channel-aware routing in manets with route handoff

Channel-Aware Routing in MANETswith Route Handoff

Xiaoqin Chen, Haley M. Jones, and Dhammika Jayalath, Senior Member, IEEE

Abstract—In wireless mobile ad hoc networks (MANETs), packet transmission is impaired by radio link fluctuations. This paper

proposes a novel channel adaptive routing protocol which extends the Ad hoc On-Demand Multipath Distance Vector (AOMDV) routing

protocol to accommodate channel fading. Specifically, the proposed Channel-Aware AOMDV (CA-AOMDV) uses the channel average

nonfading duration as a routing metric to select stable links for path discovery, and applies a preemptive handoff strategy to maintain

reliable connections by exploiting channel state information. Using the same information, paths can be reused when they become

available again, rather than being discarded. We provide new theoretical results for the downtime and lifetime of a live-die-live multiple

path system, as well as detailed theoretical expressions for common network performance measures, providing useful insights into the

differences in performance between CA-AOMDV and AOMDV. Simulation and theoretical results show that CA-AOMDV has greatly

improved network performance over AOMDV.

Index Terms—Mobile ad hoc networks, routing protocols, channel adaptive routing.

Ç

1 INTRODUCTION

WIRELESS mobile ad hoc networks (MANETs) are self-configuring, dynamic networks in which nodes are

free to move. A major performance constraint comes frompath loss and multipath fading [1]. Many MANET routingprotocols exploit multihop paths to route packets. Theprobability of successful packet transmission on a path isdependent on the reliability of the wireless channel on eachhop. Rapid node movements also affect link stability,introducing a large Doppler spread, resulting in rapidchannel variations [2].

Routing protocols can make use of prediction of channelstate information (CSI) based on a priori knowledge ofchannel characteristics, to monitor instantaneous linkconditions. With knowledge of channel behavior, the bestlinks can be chosen to build a new path, or switch from afailing connection to one with more favorable channelconditions. Several channel adaptive schemes that havebeen developed for MANETs to maintain connectionstability can be found in the literature. In [3], [4] channeladaptive schemes are implemented in medium accesscontrol (MAC) protocols; [5] considers link stability largelyin terms of longevity of a given link, termed “associativity”;a similar idea, with respect to node mobility, is consideredin [6] while [7] considers node mobility to improve path

reliability, utilizing only the naive transmission rangechannel model, not taking into account the fading char-acteristics of the wireless channel; [8] utilizes node-to-noderouting, based on the “best” node which received a giventransmission. While throughput improvements of 35 percentover traditional routing techniques are achieved, it is notclear how much delay or overhead is expended throughnode negotiation with each transmission. Signal strength asa path selection criterion, is used in [9]; [10] introducesoutage probability into both the routing and MAC proto-cols; [11], [12], [13] utilize the bit transmission rate in thenetwork layer; and [14] employs SNR to support channeladaptive routing.

In this paper, we introduce an enhanced, channel-awareversion of the AOMDV routing protocol. The key aspect ofthis enhancement, which is not addressed in other work, isthat we use specific, timely, channel quality informationallowing us to work with the ebb-and-flow of path avail-ability. This approach allows reuse of paths which becomeunavailable for a time, rather than simply regarding them asuseless, upon failure, and discarding them. We utilize thechannel average nonfading duration (ANFD) as a measure oflink stability, combined with the traditional hop-countmeasure for path selection. The protocol then uses the sameinformation to predict signal fading and incorporates pathhandover to avoid unnecessary overhead from a new pathdiscovery process. The average fading duration (AFD) isutilized to determine when to bring a path back into play,allowing for the varying nature of path usability instead ofdiscarding at initial failure. This protocol provides a dual-attack for avoiding unnecessary route discoveries, predictingpath failure leading to handoff and then bringing paths backinto play when they are again available, rather than simplydiscarding them at the first sign of a fade. Further, the sameinformation is required to determine ANFD, AFD andpredict path failure, enhancing efficiency. The overall effectis a protocol with improved routing decisions leading to amore robust network. Improvements in performance over

108 IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 10, NO. 1, JANUARY 2011

. X. Chen is with Star Intellect Pty Ltd, 6/22 Harrison Street, Mitcham,Victoria 3132, Australia. E-mail: [email protected].

. H.M. Jones is with the School of Engineering, College of Engineering andComputer Science, The Australian National University, Old EngineeringBuilding, Bldg 32, Canberra, ACT 0200, Australia.E-mail: [email protected].

. D. Jayalath is with the School of Engineering Systems, QueenslandUniversity of Technology, Room 836, Level 8, S Block, QUT GardensPoint, 2 George St Brisbane, GPO Box 2434 Brisbane, Queensland,Australia 4001. E-mail: [email protected].

Manuscript received 25 Nov. 2008; revised 25 Aug. 2009; accepted 21 Dec.2009; published online 23 July 2010.For information on obtaining reprints of this article, please send e-mail to:[email protected], and reference IEEECS Log Number TMC-2008-11-0475.Digital Object Identifier no. 10.1109/TMC.2010.144.

1536-1233/11/$26.00 � 2011 IEEE Published by the IEEE CS, CASS, ComSoc, IES, & SPS

AOMDV are around 25 percent for standard networkperformance measures. We call this protocol Channel-AwareAOMDV (CA-AOMDV). Note that this protocol is intendedto improve on AOMDV in conditions where the channel canbe reasonably allowed for. In conditions of high channelvariability, there is little sense in even attempting channelprediction and other performance improvement methodol-ogies will need to be utilized.

Further, we provide a detailed theoretical analysis ofthe lifetimes of both protocols and expressions forperformance with respect to routing control overhead,throughput, and packet delivery ratio. We derive a newexpression for the expected lifetime of a live-die-live-againsystem of multiple fading paths. These expressions allowus to show the superiority of CA-AOMDV over AOMDV,verified by simulations. Theoretical and simulation resultsare well matched.

In Section 2, we review AODV and AOMDV. Statisticalproperties of the mobile-to-mobile channel model arediscussed in Section 3. The route discovery and routemaintenance processes in CA-AOMDV are presented inSection 4. A detailed theoretical analysis is presented inSection 5 while simulation results are discussed in Section 6and conclusions in Section 7.

2 REVIEW OF AODV AND AOMDV

Transmissions via unreliable wireless connections canresult in large packet losses. Thus, it makes sense toconsider routing protocols which adapt to channel varia-tions. We propose a channel-aware routing protocol whichextends the Ad hoc On-Demand Multipath Distance Vector(AOMDV) routing protocol [15]. We call it CA-AOMDV.AOMDV is, itself, an extension of the Ad hoc On-DemandDistance Vector (AODV) routing protocol [16]. In thissection, we review the details of these two predecessorprotocols that are useful to our discussion in this paper.

2.1 AODV

AODV is a single-path, on-demand routing protocol. Whena source node, ns, generates a packet for a particulardestination node, nd, it broadcasts a route request (RREQ)packet. The RREQ contains the following fields:

<source IP address,

source sequence number,

broadcast ID,

destination IP address,

destination sequence number,

hop-count>,

where the source and destination IP addresses remain constantfor the lifetime of the network, source sequence number is amonotonically increasing indicator of packet “freshness,”destination sequence number is the last known sequencenumber for nd at ns and hop-count is initialized to zero andincremented at each intermediate node which processes theRREQ. A RREQ is uniquely identified by the combination ofsource sequence number and broadcast ID. An intermediatenode only processes a RREQ if it has not received a previouscopy of it. If an intermediate node has a route to nd withdestination sequence number at least that in the RREQ, it returns

a route reply (RREP) packet, updated with the informationthat it has. If not, it records the following information: sourceIP address, source sequence number, broadcast ID, destination IPaddress and expiration time for reverse path route entry, andforwards the RREQ to its neighbors.

Like the RREQ, a RREP is only processed on firstsighting and is discarded unless it has a greater destinationsequence number than the previous RREP or the samedestination sequence number but a smaller hop-count. TheRREP packet contains the following fields:

<source IP address,

destination IP address,


hop-count,

route expiration time>.

The route expiration time is the time after which the route isconsidered to have expired and a new route discoveryprocess must be undertaken. ns sends packets via the firstpath it hears about. If it receives a later RREP which haseither fresher information or a shorter hop-count, it swaps tothat, discarding the original route information.

