<![cdata[improving source routing reliability in mobile ad hoc networks]]>

12
Improving Source Routing Reliability in Mobile Ad Hoc Networks Song Guo, Student Member, IEEE, Oliver Yang, Senior Member, IEEE, and Yantai Shu, Member, IEEE Abstract—In this paper, we propose a novel on-demand routing protocol called Backup Source Routing (BSR) to establish and maintain backup routes that can be utilized after the primary path breaks. The key advantage of BSR is the reduction of the frequency of route discovery flooding, which is recognized as a major overhead in on-demand protocols. We define a new routing metric, called the route reliability, and use it to provide the basis for the backup path selection. We use a heuristic cost function to develop an analytical model and an approximation method to measure this metric. Various algorithms for our BSR protocol in the route discovery phase and route maintenance phase have been designed based on this cost function. Extensive simulations demonstrated that our routing strategy has two interesting features: 1) In less stressful situations of lower mobility, BSR has similar performance to DSR. 2) In more challenging situations of high mobility, BSR can improve the performance significantly. Index Terms—Mobile communication system, wireless communication, routing protocol, algorithm/protocol design and analysis, multihop ad hoc network. æ 1 INTRODUCTION W ITH no preexisting fixed infrastructure, ad hoc net- works are gaining increasing popularity nowadays because of their ease of deployment and usability anytime and anywhere. Such networks not only play an important role under the disorganized or hostile setting, but they possess obvious potential applications in all traditional areas of interest for mobile computing. Since nodes in the network move freely and randomly, routes often get disconnected. The major challenge is therefore to imple- ment routing protocols that must respond to changes in the network topology in order to maintain and reconstruct the routes in a timely manner as well as to establish reliable routes. A working group called “MANET” has been formed by the Internet Engineering Task Force (IETF) to study related issues and to stimulate research. Extensive research efforts, e.g., [1] have been devoted to the design of routing protocols for MANET. Various routing strategies have been designed to address the problem of choosing the most reliable paths and minimizing the reactive cost to topological changes [2], [3], [4], [5], [6], [7], [8], [9], [10], [11], [12]. One of the earliest works is the development of Associativity Based Routing (ABR) [2]. It is a protocol that uses the connection-oriented packet forwarding approach. Although the resulting path does not necessarily result in the smallest possible number of hops, the path tends to live longer than other routes, and normally requires fewer route reconstructions. Conse- quently, a higher throughput is achieved. Signal Stability Adaptive Routing (SSA) [3] follows a similar approach by distinguishing strongly connected from weakly connected links. As a variation of signal strength measurements, the Route-lifetime Assessment Based Routing (RABR) [4] tries to predict the time when the received signal strength falls below a critical threshold using a measured value of average change in received signal strength. All of these [2], [3], [4] are based on signal strength. A prediction method for link reliability [5] is based on distance measurements between mobile devices. A refine- ment in [6] takes possible changes in speed or direction of motions into account. The distance between connection peers may be acquired with the help of GPS receivers or signal strength measurements. A major problem of this approach is that the distance of a receiver is only a very vague hint on link availability. In realistic environments, the coverage area of a radio transmission hardly ever has a circular shape and is subject to strong fluctuations. A further improvement based on the availability of GPS measurements has been suggested in [7]. However, it still suffers from the same problem. An approach often used to obtain a better reliability in varying network conditions is the use of “multipath routing” [8], [9], [10], [11], [12], [13], [14], [15]. The applications of such technique seem natural, since multi- path routing diminishes the effects of unreliable wireless links and the constantly changing topology. The on-demand multipath routing scheme is presented in [8] as an extension for the popular dynamic source routing DSR [16]. Unlike the traditional DSR protocol, which has an option to maintain multipath routes in a trivial manner without any regard to their ultimate usefulness, it takes advantage of disjoint paths, in which all intermediate nodes are equipped with a disjoint alternate route, so that in-transit data packets no longer face any route loss. Its performance surpasses that of the DSR especially under the environment of higher mobility. However, in order to support an alternate disjoint route to the destination for each intermediate node in the 362 IEEE TRANSACTIONS ON PARALLEL AND DISTRIBUTED SYSTEMS, VOL. 16, NO. 4, APRIL 2005 . S. Guo is with the University of Ottawa, 800 King Edward, Ottawa, Ontario, Canada K1N 6N5. E-mail: [email protected]. . O. Yang is with the University of Ottawa, 161 Louis Pasteur, CBY-A514, Ottawa, Ontario, Canada K1N 6N5. E-mail: [email protected]. . Y. Shu is with Tianjin University, North 5 Village 21-2-201, Tianjin, P.R. China, 300073. E-mail: [email protected]. Manuscript received 4 Mar. 2003; revised 16 Mar. 2004; accepted 29 Aug. 2004; published online 23 Feb. 2005. For information on obtaining reprints of this article, please send e-mail to: [email protected], and reference IEEECS Log Number TPDS-0023-0303. 1045-9219/05/$20.00 ß 2005 IEEE Published by the IEEE Computer Society

Post on 15-Dec-2016

213 views

Category:

Documents


1 download

TRANSCRIPT

Improving Source Routing Reliabilityin Mobile Ad Hoc Networks

Song Guo, Student Member, IEEE, Oliver Yang, Senior Member, IEEE, and Yantai Shu, Member, IEEE

Abstract—In this paper, we propose a novel on-demand routing protocol called Backup Source Routing (BSR) to establish and

maintain backup routes that can be utilized after the primary path breaks. The key advantage of BSR is the reduction of the frequency

of route discovery flooding, which is recognized as a major overhead in on-demand protocols. We define a new routing metric, called

the route reliability, and use it to provide the basis for the backup path selection. We use a heuristic cost function to develop an

analytical model and an approximation method to measure this metric. Various algorithms for our BSR protocol in the route discovery

phase and route maintenance phase have been designed based on this cost function. Extensive simulations demonstrated that our

routing strategy has two interesting features: 1) In less stressful situations of lower mobility, BSR has similar performance to DSR. 2) In

more challenging situations of high mobility, BSR can improve the performance significantly.

Index Terms—Mobile communication system, wireless communication, routing protocol, algorithm/protocol design and analysis,

multihop ad hoc network.

1 INTRODUCTION

WITH no preexisting fixed infrastructure, ad hoc net-works are gaining increasing popularity nowadays

because of their ease of deployment and usability anytimeand anywhere. Such networks not only play an importantrole under the disorganized or hostile setting, but theypossess obvious potential applications in all traditionalareas of interest for mobile computing. Since nodes in thenetwork move freely and randomly, routes often getdisconnected. The major challenge is therefore to imple-ment routing protocols that must respond to changes in thenetwork topology in order to maintain and reconstruct theroutes in a timely manner as well as to establish reliableroutes. A working group called “MANET” has been formedby the Internet Engineering Task Force (IETF) to studyrelated issues and to stimulate research. Extensive researchefforts, e.g., [1] have been devoted to the design of routingprotocols for MANET.

Various routing strategies have been designed to addressthe problem of choosing the most reliable paths andminimizing the reactive cost to topological changes [2],[3], [4], [5], [6], [7], [8], [9], [10], [11], [12]. One of the earliestworks is the development of Associativity Based Routing(ABR) [2]. It is a protocol that uses the connection-orientedpacket forwarding approach. Although the resulting pathdoes not necessarily result in the smallest possible numberof hops, the path tends to live longer than other routes, andnormally requires fewer route reconstructions. Conse-quently, a higher throughput is achieved. Signal Stability

Adaptive Routing (SSA) [3] follows a similar approach bydistinguishing strongly connected from weakly connectedlinks. As a variation of signal strength measurements, theRoute-lifetime Assessment Based Routing (RABR) [4] triesto predict the time when the received signal strength fallsbelow a critical threshold using a measured value ofaverage change in received signal strength. All of these[2], [3], [4] are based on signal strength.

