an be used as the intermediate node. Assume all nodes have a single interfae, and assume C and B an relay at most w bytes per seond. If node C is hosen as the intermediate node, then node C has to forward data along both routes A-C- D and E-C-F, and the throughput reeived by eah flow is at most w/2. On the other hand if node B is hosen as the intermediate node, then both routes A-C-D and E-B-F an be simultaneously used (assuming hannels used on routes A-C-D and E-B-F an be hosen to be orthogonal), and eah flow reeives a rate of w. Although

8 8 A B F Fig. 3. Impat of route seletion on effetive utilization of multiple hannels this example assumed eah node had a single interfae, similar issues arise even when multiple interfaes are available. The above senario highlights the need for the routing protool to appropriately distribute routes among nodes in the neighborhood. In the ase of single hannel networks, the throughput obtained is the same whether B or C is hosen as the intermediate node. When a single hannel is available, and say, when C is transmitting a paket along route A-C-D, B annot transmit a paket even if it is hosen as the intermediate node (as the ommon hannel is busy). Consequently, routing protools designed for single hannel networks do not need to distribute routes within a neighborhood. However, to exploit the benefit of multiple hannels, it is important for a routing protool to ensure routes are arefully distributed in the network. E C V. ASSUMPTIONS In this setion, we desribe the assumptions we make while developing our protools. A. Interfae swithing ost The ability to swith an interfae from one hannel to another is a key property we exploit to utilize all the available hannels, even when the number of interfaes available is signifiantly lesser than the number of available hannels. We assume that hannels are separated in frequeny, and swithing an interfae requires hanging the frequeny of operation. Swithing an interfae from one hannel to another inurs some delay D whih may be non-negligible. In the urrent literature, estimates for D (for swithing between hannels on the same frequeny band) with ommodity IEEE 802. hardware are in the range of a few milliseonds [37] to a few hundred miroseonds [38]. It is expeted that with improving tehnology, the swithing delay will redue to a few tens of miroseonds [26]. Protools that utilize interfae swithing need to be flexible enough to D aommodate a range of swithing delays. The routing protool may have to aount for the swithing ost while seleting routes. Interfae swithing is possible aross different frequeny bands as well. For example, wireless ards are urrently available that support both IEEE 802.a (operates on 5 GHz band) and IEEE 802.b (operates on 2.4 GHz band), and an swith between the two bands. However, with the urrently available hardware, swithing aross bands inurs a large delay, but the swithing delay is expeted to redue in the future. The arhiteture presented in this paper allows for the utilization of hannels on the same band as well as hannels on different bands. B. Orthogonal hannels IEEE 802.a offers 2 non-overlapping hannels, while IEEE 802.b offers 3 non-overlapping hannels. When a single node is equipped with multiple interfaes, it has been noted [32] that ommuniation on different interfaes using adjaent non-overlapping hannels may interfere. Thus, the number of available orthogonal hannels may be smaller than the number of non-overlapping hannels. Reently, Raniwala et al. [34] have experimentally shown that when separation between interfaes is inreased, the interferene between interfaes is redued, allowing more hannels to be used simultaneously. Furthermore, future hardware that employs better filters to redue adjaent hannel interferene may allow any pair of non-overlapping hannels to be simultaneously used. In this paper, we aount for the possibility of interferene among adjaent non-overlapping hannels in our evaluations by assuming that the number of available orthogonal hannels is smaller than the number of nonoverlapping hannels. We believe that a areful use of hannels will offer at least 3 to 5 orthogonal hannels in the 5 GHz band (IEEE 802.a). Improved hardware in the future may enable all 2 hannels to be used orthogonally. Future hanges in FCC regulations may provide for more orthogonal hannels as well. C. Problem Formulation The protools proposed in this paper are designed for a multi-hop wireless network. Nodes in the network an be mobile. We assume that the typial traffi pattern involves ommuniation between arbitrary pair of nodes. Speifially, we do not require the presene of speial gateway nodes that may be the soure or destination of all traffi in the network, although our proposal an be used in that senario as well.

