Adaptive Routing for Intermittently ConnectedMobile Ad Hoc Networks

Mirco Musolesi, Stephen Hailes, Cecilia MascoloDepartment of Computer Science, University College London

Gower Street, London WC1E 6BT, United Kingdom{m.musolesi|s.hailes|c.mascolo}@cs.ucl.ac.uk

Abstract— The vast majority of mobile ad hoc net-working research makes a very large assumption: thatcommunication can only take place between nodes thatare simultaneously accessible within in the same connectedcloud (i.e., that communication issynchronous). In reality,this assumption is likely to be a poor one, particularly forsparsely or irregularly populated environments. Moreover,asynchronous communication such as email, which isby far the pre-eminent form of networked person-to-person communication, has a natural fit to such partially-connected environments, but has been relatively littleexplored in the context of mobile ad-hoc networking. Thisis perhaps unsurprising, given the complexities involved.Indeed, the few approaches that have been described todate are simplistic or heavyweight, relying on brute forcemethods in order to achieve message delivery.

In this paper, we present the Context-Aware Routing(CAR) algorithm. CAR is a novel approach to the provisionof asynchronous communication in partially-connectedmobile ad hoc networks, based on the intelligent placementof messages. We discuss the details of the algorithm,and then present simulation results demonstrating thatit is possible for nodes to exploit context information inmaking local decisions that lead to good delivery ratiosand latencies with small overheads.

I. I NTRODUCTION

Since the earliest days of email, asynchronous com-munication has been the pre-eminent form of person-to-person communication; in comparison, relatively littleInternet traffic is generated for synchronous personalcommunication, though the balance is expected to shifta little with the increasing deployment of VoIP. Thereasons for the success of the asynchronous paradigm areclear: asynchronous communication works even whenboth parties are not simultaneously available; it is lessintrusive than synchronous communication, since recip-ients can deal with messages at their convenience; and(most importantly for us) it is less sensitive to link failurethan synchronous traffic, simply because the traffic is

deemed to have higher tolerance to delay.Ad hoc networks represent the purest form of decen-

tralised systems and, therefore, they impose many chal-lenges to cooperative communication. As a consequence,much ad hoc network research has focused on theinvestigation of fundamental algorithms for routing [1]on which almost everything else relies. However, in orderto make the problem tractable, almost all research onrouting algorithms makes the oversimplistic assumptionthat it is only meaningful to attempt to exchange mes-sages within connected clouds of nodes, in other words,that all communication is synchronous in nature.

This assumption is overly constrained if one considersthat there is a strong requirement for communication thatis asynchronous in nature, as argued above. In such acase, the delay tolerant character of the traffic allowsuseful communication to still occur by using nodesmoving between disconnected groups of nodes (clouds)to transport messages from one cloud to another. Thus,it is perfectly possible that two nodes mayneverbe partof the same connected cloud and yet may still be ableto exchange delay tolerant information by making use ofpredicted mobility patterns as an indicator of which othernodes might make good carriers for this information.

In the absence of special information, the problem ofpredicting which nodes might make good carriers in adhoc networks is a very challenging one. Likely futuremobility patterns must be inferred from past mobilitypatterns, but this alone is inadequate; parameters such asremaining battery lifetime are also key in determiningwhich potential carriers are most likely to result insuccessful delivery. In this paper, we consider what typesof information are available to nodes in deciding on acarrier. We use this analysis in the design of a Context-aware Adaptive Routing algorithm (CAR), a generalframework for the evaluation and prediction of contextinformation, aimed at achieving efficient and timelydelivery of messages. Using simulations, we explore

the performance of the CAR algorithm with respectto epidemic routing [2] and flooding. Whilst, in thedeveloped world, synchronous communication (in theform of phone and Internet) is generally cheap andeasy to come by, there are several real scenarios in lessdeveloped parts of the world in which different portionsof a logical network are physically disconnected. Thus,for example, this is the case in recent projects establishedto assist nomadic communities such as the Saamis inLapland [3] or to assist populations in rural areas ofIndia [4]. In the latter case, a number of villages eachhave their own local networking infrastructure, but thereis no interconnection between them. A bus containing awireless node travels between villages, picking up emailin one and depositing it in others on its round. Self-evidently, although there is never a direct connectionbetween sender and recipient, mail can still be delivered.

This paper is organized as follows: in Section IIwe discuss the most relevant aspects of asynchronousrouting for mobile ad-hoc networks. Section III presentsour approach. The details related to the evaluation ofcontext information are discussed in Section IV. Thedescription of the simulations carried out to evaluateCAR is provided in Section V, together with an analysisof the results. In Section VI we compare CAR withprevious work in this area and Section VII concludesthe paper, outlining possible future research directions.

II. A SYNCHRONOUSCOMMUNICATION IN MOBILE

AD HOC NETWORKS

Synchronous protocols rely on the fact that a con-nected path exists between the sender and the receiverof a message; the absence of such a path will, at best,lead to a failure indication to the originating host. Ifdelivery is important, the best that can be done is for thesender to continue to poll for the receiver. However, asin the example of the bus delivering messages outlinedin Section I, it may be the case that sender and receiverare never in the same connected cloud, so this will notachieve the wanted results.

Only a small number of approaches have been pro-posed in the field of asynchronous communication for adhoc networks [2], [5]. As described above, the challengein producing an algorithm for delivering asynchronousmessages derives from the deceptively simple questionof determining the best carrier or carriers for eachmessage. Clearly, leaving the message with the senderis inappropriate, since sender and receiver may nevermeet.

An alternative, if inefficient, solution is to spread themessages to all hosts using a form of persistent flooding.In this approach, which is more properly known asepidemic routing[2], a host floods the message it wishesto send to all hosts within its connected cloud. Eachcarrier host buffers the message and if, as a result ofmovement, they come into contact with hosts that do nothave a copy, they transfer it to them, making them newcarriers, in an analogous was to the spread of disease.Eventually, the message will reach all nodes in thesystem, provided that movement patterns allow for this.Epidemic routing is a reasonable approach when thereis no information about the likely movement patterns ofnodes in the system. In other words, when there is nobasis on which to distinguish the movement pattern ofany node from another, and the movement pattern ofeach node is individually random, the only choice aboutmessage placement is to place messages randomly or toplace them everywhere, since there is no more intelligentbasis for making a decision.