A prediction method for link reliability [5] is based ondistance measurements between mobile devices. A refine-ment in [6] takes possible changes in speed or direction ofmotions into account. The distance between connectionpeers may be acquired with the help of GPS receivers orsignal strength measurements. A major problem of thisapproach is that the distance of a receiver is only a veryvague hint on link availability. In realistic environments, thecoverage area of a radio transmission hardly ever has acircular shape and is subject to strong fluctuations. Afurther improvement based on the availability of GPSmeasurements has been suggested in [7]. However, it stillsuffers from the same problem.

An approach often used to obtain a better reliability invarying network conditions is the use of “multipathrouting” [8], [9], [10], [11], [12], [13], [14], [15]. Theapplications of such technique seem natural, since multi-path routing diminishes the effects of unreliable wirelesslinks and the constantly changing topology. The on-demandmultipath routing scheme is presented in [8] as an extensionfor the popular dynamic source routing DSR [16]. Unlikethe traditional DSR protocol, which has an option tomaintain multipath routes in a trivial manner without anyregard to their ultimate usefulness, it takes advantage ofdisjoint paths, in which all intermediate nodes are equippedwith a disjoint alternate route, so that in-transit data packetsno longer face any route loss. Its performance surpasses thatof the DSR especially under the environment of highermobility. However, in order to support an alternate disjointroute to the destination for each intermediate node in the

362 IEEE TRANSACTIONS ON PARALLEL AND DISTRIBUTED SYSTEMS, VOL. 16, NO. 4, APRIL 2005

. S. Guo is with the University of Ottawa, 800 King Edward, Ottawa,Ontario, Canada K1N 6N5. E-mail: [email protected].

. O. Yang is with the University of Ottawa, 161 Louis Pasteur, CBY-A514,Ottawa, Ontario, Canada K1N 6N5. E-mail: [email protected].

. Y. Shu is with Tianjin University, North 5 Village 21-2-201, Tianjin, P.R.China, 300073. E-mail: [email protected].

Manuscript received 4 Mar. 2003; revised 16 Mar. 2004; accepted 29 Aug.2004; published online 23 Feb. 2005.For information on obtaining reprints of this article, please send e-mail to:[email protected], and reference IEEECS Log Number TPDS-0023-0303.

1045-9219/05/$20.00 � 2005 IEEE Published by the IEEE Computer Society

primary route, intermediate nodes are not allowed to sendreply back to the source even when they have fresh routeinformation to the destination in the route cache. Thus, thecontrol overhead in the route discovery could overwhelmthe available bandwidth of the network. Moreover, it maynot always be possible for all intermediate nodes to get analternate disjoint route. As an extension of AODV [17], theAODV-BR [9] maintains multipath routes and utilizes themonly when the primary route fails. A more recent paper [10]has proposed a multipath routing algorithm called theDisjoint Path-set Selection Protocol (DPSP). It is a heuristicthat exploits the path protection technique by replacingpreviously found interlacing paths with a set of disjointreliable paths. Finally, multipath has also been used in otherproblems, but they are used either for load balancing [13],[14], [15] or for improving end-to-end delay [15] with nodiscussion on the reliability issue.

In this paper, we propose the Backup Source Routing(BSR) algorithm. Unlike previous work, BSR uses theconcept of similar paths to establish and maintain backuppaths. We believe that utilizing backup routes is beneficialin network communications, particularly in mobile wirelessnetworks where routes are disconnected frequently due tomobility and poor wireless link quality. We shall define anew routing metric, called route reliability, and develop ananalytical model to provide a theoretical basis for backuppath selection. The analytical model allows us to implementthe BSR algorithm, which extends DSR by selecting abackup path piggybacked with the primary path in theheader of data packets to achieve the most reliable routesbetween each communicating mobile node. The implemen-tation of our BSR algorithm is both practical and efficientbecause it dramatically reduces the rate of route discoveries,which is recognized as a major overhead in on-demandprotocols. Through modeling, analysis, and performanceevaluation via simulations, we show that our routingstrategy has a couple of advantages: 1) In less stressfulsituations of lower mobility, BSR has similar performanceas DSR, and 2) in more challenging situations of highmobility, BSR can improve the performance significantly.

2 SYSTEM MODEL

We shall define precisely the system model and terminol-ogy used throughout this paper.

2.1 Network Model and Terminology

An ad hoc network can be modeled by a directed graphGðV ;EÞ, where the vertices in set V represent the mobilenodes and the edges in setE correspond to the unidirectionalcommunication links. A directed path � on graph G can bedefined by a sequence of incident edges or equivalently by asequence of adjacent vertices. Let V ð�Þ denote the node set ofpath�, and letEð�Þdenote theedge setofpath�. The lengthofa path � is equal to its hop number denoted as j�j. A path issimple if it does not contain a cycle. All the paths are simplepaths in this paper. Define the rank, r�ðuÞ, to be the index ofthe node u in an increasing sequence from the source node tothe target node of the directed path �. Then, �uv ¼ ðu; . . . ; vÞ isa directed subpath of � if r�ðuÞ < r�ðvÞ. If � ¼ ðv1; v2; . . . ; vnÞand �0 ¼ ðvn; u1; u2; . . . ; umÞ, and for all i 2 f1; 2; . . . ; ng; j 2f1; 2; . . . ;mg, vi 6¼ uj, then we say � ¼ �þ �0 ¼ ðv1; v2; . . . ;vn; u1; u2; . . . ; umÞ is a concatenation of � and �0.

Given a directed graph GðV ;EÞ, �1 and �2 are bothdirected paths from a source node s to a target node t. Thelink similarity Lð�1; �2Þ of two paths �1 and �2 is thenumber of links that appear in both paths �1 and �2.Likewise, the node similarity Nð�1; �2Þ of two paths �1 and�2 is the number of nodes that appear in both paths �1 and�2. It is obvious that Nð�1; �2Þ is always greater thanLð�1; �2Þ. The paths �1 and �2 are disjoint paths if �1 and �2

are 0-link similar and 2-node similar paths that are commononly at the source and target nodes. Let L1 and L2 be thesubpaths of �1 and �2 respectively, and if L1 and L2 aredisjoint paths, L1 and L2 are called disjoint subpaths of �1

and �2. We use Dð�1; �2Þ to denote the number of disjointsubpath of �1 and �2.

2.2 Mobility Model

We shall use a variation of the best-known mobility modelcalled the Random Waypoint model [16] for its simplicity,and describe/denote it by the notation Vmin=Vmax=Tpause. Inthis model, each node starts its journey from a randomlocation to a random destination with a randomly chosenspeed between Vmin and Vmax. Once the destination isreached, another random destination is targeted after apause time Tpause. Note that the pause time also affects therelative speeds of the mobile nodes. We shall use the steady-state average speed V to measure the level of mobility in thenetwork, where V is given by [18].

V ¼ Vmax � Vmin

lnVmax � lnVmin: ð1Þ

Two interesting results could be derived from the equationabove: 1) V � ðVmax þ VminÞ=2 (equality holds when Vmax ¼Vmin), and 2) V ¼ 0when Vmin ¼ 0. The latter is known as thespeed decay problem.1 Since the most common mobilityscenario 0=Vmax=Tpause would fail due to the speed decayproblem, and since node speed is shown to be a significantfactor to affect the performance of routing protocols (whilepause time is not) [19], our simulations shall use the samemodel 0þ=Vmax=0 used in [18] with continuous motion andnonzero minimum speed in order to achieve nonzeromobility level V . We also use different maximum speeds toevaluate performance of a number of protocols with variousmobility levels V .