When an active route link breaks, a route error (RERR)packet, with sequence number incremented from thecorresponding RREP and hop-count of 1, is sent by theupstream node of the broken link to ns. Upon receipt of aRERR, ns initiates a new route discovery process if it stillhas packets to send to nd. Nodes also periodically send“hello” messages to neighboring nodes to maintain knowl-edge of local connectivity.

2.2 AOMDV

The key distinguishing feature of AOMDV over AODV isthat it provides multiple paths to nd. These paths are loop-free and mutually link-disjoint. AOMDV uses the notion ofadvertized hop-count to maintain multiple paths with thesame destination sequence number. In both AODV andAOMDV, receipt of a RREQ initiates a node route tableentry in preparation for receipt of a returning RREP. InAODV, the routing table entry contains the fields:

<destination IP address,


next-hop IP address,

hop-count,

entry expiration time>,

where entry expiration time gives the time after which, if acorresponding RREP has not been received, the entry isdiscarded. In AOMDV, the routing table entry is slightlymodified to allow for maintenance of multiple entries andmultiple loop-free paths. First, advertized hop-count replaceshop-count and advertized hop-count is the maximum over allpaths from the current node to nd, so only one value isadvertized from that node for a given destination sequencenumber. Second, next-hop IP address is replaced by a list of allnext-hop nodes and corresponding hop-counts of the savedpaths to nd from that node, as follows:

<destination IP address,


advertized hop-count,

CHEN ET AL.: CHANNEL-AWARE ROUTING IN MANETS WITH ROUTE HANDOFF 109

route list: {(next hop IP 1, hop-count 1),

(next hop IP 2, hop-count 2), . . . },

entry expiration time>.

To obtain link-disjoint paths in AOMDV, nd can reply tomultiple copies of a given RREQ, as long as they arrive viadifferent neighbors.

3 MOBILE-TO-MOBILE CHANNEL MODEL

In MANETs, potentially all nodes are in motion, so it isappropriate to use the mobile-to-mobile channel model [17] tocharacterize the channel between any two nodes. It ispractically difficult to find the relative speeds between mobilenodes, so this channel model has the advantage of onlyusing the individual node speeds. It incorporates large-scalepath loss and small-scale flat fading. For transmission overa distance, d, in the presence of flat fading, the receivedsignal power is exponentially distributed with mean G0d

��

[18], where G0 is proportional to the transmitted signalpower and � is the propagation loss coefficient, typicallybetween two and four.

3.1 Average Nonfading Duration

The average nonfading duration is affected by both thephysical propagation environment (e.g., obstacles such astrees and buildings) and the node velocities. A typicalfading waveform is shown in Fig. 1. The ANFD, �#, is theaverage length of time that the signal envelope spendsabove a network-specific threshold, Rth, and is given by

�# ¼ 1

�fTffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi2�ð1þ �2Þ

p ¼ cffiffiffiffiffiffiG0

p

Rthd�=2f0

ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi2��v2T þ v2

R

�q ; ð1Þ

where � ¼ Rth=Rrms; ðRrms ¼ffiffiffiffiffiffiffiffiffiffiffiffiffiffiG0d��p

Þ is the ratio betweenthe transmission threshold and the root-mean-square powerof the received signal, fT ¼ f0vT=c is the maximum Dopplershift of the transmitter, f0 is the transmitter signal carrierfrequency, c � 3� 108 ms�1 is the speed of electromagneticradiation (signal speed), and � ¼ vR=vT is the ratio of thereceiver velocity to that of the transmitter where vR and vTare the receiver and transmitter node velocities, respectively.

It can be surmised from (1) that the value of the ANFDis high for low-transmission threshold (low �), anddecreases with an increase of � or �. Further, increasednode mobility (captured by vR and vT ) would cause acorresponding decrease in the ANFD due to the increasedrate of signal fluctuations and that an increased linkdistance (via d) would cause a decrease in ANFD due to agreater path-loss influence.

In MANETs, choice of stable links for route establish-ment ensures reliable packet transmission. Link stability canbe represented by the distance between the nodes formingthe link, and their mobilities. Thus, any measure of howstable a link is should include these factors. The ANFD isinversely proportional to link length, d, and node velocitiesvT and vR. The ANFD of a link between two highly mobileor separated nodes will be shorter than that of a linkbetween two slow moving and/or close nodes. In short, alink with a high ANFD will have a relatively long lifetime.Thus, using the ANFD as a metric will result in choosingmore stable links. There is minimal extra calculationrequired to determine ANFD. The parameter Rrms can begarnered from received packet signal strengths, and fT canbe calculated via fT ¼ f0vT=c. Thus, to calculate �#, nodessimply need to include speed and location in the header ofeach packet.

In this paper, we assume that all nodes know theirpositions and velocities. For example, each node isequipped with a Global Positioning System (GPS) receiver.If it is not feasible to equip every node with a GPS receiverdue to availability, node energy limitations, or obstructionof positioning infrastructure, nodes can determine theirlocation by running distributed localization algorithms [19],[20] which only require a limited number of nodes to haveposition capability. Using received signal strength [21],angle of arrival [22,] or time difference of arrival [19], [23],position techniques measure the distance or angle betweena node and some known reference points, to estimatelocation. Similarly, each node is assumed to be able tomonitor its velocity by any suitable mechanism [24], [25].

3.2 Average Fading Duration

The average fading duration, ��, is the average length oftime that the signal envelope spends below Rth. Transmis-sion is not considered possible while the signal envelope fora link is below the threshold. The AFD for the mobile-to-mobile channel is given by [17]

�� ¼ e�2 � 1

�fTffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi2�ð1þ �2Þ

p : ð2Þ

The AFD metric is used in CA-AOMDV to determine for howlong a faded link will be unavailable and is recorded in theroute cache. This will be discussed in detail in Section 4.2.

3.3 Channel Prediction Using Time Correlation

A feature of CA-AOMDV is the use of channel prediction toinstigate handoff between paths, when a fade is predicted ona link on the active path, described in Section 4.2. We choosethe linear minimum mean square error (LMMSE) algorithm[26] for channel prediction. We assume a slow fading channelsuch that it is constant for a symbol duration.


Fig. 1. A typical fading waveform with threshold indicated.

Let M be the number of previously received values usedto predict, at discrete time interval nwith a discrete time stepof �t, the signal strength time intervals into the future.Then, if xðnþ Þ is the LMMSE prediction for the receivedsignal strength, xðnþ Þ, at discrete time nþ , we have

xðnþ Þ ¼ RTxxR

�1xxx; ð3Þ

where RTxx is the 1�M cross-correlation vector of x and x,

ð�ÞT denotes transpose, Rxx is the M �M autocorrelationmatrix of x ¼ ½xðn� 1Þ; xðn� 2Þ; . . . ; xðn�MÞ�T , and x isthe M � 1 vector of the M previous signal strength values.The autocorrelation matrix Rxx is given by

Rxx ¼

Rxxð1; 1Þ Rxxð1; 2Þ � � � Rxxð1;MÞRxxð2; 1Þ Rxxð2; 2Þ � � � Rxxð2;MÞ

..

. ...

� � � ...

RxxðM; 1Þ RxxðM; 2Þ � � � RxxðM;MÞ

26664

37775; ð4Þ

where element Rxxð‘;mÞ is given by [17].

Rxxð‘;mÞ ¼ Efxðn� ‘ÞxHðn�mÞg¼ �2

1J0ð2�fT ðm� ‘ÞÞJ0ð2�fRðm� ‘ÞÞ;ð5Þ

where ð�ÞH denotes Hermitian transpose, �1 ¼ Rrms=ffiffiffi2p

andJ0 is the 0th order Bessel function of the first kind. Finally,

RTxx ¼ ½Rxxð1Þ; Rxxð2Þ � � �RxxðMÞ�;

where

RxxðmÞ ¼ Efxðnþ ÞxHðnþ �mÞg¼ �2

1J0ð2�fTmÞJ0ð2�fRmÞ:ð6Þ

With the presence of Bessel functions, the LMMSEalgorithm is computationally intensive. To enhance amen-ability to MANET channel prediction, (3) can be reduced asfollows:

xðnþ Þ ¼�RTxxR

�1xx

�x ¼Wx

¼ ½wð1Þ � � �wðMÞ�½xðn� 1Þ � � �xðn�MÞ�T

¼XMi¼1

wðiÞxðn� iÞ;ð7Þ

where W ¼ ðRTxxR

�1xx Þ with elements wðiÞ, fi ¼ 1; . . . ;Mg.