The aim of our protocol is to allow nodes to makeintelligent local decisions about the choice of carriers formessages. These decisions are based on small amountsof information that are exchanged along with standardrouting tables, and they are effected by using predictiontechniques both to reduce the amount of informationneeds to be sent and to increase its utility. In the imme-diately following sections, we analyse the informationgathering, prediction and exchange mechanisms beforeproceeding to an analysis of the performance of theprotocol.

III. C ONTEXT-AWARE ROUTING FORMOBILE AD

HOC NETWORKS

The Context-Aware Routing algorithm is built onthe assumption that the only information a host hasabout its position is logical connectivity information.In particular, we assume that a host is not aware ofits absolute geographical location nor of the location ofthose to whom it might deliver the message. Althoughthis information could potentially be useful, and, indeed,we plan to examine its utility in the near future, it iscurrently unreasonable to assume the existence of GPSfor all potential application domains for this technology.Another basic assumption is that the hosts present in thesystem cooperate to deliver the message. In other words,we do not consider the case of hosts that may refuse todeliver a message or that act in a Byzantine manner.

The delivery process depends on whether or not therecipient is present in the same cloud as the sender. If

both are currently in the same connected portion of thenetwork, the message is delivered using the underlyingsynchronous routing protocol to determine a forwardingpath. In the remainder of this paper we assume that aproactive routing protocol is used (in our simulationswe employed DSDV [6]). Reactive protocols requiredifferent approaches to optimisation that would simplyconfuse the presentation and so are deemed to be outsidethe scope of this particular work.

If a message cannot be delivered synchronously1, thebest carriers for a message are those that have the highestchance of successful delivery, i.e., the highestdeliveryprobabilities. The message is sent to one or more of thesehosts using the underlying synchronous mechanism.

Delivery probabilities are synthesized locally fromcontext information such as the rate of change of connec-tivity of a host (i.e., the likelihood of it meeting otherhosts) and its current energy level (i.e., the likelihoodof it remaining alive to deliver the message). We definecontextas the set of attributes that describe the aspectsof the system that can be used to optimize the process ofmessage delivery. Since we assume a proactive routingprotocol, every host periodically sends both the infor-mation related to the underlying synchronous routing(in DSDV this is the routing tables with distances,next hop host identifier, etc.), and a list containing itsdelivery probabilities for the other hosts. When a hostreceives this information, it updates its routing tables.With respect to the table for asynchronous routing,each host maintains a list of entries, each of which isa tuple that includes the fields (destination, bestHost,deliveryProbability). In this paper, we choose to explorethe scenario in which each message is placed with only asingle carrier rather than with a set, with the consequencethat there is only a single list entry for each destination.

When a host is selected as a carrier and receives themessage, it inserts it into a buffer. The size of this bufferis important, and represents a trade-off between storageoverhead and likely performance. If the buffer overflows,messages will be lost from the system, since we assumethe existence of a single replica.

In order to understand the operation of the CARprotocol, consider the following scenario in which twogroups of nodes are connected as in Figure 1. Host

1It is worth noting that the recipient may be in the same cloudbut not reachable using synchronous routing, since the routinginformation is not available (for example because the space in therouting tables is not sufficient to store the information related to allthe hosts in the cloud or because the node has just joined the cloud).In these cases we exploit the asynchronous mechanisms.

H1 wishes to send a message toH8. This cannot bedone synchronously, because there is no connected pathbetween the two. Suppose the delivery probabilities forH8 are as shown in Figure 1. In this case, the hostpossessing the best delivery probability to hostH8 is H4.Consequently, the message is sent toH4, which storesit. After a certain period of time,H4 moves to the othercloud (as in Figure 2). Since a connected path betweenH4 and H8 now exists, the message is delivered to itsintended recipient. Using DSDV, for example, it is worthnoting thatH4 is able to send the message shortly afterjoining the cloud, since this is when it will receive therouting information relating toH8.

What we have described is the basic model behind theCAR protocol. In the following sections we will describethe details of the algorithms and techniques exploited forthe calculation of the delivery probabilities.

H1

H4 0.7

H3 0.5

H1 0.2

H2 0.3

H5

H6

H7

H8

Fig. 1. Two connected clouds, with associated delivery probabilitiesfor message transmission betweenH1 andH8

H1

H4

H3

H1

H2

H5

H6

H7

H8

Fig. 2. H4, carrying the message, joins the second cloud.

IV. PREDICTION AND EVALUATION OF CONTEXT

INFORMATION

The general problem from the point of view of thesender of a message is to find the host with the best deliv-ery probability, as calculated using the predicted valuesof a range of context attributes. Instead of using theavailable context information as it is, CAR is optimizedby using predicted future values for the context, so tohave more realistic values. The process of prediction and

evaluation of the context information can be summarizedas follows.

• Each host calculates its delivery probabilities. Thisprocess is based on theprediction of the futurevalues of the attributes describing the context (seeSection IV-B) and on thecomposition of theseestimated values using multi-attribute utility the-ory [7] (see Section IV-A). The calculated deliveryprobabilities are periodically sent to the other hostsin the connected cloud as part of the update ofrouting information.

• Each host maintains a logical forwarding tableof tuples describing the next logical hop, and itsassociated delivery probability, for all known desti-nations.

• Each host uses local prediction of delivery probabil-ities between updates of information. The predictionprocess is used during temporary disconnectionsand it is carried out until it is possible to guaranteea certain accuracy. Morever, in the case of hostswithin reach, the interval between update shipmentis based on the possibility or not to make accuratepredictions. In other words, hosts send updates onlywhen the evolution of the mobile scenario followsa certain trend. This is done by evaluating thesampled values of context information.

