mobility-assisted routing in intermittently connected mobile cognitive radio networks

Mobility-Assisted Routing in IntermittentlyConnected Mobile Cognitive

Radio NetworksJianhui Huang, Shengling Wang, Xiuzhen Cheng, Min Liu, Zhongcheng Li, and Biao Chen

Abstract—In mobile ad-hoc cognitive radio networks (CRNs), end-to-end paths with available spectrum bands for secondaryusers may exist temporarily, or may never exist, due to the dynamism of the primary user activities. Traditional CRN routingalgorithms, which typically ignore the intermittent connectivity of network topology, and traditional mobility-assisted routingalgorithms, which generally overlook the spectrum availability, are obviously unsuitable. To tackle this challenge, we proposea Mobility-Assisted Routing algorithm with Spectrum Awareness (MARSA) to select relays based on not only the probabilitythat a node meets the destination but also the chance at which there exists at least one available channel when they meet.To the best of our knowledge, this paper is the first to bring the idea of mobility-assisted routing to deal with the intermittentlyconnected attribute of mobile ad-hoc CRNs, and the first to enhance the mobility-assisted routing by considering thetemporal, spatial, and spectrum domains at the same time. Our simulation results demonstrate the superiority ofMARSA over traditional algorithms in intermittently connected mobile CRNs.

Index Terms—Intermittently connected networks, mobile ad-hoc cognitive radio networks, mobility-assisted routing

Ç

1 INTRODUCTION AND MOTIVATION

NOWADAYS, cognitive radio networking (CRN) hasreceived a lot of attention because it bears the promise

of alleviating the spectrum scarcity problem [1], [2], [3], [4],[5], [6], [7]. In this paper, we consider a mobile ad-hoc CRN,in which two nodes are able to exchange their data onlywhen they enter each other’s transmission range and thereexists at least one available channel between them. Such aCRN is termed an Intermittently Connected Mobile CRN(ICMCRN), which typically does not have temporarilystable end-to-end paths with available spectrum.

In ICMCRNs, traditional CRN routing protocols are notappropriate because they usually assume that the availabletime for CR activities over primary bands is longer thanthe time needed to undergo a round of communications.However, this assumption does not always hold trueespecially when the primary users’ utilization of the licensedspectrum is high (the highest utilization can be 85 percent [8]).In fact, if the idle and busy durations of the licensed spectrumare switched frequently within a short time, the spectrum

is hard to use even though the utilization is low. This isbecause before employing the licensed spectrum for datacommunications, secondary users (SUs) must sense thechannel and negotiate with others for possible spectrumsharing. Thus if the time of these processes is comparable tothe idle duration, the spectrum may switch to be unavail-able quickly from the idle state, largely limiting the amountof data to be exchanged.

Due to the nature of intermittent connectivity inICMCRNs, a natural way to deal with the routing problemis to adopt the mobility-assisted routing policy, i.e., store-carry-forward. However, directly applying the traditionalmobility-assisted routing mechanisms for ICMCRNs maylead to inefficient routing performance due to the impropercriterion for selecting relays. A good relay in traditionalmobility-assisted routing that only considers the chance tocontact with the destination may not be a good one inICMCRNs because if no spectrum band exists during theircontact, the contact becomes invalid.

Conclusively, for ICMCRNs, neither the traditional CRNrouting algorithms are appropriate due to the assumptionof the existence of stable end-to-end paths nor thetraditional mobility-assisted routing algorithms are suit-able because they only consider temporal and spatialdomains while the routing efficiency of ICMCRNs alsodepends on the spectrum domain. To address the routingchallenge of ICMCRNs, we propose a Mobility-AssistedRouting algorithm with Spectrum Awareness (MARSA) inthis paper, which selects relays by considering not only thechance that a node meets the destination but also thechance at which there exists at least one available channelwhen they meet. In other words, MARSA takes into accountthe information from the temporal, spatial, and spectrumdomains for routing path construction.

. J. Huang, M. Liu, and Z. Li is with the Institute of ComputingTechnology, Chinese Academy of Sciences, Beijing, China. E-mail:{huangjianhui, liumin, zcli}@ict.ac.cn.

. S. Wang is with the College of Information Science and Technology, BeijingNormal University, Beijing, China. E-mail: [email protected].

. X. Cheng is with the Computer Science, The George WashingtonUniversity, Washington DC, USA. E-mail: [email protected].

. B. Chen is with the Department of Computer Information Science,University of Macau, Macau, China. E-mail: [email protected].

Manuscript received 30 May 2012; revised 1 Nov. 2013; accepted 7 Nov.2013. Date of publication 24 Nov. 2013; date of current version 15 Oct. 2014.Recommended for acceptance by M. Thai.For information on obtaining reprints of this article, please send e-mail to:[email protected], and reference the Digital Object Identifier below.Digital Object Identifier no. 10.1109/TPDS.2013.291

1045-9219 � 2013 IEEE. Personal use is permitted, but republication/redistribution requires IEEE permission.See http://www.ieee.org/publications_standards/publications/rights/index.html for more information.

IEEE TRANSACTIONS ON PARALLEL AND DISTRIBUTED SYSTEMS, VOL. 25, NO. 11, NOVEMBER 20142956

To the best of our knowledge, this paper is the first tointroduce the idea of mobility-assisted routing to ICMCRNsto deal with the intermittently connected attribute, and thefirst to enhance the mobility-assisted routing with spectrumawareness to adapt to the routing challenge of ICMCRNs.

The rest of the paper is organized as follows. Section 2presents the related work while Section 3 describes the designmotivation. A routing framework and the correspondingalgorithm are detailed in Sections 4 and 5, respectively.Section 6 provides an analysis and the numerical validationof the proposed algorithm. Performance evaluation basedon both real world and synthetical trace data are elaboratedin Section 7. Our conclusions are drawn in Section 8.

2 RELATED WORK

In this section, we review the most related work for bothCRN routing and mobility-assisted routing.

Existing CRN routing algorithms can be classified intotwo categories: full and local spectrum knowledge-based. Forthe first category [9], [10], [11], the network topology andthe spectrum occupied by the primary users (PUs) areknown a priori. Hence routing decision without spectrumassessment is the main task of these approaches. For thesecond category, the information of spectrum occupancy isdetected in a distributed way. Such approaches usuallyemploy various cross-layer optimization methods, wherethe routing problem is considered jointly with scheduling,power control, and spectrum management, to improve thespectrum utilization [12], maximize the data rates for a setof sessions [13], maximize the network throughput [14],[15], or minimize the end-to-end delay [11], [16].

Note that these two categories of CRN routing algo-rithms both assume the existence of at least one end-to-endpath with available spectrum bands in the network. Asdescribed in Section 1, this assumption is not always valid.

Existing mobility-assisted routing algorithms can alsobe classified into two categories: deterministic and stochastic.Deterministic approaches [17], [18], [19] provide determin-istic routing decisions assuming that some kinds of networkconnectivity information is known a priori. Stochasticapproaches [20], [21], [22], [23], on the other hand, inferthe current and future network connectivity informationthrough history data and accordingly select appropriaterelays. Not surprisingly, none of the existing mobility-assisted routing algorithms considers the spectrum avail-ability in relay selection.

As claimed earlier, neither existing CRN routing schemesnor mobility-assisted routing ones address the routingchallenges in ICMCRNs, which are the focus of this paper.