The wðiÞ values can be calculated offline and stored in alookup table indexed by Doppler frequency shift and discretetime shift. The use of this low-complexity LMMSE predictionin CA-AOMDV handoff is discussed in Section 4.2.

4 CHANNEL-AWARE AOMDV PROTOCOL

As mentioned in Section 2, route discovery in AOMDVresults in selection of multiple loop-free, link-disjoint pathsbetween ns and nd, with alternative paths only utilized ifthe active path becomes unserviceable. One of the mainshortcomings of AOMDV is that the only characteristicconsidered when choosing a path is the number of hops.Path stability is completely ignored. Thus, selected pathstend to have a small number of long hops meaning thatnodes are already close to the maximum possible commu-nication distance apart, potentially resulting in frequent link

disconnections. Further, channel conditions are idealizedwith the path-loss/transmission range model, ignoring

fading characteristics inherent in all practical wireless

communication environments.In CA-AOMDV, we address this deficiency in two ways.

In the route discovery phase, we utilize the ANFD, defined

in Section 3.1, of each link as a measure of its stability. In theroute maintenance phase, instead of waiting for the active

path to fail, we preempt a failure by using channel

prediction on path links, allowing a handover to one ofthe remaining selected paths. This results in saved packets

and consequently smaller delays.

4.1 Route Discovery in CA-AOMDV

Route discovery in CA-AOMDV is an enhanced version ofroute discovery in AOMDV, incorporating channel proper-

ties for choosing more reliable paths. In Section 3.1, we

defined the ANFD for one link of a path, according to themobile-to-mobile channel model. CA-AOMDV uses the

ANFD as a measure of link lifetime. The duration, D, of apath is defined as the minimum ANFD over all of its links,

D ¼4 min1�h�H

ANFDh; ð8Þ

where h is link number, and H is number of links/hops inthe path. Before forwarding a RREQ to its neighbors, a node

inserts its current speed into the RREQ header so that itsneighbors can calculate the link ANFD using (1). The path

duration, D, is also recorded in the RREQ, updated, as

necessary, at each intermediate node. Thus, all informationrequired for calculating the ANFD is available via the

RREQs, minimizing added complexity.Similarly, to the way the longest hop path is advertized

for each node in AOMDV to allow for the worst case at each

node, in CA-AOMDV the minimum D over all paths

between a given node, ni, and nd, is used as part of the costfunction in path selection. That is,

Di;dmin ¼4

min�2path listdi

D�; ð9Þ

where path listdi is the list of all saved paths between

nodes ni and nd. The route discovery update algorithm in

CA-AOMDV is a slight modification of that of AOMDV. If aRREQ or RREP for nd at ni, from a neighbor node, nj, has a

higher destination sequence number or shorter hop-count than

the existing route for nd at ni, the route update criterion inCA-AOMDV is the same as that in AOMDV. However, if

the RREQ or RREP has a destination sequence number andhop-count equal to the existing route at ni but with a greater

Di;dmin, the list of paths to nd in ni’s routing table is updated.So, in CA-AOMDV, path selection is based on Di;dmin as

well as destination sequence number and advertized hop-count.The routing table structures for each path entry in AOMDV

and CA-AOMDV are shown in Table 1. The handoff dormant

time field in the routing table for CA-AOMDV is the amount

of time for which the path should be made dormant due to

channel fading. It is set to the maximum value of the AFDsover all links in the path. This use of handoff dormant time is

described in more detail in the next section.


4.2 Route Maintenance in CA-AOMDV

In mobile environments, it is necessary to find efficientways of addressing path failure. Using prediction andhandoff to preempt fading on a link on the active path,disconnections can be minimized, reducing transmissionlatency and packet drop rate [27], [28].

Route maintenance in CA-AOMDV takes advantage of ahandoff strategy using signal strength prediction, detailedin Section 3.3, to counter channel fading. When thepredicted link signal strength level falls below a network-specific threshold, the algorithm swaps to a good-qualitylink. The fading threshold is chosen so as to providerobustness to prediction errors. The presence of multipleusers experiencing independent channel fading means thatMANETs can take advantage of channel diversity, unlikedata rate adaptation mechanisms such as SampleRate [29].

All nodes maintain a table of past signal strengths,recording for each received packet, previous hop, signalpower, and arrival time. Ideally, there will be M packetswhere M is the required number of past samples from (3).However, this will depend on the packet receipt timescompared with the specified discrete time interval, �t. Ifpackets are received at time intervals greater than �t,sample signal strengths for the missed time intervals can beapproximated by the signal strength of the packet closest intime to the one missed. If packets are received at intervals ofshorter duration than �t, some may be skipped. Anexample of handoff in CA-AOMDV is shown in Fig. 2.The handoff process is implemented via a handoff request(HREQ) packet. The CA-AOMDV handoff scheme isdescribed below.

4.2.1 Prediction Length

The LMMSE prediction algorithm performs quite poorly ifnot matched to the current channel conditions. Therefore,the prediction length should not be too long. In CA-AOMDV, a given node may have multiple paths to thedestination, each with a different next hop node. If anintermediate node has multiple paths to the destination,upon receiving an HREQ it can immediately switch fromthe active path to a good alternative one, without furtherpropagating the HREQ. Therefore, the time needed toimplement a handoff in CA-AOMDV is the duration, interms of the discrete time interval �t, for the HREQ to bepropagated to the fading link uplink node. For example, ifni and nj are neighbors in a given path and nj predicts afade on link ‘i;j, it will generate a HREQ and forward it to

ni. Thus, a suitable prediction length in (3) corresponds tothe number of discrete time intervals, �t, for transmissionof a HREQ between nj and ni, which can be approximatedby using the data propagation time T ij from nj to ni, with

¼ round�Ti

j=�t�; ð10Þ

where “round” is the integer rounding function. In additionto choosing a threshold with a suitable error margin, asdescribed above, to enhance the robustness of the predic-tion process to errors in CA-AOMDV, the signal strength ispredicted at t0 þ and t0 þ 2 . The algorithm is detailed inthe next section.

4.2.2 Handoff Trigger

Route handoff is triggered when a link downstream nodepredicts a fade and transmits a HREQ to the uplink node.Let TR be the transmission range, assumed to be the samefor all nodes, let RðtÞ be predicted signal strength at time tand recall Rth as the fade prediction threshold. If theprediction at t0 þ is above Rth while that at t0 þ 2 isbelow, the maximum transmitter velocity vmax

T ensuringsignal strength above Rth at t0 þ , is determined using (3).If a fade is predicted at either time, the receiver checkswhether the link is at breaking point with respect todistance. The HREQ registers the following fields: source IPaddress, destination IP address, source sequence number, fadeinterval index, long term fading indicator, AFD, and vmax

T .

4.2.3 Handoff Table to Avoid Duplicate HREQs

In addition to the routing table described in Table 1, eachnode maintains a local handoff table. Each entry includes:source IP address, source sequence number, destination IPaddress, and expiration timeout. expiration timeout indicateswhen a path is expected to be available again (out of thefade) and is set to the maximum AFD of all currently fadedlinks with paths through that node to a particular ns. Notethat this is similar to the way advertized hop-count is set to themaximum number of hops for any path going through anode for a particular ns in AOMDV. Whenever a nodereceives a HREQ targeting a particular ns, it checks itshandoff table for an entry relating to that ns. The handoff


TABLE 1Comparison of Routing Table Entry

Structures in AOMDV and CA-AOMDV

Fig. 2. Handoff in CA-AOMDV. Node F has predicted a forthcomingfade for its link with node D and has generated a HREQ. Having noalternative paths to choose from, node D forwards the HREQ tonode C which may then be able to handoff to the path with node E asthe next node.

table is updated if no entry exists for that ns, if the newHREQ has a longer AFD or if the existing entry is stale dueto the expiration timeout having expired. If any unexpiredentry is found for that ns with the same or higher sourcesequence number, the HREQ is dropped.