9 9 We define the requirements of a multi-hannel, multiinterfae solution as follows: ) Improve network apaity by utilizing all the available hannels, even if the number of interfaes is smaller than the number of available hannels. The solution must be flexible enough to aommodate different number of hannels and interfaes, with hannels potentially on different frequeny bands. 2) Ensure that a network whih is onneted when using a single ommon hannel, ontinues to be onneted when multiple hannels are being used. 3) Allow implementation using existing IEEE 802. hardware. In Setion VI we present the link layer swithing protool that manages interfae swithing. This link layer protool an be used in onjuntion with ommonly used routing protools suh as AODV and DSR. However, to obtain the full benefits of using multiple hannels, existing routing protools are not suffiient (details in Setion VII). We then present a new routing protool, alled Multi-Channel Routing Protool in Setion VII that is designed speifially for multi-hannel networks, and operates in onjuntion with the swithing protool. VI. SWITCHING PROTOCOL When the number of available interfaes is smaller than the number of available hannels, an interfae assignment strategy is required to assign interfaes to speifi hannels. Furthermore, for using all the available hannels, a swithing protool is neessary to deide when to swith an interfae from one hannel to another. The swithing protool has to ensure that the neighbors of a node X an ommuniate with it on-demand, whih requires all neighbors of X to be always aware of the hannel being used by at least one interfae of X. We first identify the different interfae assignment strategies possible. We then desribe our proposal and disuss issues involved. A. Classifiation of interfae assignment strategies Interfae assignment strategies an be lassified into stati, dynami, and hybrid strategies.. Stati Assignment: Stati assignment strategies assign eah interfae to a hannel either permanently, or for long intervals of time where long interval is defined relative to the interfae swithing time. For example, [32], [33] use stati interfae assignment. Stati assignment an be further lassified into two types: ) Common hannel approah: In this approah, interfaes of all nodes are assigned to a ommon set of hannels (e.g. [32]). For example, if two interfaes are used at eah node, then the two interfaes are assigned to the same two hannels at every node. The benefit of this approah is that the onnetivity of the network is the same as that of a single hannel approah. Note that the senario where a single hannel and a single interfae is used is a speial ase of the stati, ommon hannel assignment strategy. 2) Varying hannel approah: In this approah, interfaes of different nodes may be assigned to a different set of hannels (e.g. [33]). With this approah, there is a possibility that the length of the routes between nodes may inrease. Also, unless the interfae assignment is done arefully, network partitions may arise. Stati assignment strategies are well-suited for use when the interfae swithing delay is large. In addition, if the number of available interfaes is equal to the number of available hannels, interfae assignment problem beomes trivial. Stati assignment strategies do not require speial oordination among nodes (exept perhaps to re-assign interfaes over long intervals of time) for data ommuniation. With stati assignment, nodes that share a hannel on one of their interfaes an diretly ommuniate with eah other, while others annot. Thus, the effet of stati hannel assignment is to ontrol the network topology by deiding whih nodes an ommuniate with eah other. 2. Dynami Assignment: Dynami assignment strategies allow any interfae to be assigned to any hannel, and interfaes an frequently swith from one hannel to another. In this setting, two nodes that need to ommuniate with eah other need a oordination mehanism to ensure they are on a ommon hannel at some point of time. For example, the oordination mehanism may require all nodes to visit a ommon rendezvous hannel periodially (e.g. [24]), or require other mehanisms suh as the use of pseudo-random sequenes (e.g. [26]), et. The benefit of dynami assignment is the ability to swith an interfae to any hannel, thereby offering the potential to over many hannels with few interfaes. The key hallenge with dynami swithing strategies is to oordinate the deisions of when to swith interfaes as well as what hannel to swith the interfaes to, among the nodes in the network. 3. Hybrid Assignment: Hybrid assignment strategies ombine stati and dynami assignment strategies by applying a stati assignment for some interfaes and a dynami assignment for other interfaes. Hybrid strate-

