Topology Management in CogMesh: ACluster-based Cognitive Radio Mesh Network

Tao Chen, Honggang Zhang, Gian Mario Maggio, and Imrich ChlamtacCREATE-NET

Via Solteri 38, Trento, 38100, Italy

Abstract—As the radio spectrum usage paradigm shiftingfrom the traditional command and control allocation scheme tothe open spectrum allocation scheme, wireless mesh networksmeet new opportunities and challenges. The open spectrumallocation scheme has potential to provide those networks morecapacity, and make them more flexible and reliable. However, thefreedom brought by the new spectrum usage paradigm introducesspectrum management and network coordination challenges. Inthis paper, we study the network formation problem in cognitiveradio based mesh networks. A cluster-based approach is proposedto form a mesh network in the context of cognitive radio scenario.Moreover, a topology management algorithm is developed tooptimize the cluster configuration with regard to the networktopology. The prominent feature of the proposed approach liesin the capability to adapt the cluster configuration to networkand radio environment changes.

I. INTRODUCTION

Radio spectrum usage is undergoing a paradigm shift fromthe traditional command and control allocation to the openspectrum access[1]. Cognitive radio (CR), which was firstcoined by Mitola in 1999 [2], is a promising approach toachieve open spectrum sharing flexibly and efficiently [3] [4].CR is an intelligent wireless communication system that isaware of its radio environment and is capable of adaptingits operation to the statistical variations of incoming radiofrequency (RF) stimuli [3]. The research on cognitive radio hasalready penetrated into different kinds of wireless networks,and covered a variety of fields, including spectrum sensing,channel estimation, dynamic spectrum sharing, medium accesscontrol (MAC), and routing [4]. In this paper, we focus ourstudy on the networking formation issue of a cognitive radiobased ad hoc network, in which a number of primary users andsecondary users form a mesh type network using the detectedunoccupied frequency band (i.e. spectrum holes) dynamicallyand independently. We name this kind of network as CogMesh.

Basically, a CogMesh network can be regarded as a multi-channel multi-access network, in which the available channelsof a node undergoes dynamic changes during the node’s lifetime. It is different from the traditional mesh or ad hocnetworks in the sense that it can opportunistically utilizevarious spectral holes for smooth peer-to-peer communicationsby virtue of the unique CR functionalities. Correspondingly,the topology management of a CogMesh network is affectedby two main facts: first, a common control channel may notalways exist in the whole network; and second, the networktopology changes over time according to the presence of

primary users and secondary users. Therefore, a distributedcontrol plane is required by the network. However, untilnow most proposed spectrum control protocols for ad hocopen spectrum sharing networks assume the availability ofa common control channel [4]. Zhao et al [5] observed thatthough very limited number of global common channels existin such a network, local neighbors share many channels withothers. A distributed grouping scheme is thus proposed tosolve the common control channel problem. Nevertheless, anefficient neighbor discovery process, which is critical for openspectrum access networks, is not found in [5]. Considering thenature of CogMesh, we employ a cluster based approach tosolve this problem the following reason: the nodes in CogMeshnetworks can be grouped according to the spectrum holedistribution; therefore, the cluster based approach can ease thespectrum management task.

Lots of cluster formation algorithms have been proposedso far for ad hoc networks [6]-[7]. They are different onthe criteria to select clusterheads. To utilize those approachesin CogMesh networks, there are several critical problems: 1)They usually assume a single channel radio on each node; 2)They are designed for fixed network topology, and hence lackthe ability to adapt to dynamic physical topology changes;3) Most of them only guarantee the network connectivity.The cluster configuration may not be optimized; 4) Someapproaches need the full topology of the network. Consideringthose problems, a unique approach is demanded for CogMeshnetworks. In this paper, we propose a distributed cluster-basedapproach to provide efficient communications in a large scalecognitive radio based mesh network. The proposed mechanismis able to adapt the network topology to network and radioenvironment changes.

The remaining of the paper is organized as follows. Thenetwork model and MAC protocol are described in section II.Then in section III we introduce the neighbor discoveryprocess. The topology management algorithm, which aims forthe network topology optimization, is described in section IV.The network connectivity properties under the proposed clusterformation mechanism is discussed in section V. Next, the per-formance of the proposed algorithm is studied in section VI.Finally, the conclusion is drawn in section VII.