4.2.4 Forwarding the HREQ

Any node receiving a nonduplicate HREQ checks foralternative paths to nd. If not, as for the case of node D inFig. 2, it propagates the HREQ. Otherwise, if it has one ormore “good” alternative paths to the nd, it marks the fadingpath indicated in the HREQ as dormant, setting the handoffdormant time in its routing table entry for that path to theAFD recorded in the HREQ. The HREQ is then dropped. Ifa fade is predicted on the active path, a nondormantalternative path to nd is then adopted prior to the onset oflink failure. For example, if node C in Fig. 2 receives aHREQ from node D, it marks the path with next hop ¼ D asdormant, and adopts the path with next hop ¼ E. Thedormant path is retained for use when the fade is over,reducing path discovery overhead.

4.3 Section Summary

. Route discovery: ANFD is combined with the hop-count criterion from AOMDV to serve as a metricwith which to select short but stable paths instead ofsimply choosing the shortest path. So, CA-AOMDVtakes into account stability and length to improveoverall path quality.

. Route maintenance: Assuming independently time-varying paths, CA-AOMDV uses predicted signalstrength to trigger a handoff before a fade occurs,reducing the ns-nd connection failure rate. Thebreaking link AFD is recorded, so that it maybereutilized once out of the fade.

5 THEORETICAL ANALYSIS

A framework is now presented to analyze the performancesof AOMDV and CA-AOMDV. The probability densityfunctions (PDFs) of the lifetimes of a single path and multiplepaths are derived and the performances in terms of routingcontrol overhead and network throughput are analyzed.

5.1 Network Model

Assume an N-node network uniformly distributed over anarea of side length 2S. The node density is � ¼ N=ð2SÞ2per m2. The number of connections in the network is C.Assuming all nodes in the network have equal transmis-sion range, TR, the average number of neighbors isn ¼ ��T 2

R � 1. The expected distance between a ns-nd pairfor each connection is S

3 ðffiffiffi2pþ lnð1þ

ffiffiffi2pÞÞ [30], so we

can approximate the expected number of hops H perconnection by

H ¼ Sðffiffiffi2pþ lnð1þ

ffiffiffi2pÞÞ

3TR: ð11Þ

Assume each path link has equal likelihood of being brokenat anytime. On average, the number of hops into aconnection (path) before encountering a broken link is

ðH þ 1Þ=2. The average number of active links at any giventime is CH, and the total number of network links isapproximately nN=2, so the average number of connectionsover a given link is

B ¼ 2CHðnþ 1ÞN : ð12Þ

5.2 Single Path Lifetime Statistics

Link lifetime is the length of time for which a link is “active.”It is affected by the movements of the pair of linked nodesand the channel fading characteristics. If only free-spacepropagation is considered, link lifetime is simply theamount of time for which the two nodes are within eachother’s transmission ranges. If channel fading is alsoconsidered the relationship between link lifetime anddistance is not so straightforward, as the fading componentis independent of distance.

We use the random variable (r.v.) Z‘ to represent thelifetime of a fading link, which is equal to the nonfadingduration of the link, from (1). It is most commonly modeledby an exponential distribution. We therefore assume thatthe lifetime for link ‘ has PDF, fZ‘ðtÞ ¼ ‘e�‘t.

As discussed in Section 2, the multiple cached paths at agiven ns for a given nd are link-disjoint in AOMDV. Thus,we assume that the lifetimes of the links of the multiplecached paths are independent. While this is not strictly thecase, here we are seeking general trends, and this approx-imation is, therefore, appropriate. For a path composed ofL fading links, path lifetime can be represented byZp ¼ min½Z1; . . . ; ZL�, where Z‘, f‘ ¼ 1; . . . ; Lg, representsthe exponentially distributed r.v. for the lifetime of link ‘with parameter ‘. It can be shown that the CDF FZpðtÞ of Zpis given by [31]

FZpðtÞ ¼ 1� e�ðt�L‘¼1‘Þ; ð13Þ

and the PDF,

fZpðtÞ ¼ e�tPL

k¼1kXL‘¼1

‘ ¼ pe�pt; ð14Þ

where p ¼PL

‘¼1 ‘. If all L links are independent andidentically distributed (i.i.d.) with parameter , then (14)becomes

fZpðtÞ ¼ Le�Lt: ð15Þ

Using (14), the expected path lifetime is, then

EfZpg ¼Z 1

0

tpe�ptdt ¼ 1

p¼ 1PL

‘¼1 ‘; ð16Þ

and 1=p ¼ 1=L for L i.i.d. links in a path. So, a path with Li.i.d. links has a lifetime L times less than any individuallink, on average.

5.3 Multiple Path Lifetime in AOMDV

To analyze the active lifetimes of the multiple path systemsin AOMDV and CA-AOMDV, we assume Np paths areestablished during route discovery. In order to enablemeaningful comparisons between the two schemes, we alsoassume that each path has exactly L links and that the


individual link lifetimes are exponentially i.i.d. with

parameter for Z‘; f‘ ¼ 1; . . . ; Lg.Assume that at time t, path p1 is the active path in an

AOMDV network. When p1 fails, ns attempts to adopt an

alternative path, and p1 is discarded from the cache.

Communication between ns and nd is broken when all Np

paths have failed. Let ZA be the r.v. representing the

lifetime of the multiple path system in AOMDV and let Zpibe the r.v. representing the lifetime of path pi with

parameter p ¼ L. Using (13) and (14), the probability

that all paths have failed at time t is the CDF of ZA, given by

PrfZA < tg ¼ PrfðZp1< tÞ \ ðZp2

< tÞ \ � � � \ ðZpNp< tÞg

¼YNp

i¼1

PrfZpi< tg ¼

YNp

i¼1

FZpiðtÞ

¼ ð1� e�ptÞNp :

ð17Þ

Taking the derivative of (17), gives the PDF of ZA

fZAðtÞ ¼ NppXNp�1

k¼0

ð�1ÞNp�k�1 Np � 1k

� �e�ðNp�kÞpt; ð18Þ

and the expected value of ZA is given by

EfZAg ¼Np

p

XNp�1

k¼0

ð�1ÞNp�k�1 Np � 1k

� �1

ðNp � kÞ2: ð19Þ

5.4 Multiple Path Downtime in CA-AOMDV

Handoff is adopted in CA-AOMDV to enable swapping

between alternative paths when a fade is predicted on the

active path. At time t, let path p1 be the active path. When p1

fails, it is marked as dormant for the predicted period of its

fade and p2 is made active, and so on. The difference from

AOMDV is that any path recovering from a fade can again

be considered for active use. So, with CA-AOMDV, a new

route discovery process is only required when all Np paths

are simultaneously in a fade. So, we need to know how often

this happens, or how long the multiple path system

“downtime” is for multiple reselectable paths.Now, because, in CA-AOMDV the system is active if any

path is active at anytime, but “disconnected” only if all

paths are “down” (faded) simultaneously, we start with

expressions for the system downtime and then determine

the system lifetime in terms of the system downtime. We

start at the link-level. Similarly to the link lifetime

(nonfading duration), the link downtime (fading duration)

has an exponential distribution. Remembering that all links

are i.i.d., let the r.v. Y‘ represent the downtime of link ‘,

with parameter , with CDF FY‘ðtÞ ¼ 1� e�t. Also let the

r.v. Yp represent the downtime of path p, with parameter p,

with CDF, FYpðtÞ. Finally, let the r.v. YC represent the

multiple path system downtime for CA-AOMDV, with

CDF, FYC ðtÞ. The CA-AOMDV system is down only when

all paths are down, which occurs when at least one link in

each path is down. So, the probability that the CA-AOMDV

system is down at time t is

PrfYC > tg ¼YNp

i¼1

PrfYpi> tg ¼

YNp

i¼1

1�YL‘¼1

PrfY‘ < tg !

¼YNp

i¼1

1�YL‘¼1

FY‘ðtÞ !