10 0 gies an be further lassified based on whether the interfaes that apply stati assignment use a ommon hannel approah, or a varying hannel approah. An example of hybrid assignment with ommon hannel at the MAC layer is [22], whih assigns one interfae of eah node statially to a ommon ontrol hannel, and other interfae an be dynamially swithed among other data hannels. We propose to use a hybrid hannel assignment strategy with varying hannel assignment. Hybrid assignment strategies are attrative as they allow simplified oordination algorithms supported by stati assignment while retaining the flexibility of dynami assignment. B. Assigning interfaes to hannels We propose a new hybrid assignment strategy. We assume that there are M interfaes available at eah node. The available interfaes are divided into two subsets. Fixed Interfaes: Some K of the M interfaes at eah node are assigned for long intervals of time to some K hannels, and we designate these interfaes as fixed interfaes, and the orresponding hannels as fixed hannels. Swithable Interfaes: The remaining M K interfaes are dynamially assigned to any of the remaining M K hannels (over short time sales), based on data traffi. These interfaes are designated as swithable interfaes, and the hannel to whih a swithable interfae is assigned to is alled a swithable hannel. Different nodes may assign their K fixed interfaes to a different set of K hannels. It is possible for eah node to use a different value of K and M, and it is also possible to vary K with time. To simplify rest of the disussion, we assume M = 2, K = for all nodes, i.e., there is one fixed, and one swithable interfae (although the proposed protool is appliable to any values of M and K). The main idea of the interfae assignment strategy is to reeive data using the fixed interfae. Figure 4 illustrates the protool used for ommuniation between nodes when using fixed and swithable interfaes. Assume that node A has a paket to send to node C via node B. Nodes A, B, and C have their fixed interfaes on hannels, 2, and 3 respetively. Initially, node A has its swithable interfae on hannel 3, node B has its swithable interfae on hannel, and node C has its swithable interfae on hannel 2. In the first step, node A swithes its swithable interfae from hannel 3 to hannel 2, before transmitting the paket, beause Step : A (fixed = ) B (fixed = 2) C (fixed = 3) Initially: swithable = 3 Step 2: swithable = 2 swithable = swithable = 3 swithable = 2 Fig. 4. Example of swithing protool operation with 3 hannels, 2 interfaes hannel 2 is the fixed hannel of node B. Node B an reeive the paket sine its fixed interfae is always listening to hannel 2. In the next step, node B swithes its swithable interfae to hannel 3 and forwards the paket, whih is reeived by node C using its fixed interfae. One the swithable interfaes are orretly set up during a flow initiation, there is no need to swith the interfaes for subsequent pakets of the flow (unless a swithable interfae has to swith to another hannel for sending pakets of a different flow). Fixed interfaes are assigned to a hannel for long intervals of time. The hannel to whih the fixed interfae of a node is assigned to is known by other nodes in the neighborhood (using a protool desribed later). Thus, there is no need for speial oordination between a sender and a reeiver on when to shedule transmissions to the reeiver. When a node X has to ommuniate with a node Y over hannel i, if the fixed hannel being used by X is also i, then the fixed interfae is used for ommuniation. Otherwise, the swithable interfae of X is swithed to hannel i for ommuniating with Y. Therefore, the swithable interfae enables a node X to transmit to any node Y in its neighborhood by swithing (if required) to the fixed hannel used by Y. Different nodes in the neighborhood hoose different fixed hannels, as we will elaborate later. This flexibility an be used to ensure that all hannels in the network are utilized. In summary, the proposed interfae assignment strategy has the following benefits: A sender and a reeiver do not have to synhronize for hannel swithing. Thus, the assignment strategy is designed to not require a oordination algorithm for ensuring the sender and reeiver are on the same hannel. By arefully balaning the assignment of fixed interfaes of different nodes over the available hannels, all hannels an be utilized, and the number of ontending transmissions in a neighborhood signifiantly redues.

