Available online at www.sciencedirect.com

Ad Hoc Networks 7 (2009) 63–78

www.elsevier.com/locate/adhoc

On-demand routing and channel assignment inmulti-channel mobile ad hoc networks q

Michelle X. Gong a, Scott F. Midkiff b, Shiwen Mao c,*

a Corporate Technology Lab, Intel Corporation, Santa Clara, CA 95054, USAb The Bradley Department of Electrical and Computer Engineering, Virginia Tech, Blacksburg, VA 24061, USA

c Department of Electrical and Computer Engineering, Auburn University, 200 Broun Hall, Auburn, AL 36849, USA

Received 8 December 2006; received in revised form 6 April 2007; accepted 27 November 2007Available online 5 December 2007

Abstract

The capacity of mobile ad hoc networks is constrained by the intra-flow interference introduced by adjacent nodes onthe same path, and inter-flow interference generated by nodes from neighboring paths. By assigning orthogonal channels toneighboring nodes, one can minimize both types of interferences and allow concurrent transmissions within the neighbor-hood, thus improving the throughput and delay performance of the ad hoc network. In this paper, we present three noveldistributed channel assignment protocols for multi-channel mobile ad hoc networks. The proposed protocols combinechannel assignment with distributed on-demand routing, and only assign channels to active nodes. They are shown torequire fewer channels and exhibit lower communication, computation, and storage complexity, compared with existingapproaches. Through simulation studies, we show that the proposed protocols can effectively increase throughput andreduce delay, as compared to several existing schemes, thus providing an effective solution to the low capacity problemin multi-hop wireless networks.� 2007 Elsevier B.V. All rights reserved.

Keywords: Cross-layer design; Distributed channel assignment; Multi-channel medium access control; Multi-channel mobile ad hocnetworks; Routing

1. Introduction

Despite recent advances in wireless technologies,today’s wireless links still cannot offer the compara-

1570-8705/$ - see front matter � 2007 Elsevier B.V. All rights reserved

doi:10.1016/j.adhoc.2007.11.011

q This research was supported in part by a National ScienceFoundation Integrated Graduate Education and Research Train-ing (IGERT) grant (award DGE-9987586). Part of this work hasbeen presented at the 2005 IEEE International Conference onCommunications, Seoul, Korea, May 2005 [1].

* Corresponding author. Tel.: +1 334 844 1845.E-mail addresses: michelle.gong@gmail.com (M.X. Gong),

midkiff@vt.edu (S.F. Midkiff), smao@ieee.org (S. Mao).

ble data rates as their wired counterparts. The lowthroughput problem is further aggravated inmulti-hop wireless environments due to the intra-

flow interference introduced by adjacent transmit-ting nodes on the same path and inter-flow interfer-

ence generated by transmitting nodes fromneighboring paths. Two transmitting nodes withinthe interference range will interfere with each other.In addition, cumulative interference from a largenumber of transmitters outside a node’s interferencerange will also cause low signal-to-interference-plus-noise-ratio (SINR) at the node. For instance, it has

.

64 M.X. Gong et al. / Ad Hoc Networks 7 (2009) 63–78

been shown in [2] that the maximum capacity thatthe IEEE 802.11 MAC can achieve for a chain net-work could be as low as just one seventh of thenominal link bandwidth.

We observe that all current IEEE 802.11 physical(PHY) standards divide the available frequency intoseveral orthogonal channels, which can be usedsimultaneously within a neighborhood. Therefore,increasing capacity by exploiting multiple channelsbecomes particularly appealing. In fact, such band-width aggregation has been widely used in infra-structure-based WLANs, where high-end accesspoints are equipped with multiple interfaces thatoperate on different channels simultaneously [3]. Insuch networks, non-overlapping channels are dis-tributed among different access points at the net-work planning stage [4]. However, IEEE 802.11WLANs that operate in ad hoc mode rarely usemultiple channels simultaneously. This is partlybecause that the IEEE 802.11 MAC is not designedto operate with multiple channels, resulting in awaste of precious network resources. As an exam-ple, an ad hoc network based on the IEEE802.11a technology utilizes only one out of 12 avail-able orthogonal channels, wasting more than 90%of the potentially available spectrum.

Consequently, there has been substantial interestin multi-channel MAC schemes that can achievehigher throughput by exploiting multiple availablechannels [3,5,6]. Some of the early works, e.g.[7,8], assume that every node has its own uniquechannel. Therefore, no channel assignment or selec-tion is needed. However, in reality, the number ofchannels is limited and has to be carefully assignedto each node, in order to avoid contention and col-lisions and to enable optimal spatial reuse of avail-able channels. Many channel assignment problemshave been proven to be NP-complete and, thus,computationally intractable [4,9,10]. There existonly a few heuristic solutions, which have good per-formance under certain environments, for instance,in a static wireless network. However, these heuris-tic schemes suffer from inefficiency when employedin the mobile ad hoc environment [9,11].

In this paper, we present three principles fordesigning efficient distributed channel assignmentschemes. First, to reduce the complexity of the chan-nel assignment algorithm, channel assignment androuting should be jointly designed. This ‘‘cross-layer” design approach is motivated by the fact thatboth the channel assignment algorithm and the adhoc routing algorithm are invoked when there is a

change in the network topology. Exploring thisdesign principle can potentially reduce the complex-ity of channel assignment algorithms. Second, chan-nels should be assigned only to active nodes. This‘‘on-demand” channel assignment principle is moti-vated by the fact that only nodes on active routesneed valid channels. Some existing channel assign-ment schemes assign channels to all nodes in thenetwork, regardless of whether they are active ornot, thus requiring a large number of orthogonalchannels. If this on-demand assignment principleis implemented, fewer channels (i.e. fewer resources)may be required in the network to achieve a compa-rable performance.

Third, the capacity of mobile ad hoc networks canbe adversely affected by both ‘‘hidden terminals” and‘‘exposed terminals.” The hidden terminal problemoccurs when transmitters outside of radio range ofeach other transmit at the same time, causing a colli-sion at the receivers [10]. The exposed terminal prob-lem occurs when a node is prevented from sendingpackets to other nodes due to a neighboring transmit-ter, even though this transmission will not causeinterference. In addition, cumulative interferencegenerated by nodes two or more hops away may alsoadversely affect channel utilization and networkcapacity. Thus, to improve network performance,distinct channels should be assigned such that hiddenterminals, exposed terminals, and cumulative inter-ference can be avoided as much as possible.

We present a new channel assignment protocol,named Channel Assignment Ad hoc On-demandDistance Vector routing (CA-AODV), that imple-ments these design principles. In CA-AODV, chan-nel assignment is combined with the AODV routingprotocol and is performed in a cross-layer and on-demand fashion. Specifically, channel assignmentis performed during the route discovery phase andchannel information is piggybacked in the routingcontrol messages. CA-AODV assigns different chan-nels for neighboring nodes within a k-hop regionalong the same path, thus allowing concurrenttransmission on neighboring links along the pathand effectively reducing the intra-flow interference.We also present two extensions to the CA-AODVprotocol, namely, the Enhanced 2-hop CA-AODV(E2-CA-AODV) protocol and the Enhanced k-hopCA-AODV (Ek-CA-AODV) protocol. In additionto intra-flow interference, these two extensions alsoaim to minimize inter-flow interference by assigningorthogonal channels to active nodes within a k-hopneighborhood, where k P 2. With such channel

Table 1Notation

Symbol Comments

C Set of channelsA Set of available channelsV Set of nodes in the networkE Set of edges that represent radio linksv 2 V A node in the networknkðvÞ Number of k-hop neighbors sharing the same

channel with node v

V t � V Set of active transmittersV r � V Set of receiversvt

j 2 V t An active transmittervr

i 2 V r An active receivervt

T ðiÞ Desired transmitter for receiver vri

P ðvtj; v

ri Þ Received power at vr

i from transmitter vtj

SðvtT ðiÞ; v

tjÞ Cross-correlation between the channels used by the

two transmittersb SINR threshold for successful receptionP N Power level of additive white Gaussian noise

M.X. Gong et al. / Ad Hoc Networks 7 (2009) 63–78 65

assignment, more concurrent transmissions areachieved for nodes along various routes within thek-hop neighborhood, thus effectively improvingthe throughput and delay performance. Simulationresults in mobile ad hoc networks show that the per-formance of E2-CA-AODV approaches that of amulti-channel scheme with an unlimited number ofchannels (i.e., the ideal case with unlimited amountof resources). In addition, the proposed protocolsexhibit lower complexity than many existing cen-tralized and distributed approaches.