¼ ½1� ð1� e�tÞL�Np :

ð20Þ

Then, the CDF of YC is given by

FYC ðtÞ ¼ 1� ½1� ð1� e�tÞL�Np : ð21Þ

Then, taking the derivative of (21), the PDF of the CA-

AOMDV multiple path system downtime, YC , with i.i.d.

links is given by (22) and the expected downtime of the CA-

AOMDV multiple path system is given by (23).

fYC ðtÞ ¼ NpLXNp�1

k¼0

ð�1ÞNp�k�1 Np � 1

k

� �

XLðNp�kÞ�1

i¼0

ð�1ÞLðNp�kÞ�1�i

LðNp � kÞ � 1

i

� �e�½LðNp�kÞ�i�t; ð22Þ

EfYCg ¼NpL

XNp�1

k¼0

ð�1ÞNp�k�1 Np � 1

k

� �

XLðNp�kÞ�1

i¼0

ð�1ÞLðNp�kÞ�1�i LðNp � kÞ � 1

i

� �1

ðL½Np � k� � iÞ2: ð23Þ

5.5 Multiple Path Lifetime in CA-AOMDV

In a fading channel, the AFD is given by the quotient of the

probability of a fade and the level crossing rate (LCR),

where the LCR is the rate in times per second that the signal

amplitude in the channel crosses a given threshold in the

positive going direction. Similarly, the ANFD is given by

the quotient of the probability of not being in a fade and the

LCR. Then, the ANFD in terms of the AFD, is

ANFD ¼ Prðnot in fadeÞLCR

¼ 1� Prðin fadeÞPrðin fadeÞ AFD: ð24Þ

For a multiple path system, the AFD is analogous to the

system downtime and the ANFD is analogous to the system

lifetime. So, for CA-AOMDV, we substitute (23) for the AFD

in (24). We need to determine an expression for the

probability that the CA-AOMDV system is “in a fade,” or

“down.” Now, the probability that the CA-AOMDV multi-

ple path system is down is the probability that all paths are

down. The probability that any particular path is down is

the probability that at least one link in the path is

undergoing a fade. In [32], it was shown that the signal

envelope of a mobile-to-mobile signal has a Rayleigh

distribution. Recall the Rayleigh parameter, �, from

Section 3.1. The probability that the CA-AOMDV system,

with Np paths each with L i.i.d. links, is in a fade is given by


Prðsystem in fadeÞ ¼YNp

i¼1

1�YL‘¼1

e��2

!¼ 1� e�L�2� �Np

:

ð25Þ

Then, letting ZC be the r.v. representing the CA-AOMDVsystem lifetime and combining (23), (24), and (25), theaverage system lifetime for CA-AOMDV is given by

EfZCg ¼1� ð1� e�L�2ÞNp

ð1� e�L�2ÞNpEfYCg: ð26Þ

This new theoretical result allows us to compare thetheoretical lifetimes of multiple path systems with andwithout reuse.

5.6 System Lifetime Comparison

In Figs. 3 and 4, we show the ratio of the multiple pathsystem lifetimes in CA-AOMDV to AOMDV with fourpaths and varying numbers of links per path, and varyingnumbers of paths with three links, as it varies with linkRayleigh fading threshold parameter �. The ratios arecalculated from dividing (26) by (19). Note that the morepaths there are, the better the performance of either systemand the more links there are, the worse the performance.The ratio of system lifetime for CA-AOMDV over AOMDVincreases exponentially as the fading threshold, �, de-creases. Recall that � is the ratio of the signal fadingthreshold to the link RMS signal strength. So, the higher thevalue of �, the higher the signal strength must be for the linkto be considered not in a fade. The ratio improves with adecrease in number of paths and a decrease in number oflinks per path. That is, while the actual lifetimes improvewith more paths and fewer links, the advantage of CA-AOMDV over AOMDV improves with fewer paths andfewer links. This makes sense because taking the channelinto account, as in CA-AOMDV, becomes more complexand less effective with more channels (links) to consider.

Fig. 5 shows the theoretical and simulated multiplepath system lifetimes for CA-AOMDV and AOMDV forNp ¼ 2 paths and L ¼ 2 links. The theoretical values arefrom (19) and (26). The simulation values were generated

in Matlab using a scenario with 20 scattering pointsoutside a network area of 1;000� 20 m2 with a transmis-sion wavelength of 0.3 m (�1 GHz). Adding the signalsfrom each scattering point at each point of interest in thenetwork results in a Rayleigh distributed wavefield. Theresolution of points at which signal strength values weretaken was 1 mm. At low-lifetime distances, the simulationresults are limited by this resolution. It can be seen that thesimulated and theoretical results are in excellent agree-ment, particularly, for lower values of �. CA-AOMDVoutperforms AOMDV up until � � 0:7, after which timeAOMDV performs better in theory, with no distinction inthe simulation results. Further, if we look at the actuallifetime values, by this point they are, at most, 8 cm which,even at 20 km/hr is only 14 ms. Such small lifetimes leaveboth schemes requiring further enhancements such assignal coding. Note these results also do not take intoaccount the fact that AOMDV chooses paths based only onhop-count, whereas CA-AOMDV includes link quality.


Fig. 3. Ratio of multiple path system lifetime for CA-AOMDV to AOMDVfor increasing values of link fading threshold parameter �.

Fig. 4. Ratio of multiple path system lifetime for CA-AOMDV to AOMDVfor increasing values of link fading threshold parameter �.

Fig. 5. Theoretical and simulated multiple path system lifetimes, inmeters, for Np ¼ 2 and L ¼ 2 for CA-AOMDV and AOMDV.

Taking these points into account would improve thesystem lifetime performance of CA-AOMDV even furtherover that of AOMDV.

Note that 1=‘ for the AOMDV results in Figs. 3, 4, and 5can be calculated for a given value of � by equating it withthe right-hand side of (1) and 1= for the CA-AOMDVresults can be calculated for a given value of � by equating itwith the right-hand side of (2). For these figures, � was setto 1, a reasonable assumption, as was fT . Changing fT hasno effect on Figs. 3 and 4 and has a simple scaling effect onthe results in Fig. 5.

5.7 CA-AOMDV/AOMDV Performance Analysis

We now determine expressions for routing control overheadand packet delivery ratio. If �d is the average delay for aroute discovery and EfZg is the average path system lifetime,from (19) or (26), then (EfZg þ �d) represents the averagetime between two successive route discoveries [31] per ns-ndpair. Let � be the number of route discoveries per second,

� ¼ 1

EfZg þ �d: ð27Þ

In CA-AOMDV and AOMDV, when ns needs a path to nd, itbroadcasts a RREQ. nd, or any intermediate node which hasa fresh enough path to nd, feeds a RREP back to ns toestablish the path. We can approximate �d as

�d ¼ HðtQ þ tP Þ; ð28Þ

where tQ is the one-hop propagation time of a RREQ,and tP is the one-hop propagation time of a RREP. WithC connections in the network, each lasting an average ofT seconds, the average number of route discoveryprocesses over time T is equal to Nrd ¼ CT�.

5.7.1 Routing Control Overhead (�)

In a multiple path system, the routing control overheadincludes replacing the failed path, �r, and route discoveryoverhead, �d, when all alternative paths are broken. First, weconsider the routing control overhead introduced by routediscovery. The RREQs are flooded into the N-node network,and Np RREPs are unicast from nd to ns, where Np is thenumber of multiple paths established during route discov-ery. So, the routing control overhead over a time T is given by

�d ¼ nRREQ þ nRREP ¼ CT�ðN þNpHÞ; ð29Þ

where nRREQ is the number of RREQs, and nRREP is thenumber of RREPs, generated during the route discovery.

Now we consider the control overhead for pathreplacement, �r. In each source-destination pair, there areNp alternative paths, so, there can be at most Np pathrepairs. In the event of a path disconnection, the upstreamnode of the broken link sends a RERR to ns. If the upstreamnode has no alternative paths to nd, and it is located closerto nd than ns, it broadcasts a RREQ for nd to salvage thedata packets. (If the broken link is closer to ns than nd, anyenroute packets are discarded.) Assuming that the linkfailure is equally likely to happen in any link of the path,the maximum number of path replacements is CT�ANp=2,where �A is the AOMDV route discovery frequency.