3 DESIGN MOTIVATION

In this section, we present the mobility traits of the nodesand the spectrum exploited by MARSA, which are obtainedby analyzing the real world phenomenon and trace data.

3.1 Node Mobility Exhibits Spatial and TemporalRegularity

In real world, the mobility of nodes (such as human beingsor vehicles) is usually highly motivated. For example, people

often visit their work places at fixed time; Students usually goto/get off classes according to predefined schedules; Busesand airplanes arrive in certain places at certain times undernormal circumstances. Hence, the mobility of nodes exhibitsspatial and temporal regularity. This trait is verified by thedata [24] obtained through tracking the trajectories of 100,000anonymized mobile phone users for a six-month period,where the nodes are found at their first two preferred placeswith a probability over 40 percent. A three-month record[25] capturing the mobility patterns of 50,000 nodes chosenfrom about 10 million anonymous mobile phone usersalso demonstrates that human beings usually stay in theirhighly visited places at an hourly regularity.

3.2 Spectrum Mobility Shows RegularityMany real world datasets [26], [27] demonstrate that theactivities of PUs are not random. For example, the dataset[27] based on the billing information of one Europeanmobile phone operator indicates that the change trends ofthe average number of calls by hour and the average callduration in each weekday within a week are very similar.The dataset [26] from a US CDMA-based cellular operatorshows that the distribution of the call inter-arrival time iswell described by an exponential distribution in more than90 percent of the hours for most cells. The approximatelyregular PU behaviors result in the approximate regularity of themobility of the spectrum that could be available for SUs.

The mobility regularity of the nodes and the availablespectrum stirs us to design an ICMCRN routing frameworkand a relay selection algorithm, which will be presented inthe next two sections.

4 THE FRAMEWORK FOR MARSABased on the observations from the real world data, theregularities of the nodes and the spectrum are tightlyrelated to the places people visit. Hence in MARSA, thenetwork ðWÞ is partitioned into multiple zones ðZiÞ satisfying[Zi ¼ W and \Zi ¼ ;. Zones can have regular or non-regularshapes, which are determined by their usage purposes such asresidences, office buildings, and shopping centers. Somerelated definitions are presented as follows.

Definition 4.1 (Home). The home of a node j, denoted by Hj,is a set of zones it usually visits.

There exist two simple strategies to determine the homeof a node: 1) a node can statically configure the zones itusually visits as its home; and/or 2) it can dynamically adda zone to its home once the visiting frequency of the zone islarger than a given threshold. Similarly, a zone can be deletedfrom a node’s home either statically or dynamically.

Definition 4.2 (Neighbor Set). The neighbor set of node j,denoted by Nj, is a set of nodes that can communicate directlywith j if there exists at least one available channel.

We assume that any two nodes located at the same zonecan communicate directly with each other if at least one

HUANG ET AL.: MOBILITY-ASSISTED ROUTING IN MOBILE COGNITIVE RADIO NETWORKS 2957

channel is available. Based on this assumption, all the nodescovered by the zone where j resides belong to Nj. Moregenerally, Nj includes the nodes in a neighboring zone thatcan communicate with j directly if spectrum is available.Note that j 62 Nj.

Definition 4.3 (Destination Zone). The zone the destinationd currently resides is the destination zone, denoted by Zd.

Definition 4.4 (Spectrum Availability). The spectrumavailability is the probability that there exists at least oneavailable channel for SUs to communicate.

In MARSA, the time for each zone is divided into equallength cycles. Within each cycle, it is further divided intoequal length intervals. The length of an interval depends onthe variation of the available spectrum: if the variation isbig, it should be reduced; otherwise, it should be enlarged.Due to the daily regular behaviors of human beings, a nodeis supposed to periodically appear in its home zones atfixed interval(s) during each cycle, as illustrated in Fig. 1.

Moreover, the visit of a node is described based on theintervals. If a node resides in a zone for n intervals, we sayit has n visits to that zone. For example, if a node visits azone from 3:30 to 4:50, we say it has two visits to that zone ifthe length of an interval is one hour, and the visitingperiods are from 3:30 to 4:00 and from 4:00 to 4:50.

Definition 4.5 (Highly Available Visit). A highly avai<lable visit is one during which the spectrum availability islarger than a given threshold Th.

Note that here we care about whether there exists anavailable channel, not a particular channel, because weassume that the cognitive radios are fragile and they areable to operate on any available channel of the spectrum.

We require that two nodes exchange their routinginformation whenever they contact. The informationcarried by each node i is represented as three-tuplesfZj; T ij ; P i

jg, with one for each home zone, where Zj 2 Hi,

T ij ¼ ftij1; tij2; . . .g is the set of starting time at which i

usually visits Zj, and Pij ¼ fpij1; pij2; . . .g, with each pijk

representing the spectrum availability in the visitinginterval starting from tijk. The above requirement is realisticbecause it basically accords with the results in [25] that thepotential predictability of the user mobility in the temporaland spatial domains is about 93 percent. Because thespectrum mobility is regular in reality [26], [27], thespectrum availability in each zone of each interval can bededuced by analyzing the long-term statistical data. It alsoneeds to consider the sharing of the spectrum among

multiple SUs: the more number of SUs competing for thesame channel, the lower the spectrum availability. This is aMAC layer issue, and we leave it to our future research.For simplicity, in this paper we simply assume that thereexists a centrally-maintained database that can period-ically broadcast the available spectrum information. Themethod of deploying the centrally-maintained databaseto get available channel information is widely used inCRNs such as [28], [29]. Thus each node can be aware ofthe spectrum availability of its home zones during thevisiting intervals.

5 DESCRIPTION OF MARSAThis section details the MARSA algorithm. Its pseudocodeis given in Algorithm 1, where the function sendðB; iÞ sendsdata B to node i. Because each node has CR capability, anynode j currently holding the data senses the spectrumperiodically, as suggested by [11], [30]. If at least onechannel is available, node j needs to update its neighbor setNj. Each node can discover its neighbors through beacontransmission or a HELLO protocol [31].

In this algorithm, when the destination d is a neighbor ofthe source or a relay, the data is delivered to d directly(Lines 4-6). For any node j currently carrying the data,when it meets a node i that is more optimal than itself andis also the most optimal one among its neighbors, ittransmits the data to node i, which acts as the new relay(Lines 7-12). Obviously, if the source does not meet a nodethat is more optimal than itself, it acts as the relay.

Algorithm 1 Pseudocode of MARSA

1: repeat2: if at least one channel is available then3: Update Nj if needed; . Node j may be the source

or a relay.4: if d 2 Nj then . d is the destination;5: sendðB; dÞ; return6: end if7: for each node i 2 Nj do8: exchange routing information with i;9: if Oði; jÞ ¼ True ^ for 8k 2 Nj n fig, Oði; kÞ ¼ True

then . If i is better than j and j’s neighborhood;10: sendðB; iÞ;11: end if12: end for13: end if14: until the data expires;

Fig. 1. Node periodically appears at one of its home zones at the second interval in each cycle.


Algorithm 2 Function Oði; jÞ1: Ai fm j 0 � tidm � tc � tlg; Aj fn j 0 � tjdn � tc � tlg. Compute the set of visits whose starting time is beforethe lifetime of the data;

2: Bi fm j pidm 9 Th; m 2 Aig; Bj fn j pjdn 9 Th; n 2 Ajg. Compute the set of highly available visits.