11 The protool (desribed next) an easily sale if the number of available hannels inreases. C. Swithing Protool: Fixed interfae assignment Fixed interfae assignment involves two omponents - hoosing the hannel to be assigned to the fixed interfae, and informing neighbors about the hannel being used by the fixed interfae. The interfae assignment protool has to ensure that fixed interfaes of nodes in a neighborhood are distributed aross different hannels. For example, suppose a node A uses hannel for the fixed interfae. Then, all transmissions direted to A will be on hannel. For balaning the usage of hannels, it is benefiial if other nodes in the neighborhood use a different hannel for their fixed interfae. In general, all nodes within the interferene range of a node an interfere with reeption on its fixed hannel, and it is important to balane the number of nodes that use eah hannel as their fixed hannel. We propose a loalized protool for fixed interfae assignment. Eah node maintains a NeighborTable ontaining the fixed hannels being used by its neighbors. Nodes also maintain a ChannelUsageList ontaining a ount of the number of nodes using eah hannel as their fixed hannel. Initially, a node hooses a random hannel for its fixed interfae. Periodially, eah node broadasts a Hello paket on every hannel. The hello paket ontains the fixed hannel being used by the node. When a node reeives a hello paket from a neighbor, it updates its NeighborTable and ChannelUsageList. Information about the fixed hannel used by neighbors an also be obtained by snooping on the route disovery pakets (the ontents of route disovery paket are desribed in Setion VII). Eah node periodially onsults its ChannelUsageList (the period hosen is large relative to paket transmission time). If the number of other nodes using the same fixed hannel as itself is large, then a node with some probability p hanges its fixed hannel to a less used hannel, and transmits a hello paket informing neighbors of its new fixed hannel. The probabilisti approah is used to avoid frequent hange of fixed hannels by multiple nodes. The ChannelUsageList maintained by a node only traks the nodes present within its ommuniation range. Nodes outside the ommuniation range an be aounted for if the Hello paket also inludes ChannelUsageList, thereby exhanging two-hop information, though at the ost of inreased hello paket size. The frequeny of hello paket exhange depends on the magnitude of average node mobility. A node moving into a new neighborhood annot ommuniate with its neighbors until it has exhanged hello pakets with them to learn about the fixed hannels being used by neighbors. Hello paket exhange is used by many routing protools (suh as AODV) as well, and with moderate degrees of mobility, the overhead of hello paket exhange is not expeted to be large. We do not use hannel load information to swith fixed hannels. Using hannel load may be benefiial if the load in the network does not hange frequently. On the other hand, if the load in the network hanges frequently, say when there are many short-lived flows, it may lead to frequent and unneessary hannel swithing. For example, HTTP transfers often are less than a seond, and if suh short-lived flows dominate the network traffi, then it may lead to frequent hannel swithing. Basing fixed hannel swithing deisions on the network topology requires swithing only when the topology hanges, whih is of the order of tens to hundreds of seonds even with moderate mobility. Hene, we have hosen to swith fixed hannels based on the number of nodes using a hannel in a given neighborhood. D. Swithing Protool: Managing swithable interfae The swithable interfae of a node X is used to transmit data whenever the fixed hannel of the destination is different from the fixed hannel of X. One issue to be resolved is how frequently to swith hannels. For example, onsider a stream of pakets at a node X where the even-numbered pakets are to destination A, and the odd numbered pakets are to destination B, with A and B on different fixed hannels. One possibility is to alternately swith between hannels for forwarding eah paket. However, suh frequent swithing may be very expensive when the swithing delay is large. Another possibility is to swith over longer intervals of time, thereby amortizing the ost of swithing among multiple pakets. Thus, a poliy is needed to deide when to swith an interfae, and what hannel to swith the interfae to. Eah hannel is assoiated with a paket queue, as shown in Figure 5. Based on the above disussion, we propose to transmit at most BurstLength queued pakets on one hannel, before swithing to another hannel (only if there are pakets for some other hannel). In addition, the swithable interfae stays on a hannel for at most MaxSwithTime seonds, before swithing to another hannel (again, swithing happens only if there