3 BACKUP ROUTE AND ROUTING METRICS

In on-demand ad hoc routing protocols, since the routediscovery is only triggered by the need for sending a datapacket for which the route is unknown, a source must waitfor the route discovery to complete before sending it. Once apath has been discovered and used, the longer it lasts, theshorter the latency in data packet delivery, and the lessfrequently a rerouting process needs to be initiated. Thererouting frequency is recognized as a major overhead inon-demand protocols and, thus, critically affects thecommunicating performance, especially in a stressfulenvironment. Therefore, we would like to reduce it by

GUO ET AL.: IMPROVING SOURCE ROUTING RELIABILITY IN MOBILE AD HOC NETWORKS 363

1. When Vmin ¼ 0, the longer we run the simulation, the further ourresults would deviate from the real performance because the level ofmobility continuously decreases as the simulation evolves.

maintaining a backup path to the same destination so thatwhen the primary path breaks, it may be used withoutwaiting for a fresh search to be completed.

3.1 Backup Route

Normally, the primary path must have the minimal end-to-end delay. In order to keep a short delay on the backuppath, we restrict the backup path to have the shared nodeswith the primary path in the same order as explained in thefollowing definition.

Definition 1. A simple path �0 is called a backup path of theprimary path �, if � and �0 have the common source and targetnodes, and satisfy the condition, 8u; v 2 V ð�Þ [ V ð�0Þ; r�ðuÞ <r�ðvÞ () r�0 ðuÞ < r�0 ðvÞ.

Definition 2. A path-pair (�; �0) is called a backup route froma source node s to a target node t, if � is a primary path fromnode s to node t, and �0 is a backup path of �.

In order to avoid confusion, we would like to point outthat a backup path is only a single path, while a backup routeis a path-pair consisting of a primary path and a backup path.Fig. 1 illustrates an example of a backup route ð�1; �2Þ froma source node s to a target node t, where Lð�1; �2Þ ¼ 1,Nð�1; �2Þ ¼ 5, and Dð�1; �2Þ ¼ 3.

3.2 Routing Metrics

A routing algorithm uses routing metric to determine if oneroute is better than another. As mentioned before, the end-to-end delay will be used as a routing metric to find aprimary path with the minimal end-to-end delay. Thestrategy of using a backup route is to reduce the frequencyof initiating route discovery, thus prolonging the lifetime ofthe route discovered.

Definition 3. The reliability of a backup route ð�; �0Þ is definedas the mean value E½T�;�0 � of the lifetime T�;�0 (a randomvariable) of the backup route ð�; �0Þ.

When pursuing this goal in the design of reliable routingprotocol, we shall use this new routing metric for theselection of backup paths. Given a primary path �, thisrouting metric is to select a backup path �0 to construct thelongest-lived backup route ð�; �0Þ such that its mean lifetimeE½T�;�0 � is maximized.

Based on a simple mobility scenario V =V =0, which canbe obtained from (1) by setting Vmin ¼ Vmax, the lifetime foreach communication link L in the network can beapproximately represented by an independent and identical(iid) exponential random variable XL [20]. Without loss ofgenerality, the mean value E½XL� of the link lifetime can benormalized to one. It is easy to show that the lifetime for apath P consisting of n serial wireless links is also an

exponentially distributed random variable, denoted as XP ,with a mean value of 1=n. Finally, the distribution of thebackup route lifetime T�;�0 ; �; �

0 2 E, can be obtained inTheorem 1 below. Note that all the derivation in the proof ofTheorem 1 and the following analysis are based on thecharacterization of link lifetimes given in [20], and moreaccurate lifetime distribution would be adopted into ouranalytical model.

Theorem 1. Let Li denote the common directed links of ð�; �0Þ,where i ¼ 1; 2; . . . ; � ¼ Lð�; �0Þ, and let ðPj; P

0jÞ denote

disjoint subpaths of ð�; �0Þ, where Pj and P 0j have the length

of nj and n0j; j ¼ 1; 2; . . . ; d ¼ Dð�; �0Þ, respectively. Then, the

cumulative distributed function of T�;�0 is given by:

FT�;�0 ðtÞ ¼ 1� e��t �d

j¼1e�njt þ e�n0

jt � e�njt�n0jt

� �: ð2Þ

Proof. It is a straightforward exercise to show that for anybackup route ð�; �0Þ, � ¼ fL1; L2; . . . ; L�; ðP1; P

01Þ; ðP2;

P 02Þ; . . . ; ðPd; P

0dÞg is a partition of Eð�Þ [Eð�0Þ. Consider-

ing all the exponentially distributed random variables ofXLi

;XPj;XP 0

jði ¼ 1; 2 . . .�; j ¼ 1; 2 . . . dÞ, we obtained the

cumulative distributed function of T�;�0 as follows:

FT�;�0 ðtÞ¼P ½T�;�0 � t� ¼ P ½MinðXL1; . . . ; XL�

;MaxðXP1; XP 0

1Þ;

. . . ;MaxðXPd;XP 0

dÞÞ � t�

¼ 1���i¼1P ½XLi

> t��dj¼1P ½MaxðXPj

;XP 0jÞ > t� ¼

1���i¼1ð1� P ½XLi

� t�Þ�dj¼1ð1� P ½MaxðXPj

;XP 0jÞ � t�Þ

¼ 1���i¼1ð1� FXLi

ðtÞÞ�dj¼1ð1� FXPj

ðtÞFXP 0j

ðtÞÞ

¼ 1���i¼1e

�t�dj¼1ð1� ð1� e�njtÞð1� e�n0

jtÞÞ

¼ 1� e��t�dj¼1ðe�njt þ e�n0

jt � e�njt�n0jtÞ:

ut

Since E½T�;�0 � ¼R10 tdFT�;�0 ðtÞ, substituting by (2) and

using the definition, one obtains the reliability of thebackup route. In addition to the mathematical formulationabove, we can also make an intuitive observation that abackup route ð�; �0Þ would have a better performance(longer lifetime), if both � and �0 have shorter lengths, and�0 is less link-similar to but having more disjoint subpathswith its primary path �. Equation (2) allows us to perform aMonte Carlo simulation to take a closer look at theprobabilistic reason behind this observation. Through thissimulation study, we can deduce a heuristic cost functionCð�; �0Þ in (3) which can relate to the mean lifetime and,therefore, estimate the backup route reliability very closely(by using linear regression given in (4)).

Cð�; �0Þ ¼Minð3j�j; j�j þ Lð�; �0Þ þ j�0j=Dð�; �0ÞÞ ð3ÞEð�; �0Þ � �=Cð�; �0Þ þ �: ð4Þ

One can see that j�j contributes to the selection of theprimary path since Cð�; �0Þ is smaller when its primary pathis shorter. Likewise, the items Lð�; �0Þ and j�0j=Dð�; �0Þcontribute to a higher reliability via the cost function if itsbackup path is less link-similar to but having more disjointsubpaths with its primary path. We can use a linear

364 IEEE TRANSACTIONS ON PARALLEL AND DISTRIBUTED SYSTEMS, VOL. 16, NO. 4, APRIL 2005

Fig. 1. Backup route ð�1; �2Þ, �1 ¼ ðs; b; c; e; g; i; j; tÞ, and �2 ¼ ðs; a; b; c; d;f; h; i; k; tÞ.

regression to fit a line by applying the least squares errorsmethod in a set of two-dimensional points (E½T�;�0 �,1=Cð�; �0Þ) generated by (2) and (3).

Each point in Fig. 2 corresponds to a possible backuproute. Since there are a finite number of j�j, j�0j, Lð�; �0Þ, andDð�; �0Þ (all integer values) for a given range of reliability(e.g., above 0.2), Fig. 2 represents the complete enumerationof all possible backup routes. We can consider that (3) and(4) provide a very good approximation for backup routereliability since the ACC (absolute correlation coefficient)value is 0.968 from our numerical results. Although (4) wasobtained from the special model V =V =0, we shall see in oursimulation experiments later that this empirical equation isstill valid for more general mobility scenarios.