In the remainder of this section, we will analyze moreclosely how delivery probability information is predicted,spread in the system, maintained, and evaluated.

A. Local evaluation of context information

There are several techniques that can be used tocombine and evaluate the multiple dimensions of contextin order to decide which nodes are the best candidates forcarrying a particular message. The simplest is to allowapplication developers to define a static hierarchy amongthe predicted context attributes [8].

A possible alternative to this method is to use goal pro-gramming, exploiting the so-calledpreemptive method-ology. With respect to a single attribute, our goal is tomaximize its value. The optimization process is basedon the evaluation of one goal at a time such that theoptimum value of a higher priority goal is never degradedby a lower priority goal [9]. However, in general, thedefinition of static priorities is inflexible. For morerealistic situations, we expect to need to attempt simul-taneous maximization of a range of different attributes,as opposed to using a predefined hierarchy of priorities.

1) Significance-based evaluation of context-aware in-formation: The priority based technique just mentioned

seems too simplistic because, in general, our decisionproblem involves multiple conflicting objectives [7]. Forexample, if we wish to determine which host has thebest delivery probability, considering both the batteryenergy level and the rate of change of connectivity, itmay happen that the host characterized by the highestmobility has scarce residual battery energy and viceversa. In general, maximization across all parameterswill not be possible and, instead, we must trade off theachievement of one objective (i.e., the maximization ofa single attribute) against others.

The context information related to a certain host canbe defined using a set of attributes(X1, X2, ..., Xn).Those attributes denoted with a capital letter (e.g.,X1)refer to the set of all possible values for the attribute,whereas those denoted with a lower case letter (e.g.,x1) refer to a particular value within this set. In theremainder of this section we will use the classicalnotation of utility theory. Our goal is to allow each hostlocally to associate a utility functionU(x1, x2, ..., xn),representing the delivery probability, with every otherhost. We use the following definitions:

Definition 1: Given a set of attributesX1, X2, ..., Xn

partitioned into two complementary setsY =(X1, X2, ..., Xs) andZ = (Xs+1, Xs+2, ..., Xn), we saythaty’ is conditionally preferred or indifferent toy givenz if and only if

(y’ , z) � (y, z)

Definition 2: The set of attributesY is preferentiallyindependent of the complementary setZ if and only iffor somez’

[(y’ , z’) � (y, z’)] ⇒ [(y’ , z) � (y, z)],∀z, y, y’

To understand this definition, consider the case ofthree attributesX1, X2, X3: two attributesX1 and X2

are preferentially independent of a third attributeX3 ifthe preference between(x1, x2, x3) and(x′1, x

′2, x3) does

not depend on the particular valuex3 for the attributeX3.

Definition 3: The attributesX1, X2, ..., Xn are mutu-ally preferentially independent if every subsetY of theseattributes is preferentially independent of its complemen-tary set of attributes.

Given these definitions, an interesting result of the multi-attribute decision theory is the following theorem demon-strated by Debreu in 1960 [10].

Theorem 1:Given attributesX1, X2, ...Xn, an addi-tive function of the following form exists if and only if

the attributes are mutually preferentially independent

U(x1, x2, ..., xn) =n∑

i=1

Ui(xi)

whereUi is a utility function overXi.Thus, in the case of mutually preferentially indepen-

dent attributes, that is to say those characterized by thesame degree of significance, the sum of the attributesis adequate as a means of combining those attributes.However, the case of attributes that have different relativeimportance is more interesting. In this case, we use thetheory of goal programming, a branch of mathematicsthat has been studied since 1960 in the operation re-search community. More specifically, we use theweightsmethod in order to find the host that has the highestprobability of delivering the message.

Our aim is to maximize each attribute, in other words,to choose the host that presents the best trade-off be-tween the attributes representing the relevant aspectsof the system for the message delivery. Analytically,consideringn attributes, the problem can be reformulatedin terms ofn goals where each goal is given as

Maximise{Ui(xi)}, i = 1, 2, ..., n

The combined goal function used in the Weights methodcan be defined as

Maximise{f(U(xi)) =n∑

i=1

wiUi(xi)}

where w1, w2, ...wn are significance weightsreflectingthe relative importance of each goal.

It is worth noting that, in our case, the solution is verysimple, since it consists in the evaluation of the functionf(U1, ..., Un) using the values predicted for each hostand in the selection of the hosti with the maximumsuch value.

2) Autonomic adaptation of the utility evaluationfunction: As it stands, the utility function weights arefixed in advance, reflecting the relative importance of thedifferent context attributes. However, such a formulationis still too static, since it fails to take into account thevalues of the attributes. Thus, for example, a small dropin battery voltage may be indicative of the imminentexhaustion of the battery; consequently, it would beuseful to reduce the weight of this attribute nonlinearlyto reflect this.

In general, we wish to adapt the weights of eachparameterdynamicallyand in ways that are dependenton the values of those parameters. In other words, weneed a runtime self-adaptation of the weightings used

for this evaluation process that could be categorized as atypical autonomic mechanism [11]. A simple solution tothis problem is the introduction of adaptive weightsai

into the previous formula, in order to modify the utilityfunction according to the variation of the context.