II. SYSTEM MODEL

The CogMesh network exploits spectrum holes for com-munication. The spectrum holes are white or gray spaces

which are free of primary users or partially occupied bylow-power interferers [8]. Most of nodes in the network aresecondary users equipped with cognitive radio modules, whichare capable of detecting and utilizing spectrum holes efficientlyin a distributed way. A given number of spectrum holesare available for the whole network and can be identifiedby each node after effective spectrum sensing and channelstate estimation. Upon spectrum holes channels are built andidentified by their unique channel IDs. In this sense, theproblem studied in this paper becomes a dynamic spectrumaccess (DSA) problem. Note that DSA can be treated as afunctionality or an element of CR.

For each node its available spectrum holes depend on itslocation. We assume the spectrum holes of a node changein time but with a relatively slow rate, and the CR nodesmove only at a slow speed. The network topology, therefore,is relatively dynamic with stable status. Symmetric links areassumed in the network. A typical CogMesh network isillustrated in Fig. 1, in which a grid means a region sharingsame spectrum holes. Although this kind of model is notrealistic in the physical world, it provides adequate abstractto study network formation issues at the link layer.

{ 1,2 ,3 }

1

1

1

{1 ,3}

{1 ,2 ,3}{ 1,3 }

{2 ,3}{3 }

12

{ 2,3 }

{2 ,3}

{ 1,3 }21

2

23

P rim ary us erso n c hannel nC lus terhead ,mas ter c hannel nG atewayO rd inary no de

{1 ,2,3} C hannel lis to f a no de

n

m

Fig. 1. A CogMesh network.

The CogMesh network is constructed by local nodes form-ing 1-hop clusters, and the clusters interconnected to a singlenetwork. An 1-hop cluster here is characterized by the prop-erties that the ordinary nodes of the cluster is 1-hop awayfrom the clusterhead, and the cluster members locally share acommon control channel.

In the CogMesh network, the cluster formation and inter-cluster connection are performed distributively based onnodes’ neighbor information. Traditional approaches rely on acommon control channel for neighbor discovery [9]. Howeverit is not the case in our network since no common channelis always available due to the dynamic network changes. Tosolve this control channel problem, we propose a new MACprotocol.

Corresponding to the CogMesh network, channel accesstime in each cluster is divided into a sequence of superframes,whose structure is shown in Fig. 2. Except the data period,all other periods are related to the cluster formation. Thebeacon period issued by the clusterhead conveys the clusterID and cluster control information. The cluster ID is the

node ID of the clusterhead. Here we assume each nodehas a unique ID, which can be the MAC address of thenode. The neighbor broadcast period (NBP) is used by eachcluster member to broadcast its node ID, cluster ID, clustersize, and 1-hop neighbor list in an allocated mini-slot. Aneighbor list includes the node IDs of neighbors, their clusterIDs, the size of clusters, their master channels, and availablechannels of each neighbor. We define the master channel as thecontrol channel of a cluster. The internal random access period(RAP) is a slotted period for cluster members exchangingcontrol messages. The public RAP is used for clusterheadsexchanging inter-cluster control messages, such as neighborlist exchanging. As in IEEE 802.22 [10], there are one orseveral spectrum detection periods inserted in a superframe,used for nodes detecting spectrum holes.

B ea c onN eighb orBro ad ca s t

Pe riodD a ta Pe rio d

P riv ateR ando mA c c es s

Pu blicRa nd o mA c ce ss

Su pe rfra meS pe ctrum D ete ction Pe rio d

Fig. 2. Superframe structure.

III. NEIGHBOR DISCOVERY AND INITIAL CLUSTER SETUP

In our approach, the neighbor discovery and cluster forma-tion processes are combined together through the life time ofthe network. At the bootstrap phase of the CogMesh network,a given number of nodes need to form a network. We call thephase those nodes form clusters and make inter-connectionas the initial cluster setup (ICS) phase. In this phase, eachnode listens and detects one of its spectrum holes during agiven period of time, waiting for beacons on that spectrumhole. Usually, the node orders its channels with frequency andstarts its detecting process from the lowest one.

At the ICS phase, three cases occurs during a listeninginterval: no message comes; a beacon comes in the listeninterval; or neighbor messages come but no beacon comes.In the first case, the node forms a cluster on the listeningchannel and becomes the clusterhead. In the second case, thenode requests to join the cluster through the public RAP ofthe cluster. In the third case, the node detects neighbor clustersbut it is 2-hop away from clusterheads. The node exchangesneighbor information with the detected neighbor through thepublic RAP of that neighbor cluster, and shifts to next channelto listen. If a node can not find a cluster to join, it starts a newcluster on a randomly chosen channel. A clusterhead discoversits 1-hop, 2-hop, or 3-hop away neighbor clusters from thecollected neighbor information and makes interconnection bychoosing gateway nodes. The ICS phase stops when all initialnodes join clusters and clusters form interconnections.