3: if Bi 6¼ ; ^Bj 6¼ ; then4: if minftidm j m 2 Big G minftjdn j n 2 Bjg then5: return True6: else7: return False8: end if9: end if10: if Bi 6¼ ; ^Bj ¼ ; then11: return True12: end if13 if Bi ¼ ; ^Bj 6¼ ; then14: return False15: end if16: if Bi ¼ ; ^Bj ¼ ; then17: if Ai ¼ ; then18: return False19: else . Ai 6¼ ;.20: if Aj�; _ ðAj 6¼ ; ^ minftidm j m 2 Aig G

minftjdn j n 2 AjgÞ then21: return True22: else23: return False24: end if25: end if26: end if

The functionOði; jÞ judges whether node i is more optimalthan j to act as a relay. If yes, it returns True; otherwise False.The process of FunctionOði; jÞ is given in Algorithm 2, wheretc is the current time and tl is the remaining lifetime of thedata. For two nodes both having at least one highly availablevisit toZd, or both only having at least one visit rather than thehighly available one(s) to Zd, if the earliest visit of one node isearlier than that of the other, it is more optimal (Lines 3-9 and20-24); the one that has at least one highly available visit ismore optimal than the other that does not have (Lines 10-15); ifall visits of node i are after the data expiration time, node j ismore optimal than node i even if all visits of node j are alsoafter the data expiration time because j currently holds thedata (Lines 17-18); if node i has visits before the data expiresand node j does not have, node i is more optimal; otherwisenode j is more optimal (Lines 20-24).

MARSA is potentially based on the following assumption:the destination stays atZd during the process of data delivery.This assumption is also adopted by the mobility-assistedrouting in [32]. However, if it is not true, MARSA can still workby changing its way of usage. For example, the data can bedelivered to all the home zones of the destination in light of therelay selection policy of MARSA. Moreover, if the destinationis not located at any zone of its home, the data can be deliveredto a static node in each home zone of the destination, fromwhich the data can be retrieved when the destination comes

back. In addition, if the source knows the mobility regularity ofthe destination in temporal and spatial domains, throughmodifying MARSA slightly, the data can be directly for-warded to the zone the destination currently resides.

When designing MARSA, we do not consider the impactof the available spectrum duration and the contact durationon the data delivery. For simplicity, we assume that once achannel is available when two nodes meet, the data can bedelivered successfully. We will consider the availablespectrum duration and the contact duration for ICMCRNrouting in our future research. In addition, though we do notconsider the limitation of the buffer length in MARSA, wediscuss its impact on the routing performance in Section 7.

Our relay selection policy cares about the characteristics ofthe node itself, i.e., at what time a node visits which place,rather than the nodal relationships, i.e., at what time a nodecontacts which node. This makes our scheme scalable andresults in a low overhead because the volume of informationeach node carries, maintains, and exchanges with others doesnot increase with the increase of the number of nodes in thenetwork. In addition, each node is in charge of its actionsbased on the local information. Hence, MARSA is distributedin nature and does not cause any communication bottleneck.

6 THEORETICAL ANALYSIS AND NUMERICALVALIDATION

6.1 Theoretical AnalysisTo analyze MARSA, we model the data delivery by anabsorbing discrete-time Markov chain, where the successfuldata delivery to the destination is considered by an absorbingstate denoted by Q. Besides the absorbing state Q, there existother states, with each representing the data arriving at onetype of nodes, which will be introduced in the next.

Assume that the set of nodes in the whole network isdenoted by U and kUk is the potency of the set U , which isalso the total number of nodes in the network. LetS1 ¼ fiji 2 U ^Ai ¼ ;g be the set of nodes that could notvisit Zd before the data expires. Denote by S1 the staterepresenting the data being held by any node in S1. LetS2 ¼ fiji 2 U ^Ai 6¼ ; ^Bi ¼ ;g be the set of nodes that canvisit Zd before the data expires but do not have any highlyavailable visit to Zd, and S3 ¼ fiji 2 U ^Bi 6¼ ;g be the set ofnodes with highly available visit(s) toZd before the data expires.

Denote by S2;j and S3;j the sets of nodes in S2 and S3 thatwill visit Zd after j ðj 2 f0; 1; . . . ; E � 1gÞ intervals from thecurrent interval, where E is the total number of intervals inone cycle. Let S2;j and S3;j be the states representing thedata being held by any node in S2;j and S3;j, respectively.As time goes by, the time span that a node will visit Zd fromthe current interval changes. Hence, the nodes in the setsS2;j and S3;j ðj 2 f0; 1; . . . ; E � 1gÞ may change, making thestate transition matrix change at different intervals. As aresult, the process of data delivery in MARSA is actuallymodeled by a nonhomogeneous Markov chain. Let Y ¼S1 [ S2;0 [ S2;1 [ . . . [ S2;E�1 [ S3;0 [ S3;1 [ . . . [ S3;E�1.

Definition 6.1 (Optimal Relation �). Denote by � the opti<mal relation among the states in Y. For two states S and R,S � R if and only if for 8i 2 R and 8j 2 S, Oði; jÞ ¼ True.


Lemma 6.1. � has the property of irreflexivity in the sensethat for 8S 2 Y, S 6� S.

Proof. According to Algorithm 2, for 8i 2 S, Oði; iÞ ¼ Falseand hence S 6� S. Therefore � has the property ofirreflexivity. Ì

Lemma 6.2. � has the property of asymmetry in the sense thatfor 8S;R 2 Y, S � R) R 6� S.

Proof. According to the definition of �, for 8i 2 R and8j 2 S, if S � R, Oði; jÞ ¼ True. On the other hand,Oði; jÞ ¼ True implies that Oðj; iÞ ¼ False fromAlgorithm 2. Hence the lemma holds. Ì

Lemma 6.3. � has the property of transitivity in the sense thatfor 8R; S; T 2 Y, R � S ^ S � T ) R � T .

Proof. According to the definition of �, for 8i 2 R, 8j 2 S,a n d 8k 2 T , i f R � S, Oðj; iÞ ¼ True; i f S � T ,Oðk; jÞ ¼ True. On the other hand, Algorithm 2 indicatesthat if Oðj; iÞ ¼ True and Oðk; jÞ ¼ True, Oðk; iÞ ¼ True.Hence for 8i 2 R and 8k 2 T , Oðk; iÞ ¼ True. ThusR � T . Ì

Theorem 6.1. � is a strictly partial order relation on theset Y.

Proof. According to Lemmas 6.1, 6.2, and 6.3, � has theproperties of irreflexivity, asymmetry, and transitivity.Hence, it is a strictly partial order relation on the set Y. Ì

In the following, we will calculate the transition proba-bilities among different states to obtain the average deliverylatency and average delivery ratio of MARSA.