Maximise{f(U(xi)) =n∑

i=1

ai(xi)wiUi(xi)}

ai(xi) is a parameter that may itself be composite. Forour purposes, we define it to have three important aspectsthat help to determine its value, though the model couldeasily be expanded to incorporate other aspects deemedto be of importance:

a) Criticality of certain ranges of values,arangei(xi)

b) Predictability of the context information,apredictabilityi

(xi)c) Availability of the context information,

aavailabilityi(xi)

We now compose theai weights as factors in thefollowing formula:

ai(xi) = arangei(xi) ·apredictabilityi

(xi) ·aavailabilityi(xi)

a) Adaptive weights related to the possible rangesof values assumed by the attributes :We can modelthe adaptive weightsarange(xi) as a function whosedomain is[0, 1]. For example, with respect to the batteryenergy level (modeled using the percentage of residualbattery energy), we would use a monotonically decreas-ing (though not necessarily linear) function to assign adecreasing adaptive weight that is, in turn, used to ensurethat the corresponding utility function decreases as theresidual energy tends towards zero.

b) Adaptive weights related to the predictabilityof the context information:In general, it is possibleto exploit different statistical attributes for the analysisof time series [12]. One could, for example, use theautocorrelation functionto describe the degree of as-sociation between values of the time series at differentlags2. In short, this gives a measure of the predictabilityof the context information. Furthermore, there are clearguidelines for adapting the use of the autocorrelationfunction for non-stationary data with both trends andseasonal variations.

In building the autocorrelation function, we first needto consider the auto-covariance of the system: given atime series characterized by the meanµt at the timet,

2This is a simplification by assuming independent attributes. Ifthis is untrue, then one might wish to use cross correlation insteadof simple autocorrelation here.

the auto-covarianceCov(Xt, Xt+k) of the time series{Xt} at lagk is defined as follows

Cov(Xt, Xt+k) ≡ E[(Xt − µt)(Xt+k − µt+k)] =

=

n−k∑t=1

(xt − µt)(xt+k − µt+k)

n

The lag represents the time difference (in terms of thenumber of samples) between the two instants beingconsidered. The variance of then samples of the timeseries can be expressed as follows

σ2(Xt) ≡

n∑t=1

(xt − µt)2

n

Therefore, we use theautocorrelation coefficientρk,at lagk defined as follows

ρk ≡Cov(Xt, Xt+k)

V ar(Xt)

that can also be expressed as

ρk =

n−k∑t=1

(xt − µt)(xt+k − µt+k)

n∑t=1

(xt − µt)2

It is worth noting that is possible to prove that

0 ≤ |ρk| ≤ 1

The absolute value ofρk is exactly 1 for a perfectautocorrelation, whereas an autocorrelation coefficientclose to zero (either positive or negative) indicates littleor no correlation between two samplesXt andXt+k. Inthe case of a so-called random series, for a large numbern of samples, the value ofρk is approximately equal to0. We therefore determine parameterapredictabilityi

thus:

apredictabilityi= |ρk|

An interesting issue is the choice of the value of thelag k. It is possible that autocorrelation signals will driftslowly over time and, consequently, the value ofk willalso need to change to reflect this. However, we expectthe underlying processes that determine the nature of theoriginal signal to change slowly if at all.

Thus, in order to adapt the lag value to retain astrongly correlated signal, we adopt a very simple adap-tive technique. At the initial instantt0, k is set to1. This is increased, up to a value ofkMAX , if theautocorrelation coefficient is below a given lower boundthresholdρstrongCorrLB. The process wraps on reaching

kMAX , setting the value ofk back to unity in order toensure that the entire space is searched. If, on the otherhand, the autocorrelation coefficient exceeds an upperbound thresholdρstrongCorrUB, k is decreased until itreaches the value 1.

We can summarize these concepts using the followingupdate equation for the lagk:

k(t+1) =

1 if t = t0 or k(t) = kMAX

k(t) + 1 if ρ(t) ≤ ρstrongCorrLB , k(t) < kMAX , t 6= t0k(t)− 1 if ρ(t) > ρstrongCorrUB , k(t) ≥ 1, t 6= t0k(t) otherwise

c) Adaptive weights related to the availability of thecontext information:It is unreasonable to assume that allcontext attributes have the same degree of availability.Thus, we expect to have a time-varying set of attributesavailable whose values are known or predictable. At-tributes may drop out of this set if meaningful values canno longer be predicted for them, since the information onwhich the prediction would have been based is too old.The simplest approach to this problem is to ensure thatmissing context information carries an adaptive weightai equal to 0:

aavailabilityi=

{1 if the context information is currently available0 if the context information is not currently available

Formally, to date, we have implicitly assumed thata static set of attributes will be defined. However, itis worth noting that, using this approach, we can dy-namically incorporate new attribute values, simply byassuming that they were always there, but had zeroweight for aavailabilityi

.d) Automatic adaptation of the refresh period of

routing tables and context information:In wired net-works, routing table state update is often done on anunvarying regular basis as well as on a by-need basis.However, this approach is wasteful in mobile ad hocenvironments. Thus, we consider how to adapt the rateof context information dissemination by noting that wealready know that such information is predicted byrecipients and that such predictions are likely to bemost accurate when the signal on which they are basedis most predictable. Thus, a possible function for thedetermination of refresh time is given by:

t(x1, x2, ..., xn) = cn∑

i=1

|ρki|

where c is a constant of proportionality.There are several possible extensions of this model.

For example, one might wish to take account of theabsolute value of a parameter in determining updaterates. Thus, for example, as battery energy levels decline,

one might wish to update information increasingly lessfrequently despite the consequent unpredictability at theother end, in order to conserve remaining energy. Ifinformation at the recipients becomes totally outdated,thenaavailabilityi

will be set to zero for all our attributesand the result is that we will not be likely recipients ofmessages to transfer, which is in line with the behaviorwe would expect. Thus, we could replace the simpleconstant in the above equation with a generic functionof values of individual attributes. Likewise, we couldobtain a more refined model associating different weightswith the autocorrelation coefficient for each attribute ina way similar to that applied previously for composingthe utility functions for evaluating which host has thebest message delivery probability.

B. Prediction of the context information attributes usingKalman filters

Kalman filter prediction techniques [13] were orig-inally developed in automatic control systems theory.These are essentially a method of discrete signal pro-cessing that provides optimal estimates of the currentstate of a dynamic system described by astate vector.The state is updated using periodic observations of thesystem, if available, using a set ofprediction recursiveequations.