3.3 Backup Source Routing: An Overview

Wepropose a distributed routing protocol, called the BackupSource Routing (BSR). It extends DSR by selecting a backuppath to achieve the most reliable routes between eachcommunicating mobile node. The backup path informationis piggybacked with the primary path in the header of datapackets, which can be useful in case the primary paths break.The backup route can help by minimizing route recoveryprocess and control message overhead. Like DSR, the set ofroutes are discovered on demand in BSR. BSR consists of twophases: 1) Route Discovery and 2) RouteMaintenance. RouteDiscovery is only invoked when needed, and Route Main-tenance operates onlywhen the route is used actively to sendindividual packets.

In the route discovery phase, BSRmodifies the forwardingmechanism of Route Request (RREQ) messages used in DSRprotocol. We do this because in the route discovery phase ofDSR, a large number of duplicate RREQmessages are usuallydropped,making itdifficult to establishbackuproute for eachintermediate node. BSR also employs a different mechanismto expire stale routes, which will benefit the routingperformance in higher mobility situations.

In the route maintenance phase, when the existingprimary path is broken due to the node mobility, insteadof simply sending an error message back to the source inDSR, BSR uses the backup path immediately to prolong theconnection duration between the source and the target.Only if both the primary path and backup path fail, an errormessage will be sent back to the source to initiate a newroute discovery. Since the topology is continuously chan-

ging, BSR will also dynamically repair backup routes,which otherwise become less optimal or even invalid.

4 ON-DEMAND ROUTE DISCOVERY

BSR constructs the backup route using request-reply cycles.When the source needs a route to the destination but noroute information is known, it propagates the ROUTEREQUEST (RREQ) messages to the network. Severalduplicates may traverse different routes to reach anintermediate node. From this information, a node wouldselect the primary path and the backup path, and recordthem in its routing cache. Finally, the destination nodeconstructs its backup route and sends ROUTE REPLY(RREP) messages back to the source via the chosen routes.

4.1 Data Structure

A simplified data structure for RREQ control message canbe ½�; s; t; RID; ttl�, where � ¼ ðs; . . . ; pÞ is a source routefrom the source node s forwarded by node p, t is the targetnode, RID is the request ID, and ttl is the time-to-live of themessage. The value of ttl may be used to limit the numberof intermediate nodes to forward that copy of RREQ. As theRREQ is forwarded, this value is decremented, and theRREQ message is discarded if ttl reaches zero beforefinding the target. Each node i would update its route cacheentry ð�iðnÞ; �0

iðnÞÞ once it receives a RREQ message. Here,�iðnÞ and �0

iðnÞ denote the primary path and backup path tonode n in the route cache of node i, respectively. Node i alsorecords the most recently received request ID, RIDiðnÞ,from the RREQ message that has the destination of node n.

4.2 Forwarding the RREQ

The main goal of BSR in the route discovery phase is toequip each intermediate node in the backup path selectionwith the most reliable (lowest cost) backup route. Astraightforward approach is to take advantages of the BSRflooding technique: Instead of dropping every duplicateRREQs, each intermediate node would forward packets toall its one-hop neighbors by appending its own address tothe source route in the RREQ only if all of the followingconditions (denoted by “Cx”) are met:

(C1): The node is not the destination node of the RREQmessage.

(C2): The node is not listed in the source route (in order toavoid loop).

(C3): The time-to-live value is greater than zero.

In this way, every possible path from the source willarrive at the destination, and the most reliable backup routeis therefore constructed by choosing the shortest-delay(primary) path and a backup path �0 from all thecandidates such that Cð�; �0Þ is minimized. Recall thatrouting overhead is one of the major concerns in MANETrouting protocols. Therefore, we shall not use floodingmechanism indiscriminately, or there would be potentiallyOðn!Þ RREQ packets sent in a network with n nodes. Wepropose a local search of a small number of backup paths,instead of making an expensive exhaustive search of thepaths obtained from BSR-flooding. Each intermediate nodewould make an “intelligent” hop-by-hop path selection

GUO ET AL.: IMPROVING SOURCE ROUTING RELIABILITY IN MOBILE AD HOC NETWORKS 365

Fig. 2. Backup route reliability approximation.

based on its cost function to guide the search along (or closeto) the best candidate backup paths. To achieve this goal inan on-demand routing scheme, we have designed adistributed BSR-LCS (Lower Cost Search) algorithm tosearch a backup route from a source node s to a destinationnode t. Compared to BSR-flooding, our BSR-LCS algorithmrequires an additional condition.

(C4): Intermediate nodes are allowed to forward RREQsonly if the path in the duplicate packet can produce anew backup route whose cost function has a lower valuethan that of its backup route found previously.

The algorithm description in pseudocode is given below.It will be invoked once any node i receives a RREQ message½ðs; . . . ; pÞ; s; t; RID; ttl�.

BSR-LCS Algorithm

Step 1) If it is a new RREQ message, i.e., RID > RIDiðsÞ,then update the corresponding route cache as

�iðsÞ ¼ ðs; . . . ; p; iÞ and �0iðsÞ ¼ null, set

RIDiðsÞ ¼ RID, and go to Step 3).

Step 2) Otherwise, if the duplicated RREQ message cangenerate a more reliable backup route, i.e.,

ð�iðsÞ; ðs; . . . ; p; iÞÞ is a backup route and

Cð�iðsÞ; ðs; . . . ; p; iÞÞ is less than Cð�iðsÞ; �0iðsÞÞ, then

update the backup path in the corresponding route

cache �0iðsÞ ¼ ðs; . . . ; p; iÞ and go to Step 3), else set

ttl ¼ 0.

Step 3) Check if ttl ¼ 0, if so, drop this RREQ message and

terminate the algorithm.Step 4) Otherwise, if the node receiving the RREQ message

is not the destination, local broadcast RREQ

½ðs; . . . ; p; iÞ; s; t; RID; ttl� 1�, else unicast RREP back

to source node s.

Since the forwarding condition (C4) only limits a smallnumber of backup paths to arrive at the destination, ouralgorithm does not guarantee the most reliable backuproute for the route request. The way of forwardingduplicate packets at each intermediate node, however,helps to decrease the value of the backup route costfunction so that the resulting backup route equipped foreach node will be eventually close to optimality. We shallobserve from our simulation experiments in Section 6.2 thatour RREQ forwarding algorithm not only reduces thecontrol message overhead significantly (i.e., the totalnumber of RREQs sent in an n-node network drops fromOðn!Þ to OðnÞ), but it also equips each intermediate nodewith a near optimal backup route.

As the network operations progress, each node in thenetwork would have learned some routing information inits cache sooner or later. This knowledge would benefit theroute discovery by reducing the latency of RREP anddecreasing the overhead, both achieved by preventing moreRREQs. Different from the “Reply from Cache” operation inDSR, an intermediate node may construct a backup route(not a single path) from the source to the destination byconcatenating the route from itself to the source, and theroute from itself to the destination using information in itsroute cache. We note that such a node must verify, beforetransmitting an RREP back to the source, that the resulting

path pair is a backup route (thus, satisfying Definition 1).Otherwise, it must not reply to such an RREP from its cache.

Finally, it remains to show that the backup routesdiscovered by BSR-LCS algorithm are loop-free. This isstipulated in Theorem 2 as follows.

Theorem 2. Alternate routes adopted from the BSR RouteDiscovery Phase are loop free.