In addition to the distributed channel assignmentprotocols, we also develop a transmitter-basedmulti-channel MAC (MC-MAC) protocol thatextends the benefit of channelization to multi-hopmobile ad hoc networks. Because multiple hopinformation or the whole network topology can bevisible to the routing layer, MC-MAC can benefitfrom the combined routing and channel assignmentscheme and offers improved network performance.

The remainder of this article is organized as fol-lows. We formulate the channel assignment problemin Section 2. Table 1 summarizes the notation usedin this paper. In Section 3, we present CA-AODVand its extensions. The MC-MAC protocol isdescribed in Section 4. We present simulation resultsin Section 5, and discuss related work in Section 6.Section 7 concludes this paper.

2. Problem statement

An ad hoc network can be modeled as a graphG ¼ fV ;Eg, where V is the set of nodes and E is

the set of edges that represent wireless links. Weassume that nodes use omnidirectional antennasand radio links are bidirectional. A link is assumedto exist between two nodes if and only if the twonodes are within each other’s radio range.

The interference range is defined to be the k-hopneighborhood of a node. Interference can be signif-icantly reduced if nodes within the k-hop neighbor-hood are assigned different orthogonal channels.The k-hop neighbors of a node v is the set

NkðvÞ ¼ fw 2 V j hðv;wÞ 6 kg; ð1Þ

where hðv;wÞ is the hop distance from v to w, i.e. theminimum number of hops of any path from v to w.Note that N 1ðvÞ is the set of directly connectedneighbors of node v. Let C denote the set of chan-nels in the network. We further define V t � V tobe the set of active transmitters and V r � V the setof active receivers. Let vr

i 2 V r be an active receiverand vt

j 2 V t be an active transmitter. If vtj is trans-

mitting on the same channel on which vri is receiv-

ing, but vtj is not the intended transmitter to vr

i ,then transmitter vt

j will interfere with receiver vri .

The power associated with the interference,P ðvt

j; vri Þ, is the received power at node vr

i , which isa function of the effective transmitting power ofnode vt

j, the distance between nodes vtj and vr

i , andchannel conditions, e.g. path loss and fading [12].

If k is set to an appropriate value and jCj is suf-ficiently large, hidden nodes and exposed nodes canbe largely avoided and harmful interference can bemitigated. In this case, distributed channel assign-ment algorithms should assign distinct channels toany nodes within a k-hop neighborhood. However,in many cases, the number of available channelsmay be less than the number of nodes in a k-hopneighborhood. Therefore, two or more nodes mayneed to share the same channel. To balance thenumber of nodes sharing the same channel, themain design objective of a distributed channelassignment algorithm is to minimize the maximumnumber of nodes sharing the same channel with adesignated node vt

j 2 V t within its k-hop neighbor-hood, i.e.

Minimize maxvt

j2V t

fnkðvtjÞg ð2Þ

Subject to

P ðvtT ðiÞ; v

ri Þ

Pvt

j2V tnvtT ðiÞ½P ðvt

j; vri Þ � Sðvt

T ðiÞ; vtjÞ� þ P N

P b; ð3Þ

66 M.X. Gong et al. / Ad Hoc Networks 7 (2009) 63–78

where P N is the power level of additive white Gauss-ian noise (AWGN) noise and b is the minimumSINR required for a successful packet reception.The term Sðvt

T ðiÞ; vtjÞ indicates the adjacent channel

interference between the channels used by the de-sired transmitter vt

T ðiÞ and transmitter vtj. We have

SðvtT ðiÞ; v

tjÞ ¼ 0 if vt

T ðiÞ and vtj use strictly orthogonal

channels; SðvtT ðiÞ; v

tjÞ ¼ 1 if vt

T ðiÞ and vtj use the same

channel. Generally we have 0 < SðvtT ðiÞ; v

tjÞ < 1.

A channel assignment protocol should distributeavailable channels with any pre-defined value of k insuch a way that the maximum number of transmit-ters that share the same data channel is minimized.Meanwhile, the same set of channels should bereused in such a way that the accumulated interfer-ence generated on any particular data channel isbelow a certain threshold. Because the channelassignment problems are shown to be NP-complete,there exists no polynomial algorithms for generalmobile ad hoc networks. Thus, efficient heuristicalgorithms, including both centralized [4] and dis-tributed [9,10], become particularly appealing. Inaddition, since it is very challenging to obtain globalstatistics in a mobile ad hoc network, a realisticchannel assignment protocol should be distributed,i.e. it should operates with local information only.In the following section, we present such distributedheuristic algorithms for the formulated problem,which provide near-optimal solutions.

3. Distributed channel assignment protocols

In this section, we first describe the three designprinciples on distributed channel assignment andthen introduce three protocols that implement thesedesign principles. It is assumed that channel assign-ment is transmitter-based, meaning that distinctchannels are assigned to different transmitters.1 Ithas been shown that transmitter- and receiver-basedchannel assignment problems are essentially equiva-lent [9]. Therefore, the proposed design principlesapply to receiver-based protocols and the proposedprotocols can be easily modified to assign channelsbased on receivers as well.

1 Note that actually the channels are assigned to the links.When a transmitter has a frame to send, it will notify thecorresponding receiver to tune its transceiver to the assignedchannel, in order to receive the frame (see Section 4). That is, thechannel is actually assigned to this link from the transmitter tothe receiver.

3.1. Design principles

The first and primary principle is ‘‘cross-layer”

design, where channel assignment is jointly consid-ered with routing. This is motivated by the fact thatboth channel assignment and routing will beinvoked when there is a topology change. In addi-tion, piggybacking channel information in routingcontrol messages can greatly reduce the communi-cation overhead of channel assignment protocols.For instance, an existing channel assignment algo-rithm has a communication complexity ofOðd2 � jV jÞ, where d is the maximum number ofone-hop neighbors that a node can have and jV j isthe total number of nodes in the network [11]. Sucha complexity implies that whenever there is a topol-ogy change, up to Oðd2 � jV jÞ messages will beexchanged in the network. Similarly, a recently pro-posed channel assignment protocol has a complex-ity of OðKjV j3 log mþ m2Þ [13], where m is thetotal number of radio connections in the networkand K is the minimum number of neighbors that anode has. Such high complexity makes it difficultto implement these protocols in a mobile ad hoc net-work environment where topology is constantlychanging. By exploiting the cross-layer design prin-ciple, a combined Channel Assignment and AODV(CA-AODV) algorithm can reduce the communica-tion complexity to Oð1Þ, since all channel informa-tion is carried in routing control messages (seeSection 3.2).

The second design principle states that channelsshould be assigned only to active nodes. Before anode acquires a valid route, it cannot transmit orreceive data packets and, thus, does not need a chan-nel. We call this type of node an ‘‘inactive” node, anda node in a valid route an ‘‘active” node. Most exist-ing channel assignment schemes assign channels to all

nodes in a wireless network, regardless whether theyare active or inactive [9–11,13]. If not all nodes in anetwork are active at the same time, such schemesmay require more wireless channels than necessary.By assigning channels on-demand to only activenodes, we can potentially reduce the number of chan-nels required in a wireless network, since the numberof required wireless channels is generally propor-tional to the number of active nodes rather than thetotal number of nodes in the entire network.