Kalman filter theory is used in CAR both to achievemore realistic prediction of the evolution of the con-text of a host and to optimize the bandwidth use. Asdiscussed above, the exchange of context informationthat allows the calculation of delivery probabilities isa potentially expensive process, and unnecessarily sowhere such information is relatively easily predictable. Ifit is possible to predict future values of the attributes de-scribing the context, it is possible to update the deliveryprobabilities stored in the routing tables, even if freshinformation is unavailable. Fortunately, it is possibleto express this prediction problem in the form of astate space model. We have a time series of observedvalues that represent context information. From this it ispossible to derive a prediction model based on an innerstate that is represented by a set of vectors, and to add tothis both trend and seasonal components [14]. It is worthnoting that one of the main advantages of the Kalmanfilter is that it does not require the storage of the entirepast history of the system, making it suitable for a mobilesetting in which memory resources may potentially bevery limited. In view of the fact that we use existingresults, we do not present the mathematical aspects ofthe application of state space models theory and Kalman

filter time series analysis in this paper; however, theinterested reader can find these in [15].

The use of prediction is complicated by the fact thatthe information on which it relies for its accuracy travelsacross networks. In mobile settings, bit error rates arerelatively high, and so the loss of messages is moreprobable than for wired settings. If context informationis exchanged only when significant, then its loss hasa greater effect. The tradeoff between loss and theadditional overhead needed for redundant transmissionof context is a complex study in coping with uncertaintyand is outside the scope of this paper.

V. SIMULATION AND RESULTS

We evaluated the CAR algorithm by using the Om-Net++ discrete event simulator [16]. In order to obtaincredible results and to test the peculiar characteristicsof our protocol, it was also necessary for us to developa new group mobility model, that will be presented inSection V-B.

A. Description of the simulation

1) CAR Simulation:For reasons of space and in orderto allow for fair comparison with existing research, wereport results based on simulations that use only partof the full generality of the CAR algorithm. Thus, wesimulated the CAR model using a utility function basedon the evaluation of two attributes: (i) the change rateof connectivity and (ii) the probability of being locatedin the same cloud as the destination. We made theassumption that these factors have the same relevance,so assigned them the same weights in the evaluation ofthe overall utility (i.e.,wi = 0.5). Moreover, we alsoassumed that all the possible values in the range hadthe same importance (i.e,arangei

(xi) = 1) and that thethe values of attributes are always available during thesimulation (i.e.,aavailabilityi

(xi) = 1).The change rate of connectivity attribute is locally

calculated by examining the percentage of a node’sneighbors that have changed their connectivity status(connected to disconnected, or vice versa) between twoinstants. The co-location attribute measures the percent-age of time that two hosts have been in reach. To cal-culate it, we periodically run a Kalman filtering process,assuming that the value is 1 if the host is currently inreach or 0 if not. Clearly, the resultant predicted valueswill be in the range[0, 1] and they will directly expressan estimation of the probability of being in reach of thehost in the future.

We implemented a simplified version of the DSDVprotocol [6] in order to simulate and test the synchronousdelivery in connected portions of the network, as de-scribed in Section III.

Each host maintains arouting and context informationtable used for asynchronous and synchronous (DSDV)routing. Each entry of this table has the followingstructure:

(targetHostId, nextHopId, dist, bestHostId, delProb)

The first field is the recipient of the message, thesecond and the third are the typical values calculatedin accordance with the DSDV specification, whereas thefourth is the identifier of the host with the best deliveryprobability, the value of which is stored in the last field. Itis worth noting that all the autonomic mechanisms, suchas the variable refresh period of routing tables, describedpreviously, were implemented.

We also simulated flooding and the epidemic protocolsin order to provide comparators for the performance ofthe CAR solution.

2) Flooding simulation:We elected to compare ourapproach with flooding. This decision may seem strange,since flooding only works in a fully connected envi-ronment. However, since communications patterns arerandom in the simulations, many messages will be passedbetween hosts that are in connected portions of thenetwork, even when assessing the performance of theepidemic algorithm and of the CAR algorithm. In orderto see the difference in delivery rates that result fromthe algorithms’ ability to handle partial connectivity, itis therefore essential to compare against a synchronousprotocol with optimum delivery ratio.

3) Epidemic routing simulation:The implementationof the epidemic protocol follows the description pre-sented in [2]. The only assumption made by the authorsis a periodic pair-wise connectivity, since the protocolrelies on the transitive distribution of messages fordelivery. When two hosts become neighbors (in otherwords, they are within each other’s radio range), theydetermine which messages each possesses that the otherdoes not, using summary vectors that index the list ofmessages stored at each node; they then exchange them.Each message is characterized by a unique messageidentifier and a hop count value; the latter determines themaximum number of possible exchanges of a message.Higher hop count values reduce the delivery latency,but, at the same time, increase the quantity of resources(memory, battery, bandwidth) consumed in this process.The simulation prepared by the authors shows that the

algorithm achieves total delivery of messages sent after alimited period of time, but at the cost of very substantialoverheads. The epidemic approach represents the clas-sic example of an asynchronous protocol and thereforeprovides the ideal comparator.

4) Simulation system parameters:We evaluated theperformance of each protocol sending 100 messages witha simulation time equal to 300 seconds. The messageswere sent after 40 seconds, in order to allow for thesettling of initial routing table exchanges, and the inter-vals between each message were modeled as a Poissonprocess, withλ = 5s−1, and the consequence that allmessages are sent in about 20 seconds. The sender andreceiver of each message are chosen randomly.

In the CAR simulation, each message has a fieldthat is similar to atime to livevalue that is decreasedeach time that the message is transferred to another host(the initial value being 15). Moreover, in this case, wealso introduced asplit horizon mechanism to preventmessages from being retransmitted unnecessarily. Thebuffer for each node was set to 20 messages, unlessotherwise specified. Table I summarizes the simulationparameters.