Proof. From Step 2) in the RREQ forwarding algorithm inSection 4.2, we can conclude that any nonempty routecache entry obtained from BSR Route Discovery Phasecontains a backup route. Therefore, we only need to provea backup route ð�; �0Þ is loop free.We assume there exists aloop C � Eð�Þ [ Eð�0Þ. Without loss of generality, let C ¼ðu1; u2; . . . ; uk; ukþ1 ¼ u1Þas shown inFig. 3, andCu1u2 � �;

Cu2u3 � �0; . . . ; Cuk�1uk � �;Cuku1 � �0. From Definition 1,r�0 ðu2Þ < r�0 ðu3Þ ) r�ðu2Þ < r�ðu3Þ and r�ðu1Þ < r�ðu2Þ,we have r�ðu1Þ < r�ðu3Þ. In the same way, we derive that

r�ðu1Þ < r�ðu2Þ < r�ðu3Þ < . . . < r�ðukÞ < r�ðukþ1Þ ¼ r�ðu1Þ;

which is definitely false. Therefore, it is loop free for theroutes discovered by BSR. tu

4.3 Backup Route Selection

In our algorithm, each intermediate node i selects twopaths: a backup path and a shortest-delay primary path.The shortest delay path is the path taken by the first RREQto reach a node. We use the shortest delay path to minimizethe route acquisition latency required by on-demandrouting protocols. Upon receiving this first RREQ, theintermediate node records the entire path as its primarypath. Then, from all subsequent duplicate packets (with thesame request ID) received, each intermediate node selectsthe path that gives the minimal value of the cost functiondefined in (3), and records this path as its backup path.When the target node receives an RREQ, it should wait for ashort period in order to receive more subsequent duplicatepackets, so that it can construct the backup route beforereturning the accumulated route record to the source in anRREP. Note that the above operation assumes a bidirec-tional wireless environment.2 If this is not available, thetarget node may have to go through a second RouteDiscovery and, therefore, producing more overheads. Weuse i ! j to denote that node i has a route cache entry tonode j, and use i $ j if both i ! j and j ! i are satisfied.For each intermediate node i, let � be �iðsÞ and �0 be �0

iðsÞ.Under the symmetric links assumption, the route cacheentries (of both primary path � and backup path �0) for use

366 IEEE TRANSACTIONS ON PARALLEL AND DISTRIBUTED SYSTEMS, VOL. 16, NO. 4, APRIL 2005

2. MAC protocols such as IEEE 802.11 usually require a bidirectionalframe exchange as a part of the MAC protocol.

Fig. 3. Illustration for proof of Theorem 2.

in sending subsequent packets can be set according to thefollowing rules.

Rule 1. If i 6¼ t, then i ! j; j 2 V ð�Þ \ V ð�0Þ, after forward-ing the RREQ message. The route cache of node i for theentry j is set as �iðjÞ ¼ �ij and �0

iðjÞ ¼ �0ij.

Rule 2. If i ¼ t, then j $ k; j; k 2 V ð�Þ \ V ð�0Þ, after forward-

ing the RREQ message and executing the route selection

method. Without loss of generality, let r�ðjÞ < r�ðkÞ. Theroute cache of node j for the entry k is set as �jðkÞ ¼ �jk and

�0jðkÞ ¼ �0jk. In this case, �kðjÞ ¼ ð�jðkÞÞR and �0

kðjÞ ¼ ð�0jðkÞÞR, where ð:ÞR is a reverse route of ð:Þ.After the Route Discovery phase, data packets can now

be delivered using the union of the primary path andbackup path. The primary path should always be the firstchoice, and the backup path is used only when the existingprimary path breaks. Fig. 4 is an example explaining howthe route selection algorithm uses Rules 1 and 2 to buildbackup routes (columns �sðtÞ and �0

sðtÞ of Table 1) for s-tpairs (columns 1 and 2 in Table 1) in the network. Table 1also shows other information such as the primary path�sðtÞ, backup path �0

sðtÞ and then heuristic cost C in therouting cache during the algorithm.

More generally, when a node forwards a data packet, itmay cache the backup route in the packet using Rule 3below. All these would allow a much faster reaction torouting changes because a node with backup routes to adestination can try another cached route right away as soonas the one it has been using fails. Let � ¼ �sðtÞ and�0 ¼ �0

sðtÞ.Rule 3. When a node i is forwarding a data packet with a

backup route ð�sðtÞ; �0sðtÞÞ, and if i 2 V ð�Þ \ V ð�0Þ, then

i ! j; j 2 V ð�Þ \ V ð�0Þ. The route cache of node i for the

entry j is set to �iðjÞ ¼ �ij and �0iðjÞ ¼ �0ij if r�ðiÞ < r�ðjÞ, or

�iðjÞ ¼ ð�jiÞR and �0iðjÞ ¼ ð�0jiÞ

R otherwise.

4.4 Rerouting and Adaptive Cache TimeoutMechanism

Rerouting is triggered at a source node when the backuproute to the destination fails and no route information isknown in its cache. It can also be done periodically after thetimeout expired for the cached backup route in order tobenefit the overall performance by excluding the unfortu-nate use of stale routing information. It is known that routecaching times affect the performance of routing protocols intwo ways: 1) caching too long causes bad routes to beselected, but to be invalidated later, and 2) caching too shortcauses good routes to be purged from the cache, but to berequired shortly afterward. Both cases result in extralatency since a new discovery has to be attempted. Unlike

the traditional way of the cache operation “Link-MaxLife”used by DSR, we have applied an adaptive timeoutmechanism by utilizing (3) and (4) to predict the timewhen cached information becomes stale. For an adaptivetimeout scheme, this timeout value at time t is determineddynamically by the caching node according to the pastbehavior of backup route reliability observed back to thetime t� ", where " is a constant. The value of timeout for abackup route at time t can be approximated by thereliability T�;�0 of backup route ð�; �0Þ using (5) below.

T�;�0 ¼ �t=Cð�; �0Þ þ �t; ð5Þ

�t ¼n � �n

i¼1ti=ci � �ni¼11=ci � �n

i¼1ti

n � �ni¼1ð1=ciÞ

2 � ð�ni¼11=ciÞ

2; ð6Þ

�t ¼1

n�n

i¼1ti ��t

n�n

i¼11=ci: ð7Þ

Here, �t and �t are the adaptive parameters given in (6) and(7), which are computed online as time evolves. Thereliability ti of any backup route ð�; �0Þ is obtained byaveraging the duration from the time when the backuproute ð�; �0Þ was found by the BSR-LCS algorithm to themoment when the source node received an error messageindicating that the backup route was disconnected, or whenthe cache timer expired. Its corresponding cost Cð�; �0Þ isdenoted by ci. Variable n is the total number of pairs ðti; ciÞmeasured in the duration ½t� "; t�.

5 DYNAMIC ROUTE MAINTEMANCE

5.1 Repairing a Broken Backup Route

A link of a route can be disconnected because of mobility,congestion, or packet collision. It is important to recoverbroken routes immediately to maintain effective routing. Inour BSR algorithm, when a node fails to deliver the datapackets to thenext hopusing its primarypath, it tries touse itsbackuppath (piggybacked in its header) to deliver the data tothe destination. For example, if the link ðS;AÞ in Fig. 4 isbroken, the packet with the backup path ðS;B;C;D; T Þ willfind an alternate route ðS;B;CÞ for transmission. When thepacket is delivered to node C, it will be forwarded along theprimary path ðC;E; T Þ again if it is available. Sometimes,packets have to be delivered in a backtracking fashion. Forexample, in order to use backup path after the link ðE; T Þwasbroken in Fig. 4, the packets need to be sent back to C beforeusing ðC;D; T Þ.

In contrast to the salvagingoperationused inDSR, ourBSRuses the “backup route salvaging” procedure when an inter-

GUO ET AL.: IMPROVING SOURCE ROUTING RELIABILITY IN MOBILE AD HOC NETWORKS 367

Fig. 4. Example to illustrate the Route Discovery phase of Backup

Source Routing Protocol. (The directed links in bold face represent the

shortest delay tree from S.)

TABLE 1The Routing Cache Constructed after RREQ Propagation andRoute Selection of the Route Discovery Phase in All the Nodes

mediate node forwarding a packet discovers that the target isunreachable either through its primary path or its backuppath. It examines its route cache for other routes to the samedestination. If a route exists and the node can reconstruct avalid backup route with the routes from its cache, the nodereplaces the brokenbackup route, and retransmits thepacket.Otherwise, the node drops the packet. In the former case, thenode attempting to perform salvaging returns an RERR(Route Error) to the source of the data packet with thereconstructed backup route, but for the latter case, it simplyreturns an RERR to the source. Note that the node sending anRERR message in the upstream direction of the route is stillusing the reverse backup route information contained in thedata packets in order to reduce theprobability of losingRERRmessages. The RERR message contains the backup route tothe source node, and the immediate upstream and down-stream nodes of the broken link. Upon receiving these RERRmessages, the source node reconstructs the backup route tothe target node using the updated routing information. Ifthere is not enough information, it initiates a new routerecovery process.