6.1.1 Probabilities of Transferring to a State � theCurrent One

Obviously in MARSA, S1 � S2;E�1 � . . . � S2;0 � S3;E�1 �. . . � S3;0. The main idea of MARSA is to continuously finda relay that is more optimal than the current one. In ouranalysis, different states represent different cases with eachindicating that the data arrives at one type of nodes.Moreover, the optimal relation ð�Þ among the states is astrictly partial order relation on the set Y. Hence, a nodei 2 R cannot deliver the data to another node j 2 T ifT � R. As a result, the probability of transferring to a state�the current one is zero except for two special cases: R ¼ S2;0

or S3;0.When the current state is S2;0 ðS3;0Þ, the relay is visiting

Zd. In this case, if there is no available spectrum in thecurrent interval and the relay is going to leave Zd in the nextinterval, the system transfers to a state S2;j ðS3;jÞ,j 2 f1; . . . ; E � 1g, which is � the current one. As a conclu-sion, the probabilities of transferring fromR to T ðT � RÞ canbe calculated as:

PR!T ðnÞ

¼0 R 6¼S2;0; S3;0

ð1��nÞPr�Zd;j R¼S2;0; T ¼S2;j; j2f1; . . . ; E�1gð1��nÞPr�Zd;j R¼S3;0; T ¼S3;j; j2f1; . . . ; E�1g:

8><>:

In the above formula, �n is the spectrum availability atZd in the nth interval; Pr is the average value of the

probabilities that the home nodes choose to leave Zd, whichis used to approximate the probability that the currentrelay chooses to leave Zd; �Zd;j is the probability that a nodevisits Zd after j intervals, which is used to approximate theprobability that the current relay visits Zd again after jintervals. Because node mobility demonstrates temporalregularity as stated in Section 3, the time at which peopleusually visit one place can be estimated to get �Zd;j.

6.1.2 Probabilities of Transferring to a State That isMore Optimal Than the Current One

The following definition is needed to calculate the transi-tion probabilities for this case.

Definition 6.2 (Predecessor State). The predecessor state ofS is denoted by S�, which satisfies: S� � S ^ 8R �S ! R � S�.

According to the above definition, S1 has no predecessorstate;S�2;jð or S�3;jÞ ¼ S2;jþ1ð or S3;jþ1Þ, where j ¼ 0; 1; . . . ;E � 2;S�3;E�1 ¼ S2;0; and S�2;E�1 ¼ S1.

In ICMCRNs, whether a node can deliver a data toanother node depends on the spectrum availability in boththe temporal and the spatial domains. For simplicity we use~� to represent the spectrum availability for the communica-tions between two nodes at different non-absorbing states,which is the average value of the spectrum availabilities inall zones except Zd at all time.

We assume that the number of nodes met by a relaywithin an interval follows a Poisson distribution with anexpectation �. Thus, a relay has the probability of e��k

k! tomeet k ðk ¼ 0; 1; 2; . . .Þ nodes within an interval. Among thecontacted nodes, the probability that one belongs to a stateT is kT k

ðkUk�1Þ, where kT k is the potency of the set T , denotingthe number of nodes in this set. Therefore, when thecurrent relay meets k nodes, the probability ð#k;T Þ that thestate T is the optimum state of the contacted nodes is:

#k;T ¼

kTkkUk�1

� �kT ¼ S1

1�PT �S

kSkkUk�1

� �k�

PG�T

kGkkUk�1

!k

o:w:

8>>><>>>:

For the nodes that will visit Zd after j intervals, i.e., thenodes in S2;j and S3;j, where j 2 f1; . . . ; E � 1g, eventhough they cannot deliver the data to other nodes thatare more optimal than themselves, a more optimal state, i.e.S2;j�1 and S3;j�1, will be automatically entered in the nextinterval. Hence, the probability (PR�!RðnÞ, whereR 2 Y n S2;E�1; S3;E�1) of transferring to any state exceptS2;E�1 and S3;E�1 from its predecessor state is:

PR�!RðnÞ ¼ 1� ~�þ ~�X1k¼0

e��k

k!#k;R�

¼ 1� ~�þ ~�e�� e� 1�P

R��TkT kkUk�1

� ��

�e�P

G�R�kGkkUk�1

� �!:

When the optimum state of the nodes contacted by thecurrent relay a 2 S2;j is S2;j�m, where j ¼ 2; . . . ; E � 1,


m ¼ 1; 2; . . . ; E � 2, j�m 9 0, or S3;i, where i ¼ 1; . . . ; E � 1;and there exists at least one available spectrum band at thecurrent interval, the state will transfer from S2;j to S2;j�m�1

or S3;i�1 at the next interval. Similarly, when the optimumstate of the nodes contacted by the current relay a 2 S3;j

i s S3;j�m, w h e r e j ¼ 2; . . . ; E � 1, m ¼ 1; 2; . . . ; E � 2,j�m 9 0, and there exists at least one available spectrumat the current interval, the state will transfer to S3;j�m�1 atthe next interval. Hence, the probability (PT!RðnÞ, whereT � R, T 6¼ R�, T 2 Y n S2;0; S3;0, R 2 Y n S2;E�1; S3;E�1) oftransferring to a more optimal state is:

PT!RðnÞ ¼ ~�X1k¼0

e��k

k!#k;R�

¼ ~�e�� e� 1�P

R��TkT kkUk�1

� �� e

�P

G�R�kGkkUk�1

� � !:

Obviously, the state transition probabilities from thestate S2;0 to other more optimal ones are zero. Because S3;0

is the optimum state in Y, there is no transition probabilityfrom this state to another more optimal one. Moreover, nostate can transfer to S2;E�1 and S3;E�1 except S2;0 and S3;0.Hence the probabilities PT!S2;E�1

ðnÞ and PT!S3;E�1ðnÞ, where

T 2 Y n S2;0; S3;0, are also zero.

6.1.3 Probabilities of Transferring to the AbsorbingState

In MARSA, the data may be delivered to the destinationonly when the relay visits Zd. Hence we have:

PT!QðnÞ ¼ �n T ¼ S2;0; S3;0

0 o:w:

n6.1.4 State Self-Transition ProbabilitiesFor the relay belonging to the state S1, if there is noavailable spectrum or if there exists at least one availablespectrum band but no more optimal node is contacted, thestate should keep unchanged. For the relay belonging toS2;0 or S3;0, if there is no available spectrum and the relaychooses to reside in Zd in the next interval, the state will beself-transferred. For the relay belonging to other states, the

state self-transition probabilities are zero because eventhough the relay cannot deliver the data to other nodes thatare more optimal than itself, the system will automaticallyenter a more optimal state. Thus the state self-transitionprobabilities can be calculated as:

PT!T ðnÞ ¼1� ~�þ ~�

P1k¼0

e��k

k! #k;S1T ¼ S1

ð1� �nÞð1� PrÞ T ¼ S2;0; S3;0

0 o:w:

8<:

whereP1

k¼0e��k

k! #k;S1¼ e�ð kS1k

kUk�1� 1Þ.

6.1.5 Average Delivery Latency and Delivery RatioLet � ¼ ða1; a2; . . . ; a2Eþ1Þ be a row vector of size 2E þ 1 witheach entry recording the probability at which the data deliveryprocess starts at a state in Y. Let WðnÞ be the matrix with a sizeof ð2E þ 1Þ � ð2E þ 1Þ representing the probabilities withwhich the data is transferred among the non-absorbingstates at the nth interval, i.e. WðnÞ ¼ fPR!SðnÞgð2Eþ1Þ�ð2Eþ1ÞðR;S 2 YÞ. Also let ! be a column vector of size 2E þ 1representing the probabilities PR!QðnÞ ðR 2 YÞ with whichthe data is transferred from the non-absorbing states to theabsorbing state at the nth interval.