Finally, distinct channels should be assigned insuch a way that hidden nodes, exposed nodes andinterference can be avoided as much as possible.Many existing channel assignment schemes are

M.X. Gong et al. / Ad Hoc Networks 7 (2009) 63–78 67

designed to solve the hidden node problem [9–11].However, the exposed node problem can alsoreduce channel utilization. Interference can alsocorrupt data packets and reduce throughput. Wepropose to assign distinct channels to any nodeswithin a k-hop neighborhood, where k is a designparameter that provides a suitable tradeoff betweenthe number of required channels and the achievedinterference level, and between the system perfor-mance and control overhead. Therefore, concurrenttransmissions in the k-hop neighborhood can beachieved to improve the network performance.

3.2. Distributed channel assignment protocols

The proposed design principles can be applied toboth reactive routing algorithms, such as Ad hoc OnDemand Distance Vector (AODV) [14] protocol,and proactive routing protocols, such as the Opti-mal Link State Routing (OLSR) protocol [15]. Inthe following, we will use AODV as an example todemonstrate the three design principles. AODV isa reactive routing protocol, which means nodes donot maintain up-to-date routes to all destinationsat all times. Instead, a node initiates a route discov-ery procedure by broadcasting a Route Request(RREQ) message only when it has packets for thedestination node and it has no valid route to thedestination. Upon receiving the RREQ message,the destination node or an intermediate node thathas a valid route to the destination will send backa Route Reply (RREP) message to the source node.The RREP sets up a path from the source to thedestination as it is forwarded back to the sourcenode.

Throughout this paper, we assume that eachnode is equipped with two transceivers: one for datatransmissions and the other for control messages.Control messages, such as RREQ and RREP, aresent on the control channel that is shared by allthe nodes in the network. Data packets may be senton different data channels that have been assignedto active nodes in the network. In the following,we present a combined channel assignment andAODV scheme, namely, Channel AssignmentAODV (CA-AODV), and its two extensions:Enhanced 2-hop CA-AODV (E2-CA-AODV) andEnhanced k-hop CA-AODV (Ek-CA-AODV).

3.2.1. Combined channel assignment and AODV

CA-AODV, like AODV, operates in two phases:route discovery and route reply. During the route

discovery phase, channel information about anode’s k-hop neighbors along the same route is car-ried by the broadcast RREQ message. Any nodethat receives the RREQ message updates its next-hop table entries with respect to preceding nodesin the path back to the source. Each table entry con-sists of both the route and the indices of channelsthat have been taken, so far, by the node’s k-hopneighbors on the same route. If the node has nochannel assigned to it, it updates its available chan-nel set, denoted by A � C, by marking the channelstaken by the preceding k (or fewer if the route is notk hops long) nodes on the path as unavailable.Then, it randomly picks a channel from the set ofavailable channels A. Furthermore, this node willassociate a timer with these updates. If this nodedoes not receive an RREP before the timer expires,these updates will be restored to the original states,since this node will not be included in that path.

During the route reply phase, channel informa-tion is carried in the unicast RREP message. Uponreceiving the RREP packet, each node along theroute updates its next-hop table entries, as well asthe index of the channel to be used for this link.After the route has been established, each nodealong the route should have a channel that is differ-ent from any of its k-hop neighbors on the sameroute. As in AODV, a route expires if it is not usedor reactivated for a certain period of time. The entrythat corresponds to this route will then be deletedfrom the routing table and the channels assignedto this route will become available for re-assignment.

A flowchart that describes the operation ofCA-AODV is shown in Fig. 1. Figs. 2–5 describeits four main procedures, i.e. sendRREQ(), recvR-REQ(RREQ), sendRREP() and recvRREP(RREP).sendRREQ() and recvRREQ(RREQ) are invoked inthe route discovery phase, while sendRREP() andrecvRREP(RREP) are invoked in the route replyphase. In recvRREQ(RREQ) and recv(RREP), get-

NeighborInfo() retrieves channel information fromRREQ or RREP messages received from neighboringnodes, while randomChannel(A) returns a channelindex that is uniformly chosen from the availablechannel set A. As shown in Fig. 5, a node willupdate its available channel set A upon receivingan RREP. If a channel conflict is found, the noderandomly chooses a channel for itself from setA. This node will then forward the RREP carryingits own channel and its upstream neighbors’channels.

init

recvRREPrecvRREQ

wait

Channel conflict ?

Update set A

sendRREQ

Update set A

Pick a random channel

Know a route to destination ?

sendRREP

Update myChannel

Yes

YessendRREP

No

No

Am I the source ?Yes

NoRouteestablished

Am I an active node ?

Yes

No

Keep my channel

Fig. 1. Algorithm flowchart of CA-AODV.

Fig. 2. Procedure sendRREQ() in CA-AODV.

Fig. 3. Procedure recvRREQ(RREQ) in CA-AODV.

Fig. 4. Procedure sendRREP() in CA-AODV.

Fig. 5. Procedure recvRREP(RREP) in CA-AODV.

68 M.X. Gong et al. / Ad Hoc Networks 7 (2009) 63–78

CA-AODV introduces very low control over-head, since the channel information is completelypiggybacked in routing control messages (e.g., as alist of channel IDs or in the form of a ‘‘bit-map”).However, CA-AODV only considers ‘‘intra-path”contention. It does not explicitly attempt to reduce

collisions and interference introduced by neighbor-ing paths. It can be used when ‘‘intra-path” conten-tion is the dominant limiting factor, e.g. in sparsenetworks with a few long lasting multi-hop sessions.In the rest of this section, we present two extensionsof CA-AODV that can effectively handle both‘‘intra-path” and ‘‘inter-path” contentions. Weshow that with a small increase in control overhead,the performance of CA-AODV can be furtherimproved.

3.2.2. Enhanced 2-hop channel assignment and

AODV (E2-CA-AODV)To improve the performance of CA-AODV while

still maintaining low control overhead, an extensionof CA-AODV, i.e. E2-CA-AODV, is proposed.Unlike CA-AODV that has constant communica-tion overhead, E2-CA-AODV has a linear commu-nication overhead, i.e. OðdaÞ, where da is themaximum number of active neighbors that a nodecan have. E2-CA-AODV seeks to assign distinct

Fig. 6. Procedure sendHello() in E2-CA-AODV.

Fig. 7. Procedure recvHello(HELLO) in E2-CA-AODV.

M.X. Gong et al. / Ad Hoc Networks 7 (2009) 63–78 69

channels to active nodes within any two-hop

neighborhood.In addition to RREQ and RREP messages, E2-CA-

AODV uses another AODV routing control mes-sage, the HELLO message, for channel assignment.As in AODV, HELLO messages are broadcast peri-odically among one-hop neighbors. If a node isactive, it will indicate its assigned channel and aNodeNumber in its HELLO messages. Each time anode chooses a data channel, it updates its Node-

Number by generating a random number in ½1;M �,where M � 1 in order to minimize the chance thattwo or more neighboring nodes choose the sameNodeNumber. In addition, HELLO messages alsocarry the channel and NodeNumber information ofa node’s active one-hop neighbors. Upon receivingHELLO messages from one-hop neighbors, a nodewill learn the channel assignments within its 2-hopneighborhood, and update its available channel setA by removing the channels taken by active neigh-bors. The NodeNumber is used for resolving channelconflicts. If two or more active nodes choose thesame channel, the node with the smallest NodeNum-

ber will retain its channel while the other nodes shallrandomly pick another data channel from A. Notethat NodeNumber is randomly generated each time achannel is chosen. So every node will have a fairchance of winning a channel when involved in a col-lision, while schemes using static node IDs willalways favor a node with a lower (or higher) ID.Finally, the node that updates its own data channelwill inform its neighbors in the next HELLO

message.The other operations of E2-CA-AODV, such as

Route Request and Route Reply, are similar tothose of CA-AODV. As discussed, E2-CA-AODVpiggybacks channel assignment information in threetypes of routing control messages, i.e. RREQ, RREP,and HELLO. Therefore, E2-CA-AODV has twoextra procedures, i.e. sendHello() and recv-Hello(HELLO), as shown in Figs. 6 and 7, for han-dling the Hello messages. In recvHello(HELLO),the getActiveNeighborInfo(HELLO) functionretrieves channel information from received HELLO

messages.Because E2-CA-AODV exchanges channel

information among two-hop neighbors, it cansuccessfully mitigate both intra-flow and inter-interference with the 2-hop neighborhood, result-ing in lower interference and more concurrenttransmissions. We will present its performance inSection 5.