12 2 QUEUES FIXED D 2 E N SWITCHABLE A Fig C B Need for a diversity ost metri Fig. 5. Illustration of queues assoiated with interfaes are pakets for some other hannel). The two onditions in onjuntion ensure that the extra lateny introdued by the swithing protool is bounded by MaxSwithTime, while the swithing ost is amortized among up to BurstLength pakets. The parameters BurstLength and MaxSwithTime an be suitably set to trade-off lateny with throughput. Furthermore, to ensure fairness and to prevent starvation, when swithing hannels, the swithable interfae is set to the hannel having the oldest data paket in its queue. In single hannel networks, a paket broadast on the hannel an be reeived by all neighbors of the transmitter. However, when multiple hannels are being used, a paket broadast on a hannel is reeived only by those nodes listening to that hannel. Many higherlayer protools (e.g., routing protools) require broadast pakets to be reeived by all nodes in the neighborhood. Suh neighborhood broadast is supported in our ase by transmitting the broadast paket separately on every hannel. A opy of the broadast paket is added to eah hannel s queue, and sent out when that hannel is sheduled for transmission by the swithing protool. VII. MCR: MULTI-CHANNEL ROUTING PROTOCOL In this setion, we desribe the details of the proposed on-demand Multi-Channel Routing (MCR) protool that operates over the proposed swithing protool. Popular on-demand routing protools suh as AODV and DSR use the shortest-path metri for route seletion. Shortestpath metri assigns a unit ost for eah hop in a route, and does not distinguish between a route that uses many hannels, and a route that uses few hannels. In the proposed arhiteture, the routing protool has to onsider the ost of interfae swithing as well. MCR protool uses a new routing metri that onsiders the hannel diversity, and interfae swithing osts, in addition to the number of hops in a route. A. Measuring hannel diversity ost We design a hannel diversity ost metri that assigns smaller ost to routes using many hannels (alled hannel diverse routes) than routes using few hannels. Figure 6 illustrates the need for onsidering diversity ost. In the figure, node A is setting up a route to node C, and there are two possible routes: A-B-C, and A-D-E-C. Eah link is labeled with the hannel used to transmit along that link (the hannel used on a link is the fixed hannel that is being used by the destination of the link). Assume that eah link an support a maximum data rate w. When the shortest-path metri is used (as is the ase in DSR and AODV), route A-B-C is preferred, as it requires fewer hops than A-D-E-C. However, both links on route A-B-C use hannel 3, and at any time only link A-B or link B-C an be ative, resulting in a maximum endto-end throughput of w/2. On the other hand, links on route A-D-E-C use different hannels, allowing all links to be ative simultaneously, resulting in a maximum endto-end throughput of w. Using the proposed swithing protool, when route A-D-E-C is used, the swithable interfae of A is set to hannel, D to hannel 2, and E to hannel 3 for sending the first paket, and interfae swithing is not required for subsequent pakets (unless a swithable interfae has to swith to another hannel for sending pakets of a different flow). The availability of two radios allows multiple hannels to be used by the swithing protool, provided the routing protool arefully selets the routes. We develop the diversity ost metri by noting that a link i in a route is interfered by other links in the route that use the same hannel, and are in its interferene range. The interferene range is typially assumed to be a small multiple (say, 3) of the ommuniation range. We define a parameter alled InterfereneLength (IL) (set to 3 in simulations) that is used to identify whih links along a route interfere. The i th link is onsidered to interfere with k th link on a route, for i+ k (i+il) (we do not onsider k i to avoid ounting links twie), if links i and k use the same hannel. Suppose the hannel being used by link i is C(i), and suppose there