Denote by X the time span before transferring into theabsorbing state Q. According to [33], the distribution of Xconforms to the discrete phase type distribution. Thus theprobability distribution PfX ¼ �g is:

PfX ¼ �g ¼1�

P2Eþ1k¼1 ak � ¼ 0

�! � ¼ 1�Q��1

i¼1 WðiÞ!; WðiÞ ¼ WðiþEÞ � � 2:

8<:

Hence, the average delivery latency of MARSA, i.e., theexpected time EðXÞ to the absorbing state, is:

EðXÞ ¼PfX ¼ 1g þX1�¼2

�PfX ¼ �g

¼�!þX1�¼2

��Y��1

i¼1

WðiÞ!:

Assume that the lifetime of a data is D. Denote by EðXDÞthe average delivery latency before the lifetime expires.Then, EðXDÞ ¼ �!þ

PD�¼2 ��

Q��1i¼1 WðiÞ!.

The average delivery ratio �D before the lifetime expirescan be calculated by:

�D ¼ PfX � Dg ¼ 1� �YDi¼1

WðiÞe:

In the above equation, e is a column vector with a size of2E þ 1 and each element being 1. then the complexities ofcomputing the average delivery latency and the averagedelivery ratio before the lifetime expires are both OðE3Þ.

6.2 Numerical ValidationTo validate the theoretical analysis, we construct asimulated network containing 10 � 10 zones. Each zoneat each interval may have one of two states: an idle stateand a busy state. We randomly assign the idle and busystates to different zones at different intervals. According to [8],temporal and geographical variations in the utilization of theassigned spectrum range from 15 percent to 85 percent. Hence

Fig. 2. Validation of the average delivery latency.


we set the spectrum availability to be a random numberranging from 15 percent to 30 percent for the busy state whilethat ranging from 70 percent to 85 percent for the idle state.Whether a data can be delivered successfully in a zone duringan interval depends on the spectrum availability at that zoneduring that interval. we set Th ¼ 0:5 in this simulation, whichis justified in Section 7.1.

There are 200 mobile nodes and 100 stationary nodes, withone for each zone. Each mobile node moves in light of thetime-variant community mobility (TVCM) model [34],where each node is randomly allocated some home zonesas its community and a cycle is a day including 24 intervals/periods, namely 24 hours. If a node is supposed to visit somezone at some interval, it has a larger probability (default valueis 0.8) to visit the desired zone. A node moves according to therandom direction mobility model within a zone.

We select all mobile nodes as sources and 16 stationarynodes as destinations residing at the zones crossing at the1st, 4th, 7th, and 10th rows and columns of the 10 � 10 gridnetwork. We test the delivery ratio and the delivery latencyof each communication pair. Each test is repeated 50 timeswith different random seeds for statistical confidence. Theaverage rate at which each node contacts with others ð�Þ isset according to the statistical values of the simulation. Inour theoretical analysis, for each destination, we determinethe type of the nodes at the beginning of the whole simulation

to get � and that at the beginning of each interval to get thestate transition matrix, and finally to calculate the averagedelivery ratio and latency.1

Due to page limitation, we only report the simulationresults when the destination zone ID ¼ 0; 3; 6; 9; 30; 33. Fig. 2shows the results of simulation and theoretical analysis on theaverage delivery latency when the date lifetime D ¼ 2000.Fig. 3 reports the results on the delivery ratio when the datalifetime varies. From both figures, we observe that the trendsof these two sets of results match each other.

7 PERFORMANCE EVALUATION

In this section, we evaluate the performance of MARSA.For this purpose we develop a stand-alone, discrete eventsimulator using both the real world data and the syntheticdata as the inputs. The design premise of existing CRNrouting schemes does not hold true in ICMCRN settingsbecause they assume that there exists at least one end-to-end path with available spectrum bands in the network.Hence in this paper, we only compare our routing schemewith its simplified versions and several traditional mobil-ity-assisted routing algorithms, which are all introduced inthe following:

1. When calculating the average delivery latency, the maximumvalue of i is set to be 2000 because

Q2000i¼1 WðiÞ approaches to zero.

2. The source-destination pair that is hard to communicatesuccessfully usually needs more relays and has a longer deliverylatency.

Fig. 3. Validation of the delivery ratio, Destination zone ID ¼ (a) 0# (b) ¼ 3# (c) ¼ 6# (d) ¼ 9# (e) ¼ 30# (f) ¼ 33#.


The Epidemic AlgorithmIn the epidemic algorithm, a node copies the data to everynewly encountered node that does not have the data ifthere exists an available channel.

The Basic AlgorithmIn the basic algorithm, if the source locates a neighbor-ing node that is more optimal than not only itself but also itsother neighbors, the source transfers the data to thatneighbor. Then the data will not be delivered to any othernode except the destination. In other words, the basicalgorithm is a single-copy single-relay version of MARSA.

The Time-First AlgorithmThere are two versions of the time-first algorithm: the single-copy single-relay (SCSR) version, which selects as the relaythe neighbor of the source that arrives in the destination zoneat the earliest time among all the neighbors of the source,and the single-copy multiple-relay (SCMR) version, whichcontinuously selects a node as the relay whose arrival timeto the destination zone is earlier than that of the current relayand its neighbors. If such a node is found and there exists atransmission opportunity, the current relay delivers the datato the new relay and then abandons the data. Obviously,the SCSR version is a variant of the basic algorithm while theSCMR version is a variant of MARSA. Neither of these twotime-first variations considers the spectrum availability whenmaking the routing decision.

The MobySpace AlgorithmThe main idea of MobySpace [22] is to forward the data to anode having a mobility pattern that is more similar to that ofthe destination. The similarity of two nodes’ mobilitypatterns is measured by the Euclidean distance of theMobyPoints. The MobyPoint of a node k is defined to bemk ¼ ðc1k; . . . ; cnkÞ, where cik is the probability that nodek visits location i, and n is the total number of locationsin the network.

Note that it is not meaningful to compare the averagevalues of the number of relays and the delivery latencieswhen the successful communication pairs of the comparedalgorithms are not the same. This is because the commu-nication pairs that succeed in one algorithm but fail inanother may worsen the average values of these twometrics in the algorithm with a higher delivery ratio.2

Hence in the following, we only count the communicationpairs that succeed in all algorithms for these two perfor-mance metrics.

7.1 Performance Evaluation Over the Real-WorldTrace Data

In this subsection, we evaluate our algorithms using theNational University of Singapore (NUS) student contacttrace data [35] as the input.

The NUS trace data contains the information regardingthe contact patterns among students of the NUS campusover large time scales through recording the class sche-dules and the student enrollment for each class. In [35], thistrace data is simplified from three aspects: 1) two studentsare in contact with each other iff they are in the sameclassroom at the same time; 2) contact sessions start on anhour and end on an hour, which means that ‘‘hour’’ is theunit of time for the contact duration; and 3) only thecontacts that take place during the class hours are used. Inour simulation, we adopt the same simplification to dealwith the NUS trace data.

Note that this trace data does not have any informationregarding the classroom assignments, which is importantto MARSA. To solve this problem, we allocate differentclassrooms to different classes according to the followingcriteria: 1) the classes whose durations overlap cannot beallocated to the same classroom; and 2) using as fewclassrooms as possible to hold as many classes as possible.

In our simulation study, each classroom is represented asa zone and the spectrum availability of each zone at eachtime slot is set according to the policy described in Section 6.