3.2.3. Enhanced k-hop channel assignment and

AODV (Ek-CA-AODV)

Because CA-AODV only considers collisions andinterference on the same route, it does not performwell when multiple active nodes on different routesco-locate in the a k neighborhood. E2-CA-AODVseeks to both avoid collisions and mitigate intra-flow interference in the two-hop neighborhoodwhile maintaining low control overhead. In orderto further reduce cumulative interference from trans-mitters beyond the two-hop neighborhood, we pro-pose Ek-CA-AODV that considers both collisionsand interference within a k-hop neighborhood(where k P 2), at the cost of higher communicationoverhead.

Ek-CA-AODV introduces an extra controlmessage called ChannelTaken. Most of the

Fig. 9. Procedure recvChannelTaken (ChannelTaken) in Ek-CA-AODV.

70 M.X. Gong et al. / Ad Hoc Networks 7 (2009) 63–78

operations of Ek-CA-AODV are similar to thoseof E2-CA-AODV, except for the operation involv-ing the extra control message ChannelTaken. Ifa node on an established route detects that a newroute in the neighborhood is being set up, it shallbroadcast a ChannelTaken message that carriesits own channel index. The TTL of the Channel-Taken message is set to k to ensure that theChannelTaken message be broadcast only tothe current node’s k hop neighbors. Upon receiv-ing a ChannelTaken message, each node willupdate its next-hop neighbor table and the avail-able channel set A. If a channel conflict is detectedby a node that has not yet on an established route,this node shall set a channelConflict flag.Once receiving a RREP message, a node checks tosee whether the channelConflict flag is set.If so, the node will randomly pick another channelfrom the channel set A. Through ChannelTaken

messages, channels taken by nodes on establishedroutes can be conveyed to other nodes in the net-work, up to k hop away. Therefore, conflictingchannels within the k-hop neighborhood can belargely avoided, provided that the number of avail-able channels is sufficiently large. The two addi-tional procedures for handling ChannelTaken

messages, i.e. sendChannelTaken() and recvChan-nelTaken(), are shown in Figs. 8 and 9.

To allow sufficient time for ChannelTaken

messages to propagate to all nodes within the k-hop range, the destination node or a node that hasa valid route to the destination should wait for aperiod of time, denoted by W t, before returningthe RREP message. The value of W t provides atradeoff between route discovery delay and the cor-rectness of channel assignment information. Wechoose W t to be a function of both k and tp, i.e.

W t ¼ a� k � tp; ð4Þ

where 1 6 a 6 2 is a constant and tp is the per hoppropagation delay.

Parameter a is used to accommodate variationsin tp. Usually a large value should be used if the net-work is dense and traffic load is high; while smallvalues should be used for low-load sparse networks.

Fig. 8. Procedure sendChannelTaken() in Ek-CA-AODV.

With the assumption that a transceiver and channelis specifically used for control messages, the delayfor the ChannelTaken message to propagate tothe entire k-hop neighborhood will only be affectedby the control traffic load. Given a traffic load andnetwork density, we should choose an a value thatis sufficiently large so that ChannelTaken mes-sages can propagate to the entire k-hop neighbor-hood, while it should also be small so as tominimize the route discovery delay.

Because a ChannelTaken message is relayed toan active node’s k-hop neighbors, in the worst case,every active node in the network may send a Chan-nelTaken message to its k-hop neighbors. Thus,the communication overhead of Ek-CA-AODV isOðjV tj � jV kjÞ, where jV tj is the number of activenodes in the network and jV kj is the number ofnodes in a k-hop neighborhood. Both the computa-tion overhead and the storage overhead of Ek-CA-AODV are OðjV kjÞ since each node only needs toprocess and store channel information of its k-hopneighbors, denoted by jV kj. Due to its higher com-munication overhead, Ek-CA-AODV is best suitedfor a mobile ad hoc network with low mobility,where link breakage is not so often and channelsdo not need to be frequently re-assigned to nodes.However, Ek-CA-AODV has its strength in the casewhen interference range is more than 2-hops andwhen cumulative interference needs to be consid-ered, as compared to the previous two schemes.

3.3. Protocol analysis

In this section, we first derive an upper bound onthe number of distinct channels required by Ek-CA-AODV (k P 2). We then prove the correctness ofEk-CA-AODV. The proof for CA-AODV is similarand is omitted for brevity.

Proposition 1. To assign distinct channels to anynode within a k-hop range, the number of channels

required has an upper bound of nr � ðk þ 1Þ, where nr

D B RTS

S CTS

S DATA

S ACK

Node A

Node B

Nodes that have a channel conflict with Node A

Common control channel Node A’s data channel

NAV(RTS) NAV(DATA) D

DNAV(CTS)

Defer AccessB: backoff D: DIFSS: SIFS

Fig. 10. The four-way handshake procedure of MC-MAC.

M.X. Gong et al. / Ad Hoc Networks 7 (2009) 63–78 71

is the number of active routes that lie within the k-hop

range of any node in the network.

Proof. This proposition can be proven byinduction.

(i) Base Case: If there is only one route in the net-work, it can be easily shown that the numberof required distinct channels is k þ 1.

(ii) Induction Step: Assume that when there are n

active routes within a k-hop range, therequired number of channels is n � ðk þ 1Þ. Ifthere are nþ 1 active routes within a k-hoprange, the ðnþ 1Þth route can be assignedk þ 1 new channels that are different fromany of the previous n � ðk þ 1Þ channels. Thetotal number of channels needed for nþ 1active routes is then ðnþ 1Þ � ðk þ 1Þ.

Therefore, the proposed algorithms need at mostnr � ðk þ 1Þ distinct channels. h

Note that this upper bound is achieved whenthere is no common links among active routes. Ifthat is the case, channels are assigned to each activeroute independently and each route requires k þ 1channels. If active routes in the network share links,the network requires fewer channels than n � ðk þ 1Þ.

Proposition 2. After a new route has been estab-

lished, each node along the new route is assigned adistinct channel among its k-hop neighbors, provided

that both the number of available channels and W t is

sufficiently large.

Proof. Under the assumption that the channelassignment procedure is not disrupted by sudden fail-ure or malfunction of nodes, the channel informationcarried by control messages is consistent with thechannel information saved at each node. Moreover,ChannelTaken messages can propagate to allnodes within a k-hop range, provided that W t is suffi-ciently large. When a new route is established, eachnode on the route will have its k-hop active neighbors’channel information. Because a node randomly picksits own channel from the available channel set, whichdoes not contain any of its k-hop neighbors’ channels,this node must have a channel that is distinct fromany of its k-hop neighbors. h

Many existing algorithms do not explicitly con-sider the case when the number of available chan-nels is not sufficiently large [9–11]. The proposed

algorithms can effectively handle this case. A gen-eral rule would be to let the nodes using the samechannel to be as far apart from each other as possi-ble. Specifically, if k 6 2, a node with A ¼ ; shouldrandomly pick a channel from the set of least reusedchannels (in its k-hop neighborhood) in C. If k P 3,this node should randomly pick a channel from thechannels that are taken by nodes two-hop away.Thus, the algorithms ensure minimum number ofcollisions, as well as taking interference into consid-eration (see Eqs. (2) and (3)).