So, the routing control overhead due to path failureincludes the RERRs triggered by the failed link, the path

repair RREQs flooded from the fading link to nd, and theRREPs from nd to the RREQ generator. Because only linkbreaks in the latter half of a path are atoned for, RREQs areflooded into the network over an average of H=4 hops, withthe average number of flooded RREQs during each pathrepair approximately being Nrq ¼ �ðHR=4Þ2�. The RREPsare unicast from nd to the RREQ broadcasting node, whilethe RERRs are unicast from the fading link to ns. In eachpath failure, the average number of connected paths over abreaking link is B from (12). The AOMDV routing controloverhead due to path failure, �r

A, is

�rA ¼ nRERR þ nRREQ þ nRREP; ð30Þ

nRERR ¼ ½CTNp�ABðH þ 1Þ�=2; ð31ÞnRREQ ¼ ½CTNp�ABNrq�=2; ð32ÞnRREP ¼ ½CTNp�ABH�=4; ð33Þ

where nRERR is the number of RERRs due to path failure,and �A is AOMDV average route discovery frequency. Thetotal routing control overhead in AOMDV due to pathfailure in time T is, from (30), (31), (32), and (33)

�rA ¼ CTNp�AB

3H þ 2

4þNrq

2

!: ð34Þ

If the upstream node of the broken link has multiplepaths to nd, it can use an alternative path withoutpropagation of RREQs and RREPs. Under these circum-stances, (30) can be simplified to �r

A ¼ nRERR from (31).In CA-AOMDV, routing control overhead involves

HREQs, terminated at the nearest intermediate nodeshaving multiple paths to the destinations. The number ofhandoffs between two successive route discoveries isdetermined by the channel fading status. However, we canapproximate it by Np. Assuming the expected number ofhops to deliver a HREQ is I with I � H, then the CA-AOMDV routing control overhead due to path failure is

�rC ¼ nHREQ ¼ CTNp�CBI ; ð35Þ

where �C is the average route discovery frequency in CA-AOMDV. For the case where multiple paths exist in theupstream node of a broken link, the CA-AOMDV routingcontrol overhead is just one-hop HREQ propagation withI ¼ 1, that is, �r

C ¼ CTNp�CB. Then, the total routingcontrol overhead is � ¼ �d þ�r, with

�A ¼ CT�AðN þNpHÞ þ CTNp�AB3H þ 2

4þNrq

2

!; ð36Þ

�C ¼ CT�CðN þNpHÞ þ CTNp�CBI ; ð37Þ

where �A and �C are the route discovery frequencies forAOMDV and CA-AOMDV, respectively.

We showed in the previous section that the multiplepath system lifetime for CA-AOMDV was longer than thatfor AOMDV, for median to low thresholds with nopractical difference at higher thresholds. So the numberof route discoveries per second, �, is generally smaller forCA-AOMDV than AOMDV. If we also compare thedifferentiating terms in (36) and (37), it can be shownthat 3ðH þ 2Þ=4þNrq=2 > I , always. Thus, the routing


overhead for AOMDV is greater than that for CA-AOMDV as long as the signal threshold is not too high,at which point they are about equal.

5.7.2 Packet Delivery Ratio (�)

The packet delivery ratio is the ratio of the number ofsuccessfully received to the number of generated packets.The time, T , for a source-destination connection can beapproximately divided into two parts, the path connectiontime, Tcon, and the connection delay which is the time tocontend for channel access and to reconnect due to channelfading. In a network with constant bit rate sources andconstant data rate, we can model the data as being delivereduniformly over time, such that the packets delivered withinTcon are correctly received, and those delivered during thetime of connection delay are incorrectly received. Then, theratio between the path connection time, Tcon, and the totaloperation time, T , can be used to model packet delivery ratio.

The path connection delay, �, includes the delay, �r, dueto activation of an alternative path, and the delay, �d, from(28), due to route discovery when all alternative paths arebroken. In AOMDV, the delay due to alternative pathactivation is the time to propagate the RERRs back to ns. Lettr be the one-hop propagation time of a RERR. The delaydue to switching between alternative paths over a time T is,from (31)

�rA ¼ tr � nRERR ¼

TNp�AtrðH þ 1Þ2

: ð38Þ

In CA-AOMDV, the connection delay is the HREQpropagation time thI. The total connection delay is� ¼ �d þ �r. Thus, the path connection time is T � �. For802.11 DCF, there are also delays incurred for MACoverhead and retransmission. For each data transmission,the minimum channel occupation due to MAC overhead isTMAC ¼ TRTS þ TCTS þ 3TSIFS, where TRTS and TCTS are thetime consumed on RTS and CTS, respectively, and TSIFS isthe SIFS period. And, for each path failure, the delay due topacket retransmission is NRðtd þ TMACÞ, where NR is theMAC retransmission count, and td is the time for each datapacket transmission. Combining the time consumed on boththe MAC and routing layers, the delay introduced by pathfailure can be written as

�A ¼ �d�AT þTNp�Aðtr þ TMACÞðH þ 1Þ

2

þ TNp�ANRðtd þ TMACÞ þ T�ATMACH; ð39Þ�C ¼ �d�CT þ TNp�Cðth þ TMACÞI

þ T�CNRðtd þ TMACÞ þ T�CTMACH: ð40Þ

Comparison of terms in (39) and (40) shows that, wheneverroute discovery frequency �C < �A, terms 1, 3, and 4 areless for CA-AOMDV than for AOMDV. Term 3 is alsoadvantaged in CA-AOMDV because it does not have themultiplicative Np term. In term 2, however, the relativevalues depend on the actual value of I , which has beendefined as I < H, the relative values of tr and th and therelative values of �C and �A. Generally, AOMDV spendstime on repairing path failure, which is preempted byadopting a handoff scheme in CA-AOMDV. The absence of

handoff in AOMDV also increases the MAC delay due topacket retransmission in the event of path failure.

When IEEE 802.11 DCF is employed, we should alsoconsider the throughput degradation due to channel accesscontention. Based on a random network traffic model, themaximum number of nodes in the network which might beinvolved in packet relaying is

Na ¼ HC; HC < N ;N; HC N:

ð41Þ

Then, the probability that a node has packets to transmit isNa=N . In networks with IEEE 802.11 DCF, nodes withineach other’s transmission ranges cannot transmit at thesame time. Under heavy traffic conditions (every nodealways has packets to transmit), 802.11 DCF provides longterm per packet fairness in single-hop networks [33], thuseach node in the shared wireless channel has a probabilityof 1=ðnþ 1Þ of occupying the channel, where n is thenumber of node neighbors. Combining this with theprobability that a node has packets to transmit, the averagenode transmission probability is Na=ðnþ 1ÞN . For atransmitting node on an active path, the probability that itcan occupy the channel, or the probability that the channelwon’t be occupied by any of its n neighbors, is

q ¼ 1� Nan

ðnþ 1ÞN : ð42Þ

Because each node on a path suffers the same averagethroughput degradation due to channel access contention,the time available for data transmission in a path is ðT � �Þq.The achievable packet delivery ratio is � ¼ ðT � �Þq=T .For AOMDV and CA-AOMDV, respectively, the packetdelivery ratio in the network is

�A ¼ðT � �AÞq

T; �C ¼

ðT � �CÞqT

; ð43Þ

where �A and �C are given in (39) and (40), respectively.Therefore, whenever �A > �C the packet delivery ratio ofCA-AOMDV will be greater than that of AOMDV.

6 SIMULATION RESULTS

For simulations, we used network simulator ns-2.34 [34],implementing the mobile-to-mobile channel as in [35], [36],Doppler frequency controlled by transmitter and receivermovements, “hello” packet interval of 1,000 ms andphysical layer parameters of the Lucent WaveLAN wirelessnetwork card [34], with the random waypoint [37] mobilitymodel. Constant Bit Rate (CBR) sources are used with theIEEE 802.11 DCF MAC protocol.

6.1 Evaluation of ANFD in Routing

We evaluate the throughput performance of a single linkunder different ANFDs. The simulated network has an areaof 1,000 m� 1,000 m, channel bandwidth of 2 Mb/s and 50 srunning time. Nodes A and B are placed 70 m apart, movingin parallel directions. Node A moves with a fixed speed ofvA ¼ 1 m/s, while node B moves with speeds of 1, 1.5, 2, 2.5,and 3 m/s, respectively, with � ¼ 1; 1:5; 2; 2:5; 3 accordingly.By changing the speed of node B, we change the ratio of


their speeds, thus varying the ANFD value for the linkaccording to (1). The data packets are 512 bytes, transmittedat a rate of 40 packets per second. The network throughputversus velocity ratio, �, is shown in Fig. 6. Link throughputis inversely proportional to �. With increased �, nodemobility is increased, thereby decreasing link ANFD. Asexpected, lower mobility achieves higher network through-put, so utilizing the ANFD in the routing metric can help toimprove network throughput.

6.2 Varying Node Mobility

We now compare AOMDV and CA-AOMDV with respect tonode mobility. The simulated network areas were 2,200 m �600 m and 2,800 m� 600 m, 2 Mb/s channel bandwidth, 100 srunning time, 100 uniformly distributed nodes moving atmaximum speed in random directions with 20 connections.Maximum node speed was increased from 1 to 10 m/s. The512-byte CBR sources were fixed at 5 packets/s.