Fig. 4. Impact of Th on the delivery ratio.

Fig. 5. Impact of Th on the number of relays.

Fig. 6. Impact of Th on the average delivery latency.


2. According to the definition of home, each node’s home isset to the classroom(s) where the node has class(es) becauseit visits the classroom(s) frequently and regularly.

The data lifetime is set to the simulation duration.According to the NUS trace data, taking all classes at leastonce needs 76 hours. Hence, it is better to set the data lifetimelarger than 76 hours. In the following simulation, we setthe data lifetime D to be an integral multiple of 80 hours.

Though Algorithm 1 assumes that the destination staysat Zd during the process of data delivery, as we claimed inSection 5, MARSA can still work when the above assump-tion is not true by changing its way of usage. Hence, in thefollowing we evaluate the performance of MARSA whenthe destinations are static and mobile.

7.1.1 The Destinations are StaticWe assume there exist some static nodes in each classroomserving as the possible destinations. We randomly selectthe sources from the set of mobile nodes (the students from

the raw trace data) and the destinations from the set ofstatic nodes to form 100 source-destination pairs.

The threshold Th is a key parameter of MARSA and thebasic algorithm, which greatly affects the quality of the relays.In the following, we evaluate the impact of the threshold Thon the routing efficiency of these two algorithms.

Fig. 4 reports the delivery ratio of the basic algorithmand MARSA for Th ¼ 0:2, 0.5, and 0.8 when the data lifetimeD varies. Figs. 5 and 6 respectively report the average valuesof the number of relays and the average delivery latency ofthese algorithms when Th ¼ 0:2, 0.5 and 0.8. In this study weset D large enough to ensure that a source-destination paircan communicate successfully or unsuccessfully (due to thelack of appropriate relays arriving at the destination zonebefore the data expires) in all tests. Thus, we can evaluatethe average relay latency of both algorithms when theirsuccessful communication pairs are the same.

From Figs. 5 and 6, we observe that the delivery ratioand the delivery latency of both algorithms are the best

Fig. 7. Delivery ratio comparison using the real world trace when the destinations are (a) static. (b) Mobile.

Fig. 8. Delivery latency comparison using the real world trace when the destinations are (a) static. (b) Mobile.


when Th ¼ 0:5. This can be justified as follows. When Th islow ðTh ¼ 0:2Þ, the nodes that visit the destination zone atthe time when the destination zone is in the busy state canalso be selected as the relays; when Th is high ðTh ¼ 0:8Þ,even though some nodes visit the destination zone at thetime when the zone is in the idle state, they cannot beviewed as the ones with highly available visits to thedestination. However, the threshold Th ¼ 0:5 can achieve agood trade-off between the effect of the spectrum avail-ability and that of the visit time when selecting a relay.

In fact, we have tested the above metrics underTh ¼ 0:1 � 0:9 and found that the performances whenTh ¼ 0:1 and 0.9 are the worst. The reason is that a low ora high Th may reduce the effect of the spectrum availabilitywhen selecting a relay, which may damage the quality ofthe relays, decreasing the routing efficiency. We also havefound that the performances are almost the same whenTh ¼ 0:3 � 0:7. This is because in our simulation setting,a zone has only two states: busy and idle3; and in the busystate, the spectrum availability ranges from 15 percent to30 percent while in the idle state, it ranges from 70 percentto 85 percent. Therefore, when the threshold is set to 0.3� 0.7,all visits to the zone in the idle state can be viewed as highlyavailable ones. In other words, no visit to the zone in the busy(idle) state will be taken by mistake as a highly available one(regular one). Because the effects when Th ¼ 0:3 � 0:7 arealmost the same, we select the threshold from this rangerandomly. In the following study we set Th ¼ 0:5.

Figs. 7a and 8a respectively report the delivery ratios andthe delivery latencies of these six algorithms. According tothese results, we conclude that the performance of theepidemic algorithm are the best and that of MARSA is thesecond best. By considering the spectrum availability,the basic algorithm outperforms not only the SCSRversion but also the SCMR version of the time-firstalgorithm, which do not consider the spectrum availabil-ity, even though the basic algorithm only employs at mostone relay to deliver the data. The delivery ratio of theMobySpace algorithm is the lowest when D ¼ 80 while itis larger than that of the time-first algorithm whenD ¼ 160, and is even better than the basic algorithmwhen D ¼ 240; however, its delivery latency is the worstamong all algorithms.

The number of relays utilized by the six algorithms arereported in the first row of Table 1. The results indicate thatthe number of relays in the epidemic algorithm is the

highest, and that of the MobySpace is the second highest,while others involve a relatively small number of relays.

7.1.2 The Destinations are MobileTo support mobile destinations, MARSA and its variants,namely the Basic, TF-SCSR4, and TF-SCMR change theirusage ways: the source sends to each home zone of adestination one copy of the data so that once thedestination comes back to one of its home zones, it willreceive the data. We randomly select the sources anddestinations from the set of mobile nodes (the studentsfrom the raw trace data) to form 100 source-destinationpairs. We test the routing metrics of the communicationsbetween each source-destination pair. Each test is repeated50 times with different random seeds for statisticalconfidence. In this simulation, we still set Th ¼ 0:5 becausethe rule of assigning the spectrum availability to each zoneis not changed.

Figs. 7b and 8b respectively report the delivery ratios andthe delivery latencies of these six algorithms. The number ofrelays are reported in the second row of Table 1. Accordingto these results, we observe that the delivery ratio andlatency of the epidemic algorithm are still the best by payingthe highest cost because the number of relays it needs is themaximum; the basic algorithm outperforms TF-SCSRbecause it considers not only the time and space domainsbut also the spectrum domain in routing decision; theperformance of MARSA is the second best by paying amodest cost because it considers the spectrum availabilityand the power of multi-relay multi-copy at the same time.Though MobySpace has the worst performance, the numberof relays it employs is the minimum and only one data copy isneeded to support the mobile destination case.

Finally, we test the impact of the buffer size on therouting performance of MARSA and its variants. We as-sume the buffer size of each node is 10 (data units) and letmultiple sources send data to multiple destinations(including mobile and static) at the same time. We foundthat the routing performances of MARSA and its variantsdo not change much when the number of source-destination pairs is under 20,000 and the routing perfor-mances begin to decrease when the number of source-destination pairs is greater than 20,000. The average deliverylatencies of MARSA and its variants increase by 10 percent-27 percent and the average delivery ratios decrease by 12percent-17 percent when the number of the source-destina-tion pairs is 250,000.

7.2 Performance Evaluation over Synthetic TraceIn this subsection, we evaluate MARSA when the destina-tion is either static or mobile for the synthetic trace dataintroduced in Section 6.2.

For the static destination case, we assume that thereexist some static nodes at each zone serving as possibledestinations. We randomly select the sources from the setof mobile nodes and the destinations from the set of staticnodes to form 100 source-destination pairs. For the mobile

TABLE 1Num. of Relays Comparison Using the Real World Trace

3. In reality, the channels used by the primary users usually havethe busy and idle states, resulting in the same two states of theavailable spectrum usage for the SUs in each zone at each interval,which makes the same states for each zone at each interval. Hence, ourassumption is reasonable.