The one key aspect of the simulation not yet addressedis that of the mobility model. Clearly, the random way-point mobility model, which is used extensively in suchstudies, largely for reasons of simplicity, does not accu-rately reflect human behaviour and annihilate the effectof the prediction since movement is entirely random.Consequently, we devised a new group-based mobilitymodel [17]. This is presented briefly in the followingsection.

B. Mobility model

Mobility models that assume that individuals moveindependently of one another in random ways are un-realistic in terms of the deployment scenarios for adhoc networks that are most commonly expounded. Forexample, on a battlefield, it would be indicative of a verytroubled army if each soldier were to move randomlywith respect to all others. Thus, we have developed anew model with a form of hierarchical clustering thatbetter reflects the ways in which collections of people arestructured at an organizational level and, consequently,the ways in which they move [17]. This model has beeninstantiated in a simple way for these experiments, and,as used here, is somewhat akin to those in [18], [19](however, its potential is larger as explained in [17]).Thus, we introduce the concept of a collection of nodes,

TABLE I

SIMULATION PARAMETERS

Number of hosts 16/24/32Simulation area 1 Km x 1 Km

Propagation model free spaceAntenna type omnidirectional

Transmission range 200 mNumber of clouds 4

Cloud area 200 m x 200 mNode speed 1-3 m/sCloud speed 1-2 m/s

Number of messages sent 100Max number of hops 15Message buffer size 10 to 100Routing table size 20 entries

Max distance 15

which has its own motion overlaid on a form of randommotion within the cloud.

By parameterizing this model differently, we canrepresent different archetypes: for example, one wouldexpect to use different parameters for an academic whospends her life traveling between home and the univer-sity, interacting with a very closed set of people, as op-posed to a salesman who travels much more extensivelyand interacts less discriminatingly.

A host that belongs to a cloud moves inside it towardsa goal (i.e, a point randomly chosen in the cloud space)using the standard random way-point model. When ahost reaches a goal, it also implicitly reaches a decisionpoint about whether to remain within the cloud, and, ifleaving, to where it should go. Each of these decisionsis taken by generating a random number and comparingit to a threshold (which is a parameter of the model). Itis worth noting that clouds also move towards randomlychosen goals in the simulation space.

C. Simulation Configuration

50% of the hosts are initially placed randomly in acloud, whereas the others are positioned randomly in thesimulation area. Each cloud is defined using a squaredarea with a side length of 200 m. In other words, werandomly select the point(minX, minY ) that, togetherwith the length of the side, defines the cloud area. Forthese simulations, there is only a single level of cloud.

Every host is characterized by two values,Pescape,indicating the probability of escaping from the currentcloud, andPescapeCloud describing the probability of

0

20

40

60

80

100

120

16 24 32Number of hosts

Del

iver

y ra

tio (%

)

Flooding EpidemicCAR

Fig. 3. Delivery ratio vs population density

choosing a new goal in the space between clouds.Each cloud moves with a random speed (with a value

in the range 1-2 m/s); moreover, each host moves with arandomly generated different speed (with a value in therange 1-3 m/s). It is worth noting that the movement ofa host is the result of the composition of these speeds.

In our simulation, the positions of all the hostsand clouds are updated every second. When a cloudreaches its goal, a new goal is chosen in the simula-tion space. When a host reaches its goal, a thresholdprobability PescapeThreshold is generated randomly (itsrange is clearly[0, 1]). If its Pescape is greater thanPescapeThreshold the new goal is chosen outside the cur-rent cloud, else inside. If outside, we randomly generatePescapeCloudThreshold and compare it toPescapeCloud todetermine whether or not the goal should be chosen insome other cloud or in the open space between clouds.For those hosts that are already outside a cloud, thechoice of a new goal is done in an analogous way.

D. Analysis of results

In this subsection we will analyze the results of oursimulations, comparing the performance of CAR withthe flooding and epidemic protocols. We will discuss thevariation of some performance indicators as functionsdependent on the density of hosts (i.e., the number of thehosts in the simulation area) and the size of the buffersused to store messages in both the epidemic and CAR.In Figure 3, there is a comparison between the delivery

ratios of the three protocols in each of three differentscenarios (with 16, 24 and 32 hosts). In all cases, thenumber of messages that may coexist within a node’sbuffer is unconstrained.

CAR achieves a performance between that of floodingand epidemic routing, as expected. Flooding suffers fromthe inability to deliver messages to recipients that arein other clouds when the messages are sent but is here

0

20

40

60

80

100

120

10 20 30 40 50 60 70 80 90 100

Buffer size

Del

iver

y ra

tio (%

)

EpidemicCAR

Fig. 4. Delivery ratio vs buffer size

0

10000

20000

30000

40000

50000

60000

70000

10 20 30 40 50 60 70 80 90 100

Buffer size

Num

ber o

f mes

sage

s

EpidemicCAR

Fig. 5. Number of messages vs buffer size

simply as a comparator to demonstrate the numbersof messages being delivered that cannot be delivereddirectly, because the recipient is in a cloud different fromthe cloud of the sender. The epidemic protocol can beconsidered optimal in terms of delivery ratio, simplybecause each message is propagated to all accessiblehosts, all of which have buffers large enough to holdit. In CAR, we have chosen to operate under the moststringent conditions: there is only ever a single copy ofeach message, which represents the worst case for thisprotocol. Clearly, it would be possible to trade off a smallamount of intelligent replication (to improve the deliveryratio) against an increase in overhead.

The dependency of the delivery ratios on the buffersize is similar for all the protocols (see in Figure 4 theresults for the 32 hosts scenario). Both of these demon-strate a substantial degradation of their performance asbuffer size decreases; however, this phenomenon is moreevident in the epidemic approach as a result of the degreeof replication of messages.