6.2.1 Throughput

Simulation results for network throughput are shown inFig. 7. Throughput decreases with increased node mobility,with CA-AOMDV outperforming AOMDV, particularly, inthe mid-range mobilities, with significant performanceincreases realized. At 4 m/s, CA-AOMDV provides 25.5and 12.2 percent improvements for the smaller and largernetworks, respectively. At extreme mobilities, the through-put performances vary less and the advantages of CA-AOMDV are greater with smaller network area (shorterpath lengths) as previously noted. At low mobilities, pathcharacteristics vary less quickly and the advantages ofhandoff in CA-AOMDV are less. At high mobilities, channeland path characteristics change rapidly, again mitigatinghandoff scheme advantages, and increasing signal strengthprediction efficacy.

6.2.2 End-to-End Delay

Fig. 8 shows average packet transmission delay results. CA-AOMDV outperforms AOMDV, with 24 and 28 percentimprovements at a velocity of 4 m/s, for the smaller and

larger networks, respectively. At extreme mobilities, theperformances converge.

6.2.3 Normalized Routing Control Overhead

Normalized routing control overhead is the ratio of numberof routing control packets to delivered data packets, asplotted in Fig. 9. Overhead for both protocols increases withincreasing node mobility because the more quickly chan-ging network topology increases routing update frequency.Except for extreme mobilities, CA-AOMDV maintains alower routing overhead compared with AOMDV, with a 14and 16 percent improvements at 4 m/s for the smaller andlarger networks, respectively.

6.3 Varying Traffic Load

To evaluate network traffic load performance, we fixedmaximum speed at 1 m/s, varying source packet rate from 5to 40 packets/s. All other parameters were as in the previoussection. Fig. 10 shows variation of packet delivery ratio (PDR)with increasing packet rate, while Fig. 11 shows variation ofaverage end-to-end delay with increasing packet rate.

Both protocols have decreased PDR with increasingpacket rate. For low traffic loads, increased packet rateprolongs the average end-to-end delay. After a certain


Fig. 6. Link throughput versus �, � ¼ vB=vA, vA is fixed at 1 m/s, forcommunicating nodes, A and B, moving in parallel directions, separatedby 70 m, initially.

Fig. 7. Network throughput comparison between CA-AOMDV andAOMDV with increasing node mobility.

Fig. 8. Average end-to-end delay comparison between CA-AOMDV andAOMDV with increasing node mobility.

point, the packet delays decrease with increasing packetrate. The decrease of average end-to-end delay in Fig. 11occurs because, at higher packet rate, more packets aredropped due to congestion. For both PDR and average end-to-end delay, CA-AOMDV outperforms AOMDV. For apacket rate of 20 packets/second, there is a 34.1 percentimprovement of packet delivery ratio and 18.7 percentimprovement of the average end-to-end delay.

6.4 Validation of the Theoretical Model

In this section, we validate the theoretical analysis fromSection 5. The network had 80 nodes, 10 connections, 1 Mb/schannel bandwidth, 1,000 byte packets at 4 packets/s,running for 300 seconds. Average link lifetimes were variedby changing network size, for a fixed number of nodes, andthe node velocities for a fixed transmission range. Averagepath lifetime is increased from 8 to 50 s.

The theoretical and simulated routing control overheadresults are shown in Fig. 12 and match quite well. For shortpath lifetimes, the theoretical values are higher than thesimulation values because the average distance betweennodes is long and/or they are moving quickly. Under suchconditions, long paths are difficult to establish. Thus, the

established paths in the simulated network are shorter thanthose in the theoretical model, and the probability of a pathbreaking in the simulated network is lower than in theory.For longer path lifetimes, the theoretical values are lowerthan the simulation values, which maybe due to overheadintroduced by interference and packet collisions. CA-AOMDV performs better than AOMDV for both thesimulated and theoretical results, as expected.

CA-AOMDV outperforms AOMDV for both theoreticaland simulated packet delivery ratio as shown in Fig. 13.Although the theoretical results follow the same trend asthe simulated results, the former are always higher,especially for lower average path lifetimes. The differenceis due to interference and packet collisions. In an adverseenvironment, with high frequency of broken paths, packetdrop rate and number of packet retransmissions arerelatively high, raising the amount of interference andnumber of collisions, incurring greater congestion. Thepacket drop rate can increase dramatically during conges-tion. As for previous results with respect to low and highmobility, the PDR performance of AOMDV and CA-AOMDV converge for low and high average path lifetimes.


Fig. 9. Routing control overhead (normalized with respect to delivereddata packets) comparison of CA-AOMDV and AOMDV with increasingnode mobility.

Fig. 11. Average end-to-end delay comparison between CA-AOMDVand AOMDV with increasing packet rate.

Fig. 10. Packet delivery ratio comparison between CA-AOMDV andAOMDV with increasing packet rate.

Fig. 12. Routing control overhead comparison of theoretical andsimulated results with increasing average path lifetime. Theoreticalvalues evaluated from (36) and (37).

7 CONCLUSION

A channel-based routing metric is proposed which utilizesthe average nonfading duration, combined with hop-count,to select stable links. A channel-adaptive routing protocol,CA-AOMDV, extending AOMDV, based on the proposedrouting metric, is introduced. During path maintenance,predicted signal strength and channel average fadingduration are combined with handoff to combat channelfading and improve channel utilization. A new theoreticalexpression for the lifetime of the multiple reusable pathsystem used in CA-AOMDV is derived. Theoreticalexpressions for routing control overhead and packetdelivery ratio also provide detailed insights into thedifferences between the two protocols. Theoretical analysisand simulation results show that CA-AOMDV outperformsAOMDV in practical transmission environments.

ACKNOWLEDGMENTS

Xiaoqin Chen was with the National ICT Australia andaffiliated with the Australian National University at thetime of writing.

REFERENCES

[1] W.C. Jakes, Microwave Mobile Communications, second ed. WileyInterscience, 2001.

[2] G.L. Stuber, Principles of Mobile Communication, second ed. KluwerAcademic, 2001.

[3] S. Jain and S.R. Das, “Exploiting Path Diversity in the Link Layerin Wireless Ad Hoc Networks,” Proc. Sixth IEEE Int’l Symp. Worldof Wireless Mobile and Multimedia Networks (WoWMoM), pp. 22-30,June 2005.

[4] P. Pham, S. Perreau, and A. Jayasuriya, “New Cross-Layer DesignApproach to Ad Hoc Networks under Rayleigh Fading,” IEEEJ. Selected Areas in Comm., vol. 23, no. 1, pp. 28-39, Jan. 2005.

[5] C. Toh, “Associativity-Based Routing for Ad-Hoc Mobile Net-works,” Wireless Personal Comm., vol. 4, pp. 103-139, Nov. 1997.

[6] H. Zhang and Y.N. Dong, “Mobility Prediction Model Based LinkStability Metric for Wireless Ad Hoc Networks,” Proc. Int’l Conf.Wireless Comm., Networking and Mobile Computing (WiCOM), pp. 1-4, Sept. 2006.

[7] O. Tickoo, S. Raghunath, and S. Kalyanaraman, “Route Fragility:A Novel Metric for Route Selection in Mobile Ad Hoc Networks,”Proc. IEEE Int’l Conf. Networks (ICON), pp. 537-542, Sept. 2003.

[8] S. Biswas and R. Morris, “ExOR: Opportunistic Multi-HopRouting for Wireless Networks,” ACM SIGCOMM ComputerComm. Rev., vol. 35, no. 4, pp. 133-144, Aug. 2005.

[9] R. Dube, C. Rais, K.Y. Wang, and S.K. Tripathi, “Signal Stability-Based Adaptive Routing (SSA) for Ad Hoc Networks,” IEEEPersonal Comm., vol. 4, no. 1, pp. 36-45, Feb. 1997.

[10] M. Park, J. Andrews, and S. Nettles, “Wireless Channel-AwareAd Hoc Cross-Layer Protocol with Multiroute Path SelectionDiversity,” Proc. IEEE Vehicular Technology Conf. (VTC)-Fall, vol. 4,pp. 2197-2201, Oct. 2003.

[11] X. Lin, Y.K. Kwok, and V.K.N. Lau, “RICA: A Receiver-InitiatedApproach for Channel-Adaptive On-demand Routing in Ad HocMobile Computing Networks,” Proc. Int’l Conf. DistributedComputing Systems (ICDCS), pp. 84-91, July 2002.