4. When the destination is mobile, if it has multiple home zones,multiple data copies are needed in this algorithm. However, for eachhome zone of this destination, only one copy is needed. Hence, wekeep the name of this algorithm i.e. TF-SCSR, unchanged.


destination case, we randomly select the sources and thedestinations from the set of mobile nodes to form 100source-destination pairs. We test the routing metrics of thecommunications between each source-destination pair.Each test is repeated 50 times with different random seedsfor statistical confidence.

Figs. 9 and 10 respectively report the delivery ratios andthe delivery latencies of these six algorithms when thedestinations are static and mobile. The number of relaysutilized by the six algorithms are shown in Table 2. Theresults are similar to those obtained from the real worldtrace data and can be justified based on the same reasons.

We also tested the impact of the buffer size on therouting performance of MARSA and its variants using thesynthetic trace data. The buffer size of each node is still setto 10 data units. We found the routing performances ofMARSA and its variants have little change when thenumber of source-destination pairs is under 500 andbeyond that, the routing performances begin to decrease.

The average delivery latencies of MARSA and its variantsincrease by 10 percent-15 percent and the average deliveryratios decrease by 5 percent-12 percent when the number ofthe source-destination pairs is 2000.

8 CONCLUSION

This paper is the first one to bring the idea of mobility-assisted routing to mobile ad-hoc CRNs. The proposedrouting algorithm named MARSA takes into account notonly the chance that a node meets the destination, butalso the chance of successful communications when twonodes meet. We present a detailed analysis on MARSA andemploy simulation study to validate our analysis. Theperformance of MARSA is also evaluated based on bothreal world trace data and synthetic data. By comparingwith state-of-the-art mobility based algorithms, we observethat MARSA clearly performs well in terms of deliverylatency and delivery ratio. In our future research, we will

Fig. 9. Delivery ratio comparison using the synthetic trace when the destinations are (a) static. (b) Mobile.

Fig. 10. Delivery latency comparison using the synthetic trace when the destinations are (a) static. (b) Mobile.


jointly consider relay selection and spectrum management(spectrum sensing, sharing, and mobility) for ICMCRNs asinterference does affect the effectiveness of relay selection.

ACKNOWLEDGMENTS

This work has been supported by the National Natural ScienceFoundation of China (61171014, 61272475, 61371185, 61202410,10771010 and 11301015), Beijing Natural Science Foundation(No.1132008, Stochastic Analysis with uncertainty and appli-cations in finance), PHR (No. 201006102), 111 Talent ProjectFund of BJUT, P.R. China, the Fundamental Research Fundsfor the Central Universities (2013NT57), the Project Sponsoredby SRF for ROCS, SEM, and the National Science Foundationof the US (CNS-1265311). S. Wang is the corresponding author.

REFERENCES

[1] M. Song, C. Xin, Y. Zhao, and X. Cheng, ‘‘Dynamic SpectrumAccess: From Cognitive Radio to Network Radio,’’ IEEE WirelessCommun., vol. 19, no. 1, pp. 23-29, Feb. 2012.

[2] W. Li, X. Cheng, T. Jing, Y. Cui, K. Xing, and W. Wang,‘‘Spectrum Assignment and Sharing for Delay Minimization inMulti-Hop Multi-Flow Crns,’’ IEEE J. Sel. Areas Commun., vol. 31,no. 11, pp. 2483-2493, Nov. 2013.

[3] Z. Cai, S. Ji, J. He, and A.G. Bourgeois, ‘‘Optimal DistributedData Collection for Asynchronous Cognitive Radio Networks,’’in Proc. IEEE ICDCS, June 2012, pp. 245-254.

[4] S. Ji, A.S. Uluagac, R. Beyah, and Z. Cai, ‘‘Practical Unicast andConvergecast Scheduling Schemes for Cognitive Radio Networks,’’J. Comb. Optim., vol. 10, no. 1, pp. 1-17, Jan. 2012.

[5] Z. Cai, S. Ji, J. He, L. Wei, and A.G. Bourgeois, ‘‘Distributed andAsynchronous Data Collection in Cognitive Radio Networkswith Fairness Consideration,’’ IEEE Trans. Parallel Distributed Syst.,vol. 25, no. 8, pp. 2020-2029, Aug. 2014.

[6] Z. Lin, H. Liu, X. Chu, and Y.-W. Leung, ‘‘Ring-Walk Rendez-vous Algorithms for Cognitive Radio Networks,’’ Adhoc SensorWireless Netw., vol. 16, no. 4, pp. 243-271, 2012.

[7] W. Huang and X. Wang, ‘‘Capacity Scaling of General CognitiveNetworks,’’ IEEE/ACM Trans. Netw., vol. 20, no. 5, pp. 1501-1513,Oct. 2012.

[8] Spectrum Policy Task Force ReportFCC, Washington, DC, USA,Nov. 2002.

[9] X. Huang, D. Lu, P. Li, and Y. Fang, ‘‘Coolest Path: SpectrumMobility Aware Routing Metrics in Cognitive Ad Hoc Networks,’’ inProc. IEEE ICDCS, June 2011, pp. 182-191.

[10] S. Ji, R. Beyah, and Z. Cai, ‘‘Minimum-Latency BroadcastScheduling for Cognitive Radio Networks,’’ in Proc. IEEE SECON,June 2013, pp. 389-397.

[11] W. Li, X. Cheng, T. Jing, and X. Xing, ‘‘Cooperative Multi-HopRelaying via Network Formation Games in Cognitive RadioNetworks,’’ in Proc. IEEE INFOCOM, Apr. 2013, pp. 971-979.

[12] P.L.M. Pan, C. Zhang, and Y. Fang, ‘‘Joint Routing and LinkScheduling for Cognitive Radio Networks Under Uncertain Spec-trum Supply,’’ in Proc. IEEE INFOCOM, Apr. 2011, pp. 2237-2245.

[13] Y. Shi and Y.T. Hou, ‘‘A Distributed Optimization Algorithm forMulti-Hop Cognitive Radio Networks,’’ in Proc. IEEE INFOCOM,Phoenix, AZ, USA, Apr. 2008, pp. 1292-1300.

[14] L. Ding, T. Melodia, S.N. Batalama, and J.D. Matyjas, ‘‘DistributedRouting, Relay Selection, Spectrum Allocation in Cognitive and

Cooperative Ad Hoc Networks,’’ in Proc. IEEE SECON, June 2010,pp. 1-9.

[15] W. Feng, J. Cao, C. Zhang, and C. Liu, ‘‘Joint Optimization ofSpectrum Handoff Scheduling and Routing in Multi-Hop Multi-Radio Cognitive Networks,’’ in Proc. IEEE ICDCS, Montreal, QC,Canada, June 2009, pp. 85-92.

[16] Q. Guan, F. Yu, S. Jiang, and G. Wei, ‘‘Prediction-Based TopologyControl and Routing in Cognitive Radio Mobile Ad Hoc Networks,’’IEEE Trans. Veh. Technol., vol. 59, no. 9, pp. 4443-4452, Nov. 2010.

[17] S. Jain, K. Fall, and R. Patra, ‘‘Routing in a Delay TolerantNetwork,’’ in Proc. SIGCOMM, Oct. 2004, pp. 145-158.

[18] C. Liu and J. Wu, ‘‘Scalable Routing in Delay Tolerant Networks,’’in Proc. MobiHoc, Montreal, QC, CA, Sept. 2007, pp. 51-60.