5.2 Repairing a Nonoptimal Backup Route

Due to node mobility, the backup route in the packet maynot always be optimal. When a node overhears a packet thatis not addressed to it, the node checks if the packet wasforwarded using the primary path and if the packet’sheader contains the address of this overhearing node in theunprocessed portion of the primary path. If so, the nodesends a Gratuitous RREP message to the source, giving theshorter path as the concatenation of the portion of theoriginal primary path up through the node that transmittedthe overheard packet, plus the suffix of the original primarypath beginning with the node that returns the gratuitousRREP. The backup path is not overheard since the mainrouting metric is total delay, but the backup path wouldonly be in use when the primary path failed. It is easy toverify that the shortened primary path and its originalbackup path still form a backup route. Then, upon receivingthis RREP, the source will reconstruct its route cache withthe new backup route.

5.3 Repairing Cache

When a source node receives an RERR, and decides toinitiate a new route discovery (the original backup routefailed), it propagates its next RREQ and piggybacks thereceived RERR on it. In this way, stale information in thecaches of nodes around this source node will not generateRREPs that contain the same invalid link for which thissource received the RERR. When intermediate nodes

working in a promiscuous mode overheard a normal orpiggybacked RERR message, they should delete broken linkfrom its route cache and reconstruct their backup routes.

6 PERFORMANCE EVALUATION

In our evaluation, we use a detailed simulation model basedon ns-2 [21]. A mobile ad hoc network in a physical area ofsize 1,000 meters � 1,000 meters was simulated. The radiointerface is modeled as a shared-media radio with anominal bit rate of 2 Mb/s and a nominal radio range of250 meters. The IEEE 802.11 Distributed CoordinationFunction (DCF) is used as the medium access controlprotocol. The 802.11 DCF uses Request-To-Send (RTS) andClear-To-Send (CTS) control packets for “unicast” datatransmission to a neighbor node. Data packet transmissionis followed by an ACK. “Broadcast” data packets and theRTS control packets are sent using physical carrier sensing.An unslotted carrier sense multiple access (CSMA) techni-que with collision avoidance (CSMA/CA) is used totransmit these packets. A detailed description of thesimulation environment and the models is available in [21].

In the implementations of BSR and DSR, the RREQ andRERR packets are treated as broadcast packets in the MAC.The RREP and data packets are all unicast packets with aspecified neighbor as the MAC destination. Both protocolsdetect link breakages using feedback from the MAC layer.A signal is sent to the routing layer when the MAC layerfails to deliver a unicast packet to the next hop. This isindicated, for example, by the failure to receive CTS (Clear-To-Send) after a specified number of RTS (Request-To-Send) retransmissions, or the absence of an ACK followingdata transmission.

6.1 Backup Route Lifetime Statistics underDifferent Mobility Scenarios

Recall that our cost function is derived from one singlemobility scenario V =V =0, whichmay not be able to provide acomplete picture of backup route reliability in real world. Inthis section,we shall use simulation to examine the validity ofour analytical model under a wide range of mobilityscenarios, like 0þ=Vmax=0, V =V =0þ, and 0þ=Vmax=0

þ. Thiswould also provide insight to the statistic property of ourrouting metric of backup route reliability.

Fig. 5 demonstrates the relationship between E½T�;�0 � and1=Cð�; �0Þ in a mobile ad-hoc network with 50 nodes andunder ageneralmobilitymodelVmax=Vmin=Tpause that hasbeendiscussed in Section 2.2. For each set of experiments (i.e., for aspecifiedminimumspeedVmin, amaximumspeedVmax, and a

368 IEEE TRANSACTIONS ON PARALLEL AND DISTRIBUTED SYSTEMS, VOL. 16, NO. 4, APRIL 2005

Fig. 5. The approximate linear relationship of E½T�;�0 � and 1=Cð�; �0Þ in a mobile ad hoc network with 50 nodes and under different mobility scenarios.

(a) V =V =0, (b) 0þ=Vmax=0, (c) V =V =0þ, and (d) 0þ=Vmax=0þ.

pause time Tpause), we have an initial 500 seconds ofwarm-upperiod. In order to collect as many samples of backup routelifetime as possible, each node in the networkwould initiate aroute discovery to equip every other node with a backuproute. We then measure the lifetime for each backup routediscovered until all the backup routes are broken. We repeatthis process 30 times in total. Each point in Fig. 5 is a specificbackup route found by our BSR route discovery algorithm,and the corresponding reliability is the average lifetime overall the backup routes that are discovered under the sametopology in the whole simulation period. The simulationresults verify once again the fact that node mobility is asignificant factor (as shown in Figs. 5a, 5b, and 5d), whilepause time is not (as shown in Fig. 5c, and from thecomparison of Fig. 5b with Fig. 5d) [19]. Since we have avery good fit (ACC > 0:95) for eachmobility scenario, our (3)and (4) can provide a valid approximation of our routingmetric.

6.2 BSR-LCS versus BSR-Flooding

We have carried out experiments to compare two backuproute discovery algorithms BSR-LCS and BSR-flooding overdifferent connected network topologies with 10 to 50 nodes.Two important performance metrics are considered: 1) aver-age total number of RREQmessages per route discovery, and2) the average backup route cost over all backup routes. Thereare in total 50 route discoveries initiated sequentially in thesimulation. The mean number of RREQs is therefore

averaged over all the route discoveries, and the average cost

is obtained over all backup routes discovered.We have evaluated the route discovery overhead under

different TTL values of 5, 6, and 7. For each experiment setup

(i.e., for a specified network size, a TTLvalue), Figs. 6a and 6b

depict the total number of RREQ messages per route

discovery propagated through all over the network under

BSR-LCS (in linear scale) andBSR-flooding (in semilog scale),

respectively. The experiment results in Fig. 6a show that the

REEQ overhead of our BSR-LCS algorithm is increasing in an

approximately linear manner with respect to the number of

nodes in the network. Compared to the performance of BSR-

flooding as shown in Fig. 6b, it has saved bandwidth

dramatically. We attribute this advantage to its local search

in a small number of backup paths, instead of making an

expensive exhaustive search in BSR-flooding.We would also like to verify the efficiency (how close to

the optimal cost) of the “intelligent” hop-by-hop path

selection used in our BSR-LCS algorithm. We have

evaluated the average cost under different network size

while fixing the TTL at 6. The simulation results are shown

in Fig. 7, where the darker bar on the top indicates the

difference using BSR-LCS to the optimum. We observe that

BSR-LCS performs within 8 percent and 4 percent close to

the optimum (obtained by BSR-flooding), in 20-node net-

works and 50-node networks, respectively.

GUO ET AL.: IMPROVING SOURCE ROUTING RELIABILITY IN MOBILE AD HOC NETWORKS 369

Fig. 6. Number of RREQ messages in route discovery phase. (a) BSR-LCS (in linear scale) and (b) BSR-flooding (in semilog scale).

Fig. 7. Average backup route cost in route discovery phase. (a) Twenty-node networks and (b) fifty-node networks.