4. Description of the MC-MAC protocol

The proposed channel assignment and routingprotocols have a companion multi-channel MAC(MC-MAC) protocol, which is designed to allowsimultaneous data transmissions on different datachannels. It is assumed that there is one dedicatedcontrol channel and up to N data channels in thenetwork. Each data channel is equivalent and hasthe same bandwidth. Recall that each host isequipped with two transceivers, Transceiver I oper-ating on the common control channel all the timeand Transceiver II that switches from one channelto another for data packets.

MC-MAC is a transmitter-based protocol. Asdiscussed, nodes are assigned channels by the com-bined channel assignment and routing protocols.RTS and CTS messages are sent on the controlchannel with Transceiver I, while data packets andACKs are sent on the assigned data channel byTransceiver II. When a node is ready to transmit,it will first convey its assigned data channel to thedestination node through RTS/CTS exchange. Asshown in Fig. 10, when a sender, say node A,intends to transmit, its Transceiver I will broadcastan RTS message carrying its own data channelindex, cAB. Upon receiving the RTS message, the

72 M.X. Gong et al. / Ad Hoc Networks 7 (2009) 63–78

destination, say Node B, replies a CTS message car-rying cAB from its Transceiver I and sets its Trans-ceiver II to channel cAB. After Node A receives theCTS from the control channel, its Transceiver IIswitches to the confirmed data channel cAB andstarts data transmission.

Neighboring nodes that overhear the RTS/CTSexchange but do not share the same data channelwith Node A should defer only for the duration ofthe RTS/CTS exchange. If a node is assigned thesame data channel cAB, two situations may happen,as illustrated in Fig. 10.

– If the node overhears a CTS message, it shoulddefer from using the data channel cAB until theend of the data transmission to avoid causing acollision at the receiver.

– The node that overhears only an RTS message,but not a CTS message, should first defer fromusing the control channel only for the durationof the control packet transmission. Then, it per-forms carrier sensing on the data channel cAB.If the carrier is busy, which means that the trans-mitting node has successfully acquired the med-ium, the node should defer for the duration ofthe data packet transmission. However, if thecarrier is not busy, the node can start to contendfor this channel immediately.

The sender listens on the data channel until anACK is received or a timeout occurs. If a nodereceives an RTS on the control transceiver while itsdata transceiver is busy communicating with anothernode, it returns a Negative CTS (NCTS) on the con-trol channel to the sender, indicating that a collisionhas not occurred. Thus the sender is not obliged toincrease its contention window nor to back off.

Because MC-MAC transmits control packets onthe common control channel and data packets onnon-overlapping orthogonal data channels, thetransmissions on the control channel and on differ-ent data channels can occur in parallel. This type ofparallelism is sometimes called ‘‘pipelining” [16].When nodes have different data channels from theon-going data transmissions, they need to defer onlyfor the duration of control packet transmission.Since the size of data packets is usually much largerthan that of control packets, many data transmis-sions can occur in parallel on different data chan-nels. Also note that MC-MAC solves the hiddenterminal problem as in the IEEE 802.11 MAC byusing the RTS/CTS dialog to reserve data channels.

5. Performance evaluation

In this section, we first compare the performanceof E2-CA-AODV with that of an existing channelassignment protocol, i.e. the channel assignmentscheme (CAS) [11]. We then demonstrate the capac-ity improvement of E2-CA-AODV and MC-MACover the original IEEE 802.11 MAC. All the simula-tions reported in this section are performed usingthe ns-2 simulator [17,18].

5.1. Simulation setting

We assume that 64 wireless nodes are placed ran-domly in a square area. Each node is equipped withtwo half-duplex transceivers. All nodes in the net-work share the same common control channel.There are 6 or 12 different data channels available.In all simulations, the radio range of a node is setto 250 m and the interference range is set to550 m, which is approximately twice the radio range[18]. The two-ray ground propagation model isselected and the physical channel bandwidth forall data channels and the control channel is set to2 MB/s [18]. Most current wireless LAN cards havea channel switch delay of 40–80 ls [19], and wetherefore assume a channel switch delay of 80 ls.Each simulation lasts for 7600 s, where the warm-up period is 3600 s and the effective simulation timeis 4000 s, in order to get steady state statistics [20].

Four UDP flows are generated in the network.Each UDP flow has an offered load ranging from40 KB/s to 1000 KB/s. For routing, AODV is usedwith the IEEE 802.11 MAC and CAS, while E2-CA-AODV is used with MC-MAC. We chooseCAS because it is one of a few distributed channelassignment protocols that can operate in the mobile,multi-hop environment. Additionally, unlike otherchannel assignment schemes that were proposed towork with the IEEE 802.11 MAC [3,13,21], CAScan work with the proposed MC-MAC.

Two wireless ad hoc networks are simulated: a800 m � 800 m dense network and an 1600 m by1600 m sparse network. Mobile nodes move ran-domly according to the random waypoint mobilitymodel, where the maximum speed is 5 m/s and theminimum speed is 4 m/s. The maximum pause timeis 5 s. Although we have also studied E2-CA-AODVand CAS at higher speeds, the results are not signif-icantly different. We find that the network density,on the other hand, has a much greater impact onthe performance than node speed. This coincides

0 100 200 300 400 500 600 700 800 900 10000

0.2

0.4

0.6

0.8

1

1.2

1.4

1.6

1.8

2

Offered Load Per Flow (kbps)

Agg

rega

ted

Thr

ough

put (

Mbp

s)

802.11MACUnlimited-ChannelE2-CA-AODV with MC–MAC (N=12)CAS with MC–MAC (N=12)

(a) Aggregate throughput

200 300 400 500 600 700 800 900 10000

0.2

0.4

0.6

0.8

1

1.2

1.4

1.6

Offered Load Per Flow (kbps)

Del

ay (

s)

802.11MACUnlimited ChannelE2 CA AODV with MC–MAC (N=12)CAS with MC–MAC (N=12)

(b) Average end-to-end packet delay

Fig. 11. E2-CA-AODV vs CAS in a dense network. (a) Aggre-gate throughput and (b) average end-to-end packet delay.

M.X. Gong et al. / Ad Hoc Networks 7 (2009) 63–78 73

with observations made by Bahl, Chandra, andDunagan in [19].

5.2. Performance comparison

We first compare E2-CA-AODV with threeschemes: (i) the single-channel IEEE 802.11 MAC,(ii) a scheme with unlimited number of data chan-nels, and (iii) CAS. The IEEE 802.11 MAC servesas a lower bound for our performance study. In theunlimited data channel scheme, each node has itsown unique data channel in the network and a com-mon control channel is shared by all nodes in thenetwork. This is the ideal case with unlimited net-work resource and serves as a performance upper

bound: its performance upper bounds that of anydistributed/centralized channel assignment proto-cols with a finite set of channels.

CAS assigns distinct channels, or codes, to anode and its two-hop neighbors. In CAS, each nodesends out code assignment messages (CAM) thatpropagate to its one-hop neighbors. CAMs aretransmitted under three conditions: (i) when a newnode comes up, (ii) when a node detects a changeof code by any of its one-hop neighbors, and (iii)when a node finds that one of its one-hop neighborsis no longer active. A CAM contains: (i) the addressand code of the source node, (ii) the addresses andcodes of source node’s one-hop neighbors, (iii) theacknowledgements to earlier received CAMs, and(iv) a response list of zero or more nodes which needto send an ACK for this CAM. The communicationcomplexity of CAS is Oðd2 � jV jÞ, where d is themaximum number of one-hop neighbors for anynode and jV j is the total number of nodes in the net-work [11]. In contrast, E2-CA-AODV has a linearcomplexity OðdaÞ, since channel information is pig-gybacked in routing control messages.