[19] W. Zhao, M. Ammar, and E. Zegura, ‘‘Controlling the Mobility ofMultiple Data Transport Ferries in a Delay-Tolerant Network,’’ inProc. IEEE INFOCOM, Miami, FL, USA, Mar. 2005, pp. 1407-1418.

[20] S. Wang, M. Liu, X. Cheng, and M. Song, ‘‘Routing in PocketSwitched Networks,’’ IEEE Wireless Commun., vol. 19, no. 1,pp. 67-73, Feb. 2012.

[21] S. Wang, M. Liu, X. Cheng, Z. Li, J. Huang, and B. Chen,‘‘Opportunistic Routing in Intermittently Connected Mobile P2PNetworks,’’ IEEE J. Sel. Areas Commun., vol. 31, no. 9, pp. 369-378,Sept. 2013.

[22] J. Leguay, T. Friedman, and V. Conan, ‘‘Evaluating MobilityPattern Space Routing for DTNs,’’ in Proc. IEEE INFOCOM, Apr.2006, pp. 1-10.

[23] S. Wang, M. Liu, X. Cheng, Z. Li, J. Huang, and B. Chen,‘‘HeroVA Home Based Routing in Pocket Switched Networks,’’ inProc. Int’l Conf. Wireless Algorithms, Syst., Appl., Aug. 2012, pp. 20-30.

[24] M.C. Gonzalez, C.A. Hidalgo, and A.L. Barabasi, ‘‘UnderstandingIndividual Human Mobility Patterns,’’ Nature, vol. 453, no. 7196,pp. 779-782, June 2008.

[25] C. Song, Z. Qu, N. Blumm, and A. Barabasi, ‘‘Limits of Predictability inHuman Mobility,’’ Science, vol. 327, no. 5968, pp. 1018-1021, Feb. 2010.

[26] D. Willkomm, S. Machiraju, J. Bolot, and A. Wolisz, ‘‘Primary UserBehavior in Cellular Networks and Implications for DynamicSpectrum Access,’’ IEEE Commun. Mag., vol. 47, no. 3, pp. 88-95,Mar. 2009.

[27] L. Kovanen, ‘‘Structure and Dynamics of a Large-Scale ComplexSocial Network,’’ M.S. thesis, Helsinki Univ. Technol., Espoo,Finland, Oct. 2009.

[28] FCCUnlicensed Operation in the TV Broadcast Bands, Washing-ton, DC, USA, Tech. Rep. FCC 08-260, Nov. 2008.

[29] M. Fitch, M. Nekovee, S. Kawade, K. Briggs, and R. Mackenzie,‘‘Wireless Service Provision in TV White Space with CognitiveRadio Technology: A Telecom Operators Perspective and Expe-rience,’’ IEEE Commun. Mag., vol. 49, no. 3, pp. 64-73, Mar. 2011.

[30] H. Li, X. Cheng, K. Li, C. Hu, N. Zhang, and W. Xue, ‘‘RobustCollaborative Spectrum Sensing Schemes for Cognitive RadioNetworks,’’ IEEE Trans. Parallel Distributed Syst., vol. 25, no. 8,pp. 2190-2200, Aug. 2014.

[31] P. Jacquet and T. Clausen, ‘‘Optimized Link State RoutingProtocol,’’ in Internet RFC3626, Oct. 2003.

[32] T. Spyropoulos, K. Psounis, and C.S. Raghavendra, ‘‘EfficientRouting in Intermittently Connected Mobile Networks: TheSingle-Copy Case,’’ IEEE/ACM Trans. Netw., vol. 16, no. 1,pp. 63-76, Feb. 2008.

[33] G. Latouche and V. Ramaswami, Introduction to Matrix AnalyticMethods in Stochastic Modeling. Philadelphia, PA, USA: SIAM, 1999.

[34] W. Hsu, T. Spyropoulos, K. Psounis, and A. Helmy, ‘‘ModelingTime-Variant User Mobility in Wireless Mobile Networks,’’ inProc. IEEE INFOCOM, May 2007, pp. 758-766.

[35] V. Srinivasan, M. Motani, and W.T. Ooi, ‘‘Analysis andImplications of Student Contact Patterns Derived from CampusSchedules,’’ in Proc. ACM MobiCom, Sept. 24-29, 2006, pp. 86-97.

Jianhui Huang received the PhD degree incomputer science from Xi’an Jiaotong University,China, in 2009. He is currently a ResearchAssistant at the Institute of Computing Technol-ogy, Chinese Academy of Sciences. His currentresearch interests include the mobility manage-ment, routing and cloud computing.

TABLE 2Num. of Relays Comparison Using the Synthetic Trace


Shengling Wang received her PhD in 2008 fromXi’an Jiaotong University. After that, she did herpostdoctoral research in Department of Comput-er Science and Technology of Tsinghua Univer-sity. Then she worked as an assistant andassociate professor respectively from 2010 to2013 in the Institute of Computing Technology ofChinese Academy of Sciences. Now, she is anassociate professor in College of InformationScience and Technology, Beijing Normal Uni-versity. Her current research interests include the

mobility management, routing and load balancing in wireless andmobilenetworks.

Xiuzhen Cheng received the MS and PhDdegrees in computer science from the University ofMinnesotaVTwin Cities, in 2000 and 2002, respec-tively. She is an Associate Professor at theDepartment of Computer Science, The GeorgeWashington University, Washington, D.C. Her cur-rent research interests focus on cognitive radionetworks, mobile handset networking systems(mobile health and safety), wireless and mobilecomputing, sensor networking, wireless and mobilesecurity, andalgorithmdesignandanalysis.Shehas

served on the editorial boards of several technical journals and the technicalprogram committees of various professional conferences/workshops. Shealso has chaired several international conferences. She worked as aprogram director for the USNational Science Foundation (NSF) from April toOctober in 2006 (full time), and from April 2008 to May 2010 (part time). Shereceived the NSF CAREER Award in 2004. She is a member of ACM.

Min Liu received the BS and MS degrees incomputer science from Xi’an Jiaotong Universi-ty, China, in 1999 and 2002, respectively, andthe PhD in computer science from the GraduateUniversity of the Chinese Academy of Sciencesin 2008. She is currently a professor at theNetworking Technology Research Centre, Insti-tute of Computing Technology, Chinese Acade-my of Sciences. Her current research interestsinclude Mobility Management (including handoffmanagement and location management), Wire-

less Resource Management, Delay/Disruptive-Tolerant Networks.

Zhongcheng Li received the BS degree incomputer science from Peking University, China,in 1983, and MS and PhD degrees from Instituteof Computing Technology, Chinese Academy ofSciences, China, in 1986 and 1991, respectively.From 1996 to 1997, he was a visiting professor atUniversity of California at Berkeley. He iscurrently a Professor of Institute of ComputingTechnology, Chinese Academy of Sciences. Hiscurrent research interests include next genera-tion Internet, dependable systems and networks,

and wireless communication.

Biao Chen received the BS degree in computerscience from Fudan University in China and theMS degree in mathematics and PhD degree incomputer science from Texas A&M University.After graduation, he joined the Department ofComputer Science in University of Texas atDallas as a faculty member. Currently, he is avisiting professor in the Department of Computerand Information Science of University of Macau.His research interests include distributed sys-tems, networking, and security. He is a member

of Sigma Xi and ACM.

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


mobility-assisted routing in intermittently connected mobile cognitive radio networks

Documents