6.3 Performance of the BSR Protocol

There are two main distinctive features of our BSR protocolcompared to the DSR protocol: use of backup route insteadof a single path and an adaptive route caching mechanism.In order to evaluate the performance improvements madeby these two mechanisms separately, we have comparedthe simulation results of the following protocols:

1. BSR-BR-Caching: a full version of BSR protocol thatuses both backup route and our cache timeoutmechanism described in Section 4.4,

2. BSR-BR-Noncaching: a revised BSR protocol thatonly uses backup route,

3. BSR-Caching: a revised BSR protocol that only usesthe cache timeout mechanism,

4. DSR-Caching: a full version of BSR protocol thatuses single path and the caching mechanism “Link-MaxLife,” and

5. DSR-Noncaching: a revised BSR protocol that dis-ables the caching mechanism.

The following experiments were performed in a mobilead hoc network with 50 mobile nodes under differentmobility and a fixed traffic load with 25 data sessions. Themobility model uses 0þ=Vmax=0 with different maximumspeeds (1, 5, 10, 15, 20, and 25 m/s) and a minimum speedof 1 m/s. Constant bit rate traffic sources are used. Thesource-destination pairs are chosen randomly over thenetwork, each with a traffic rate of four packets /sec. Only512-bytes data packets are used. Simulations are run for1,400 simulated seconds (including the initial 500-secondwarm-up period) for each set of experiments.

Four key performance metrics are evaluated:

1. Average reliability (route lifetime): the average intervalbetween two consecutive route discoveries in a nodeto the same target over all the sessions,

2. Packet delivery ratio: the ratio of data packetsdelivered to the destination to data packets origi-nated by the sources,

3. Average end-to-end delay of data packets: the intervalfrom the time a packet departs from the source untilthe time it is received in its entirety at thedestination, and

4. Control message overhead ratio: the ratio of controlpackets transmitted to data packets delivered.

In order to facilitate the comparison of different protocolsand to illustrate the performance improvements, for eachperformance metric, we also provide its normalized value,defined as the ratio of actual value using one protocol to thebest performance among all five protocols.

Fig. 8a shows the average lifetime of discovered routesunder different node mobility levels. We observe that theprotocols using backup route (the BSR-BR-Caching andBSR-BR-Noncaching algorithms) perform much betterthan those only using single path (the BSR-Caching,DSR-Caching, and DSR-Noncaching algorithms) for allmobility scenarios. We therefore conclude that thestrategy of using backup route can always maintain alonger valid lifetime of discovered primary and backuppaths than a single path.

To investigate the effects of caching mechanism, we havecompared the normalized average lifetime among differentprotocols. These are the cases of BSR-BR-CachingversusBSR-BR-Noncaching, and DSR-Caching versus DSR-Noncaching.The significant nonlinearity in Fig. 8b shows that the cachingmechanism helps to prolong lifetime, especially in highmobility, where the cached information may become staleshortly. We attribute the improvement to the fact that itavoids extra latencywhen bad routes are selected from cacheand then found invalid. We also note that DSR-Cachingperforms a little better than BSR-Caching. This is becauseBSR-Caching has a caching mechanism that was initiallydesigned for backup route routing, but it only uses primarypath. Thus, the lifetime cannot be estimated properly basedon a backup route cost.

Fig. 9 shows how the packet delivery ratio changes as theinstantaneous average node speed varies. Under allmobility scenarios, BSR-BR-Caching outperforms DSR-Caching, and this improvement is more noticeable in ahigher mobility rate as shown in Fig. 9b. As the mobilityincreases, the performance gained by backup routesbecomes more significant. Because BSR-BR-Caching at-tempts to use the backup path for data delivery in thepresence of primary path breaks, this protocol is able todeliver more packets to the destination than DSR-Caching.When both the primary path and the backup path break,BSR-BR-Caching can send the RERR messages more reliablyback to the source to initiate a new route discovery timely inorder to prevent further packet loss. Similar observationscan be made between protocols BSR-BR-Noncaching andDSR-Noncaching as shown in Fig. 9b. At the same time, wenote that caching mechanism also affects performance, by

370 IEEE TRANSACTIONS ON PARALLEL AND DISTRIBUTED SYSTEMS, VOL. 16, NO. 4, APRIL 2005

Fig. 8. (a) Average reliability (in second) and (b) normalized average reliability as a function of mobility rate in 50-node networks.

comparing BSR-BR-Caching and BSR-BR-Noncaching, butits effect is not as significant as the effect of backup routemechanism.

Average end-to-end delays are compared in Fig. 10. Inlower mobility, protocols using backup routes like BSR-BR-Caching and BSR-BR-Noncaching, have slightly longerdelay than those only using single path including BSR-Caching, DSR-Caching, and DSR-Noncaching, because thedelays can only be measured for those data packets thatsurvived to reach their destination, and the backup routedelivers more packets, which probably have taken longerbackup paths than primary paths. The longer delays do notrepresent their ineffectiveness since they use the sameprimary route. In higher mobility scenarios, the protocolsusing only single paths (like BSR-Caching, DSR-Caching,and DSR-Noncaching) yield a much longer delay inreconstructing routes due to more frequent route discov-eries initiated, and thus the period of time to buffer datapackets during the route recovery at source results in largerend-to-end delay.

Fig. 11 compares the performance of control messageoverhead ratio. We can see that the efficiency of allprotocols decreases because of the increase in link breakagewhen the mobile nodes are moving faster. As expected, DSRalgorithms (DSR-Caching or DSR-Noncaching) outperformBSR algorithms (BSR-BR-Caching, BSR-Caching, or BSR-BR-Noncaching) in a more stable situation, because BSR has todistribute more routing messages to collect backup routes

information for every node. However, this is only a slightimprovement. On the other hand, as mobility is increased,backup route protocols (BDR-BR-Caching and BSR-BR-Noncaching) significantly outperform single path protocols,because the latter have no backup path to use when theprimary path disconnections occur and, therefore, have toinvoke more route reconstructions.

We summarize the key conclusions of this section as

follows:

1. When considering only the effect of cachingmechanism, we have shown that protocols BSR-Caching and DSR-Caching perform very similarsince both of them use single path and a dynamiccache timeout technique.

2. When considering only the effect of using backuproutes, our comparisons between BSR-BR-Noncach-ing and DSR-Noncaching show that except in thesituations with lower mobility, where BSR-BR-Non-caching has a bit longer average end-to-end delay andhigher control message overhead ratio, BSR-BR-Noncaching provides much better performance thanDSR-Noncaching in all other cases. This corroboratesour original intention to utilize backup route that isbeneficial for a long continuous application data flowin mobile wireless networks where routes are dis-connected frequently.

GUO ET AL.: IMPROVING SOURCE ROUTING RELIABILITY IN MOBILE AD HOC NETWORKS 371

Fig. 9. (a) Packet delivery ratio and (b) normalized packet delivery ratio as a function of mobility rate in 50-node networks.

Fig. 10. (a) Average end-to-end delay and (b) normalized average end-to-end delay as a function of mobility rate in 50-node networks.

3. Both backup route and cache timeout mechanismsaffect the performance of routing protocols. This issupported through the comparison of protocols BSR-BR-Caching, BSR-BR-Noncaching, and BSR-Caching.Moreover, the backup route mechanism seems to bea more significant factor, since BSR-BR-Noncachingperforms better than BSR-Caching in most cases. Thecache timeout mechanism has a notable performanceimprovement in the situations with a high level ofmobility, while only minor effects in more stablesituations.

4. As an overall comparison after evaluating eachindividual factor, we observe that in less stressfulsituations of lower mobility, BSR has similarperformance as DSR. On the other hand in morechallenging situations of high mobility, BSR canimprove the performance significantly. These resultsprovide the insights to the relationship betweensalvaging and backup route, where the former is amain mechanism of DSR to improve reliability ofdata transmission, while the latter one is used in BSRwith a similar purpose but in an aggressive way tokeep more information.

7 CONCLUSION

We have proposed a new on-demand routing protocol BSRthat can explore the most reliable backup routes, which canbe very useful in case of a breakage in the primary path. Wehave also provided a framework for modeling the timeinterval between successive route discoveries for on-demand protocols based on a simple assumption on thelifetime of a single wireless link. Evaluation of the backupsource routing under this framework demonstrates thatthere are definite advantages to be gained from providingbackup routes. In order to use the idea practically, wepresented an approximation method to measure the metricby a heuristic cost function, which has been proved to bevalid via simulation under different mobility scenarios.Based on the cost function, we have designed variousalgorithms for our BSR protocol in the route discoveryphase and the route maintenance phase.