A node in a cluster discovers neighbor nodes and clusterson different channels by periodically probing non-master chan-nels. It shifts to a non-master channel when it is idle duringa superframe. An intelligent algorithm can be developed here

to choose the non-master channel according to the neighborinformation. For instance, if it discovers new 2-hop neighborson a non-master channel, it detects neighbors on that channelfirst.

IV. TOPOLOGY MANAGEMENT

There are several motivations for topology management inCogMesh networks. First of all, the random nature of theICS phase makes the formed clusters hardly being optimizedin line with the physical topology. The number of clusterscan be reduced while maintaining the network connectivity.As a result, the control overhead is reduced, and the spec-trum efficiency is improved. Secondly, in the cognitive radioscenario, the available channels for each node fluctuate withregard to the radio environment. Consequently, clusters needto be optimized time by time so as to adapt to the radioenvironment. Since few cluster number means few inter-clustercommunication and few hops to reach other nodes, in thispaper, reducing the cluster number becomes the optimizationgoal.

The cluster optimization problem can be considered as adominating set (DS) problem in the graph theory. The DSproblem consists of finding a subset of nodes with the follow-ing properties: each node is either in the DS, or is adjacent to anode in the DS [11]. In our network, the DS is the collection ofclusterheads. The cluster optimization problem is therefore tofind a minimal dominating set (MDS) of the CogMesh networkaccording to its physical topology. The physical topology ispresented by a graph G = (V,E), where V is the set ofnetwork nodes, and E is the set of wireless links betweennodes. The MDS problem is proven to be an NP-hard problemeven when the complete network topology is known [7].However, a sub-optimum DS can be obtained through a localminimum election of the dominators by a heuristic algorithm.The algorithm is run periodically and distributively on eachnode and only relies on the discovered neighbor informationto determine the locally optimized cluster configuration. As aresult, the collection of clusterheads is gradually converged toa sub-optimum DS.

Input: Neighbor information of node AcurCHSet = N

CH(1)A ;

Form G = (V,E)where V =

⋃CHi∈N

CH(1)A

MCHi

⋃A,

and {U, V } ∈ E if U ∈ N1V ;

[CHSet, ClusterSet] =GetClustersFromDS(G, A, |curCHSet|);

if |CHSet| < |curCHSet| thenReconfigureClusters(ClusterSet);

end

Fig. 3. Main function for cluster optimization.

When the physical topology changes due to the events suchas new nodes joining the network, nodes leaving the network,or radio environment changing, the affected nodes or clusters

TABLE INOTATION

CHi clusterhead i

MCHiall members of CHi

C(A) available channel list of node A

NkA k-hop neighbors of node A

NkA(c) k-hop neighbors of node A on channel c

NCH(k)A k-hop neighbor clusterheads of node A

| · | set size

are reconfigured to absorb the changes. The optimizationalgorithm is performed thereafter to optimize the changedphysical topology. The basic rule for the reconfiguration iswhen an affected node currently belongs to no cluster, it takesaction to associate with one cluster or start a new cluster. Inother cases, the clusterheads coordinate the changes. Note thatafter reconfiguration, gateway nodes of affected clusters mayneed renegotiation.

The algorithm is shown in Fig. 3 and works as follows.For convenience, we use the notation as shown in Table I.From the neighbor list, a node, denoted as the node A, obtainsthe node set VA, which includes all members of its 1-hopneighbor clusters and its host cluster. It is the target nodeset to be optimized. The objective is to construct clustersbased on a MDS of the graph GA = (VA, EA) so that thenumber of clusters in VA can be minimized. GA is the physicalconnection graph of the node set VA, in which EA is the setof links between the node pair of VA. Each channel betweentwo nodes has a link in EA. The MDS is obtained by aheuristic algorithm [12] as shown in Fig. 4. The algorithm

Function [CHSet,ClusterSet] =GetClustersFromDS(G, myID, CHNum)

c′ = arg maxc∈C(myID)

(|N1myID(c)

⋂V |);

CHSet ← myID;ClusterSet ← {CH:myID, Master Channel: c′,

Members: N1myID(c′)

⋂V

⋃myID};

S = V \N1myID(c′)

⋃myID;

while S �= ∅ and |CHSet| ≤ CHNum doforeach ni ∈ S do

Nbri = maxc∈C(ni)

(|N1ni

(c)⋂

S|);ci = arg max

c∈C(ni)(|N1

ni(c)

⋂S|);

endi = arg max

(i|ni∈S)(Nbri);

CHSet ← ni;ClusterSet ← {CH: ni, Master Channel: ci,

Members: N1ni

(ci)⋂

S⋃

ni};S = S\N1

ni(ci)