Fig. 11 plots the performance of the aboveschemes in the dense network, and Fig. 12 plotsthe performance of the schemes in the sparse net-work. For aggregate throughput, Figs. 11a and12a show that because of its high control overhead,CAS performs worse than E2-CA-AODV in bothdense and sparse networks. Especially in a densenetwork where a node may have many neighbors,the control overhead of CAS is so high that whenthe data rates are lower than 300 Kbps, CAS with12 data channels performs even worse than the sin-gle-channel IEEE 802.11 MAC (see Fig. 11a). Thatis, the performance degradation caused by the con-trol overhead dominates the performance gain of

using multiple channels. Similar observation canbe made in Fig. 12a for rates lower than 200 Kbps.

Figs. 11b and 12b show that the delay perfor-mance of CAS is better than that of the IEEE802.11 MAC protocol, but still not as good as thatof E2-CA-AODV. Using multiple channels, bothE2-CA-AODV and CAS can effectively reduce col-lisions and contention in the network. Thus, theend-to-end delay suffered by a data packet is greatlyreduced, as compared with the single-channel IEEE802.11 MAC.

We then compare E2-CA-AODV with a random-AODV algorithm. In the random-AODV algo-rithm, a node determines its own channel indexbased on its MAC address. Therefore, no channelassignment is needed in the random-AODV scheme,but there will be more collision and severer interference

0 100 200 300 400 500 600 700 800 900 10000

0.1

0.2

0.3

0.4

0.5

0.6

0.7

0.8

0.9

1

Offered Load Per Flow (kbps)

Agg

rega

ted

Thr

ough

put (

Mbp

s)

802.11MACUnlimited-ChannelE2-CA-AODV with MC–MAC (N=12)CAS with MC–MAC (N=12)

(a) Aggregate throughput

200 300 400 500 600 700 800 900 10000

0.2

0.4

0.6

0.8

1

1.2

1.4

Offered Load Per Flow (kbps)

Del

ay (

s)

802.11MACUnlimited-ChannelE2-CA-AODV with MC–MAC (N=12)CAS with MC–MAC (N=12)

(b) Average end-to-end packet delay

Fig. 12. E2-CA-AODV vs CAS in a sparse network. (a)Aggregate throughput and (b) average end-to-end packet delay.

0 100 200 300 400 500 600 700 800 900 10000

0.5

1

1.5

2

Offered Load Per Flow (kbps)

Agg

rega

ted

Thr

ough

put (

Mbp

s)

802.11MACUnlimited-ChannelRandom MC–MAC (N=6)Random MC–MAC (N=12)E2–CA–AODV with MC–MAC (N=6)E2–CA–AODV with MC–MAC (N=12)

(a) Aggregate throughput

200 300 400 500 600 700 800 900 10000

0.2

0.4

0.6

0.8

1

1.2

1.4

1.6

1.8

2

Offered Load Per Flow (kbps)

Del

ay (

s)

802.11MACUnlimited-ChannelRandom MC–MAC (N=6)Random MC–MAC (N=12)E2–CA–AODV with MC–MAC (N=6)E2–CA–AODV with MC–MAC (N=12)

(b) Average end-to-end packet delay

Fig. 13. Performance of E2-CA-AODV in a dense network. (a)Aggregate throughput and (b) average end-to-end packet delay.

74 M.X. Gong et al. / Ad Hoc Networks 7 (2009) 63–78

since the channels are randomly choosing withoutany coordination.

The simulation results for the dense network andthe sparse network are presented in Figs. 13 and 14,respectively. Fig. 13a shows that E2-CA-AODValways achieves a higher throughput than the ran-dom-AODV scheme in a dense network, given anequal number of channels. The performance gapbetween the two is larger as the number of availablechannels decreases. We also find that E2-CA-AODV combined with MC-MAC can have athroughput up to three times higher than that ofIEEE 802.11 MAC. Fig. 13b shows that the end-to-end delay increases for all schemes when the datarate increases. However, both multi-channelschemes with 12 data channels have delays muchlower than that of IEEE 802.11 MAC.

Fig. 14a and b shows the throughput and delayfor all schemes in a sparse network. The perfor-mance gaps between different schemes are not aslarge as those in the dense network, although E2-CA-AODV still achieves an clear improvement overthe random-AODV scheme. In a sparse network,the interference and collisions generated by neigh-boring nodes is much lower compared to those ina dense network. This is because each node hasfewer neighbors in a sparse network. Therefore,the performance gain of the multi-channel MACschemes is not as significant as in a dense network.

5.3. Remarks

Based on the simulation results presented in theprevious section, we can make several interesting

0 100 200 300 400 500 600 700 800 900 10000

0.2

0.4

0.6

0.8

1

Offered Load Per Flow (kbps)

Agg

rega

ted

Thr

ough

put (

Mbp

s)

802.11MACUnlimited-ChannelRandom MC–MAC (N=6)Random MC–MAC (N=12)E2–CA–AODV with MC–MAC (N=6)E2–CA–AODV with MC–MAC (N=12)

(a) Aggregate throughput

200 300 400 500 600 700 800 900 10000

0.2

0.4

0.6

0.8

1

1.2

1.4

1.6

Offered Load Per Flow (kbps)

Del

ay (

s)

802.11MACUnlimited-ChannelRandom MC–MAC (N=6)Random MC–MAC (N=12)E2–CA–AODV with MC–MAC (N=6)E2–CA–AODV with MC–MAC (N=12)

(b) Average end-to-end packet delay

Fig. 14. Performance of E2-CA-AODV in a sparse network. (a)Aggregate throughput and (b) average end-to-end packet delay.

M.X. Gong et al. / Ad Hoc Networks 7 (2009) 63–78 75

observations, which are discussed in the following.First, communication overhead has a profoundimpact on the performance of distributed channelassignment protocols. Both E2-CA-AODV andCAS seek to assign distinctive channels to nodesin a two-hop neighborhood. However, because E2-CA-AODV has lower communication overhead, ithas much better performance than CAS in termsof throughput and delay, and the performance gainis more significant in a dense network. On the otherhand, if the control overhead is not prohibitivelyhigh, utilizing multiple channels always gives betterperformance than the IEEE 802.11 MAC scheme.The reason is that the use of multiple channelsincreases the possibility of concurrent transmissionsin the network.

Second, in all simulated scenarios, the perfor-mance gap between the scheme with an unlimited

number of channels and E2-CA-AODV with 12data channels, is not significant. Thus, we can con-clude that due to the negligible interference gener-ated by distant nodes, a large number of datachannels is not necessary to achieve the most bene-fits of the use of multiple channels. The optimalnumber of channels should be a function of nodedensity and interference range.

Third, the performance gap between E2-CA-AODV and the random scheme decreases whenthe number of available channels increases. By intel-ligently assigning channels, E2-CA-AODV caneffectively avoid collisions and mitigate intra-flowinterference. The random scheme merely tries tomake use of all the available channels without coor-dination among the nodes and consideration of col-lision and interference. However, when there is alarge number of available channels, collisions maybe infrequent even if channels are assignedrandomly.

Fourth, the performance gains achieved by of E2-CA-AODV/MC-MAC over AODV/IEEE 802.11MAC, is not in proportion to the number of chan-nels utilized. For instance, for E2-CA-AODV/MC-MAC with 12 available data channels, the through-put gain can be up to three times, rather than 12times, as compared to the throughput achieved bythe AODV/IEEE 802.11 MAC. MC-MAC assumesthat all nodes in the network share a common con-trol channel. With an increase in the number of datachannels, the control traffic also increases and causecongestion and collisions in the common controlchannel, as more nodes try to transmit in parallel.Therefore, the time period a node can spend ontransmitting data packets is reduced, whichdegrades the performance of E2-CA-AODV. Thecontrol channel becomes the performance bottle-neck. In addition, the channel switch delay is non-negligible, which further degrades the performance.In fact, Kyasanur and Vaidya [22] show that whenthe number of transceivers is less than the numberof available channels, network capacity may evenbe worse.