The effectiveness of our BSR protocol was also validatedusing the simulation experiments. Our study indicates thatBSR outperforms DSR because backup routes provide

robustness to mobility. BSR significantly reduces the

number of route discoveries from the use of backup routes

and the cache timeout mechanism. BSR has also showed

considerably fewer packet drops, shorter end-to-end delay,

and better efficiency compared with DSR.

ACKNOWLEDGMENTS

The authors wish to thank the anonymous referees for

several constructive suggestions that have greatly improved

the quality of this paper. This work was supported in part

by the Ontario Graduate Scholarship Program of Canada

(OGS), the Natural Sciences and Engineering Research

Council of Canada (NSERC) under grant No. OGP0042878,

and the National Natural Science Foundation of China

(NSFC) under grant No. 90104015.

REFERENCES

[1] E.M. Royer and C.-K. Toh, “A Review of Current RoutingProtocols for Ad Hoc Mobile Wireless Networks,” IEEE PersonalComm., pp. 46-55, Apr. 1999.

[2] C.K. Toh, “Long-Lived Ad-Hoc Routing Based on the Concept ofAssociativity,” IETF Draft, 1999.

[3] R. Dube, C.D. Rais, K.-Y. Wang, and S.K Tripathi, “Signal StabilityBased Adaptive Routing (SSA) for Ad-Hoc Mobile Networks,”IEEE Personal Comm., pp. 36-45, Feb. 1997.

[4] S. Agarwal, A. Ahija, J.P. Singh, and R. Shorey, “Route LifetimeAssessment Based Routing (RABR) Protocol for Mobile Ad-HocNetworks,” Proc. IEEE Int’l Conf. Comm., pp. 1697-1701, 2000.

[5] D. He, J. Shengming, and R. Jianqiang, “A Link AvailabilityPrediction Model for Wireless Ad Hoc Networks,” Proc. Int’lWorkshop Wireless Networks and Mobile Computing, Apr. 2000.

[6] S. Jiang, D. He, and J. Rao, “A Prediction-Based Link AvailabilityEstimation for Mobile Ad Hoc Networks,” Proc. IEEE INFOCOM,pp. 1745-752, Apr. 2001.

[7] W. Su, S. Lee, and M. Gerla, “Mobility Prediction and Routing inAd Hoc Wireless Networks,” Int’l J. Network Management, vol. 11,pp. 3-30, 2001.

[8] A. Nasipuri, R. Castaneda, and S.R. Dan, “Performance ofMultipath Routing for On-Demand Protocols in Mobile Ad HocNetworks,” Proc. ACM MONET, pp. 339-349, 2000.

[9] S.J. Lee and M. Gerla, “AODV-BR: Backup Routing in Ad HocWireless Networks,” Proc. IEEE Wireless Comm. and NetworkingConf., pp. 1311-1316, 2000.

[10] P. Zygmunt, J. Haas, and E.G. Sirer, “Path Set Selection in MobileAd Hoc Networks,” Proc. ACM MobiHoc, pp. 1-11, 2002.

[11] S. Guo and O. Yang, “Performance of Backup Source Routing inWireless Ad Hoc Networks,” Proc. IEEE Wireless Comm. andNetworking Conf., pp. 440-444, Apr. 2002.

372 IEEE TRANSACTIONS ON PARALLEL AND DISTRIBUTED SYSTEMS, VOL. 16, NO. 4, APRIL 2005

Fig. 11. (a) Control message overhead ratio and (b) normalized control message overhead ratio as a function of mobility rate in 50-node networks.

[12] S. Guo and O. Yang, “Backup Source Routing in Wireless Ad HocNetworks,” Proc. Int’l Conf. Software, Telecomm., and ComputerNetworks (SoftCOM), pp. 295-302, Oct. 2001.

[13] L. Wang, Y. Shu, O. Yang et al., “Adaptive Multipath SourceRouting in Ad Hoc Networks,” Proc. IEEE Int’l Conf. Comm.,pp. 867-871, 2001.

[14] S.J. Lee and M. Gerla, “Split Multipath Routing with MaximallyDisjoint Path in Ad Hoc Networks,” Proc. IEEE Int’l Conf. Comm.,pp. 3201-3205, 2001.

[15] M.R. Pearlman et al., “On the Impact of Alternate Path Routing forLoad Balancing in Mobile Ad Hoc Networks,” Proc. ACMMobiCom, pp. 3-10, 2000.

[16] D.B. Johnson and D.A. Maltz, “Dynamic Source Routing in AdHoc Wireless Networks,” Mobile Computing, T. Imielinski andH. Korth, eds. Kluwer Academic, pp. 153-181, 1996.

[17] C. Perkins and E.M. Royer, “Ad-Hoc On Demand Distance VectorRouting,” Proc. IEEE Workshop Mobile Computing Systems andApplications, pp. 90-100, 1999.

[18] J. Yoon, M. Liu, and B. Noble, “Random Waypoint ConsideredHarmful,” Proc. IEEE INFOCOM, pp. 1312-1321, 2003.

[19] D.D. Perkins, H.D. Hughes, and C.B. Owen, “Factors Affecting thePerformance of Ad Hoc Networks,” Proc. IEEE Int’l Conf. Comm.,pp. 2048-2052, 2002.

[20] A. Nasipuri, R. Burleson, B. Hughes, and J. Roberts, “Performanceof a Hybrid Routing Protocol for Mobile Ad Hoc Networks,” Proc.IEEE Int’l Conf. Comm., pp. 296-302, 2001.

[21] K. Fall and K. Varadhan, ns notes and documentation, http://www-mash.cs.berkeley.edu/ns/, 1999.

Song Guo received the BS degree in computerscience from Huazhong University of Scienceand Technology, China, in 1995 and the MSdegree in electrical and computer engineeringfrom Beijing University of Posts and Telecom-munications, China, in 1998. Since 2001, hehas been a PhD student in the School ofInformation Technology and Engineering at theUniversity of Ottawa, Canada. His main re-search interests lie in mobile ad-hoc routing

protocols and algorithms, power-aware design and optimization forwireless ad hoc networks, and performance evaluation. He is a studentmember of the IEEE.

Oliver Yang received the PhD degree inelectrical engineering from the University ofWaterloo, Ontario, Canada, in 1988. He is aprofessor in the School of Information Technol-ogy and Engineering at University of Ottawa,Ontario, Canada. He has worked for NorthernTelecom Canada Ltd. and has done variousconsulting jobs. His research interests are inmodeling, analysis and performance evaluationof computer communication networks, their

protocols, services, and interconnection architectures. The CCNR Labunder his leadership has been working on various projects in the trafficcontrol, traffic characterization, switch architecture, and traffic engineer-ing issues in both wireless and photonic networks. This has beenpublished in more than 200 technical papers. Dr. Yang is also interestedin queuing theory, simulations, computational algorithms and theirapplications, such as reliability and traffic analysis. Dr. Yang is currentlythe editor of IEEE Communication Magazine. He is a senior member ofthe IEEE.

Yantai Shu received the Chinese equivalent tothe BS, MS, and PhD degrees in electronicsengineering from Tianjin University. He is aprofessor of computer science at Tianjin Uni-versity, China, serving as vice president of theuniversity from 1993-1997. His current interestsare focused on computer communication net-works, wireless networks, real-time systems,modeling, and simulation. From 1974 until1991, he was employed by the Institute of

Plasma Physics, Academia Sinica in research positions. He is a memberof the IEEE and the ACM. He has published more than 120 papers andcontributed to one book.

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

GUO ET AL.: IMPROVING SOURCE ROUTING RELIABILITY IN MOBILE AD HOC NETWORKS 373