⋂S

⋃ni;

endreturn;

Fig. 4. Function for getting clusters from DS.

takes the multiple channels of a node into account. First, acluster is formed by taking the node A as the clusterheadand the channel with maximal degree as the master channel.The 1-hop neighbors of the node A in VA are assigned tothe cluster if they share the master channel with the nodeA. The members of the new cluster are eliminated from VA.The residual nodes are handled as the following: As in a MaxDegree algorithm [11], a node with max degree on a channelis chosen to form a cluster with corresponding neighbors inorder until the number of new clusters exceeds the numberof old clusters, or all nodes join the network. Finally, the newcluster configuration comes out with the selected clusterheads,master channels and members of each cluster. If the numberof resultant clusters is smaller than the current one, the nodeA starts a negotiation process to reconfigure its surroundingclusters.

To start the negotiation process, the node A sends re-arrangement requests to the target clusterheads, indicatingthe gain that can be obtained from the rearrangement, andthe reconfiguration instruction. The gain is the total numberof clusters being reduced if the rearrangement is taken. Aclusterhead, once accepts the request, sends an acknowledge tothe node A. The node A negotiates with the target clusterheadsto complete the remaining configuration process only afterreceiving all acknowledges back. Otherwise, it cancels theprocess so as to avoid increasing the cluster number by anincomplete reconfiguration.

As seen from Fig. 4, since the algorithm running on thenode A will stop when it finds that the resultant cluster numberobtained by the algorithm exceeds its current number of 1-hopneighbor clusterheads N

CH(1)A , the time complexity to obtain

the optimized cluster configuration is in relation to NCH(1)A ,

which is usually far less than the total number of nodes in thenetwork. As the time complexity to reconfigure the clusters aswell as the gateway is all proportional to N

CH(1)A , we get the

time complexity of the algorithm equal to O(NCH(1)

A ), where

NCH(1)

A is the average 1-hop neighbor clusterhead number pernode in the network.

One thing worth mentioning is the stability and overhead ofthe algorithm. If the algorithm changes the network topologytoo often, the network may be unstable for service support.Moreover, the frequent running of the algorithm produceadditional control overhead. To solve those problems, thetime interval to activate the algorithm on each node can beproperly chosen so that a balance is achieved among theagility, stability, overhead reduction. Note that each node canhave its own time interval to perform the algorithm.

V. DISCUSSION OF NETWORK CONNECTIVITY

In this section we show that if the neighbor discovery isperfect and the nodes of the network are fully connected in thephysical topology graph, the network formed by the proposedalgorithm achieves a full connection at the link layer.

Assuming the neighbor discovery process is perfect, a nodewill finally discover all its 1-hop and 2-hop neighbors in

finite time. After the ICS phase, a node will either join acluster or form its own cluster. As a result, the collection ofclusterheads forms a DS. In this set, the maximum distancea clusterhead to another closest clusterhead is three. To provethis, assuming that distance is four, according to our clusterformation algorithm, the middle node between those twoclusterheads must belong to a clusterhead. Therefore thisclusterhead is exactly 2-hop away from those two clusterheads,which leads to a contradiction. The proof that two clusterheadscan not be 5 or more hops away can be obtained from theresult that two clusterheads can not be 4-hop away iteratively.According to the Theorem 1 in [13], which stated if theclusterheads form a dominating independent set in the currentnetwork graph, then a connected backbone is guaranteed toarise if each clusterhead establishes connections to all otherclusterheads that are at most at a distance of three, the networkafter the ICS phase is connected by clusters at the link layer.

Fig. 5. Number of clusters under different algorithms, in stationary channelscenario.

Fig. 6. Average cluster size under different algorithms, in stationary channelscenario.

Since the cluster optimization algorithm does not exclude anynode from new formed clusters or original clusters, we caneasily get that the network connectivity does not change afterthe cluster optimization process. Moreover, if the events suchas nodes leaving the network or radio environment changesdo not change the physical topology graph of the network,

after network reconfiguration, the link layer connectivity isnot affected.

VI. SIMULATION RESULTS

Simulation experiments are conducted to evaluate the per-formance of the proposed algorithm and compare with otheralgorithms. The setup of the simulation is following. As shownin Fig. 1, following the Poisson distribution, nodes of theCogMesh network are randomly placed in a 600m × 600m2-dimension square. The maximum transmission range of anode is set to 100m. The available channels of a node aredetermined by its location in the square. The square is divided

Fig. 7. Number of clusters after ICS phase, in stationary channel condition,with various spectrum holes.