6. Related work

The channel assignment problem has been widelystudied in the context of infrastructure-based wire-less networks, such as cellular networks [23] andIEEE 802.11 WLANs [4]. This class of work consid-ers the interference among access points or base sta-tions, and focuses on assigning channels to the base

76 M.X. Gong et al. / Ad Hoc Networks 7 (2009) 63–78

stations to reduce interference and accommodate agiven network traffic load. Therefore, such schemesbelong to the network planning paradigm, which isquite different from the dynamic, infrastructurelessad hoc network environment considered in thispaper.

In [9], Hu studies the problem of distributed codeassignment for CDMA packet radio networks,including ad hoc networks. Under the assumptionthat each node has a neighbor table updated by anetwork-layer routing protocol, Hu’s approachtransforms the code assignment problem into agraph theory problem. This problem is shown tobe NP-complete and fast heuristic algorithms aredeveloped. Even though the solutions proposedare sound, they have high time complexity and highcommunication overhead. The schemes do not con-sider the case when the number of codes is limitedand perfect assignment is not possible. Moreover,Hu considers only static networks in his designs.

Garcia-Luna-Aceves and Raju [11] describe adistributed code assignment scheme (CAS) thatworks in a mobile ad hoc network (see the previoussection). CAS assigns distinct channels, or codes, toa node and its two-hop neighbors. If the number ofcodes available for assignment is at leastdðd � 1Þ þ 2, where d is the maximum degree, i.e.,the maximum number of neighbors for any node,it is shown that there will be no interference afterthe algorithm converges. However, this algorithmincurs high communication and computation over-

Table 3Comparison of MC-MAC to existing multi-channel MAC protocols

Protocols Medium access Channel sel

DCA [5] CSMA/CA Per packetMMAC [25] CSMA/CA Per beaconMulti-channel CSMA [6] CSMA/CA Per packet

RICH-DP [26] Channel hopping Hopping seSSCH [19] Channel hopping Hopping seMC-MAC CSMA/CA Per route c

Table 2Comparison of CA-AODV to existing algorithms

Protocols No. Channels Commun

Centralized greedy algorithm [9] dðd � 1Þ þ 1 N/ADistributed channel assignment [11] dðd � 1Þ þ 2 d2 � jV jRandom scheme N/A Oð1ÞCA-AODV k þ 1 OðkÞE2-CA-AODV daðda � 1Þ þ 1 OðdaÞEk-CA-AODV nðk þ 1Þ OðjV tj � j

head, as well as high time and storage complexity.This algorithm does not explicitly consider the casewhen the number of available channels is less thandðd � 1Þ þ 2.

There are several recent proposals for routingprotocols that are suitable for multi-hop multi-channel wireless mesh networks [3,13,21,24]. Theapproach taken by most of these proposals is tocombine routing with intelligent multi-channelassignment, such that channel utilization is maxi-mized and the system performance can be substan-tially improved. However, because the focus ofthese approaches is routing, the performance ofthe channel assignment schemes has not been stud-ied [3,21,24]. In addition, some channel assignmentprotocols have high time complexity, e.g.OðKjV j3 log mþ m2Þ [13], where m is the total num-ber of radio connections in the network and K is theminimum number of neighbors that a node has.

Table 2 compares several heuristic algorithms interms of their communication, computation, andstorage complexity. In the table, d is the maximumnumber of one-hop neighbors for any node (i.e.degree), da 6 d is the maximum number of activeone-hop neighbors for any node, k is a constant thatrepresents a neighborhood size, jV tj is the numberof active nodes in the network, jV kj is the numberof nodes in a k-hop neighborhood, and n is the totalnumber of active routes within a k-hop range ofeach other. For CA-AODV and E2-CA-AODV,the channel assignment information is piggybacked

ection Hardware requirement Synchro. required

2 transceivers Nointerval 1 transceiver Yes

1 transmitter Nomultiple receivers

quence 1 transceiver Yesquence 1 transceiver Yeshange 2 transceivers No

. complexity Comput. complexity Storage complexity

d2 � jV j d2 � jV jd2 d2

Oð1Þ Oð1ÞOðkÞ OðkÞOðd2

aÞ OðdaÞV k jÞ OðjV k jÞ OðjV k jÞ

M.X. Gong et al. / Ad Hoc Networks 7 (2009) 63–78 77

in routing messages. In addition, channel assign-ments for each neighbor node are saved along witheach entry in the neighbor table. Therefore, there isvery little communication overhead and storageoverhead for this two protocols. Ek-CA-AODVintroduces a ChannelTaken message that isbroadcast by active nodes to nodes within a k-hopneighborhood. Thus, the communication complex-ity of Ek-CA-AODV is OðjV tj � jV kjÞ.

A significant body of prior work examines thebenefits of utilizing multiple channels [3,5,6,19,25,26]. These existing multi-channel MAC proto-cols differ in their (i) channel selection techniques,(ii) medium access control schemes, and (iii) hard-ware requirements [27]. Table 3 summarizes severalimportant features of several existing multi-channelMAC protocols. There is no general rule as to whichscheme is better than another. However, a funda-mental tradeoff exists between the hardware com-plexity and the system performance. Simplerschemes with less hardware requirements are oftencompatible with the IEEE 802.11 standard and areeasy to implement, whereas complex schemes withgreater hardware requirements often yield betterperformance. Because we combine channel assign-ment with routing protocols, the MC-MAC proto-col only needs to implement the medium accesscontrol function, which greatly simplifies the proto-col design.

7. Conclusions

We examined distributed, on-demand routingand channel assignment in multi-channel mobilead hoc networks. Three principles have been pre-sented for designing effective channel assignmentprotocols. Based on these design principles, weintroduced CA-AODV, and its two extensions,E2-CA-AODV and Ek-CA-AODV, that combinechannel assignment with on-demand routing.These protocols exhibit lower communication,computation, and storage complexity, and requirefewer channels than many existing channel assign-ment algorithms. We also introduced a companionMC-MAC protocol that works with the proposedchannel assignment protocols. Simulation resultsshow that the proposed schemes can offer animprovement up to a factor of three in throughputover the IEEE 802.11 MAC protocol. The pro-posed approach provides an effective solution tothe low throughput problem in multi-hop wirelessnetworks.

References

[1] M. Gong, S. Midkiff, S. Mao, Design principles fordistributed channel assignment in wireless ad hoc networks,in: Proceedings of the 2005 IEEE International Conferenceon Communications (ICC), Seoul, Korea, 2005, pp. 3401–3406.

[2] J. Li, C. Blake, D.D. Coute, H. Lee, R. Morris, Capacity ofad hoc wireless networks, in: Proceedings of the IEEE/ACMInternational Conference on Mobile Computing and Net-working (MobiCom), Rome, Italy, 2001, pp. 61–69.

[3] A. Raniwala, T. Chiueh, Architecture and algorithms for anIEEE 802.11-based multi-channel wireless mesh network, in:Proceedings of the IEEE INFOCOM’05, Miami, FL, 2005,pp. 2223–2234.

[4] K. Leung, B.-J. Kim, Frequency assignment for IEEE 802.11wireless networks, in: Proceedings of the IEEE VTC’03,Orlando, FL, 2003, pp. 1422–1426.

[5] S.-L. Wu, C.-Y. Lin, Y.-C. Tseng, J.-P. Sheu, A new multi-channel MAC protocol with on-demand channel assignmentfor multihop mobile ad hoc networks, in: Proc. InternationalSymposium on Parallel Architectures, Algorithms and Net-works (ISPAN), Dallas/Richardson, TX, 2000, pp. 232–237.