Figure 5 is interesting because there are two com-peting effects at work for the epidemic protocol. Whenthe buffer size is small, there is a high probabilitythat messages will be eliminated due to overflow, as

0

5

10

15

20

25

30

10 30 50 70 90 110

130

150

170

190

210

230

Delay time (sec)

Num

ber o

f mes

sage

s

EpidemicCAR

Fig. 6. Average delay vs population density

0

5

10

15

20

25

30

10 20 30 40 50 60 70 80 90 100 110 120 130 140 150 160 170 180 190 200 210 220 230 240

Delay time (sec)

Num

ber o

f mes

sage

s

162432

Fig. 7. Average delay vs population density

discussed above. Consequently, the number of messagesexchanged is also low. At the other end of the scale,as the buffer size increases to a point where it canaccommodate all the messages in the system, there is norepeated exchange of messages, so the number is alsolow. In the middle of the range, however, the buffersize is insufficient to hold all messages and there isa cycle in which messages are eliminated by bufferoverflow and then reinstated by other nodes, resultingin very high overhead. In the case of CAR, it is worthnoting that the overhead in terms of the number ofmessages exchanged is more or less constant, regardlessof buffer size, demonstrating itsscalability. CAR willalways be the limiting case for performance under thismetric because it only creates a single copy of eachmessage. Thus, even at the point where buffer sizebecomes effectively infinite, the epidemic protocol willnecessarily exchange more messages than ours, simplyas a result of the replication.

Figure 6 shows the distribution of the number ofmessages with respect to their delivery latency in the 32hosts scenario. It is possible to observe that a proportionof the messages are delivered more or less immediately,since the recipients are in the same cloud as the sender.

Another interesting comparison is showed in Figure 7:the distributions of the delivery latency in the case ofdifferent node densities are very similar.

VI. RELATED WORK AND DISCUSSION

A number of approaches have been proposed to en-able asynchronous communication in mobile ad hocnetworks.

Epidemic algorithms were first devised in the con-text of distributed database systems in an attempt toguarantee data consistency after disconnections [20].Interesting theoretical results show that, using randomdata exchanges, all updates are seen by all the hostsof the system in a bounded amount of time, given rea-sonable assumptions about connectivity. The epidemicrouting protocol [2], described earlier, that forms thebasis for much of the work in this field, applied this earlyapproach to the field of asynchronous message delivery,but in a rather naive fashion.

Chen and Murphy refined the epidemic model, pre-senting the so-called Disconnected Transitive Commu-nication paradigm [5]. Their approach is similar to ours,since it essentially argues for the use of utility functions,but it provides a general framework rather than a detailedinstantiation, and so aspects related to the compositionof calculated delivery probabilities are almost entirelymissing.

In [21], Lindgren et al. propose a probabilistic routingapproach to enable asynchronous communication amongintermittently connected clouds of hosts. Their approachis based on the fact that the exploited communicationmodel is typically transitive and, for this reason, theprobability of message delivery must be calculated ac-cordingly: in other words, if, for example, a hostHA isable to communicate withHB throughHC , the overalldelivery probability is derived by the multiplication ofthe probability thatHA becomes a neighbor ofHB, withthe probability thatHB becomes a neighbor ofHC .The calculation of the delivery probabilities is based,somewhat simplistically, on the period of time of co-location of two hosts, weighted by an aging factor thatis used to decrease the overall probability with theincreasing age of the information on which it was based.

In [22], Small and Haas describe a very interestingapplication of epidemic routing protocols to a problem ofcost-effective data collection, using whales as messagecarriers.

Sasson et al. [23] studied a possible application ofpercolation theory (that studies the probability of transi-tion between two states in fluids) to improve information

dissemination based on the flooding of messages in adhoc settings. Another interesting epidemic model for mo-bile ad hoc networks is presented in [24]; in this paperthe authors investigate the similarities between flooding-based approaches for the information dissemination inmobile ad-hoc networks and the epidemic spreading ofdiseases.

In [25], Fall proposes the Delay Tolerant Networkarchitecture to solve the internetworking issues in sce-narios where partitions are frequent and a connectedpath between message senders and receivers may benot present (such as satellite and interplanetary com-munication systems). This approach relies on routingmechanism presented in detail in [26], based on optimalor sub-optimal algorithms, according to the differentknowledge about the topology of the networks and thesampled delivery delays, to compute the best end-to-endcommunication shortest path.

Zhao et al. in [27] discuss the so-called MessageFerrying approach for message delivery in mobile ad hocnetworks. The authors propose a proactive solution basedon the exploitation of highly mobile nodes called fer-ries. These nodes move according to pre-defined routes,carrying messages between disconnected portions of thenetwork.

With respect to the existing work in this researcharea, such as [26] and [27], we have introduced ageneral framework for the evolution and the predictionof the mobile context to provide efficient and effectivecommunication mechanisms in mobile ad hoc networks.Moreover, we believe that it is possible to integrate ourtechniques with these approaches, since they addressorthogonal aspects of the problem.

It is worth noting that we used lightweight mecha-nisms, because we believe that routing algorithms thatare complex from a computational point of view areunsuitable for mobile devices, usually characterised byscarcity of resources. One example is the use of Kalmanfilter techniques that do not necessitate storing all thehistory of the evolution of the context information.

VII. C ONCLUSIONS ANDFUTURE WORK

In this paper, we presented a novel approach to thechallenge of asynchronous ad hoc routing. This is aproblem that is deceptively easy to state but that requiresthe combination of results from many fields to addressefficiently. Thus, we have designed a general and flexibleframework for the evaluation of context informationusing probabilistic, statistical, autonomic and predictivetechniques in order to optimize the consumption of the

scarce resources of mobile devices whilst retaining gooddelivery performance. Previous solutions to the problemof asynchronous routing are either unoptimized (as in thecase of the basic epidemic protocol) or are insufficientlyspecific to help in the construction of systems capableof dealing with the multi-dimensional nature of context(as in the Chen and Murphy’s approach).

In order to assess our algorithm, a new mobilitymodel [17], better reflecting the realities of human orga-nization, was developed and used in simulations that givea feel for the relative performance of CAR relative toflooding and epidemic routing. The results demonstratedthat, even without message replication, CAR performsrespectably in terms of message delivery, with very muchlower overheads than the alternatives.