[12] B. Awerbuch, D. Holer, and H. Rubens, “High Throughput RouteSelection in Multi-Rate Ad Hoc Wireless Networks,” Proc. Conf.Wireless On-Demand Network Systems (WONS), pp. 251-269, March2004.

[13] S. Zhao, Z. Wu, A. Acharya, and D. Raychaudhuri, “PARMA: APHY/MAC Aware Routing Metric for Ad-Hoc Wireless Networkswith Multi-Rate Radios,” Proc. Sixth IEEE Int’l Symp. World ofWireless Mobile and Multimedia Networks (WoWMoM), pp. 286-292,June 2005.

[14] M.R. Souryal, B.R. Vojcic, and R.L. Pickholtz, “InformationEfficiency of Multihop Packet Radio Networks with Channel-Adaptive Routing,” IEEE J. Selected Areas in Comm., vol. 23, no. 1,pp. 40-50, Jan. 2005.

[15] M.K. Marina and S.R. Das, “On-Demand Multipath DistanceVector Routing in Ad Hoc Networks,” Proc. Ninth Int’l Conf.Network Protocols (ICNP), pp. 14-23, Nov. 2001.

[16] C.E. Perkins and E.M. Royer, “Ad-Hoc On-Demand DistanceVector Routing,” Proc. IEEE Workshop Mobile Computing Systemsand Applications (WMCSA), pp. 90-100, Feb. 1999.

[17] A.S. Akki, “Statistical Properties of Mobile-to-Mobile LandCommunication Channel,” IEEE Trans. Vehicular Technology,vol. 43, no. 4, pp. 826-831, Nov. 1994.

[18] M. Haenggi, “Routing in Ad Hoc Networks—A WirelessPerspective,” Proc. First Int’l Conf. Broadband Networks (BroadNets),pp. 652-660, Dec. 2004.

[19] A. Savvides, C.C. Han, and M. Srivastava, “Dynamic Fine-GrainedLocalization in Ad-Hoc Networks of Sensors,” Proc. ACMMobiCom, pp. 166-179, May 2001.

[20] D. Niculescu and B. Nath, “Ad Hoc Positioning System (APS),”Proc. IEEE Global Telecomm. Conf. (GLOBECOM), pp. 2926-2931,Nov. 2001.

[21] C. Savarese, J.M. Rabaey, and J. Beutel, “Locationing inDistributed Ad-Hoc Wireless Sensor Networks,” Proc. IEEE Int’lConf. Acoustics, Speech, and Signal Processing (ICASSP), vol. 4,pp. 2037-2040, Dec. 2001.

[22] D. Niculescu and B. Nath, “Ad Hoc Positioning System (APS)Using AoA,” Proc. IEEE INFOCOM, pp. 1734-1743, Apr. 2003.

[23] N. Priyantha, A. Miu, H. Balakrishnan, and S. Teller, “The CricketCompass for Context-Aware Mobile Applications,” Proc. ACMMobiCom, pp. 1-14, July 2001.

[24] A.B. Mnaouer, L. Chen, C.H. Foh, and J.W. Tantra, “OPHMR: AnOptimized Polymorphic Hybrid Multicast Routing Protocol forMANET,” IEEE. Trans. Mobile Computing, vol. 5, no. 6, pp. 503-514,May 2007.

[25] Z. Zaidi and B. Mark, “A Mobility Tracking Model for WirelessAd Hoc Networks,” Proc. Wireless Comm. and Networking Conf.(WCNC), pp. 1790-1795, March 2003.

[26] L.L. Scharf, Statistical Signal Processing: Detection, Estimation, andTime Series Analysis. Addison-Wesley, 1991.

[27] W. Su and M. Gerla, “IPv6 Flow Handoff in Ad Hoc WirelessNetworks Using Mobility Prediction,” Proc. IEEE Global Telecomm.Conf. (GLOBECOM), vol. 1a, pp. 271-275, Dec. 1999.

[28] T. Goff, N. Abu-Ghazaleh, D. Phatak, and R. Kahvecioglu,“Preemptive Routing in Ad Hoc Networks,” Proc. Ann. Int’l Conf.Mobile Computing and Networks (ICMCN), pp. 43-52, July 2001.

[29] J. Bicket, D. Aguayo, S. Biswas, and R. Morris, “Architecture andEvaluation of an Unplanned 802.11b Mesh Network,” Proc. ACMMobiCom, pp. 31-42, Aug. 2005.

[30] S. Panichpapiboon, G. Ferrari, and O.K. Tonguz, “OptimalTransmit Power in Wireless Sensor Networks,” IEEE Trans. MobileComputing, vol. 5, no. 10, pp. 1432-1447, Oct. 2006.


Fig. 13. Packet delivery ratio comparison between theoretical andsimulated results with increasing average path lifetime. Theoreticalvalues evaluated from (43).

[31] A. Nasipuri and S.R. Das, “On-Demand Multipath Routing forMobile Ad Hoc Networks,” Proc. Int’l Conf. Computer Comm. andNetworks (CCCN), pp. 64-70, Oct. 1999.

[32] A.S. Akki and F. Haber, “A Statistical Model of Mobile-to-MobileLand Communication Channel,” IEEE Trans. Vehicular Technology,vol. 35, no. 2, pp. 2-10, Feb. 1986.

[33] M. Heusse, F. Rousseau, G. Berger-Sabbatel, and A. Duda,“Performance Anomaly of 802.11b,” Proc. IEEE INFOCOM,vol. 2, pp. 836-843, March 2003.

[34] The Network Simulator - ns-2, http://www.isi.edu/nsnam/ns,2008.

[35] C.S. Patel, G.L. Stuber, and T.G. Pratt, “Simulation of Rayleigh-Faded Mobile-to-Mobile Communication Channels,” IEEE Trans.Comm., vol. 53, no. 11, pp. 1876-1884, Nov. 2005.

[36] R.J. Punnoose, P.V. Nikitin, and D.D. Stancil, “Efficient Simulationof Ricean Fading within a Packet Simulator,” Proc. IEEE VehicularTechnology Conf. (VTC), vol. 2, pp. 764-767, Sept. 2000.

[37] J. Broch et al., “A Performance Comparison of Multi-Hop WirelessAdhoc Network Routing Protocols,” Proc. ACM MobiCom, pp. 85-97, Oct. 1998.

Xiaoqin Chen received the BE degree inelectrical and electronic engineering from TianjinUniversity, China, in 1996 and the PhD degree inwireless communications engineering from theAustralian National University in 2009. She wasa research fellow at the Queensland Universityof Technology, and is now a system engineer atStar Intellect Pty Ltd. Her research interestsinclude cross-layer design, channel-adaptiverouting and MAC protocol design, radio resource

management, packet scheduling, link adaptation, and network optimiza-tion for wireless networks.

Haley M. Jones received the BE (hons) degreein electrical and electronic engineering and theBSc degree from the University of Adelaide,Australia, in 1992 and 1995, respectively, andthe PhD degree in telecommunications engi-neering from the Australian National University,Canberra, in October 2002. She has been anacademic in the College of Engineering andComputer Science, The Australian NationalUniversity, since January 2002. Her previous

experience includes time in industry and working on speech coding withthe Cooperative Research Centre for Robust and Adaptive systemsfrom 1993 to 1999. Her research interests have included wirelesschannel modeling, beamforming, and channel and topology issues inMANETs. She has recently branched out into sustainable systems witha particular emphasis on the cradle-to-cradle paradigm and itsapplication to sustainable manufacturing.

Dhammika Jayalath received the BSc degreein electronics and telecommunications engineer-ing from the University of Moratuwa, Sri Lanka,the MEng degree in telecommunications fromthe Asian Institute of Technology, Thailand, andthe PhD degree in wireless communicationsfrom Monash University, Australia, in 2002. Hewas a fellow at the Australian National Universityand a senior researcher at the National CTAustralia. His research interests include the

general areas of communications and signal processing, and he haspublished significantly in these areas. His current research interestsinclude cooperative communications, cognitive radios, statistical signalprocessing, and multiuser communications. He is currently serving asthe chair of the Signal Processing and Communications Chapter of theIEEE Queensland Section. He is a senior member of the IEEE.

. For more information on this or any other computing topic,please visit our Digital Library at www.computer.org/publications/dlib.


channel-aware routing in manets with route handoff

Documents