[6] N. Jain, S.R. Das, A. Nasipuri, A multichannel CSMAMAC protocol with receiver-based channel selection formultihop wireless networks, in: Proceedings of the IEEEInternational Conference on Computer Communicationsand Networks (IC3N), Dallas, TX, 2001, pp. 432–439.

[7] E. Sousa, J. Silvester, Spreading code protocols for distrib-uted spread-spectrum packet radio networks, IEEE Trans-actions on Communications 36 (3) (1988) 272–281.

[8] M. Joa-Ng, I.-T. Lu, Spread spectrum medium accessprotocol with collision avoidance in mobile ad hoc wirelessnetwork, in: Proceedings of the IEEE INFOCOM’99, NewYork, NY, 1999, pp. 21–25.

[9] L. Hu, Distributed code assignments for CDMA packetradio networks, IEEE/ACM Transactions on Networking 1(6) (1993) 668–677.

[10] A. Bertossi, M. Bonuccelli, Code assignments for hiddenterminal interference avoidance in multihop packet radionetworks, IEEE/ACM Transactions on Networking 3 (4)(1995) 441–449.

[11] J. Garcia-Luna-Aceves, J. Raju, Distributed assignment ofcodes for multihop packet-radio networks, in: Proceedings ofthe IEEE MILCOM’97, Monterey, CA, 1997, pp. 450–454.

[12] T. Rappaport, Wireless Communications: Principles andPractice, second ed., Prentice Hall, Upper Saddle River, NJ,2002.

[13] J. Tang, G. Xue, W. Zhang, Interference-aware topologycontrol and QoS routing in multi-channel wireless meshnetworks, in: Proceedings of the ACM MobiHoc’05,Urbana-Champaign, IL, 2005, pp. 68–77.

[14] C. Perkins, E. Belding-Royer, S. Das, Ad hoc On-DemandDistance Vector (AODV) routing, IETF RFC 3661, July2003.

[15] T. Clausen, P. Jacquet, Optimized Link State RoutingProtocol, IETF RFC 3626, October 2003.

[16] X. Yang, N. Vaidya, A wireless MAC protocol using implicitpipelining, IEEE Transactions on Mobile Computing 5 (3)(2006) 258–273.

[17] VINT Group, UCB/LBNL/VINT network simulator ns.Available online: <http://www.isi.edu/nsnam/ns/>.

78 M.X. Gong et al. / Ad Hoc Networks 7 (2009) 63–78

[18] The Rice University Monarch Project, Wireless and mobilityextension to ns. Available online: <http://www.mon-arch.cs.rice.edu/cmu-ns.html>.

[19] P. Bahl, R. Chandra, J. Dunagan, SSCH: Slotted seededchannel hopping for capacity improvement in IEEE 802.11ad hoc wireless networks, in: Proceedings of the ACMMobiCom’04, Philadelphia, PA, 2004, pp. 216–230.

[20] J. Yoon, M. Liu, B. Noble, Random waypoint consideredharmful, in: Proceedings of the IEEE INFOCOM’03, SanFrancisco, CA, 2003, pp. 1312–1321.

[21] P. Kyasanur, N. Vaidya, Routing and interface assignmentin multi-channel multi-interface wireless networks, in: Pro-ceedings of the IEEE WCNC’05, New Orleans, LA, 2005,pp. 2051–2056.

[22] P. Kyasanur, N. Vaidya, Capacity of multi-channel wirelessnetworks: impact of number of channels and interfaces,Technical Report, University of Illinois at Urbana-Cham-paign, 2005.

[23] I. Katzela, M. Naghsineh, Channel assignment schemes forcellular mobile telecommunication systems: a comprehensivesurvey, IEEE Personal Communications 3 (3) (1996) 10–31.

[24] J. So, N. Vaidya, Routing and channel assignment in multi-channel multi-hop wireless networks with single-NICdevices, Technical Report, University of Illinois at Urbana-Champaign, 2005.

[25] J. So, N. Vaidya, Multi-channel MAC for ad hoc networks:Handling multi-channel hidden terminals using a singletransceiver, in: Proceedings of the ACM InternationalSymposium on Mobile Ad Hoc Networking and Computing(MobiHoc), 2004, pp. 222–233.

[26] A. Tzamaloukas, J. Garcia-Luna-Aceves, A receiver-initi-ated collision-avoidance protocol for multi-channel net-works, in: Proceedings of the IEEE INFOCOM’01,Anchorage, AK, 2001, pp. 189–198.

[27] M. Gong, S. Mao, S. Midkiff, B. Hart, Medium accesscontrol in wireless mesh networks, in: Y. Zhang, J. Luo, H.Hu (Eds.), Wireless Mesh Networking: Architectures, Pro-tocols and Standards, Auerbach Publications, New York,NY, 2006, pp. 147–182 (chapter 5).

Michelle Xiaohong Gong received herPh.D. from Virginia Tech in 2005, herM.S. from University of Hawaii in 2000,and her B.S. from Wuhan University in1996, all in Electrical Engineering. From2005 to 2007, she worked as a systemarchitect in the CTO group of theWireless Networking Business Unit,Cisco Systems. Currently, she is aresearch scientist in Networking Tech-nology Lab, Intel Corporation. She is

actively involved as a voting member in the IEEE 802.11 stan-dards body, and in particular the mesh networking Task Group.

She has authored eight pending US patents, two book chapters,and numerous papers in the areas of wireless communications

and wireless networks. Her research interests include perfor-mance analysis and algorithm design for wireless networks, witha focus on medium access control and wireless routing protocolsin wireless mesh networks.

Scott F. Midkiff received the B.S.E. andPh.D. degrees from Duke University,Durham, NC, and the M.S. degree fromStanford University, Stanford, CA, all inElectrical Engineering.

He worked at Bell Laboratories andheld a visiting position at CarnegieMellon University, Pittsburgh, PA. In1986, he joined the Bradley Departmentof Electrical and Computer Engineering,Virginia Polytechnic Institute and State

University, Blacksburg, where he is now a Professor. He is nowwith the National Science Foundation (NSF) as a Program

Director for the Integrative, Hybrid and Complex Systems(IHCS) Program of the Electrical, Communications and CyberSystems (ECCS) Division in the Directorate for Engineering(ENG).

His research interests include system issues in wireless and adhoc networks, network services for pervasive computing, andperformance modeling of mobile ad hoc networks.

Shiwen Mao received the B.S. and theM.S. degree from Tsinghua University,Beijing, P.R. China in 1994 and 1997,respectively, both in Electrical Engi-neering. He received the M.S. degree inSystem Engineering and the Ph.D.degree in Electrical and ComputerEngineering from Polytechnic Univer-sity, Brooklyn, NY, in 2000 and 2004,respectively.

He was a Research Member at theIBM China Research Lab, Beijing from 1997 to 1998. In thesummer of 2001, he was a research intern at Avaya Labs-

Research, Holmdel, NJ. He has been a Research Scientist in theDepartment of Electrical and Computer Engineering, VirginiaTech, Blacksburg, VA from December 2003 to April 2006.Currently, he is an Assistant Professor in the Department ofElectrical and Computer Engineering at Auburn University,Auburn, AL.

His research interests include cross-layer design and optimi-zation in multi-hop wireless networks, cognitive networks, andmultimedia communications. He is on the Editorial Board of theHindawi Advances in Multimedia Journal and the Wiley Inter-national Journal of Communication Systems. He is a co-recipientof the 2004 IEEE Communications Society Leonard G. AbrahamPrize in the Field of Communications Systems. He is the co-author of a textbook, TCP/IP Essentials: A Lab-Based Approach(Cambridge University Press, 2004).