In future, we will further explore the tradeoff betweenincreasing delivery ratios via replication versus mainte-nance of low overhead. Moreover, we will further ex-plore an acknowledgment mechanism in order to notifythe sender about the correct delivery of messages (andto remove them from intermediate nodes), exploitingthe same techniques as those used to deliver messages.Lastly, we intend further to investigate the applicationof mathematical models of social organization, mostnotably small world models, in assessing performanceand in optimizing the reliability of the routing algorithm.

REFERENCES

[1] C. E. Perkins,Ad Hoc Networking. Addison-Wesley, 2001.[2] A. Vahdat and D. Becker, “Epidemic routing for partially

connected ad hoc networks,” Department of Computer Science,Duke University, Tech. Rep. CS-2000-06, 2000.

[3] A. Doria, M. Uden, and D. P. Pandey, “Providing connectivity tothe Saami nomadic community,” inProceedings of the SecondInternational Conference on Open Collaborative Design forSustainable Innovation, December 2002.

[4] I. T. Union, “Connecting remote communities,”Docu-ments of the World Summit on Information Society, 2003,http://www.itu.int/osg/spu/wsis-themes.

[5] X. Chen and A. L. Murphy, “Enabling disconnected transitivecommunication in mobile ad hoc networks,” inProceedings ofthe Workshop on Principles of Mobile Computing (POMC’01),August 2001, pp. 21–23.

[6] C. E. Perkins and P. Bhagwat, “Highly dynamic Destination-Sequenced Distance-Vector routing (DSDV) for mobile com-puters,” inProceedings of thr Conference on CommunicationsArchitecture, Protocols and Applications (SIGCOMM 94), Au-gust 1994.

[7] R. L. Keeney and H. Raiffa,Decisions with Multiple Objectives:Preference and Value Tradeoffs. John Wiley and Sons, 1976.

[8] S. J. Russell and P. Norvig,Artificial Intelligence: A ModernApproach, 2nd ed. Prentice Hall, 2002.

[9] H. A. Taha, Operations Research: An Introduction. PrenticeHall, 1996.

[10] G. Debreu, “Topological methods in cardinal utility theory,”in Mathematical Methods in the Social Sciences, K. J. Arrow,S. Karlin, and P. Suppes, Eds. Stanford University Press, 1959.

[11] D. F. Bantz, C. Bisdikian, D. Challener, J. P. Karidis, S. Mas-trianni, A. Mohindra, D. G. Shea, and M. Vanover, “Autonomicpersonal computing,”IBM Systems Journal, vol. 42, no. 1, 2003.

[12] C. Chatfield, The Analysis of Time Series An Introduction.Chapman and Hall, 2004.

[13] R. E. Kalman, “A new approach to linear filtering and pre-diction problems,”Transactions of the ASME Journal of BasicEngineering, March 1960.

[14] P. J. Brockwell and R. A. Davis,Introduction to Time Seriesand Forecasting. Springer, 1996.

[15] M. Musolesi, S. Hailes, and C. Mascolo, “Predictionof Context Information Using Kalman Filter Theory,”http://www.cs.ucl.ac.uk/staff/m.musolesi/papers/kalman.pdf.

[16] A. Varga, “The OMNeT++ discrete event simulation system,”in Proceedings of the European Simulation Multiconference(ESM’2001), Prague, 2001.

[17] M. Musolesi, S. Hailes, and C. Mascolo, “An ad hoc mobilitymodel founded on social network theory,” inProceedings ofthe 7th ACM international symposium on Modeling, analysisand simulation of wireless and mobile systems. ACM Press,October 2004.

[18] X. Hong, M. Gerla, G. Pei, and C.-C. Chiang, “A groupmobility model for ad hoc networks,” inProceedings of the 2ndInternational Workshop on Modeling Analysis and Simulationof Wireless and Mobile Systems, 1999, pp. 53–60.

[19] J. Carlos Cano, P. Manzoni, and M. Sanchez, “Evaluating theimpact of group mobility on the performance of mobile adhoc networks,” inProceedings of the 2004 IEEE InternationalConference on Communications (ICC 2004), 2004.

[20] A. Demers, D. Greene, C. Hauser, W. Irish, J. Larson,S. Shenker, H. Sturgis, D. Swinehart, and D. Terry, “Epidemicalgorithms for replicated database maintenance,”ACM SIGOPSOperating Systems Review, vol. 22, no. 1, January 1988.

[21] A. Lindgren, A. Doria, and O. Schelen, “Probabilistic routingin intermittently connected networks,”Mobile Computing andCommunications Review, vol. 7, no. 3, July 2003.

[22] T. Small and Z. J. Haas, “The shared wireless infostation model-a new ad hoc networking paradigm (or where is a whale, thereis a way),” in Proceedings of MobiHoc’03, June 2003.

[23] Y. Sasson, D. Cavin, and A. Schiper, “Probabilistic broadcast forflooding in wireless mobile ad hoc networks,” inIEEE WirelessCommunications and Networking Conference (WCNC 2003),March 2003.

[24] A. Khelil, C. Becker, and J. a. Tian, “An epidemic modelfor information diffusion in MANETs,” inProceeding of thethe Fifth ACM International Workshop on Modeling, Analysisand Simulation of Wireless and Mobile System (MSWiM’02),September 2002.

[25] K. Fall, “A delay-tolerant network architecture for challengedinternets,” inProceedings of SIGCOMM’03, August 2004.

[26] J. Sushant, K. Fall, and R. Patra, “Routing in a delay tolerantnetwork,” in Proceedings of SIGCOMM’04, August 2004.

[27] Z. Venrui, M. Ammar, and E. Zegura, “A message ferryingapproach for data delivery in sparse mobile ad hoc networks,”in Proceedings of the 5th ACM International Symposiunm onMobile Ad Hoc Networking and Computing (MobiHoc’04), May2004.