Fig. 8. Number of clusters after Lowest ID algo, in stationary channelcondition, with various spectrum holes.

into 16 equal size sub-squares. Secondary users in the samesub-square share identical available channels. The availablechannels for secondary users in a sub-square are randomlypicked from a channel pool (CP). We set that each sub-squarehas at least one available channel.

The simulation consists of two scenarios in line with twochannel conditions. In the stationary channel scenario, theavailable channels of each node are fixed during the lifetime of the node. In this scenario, we employ two referencealgorithms for performance comparison. The first is the lowest

ID algorithm (Lowest ID) [11], in which the node with thelowest ID among its neighbors has the highest priority to bea clusterhead. The second is the max degree algorithm (MaxDegree) [11], in which the node with max degree among itsneighbors forms a cluster first. We name our algorithm as localminimal dominating set (LMDS) algorithm. In the dynamicchannel scenario, the available channels of each sub-squarechange periodically in a way that during two consecutivechanges, the channels for each node are frozen for affectednodes rejoining the network and being optimized by theproposed algorithm.

Fig. 9. Number of clusters after Max Degree algo, in stationary channelcondition, with various spectrum holes.

Fig. 10. Number of clusters after LMDS algo, in stationary channel condition,with various spectrum holes.

Fig. 5 and Fig. 6 show the number of clusters and aver-age cluster size obtained by different algorithms in the firstscenario. The number of channels in the CP is set to two. Itis seen from Fig. 5 that after the ICS phase, the number ofclusters is high. However, after the optimization, the numberof clusters is reduced significantly and kept below 30 evenwhen the total number of nodes becomes 210. It shows thecapability of the proposed algorithm to merge small clusters.Moreover, the proposed algorithm exhibits similar efficiencyas Max Degree, and better performance than Lowest ID.

The efficiency of three algorithms under different CP sizesis shown in Fig. 8-10. When the size of the CP is one, thenetwork reduces to a single channel network. This case hasthe smallest cluster size since a node can connect to any othernode in its transmission range. When the size of the CP isgreater than one, similar performance is shown under eachalgorithm. Fig. 7 shows that in the ICS phase, the size of theCP has small impact on the number of clusters.

Fig. 11. Cluster statistic in dynamic channel scenario, before LMDS algo.

Fig. 12. Cluster statistic in dynamic channel scenario, after LMDS algo.

Fig. 11 and Fig. 12 show the number of clusters andthe average cluster size in the dynamic channel scenario.As expected, when the network has multiple channels, thechannel changes significantly affect the cluster configuration.As seen from Fig. 11, the number of clusters dramaticallyincreases after the radio environment change. However, afteroptimization, the number of clusters is reduced to the samelevel before the radio environment change. It verifies that theproposed algorithm is able to adapt to the radio environmentchange under different channel conditions and maintain thecluster configuration to a relatively optimal level.

VII. CONCLUSION

In this paper, we study the network construction problemin the CogMesh networks. The CogMesh networks oppor-tunistically utilize the spectra resources for communication,

and thus provide unique features different from traditionalwireless mesh networks. In order to solve the common controlchannel problem, we propose a cluster-based approach for theneighbor discovery and control when global control channelsare absent, and provide a topology management algorithmfor the topology optimization. From the simulation, it showsthat the proposed algorithm is able to converge to a sub-optimal configuration in the dynamic channel change scenario.For the future work, the optimization algorithm needs to befurther elaborated. A statistic learning approach may be a goodchoice to get the algorithm quickly adapt to radio environmentchanges.

REFERENCES

[1] Federal Communications Commission, “Spectrum Policy Taks Force,”Rep. ET Docket No. 02-135, Nov., 2002.

[2] J. Mitola III and G. Maguire Jr, “Cognitive Radio: Making SoftwareRadios More Personal,” Personal Communications, IEEE [see also IEEEWireless Communications], vol. 6, no. 4, pp. 13–18, 1999.

[3] S. Haykin, “Cognitive Radio: Brain-Empowered Wireless Communica-tions,” Selected Areas in Communications, IEEE Journal on, vol. 23,no. 2, pp. 201–220, 2005.

[4] I. Akyildiz, W. Lee, M. Vuran, and S. Mohanty, “Next Genera-tion/Dynamic Spectrum Access/Cognitive Radio Wireless Networks: ASurvey,” Computer Networks, vol. 50, no. 13, pp. 2127–2159, 2006.

[5] J. Zhao, H. Zheng, and G. Yang, “Distributed Coordination In DynamicSpectrum Allocation Networks,” Dyspan 2005, Baltimore, Nov, 2005.

[6] C. Lin and M. Gerla, “Adaptive Clustering For Mobile Wireless Net-works,” Selected Areas in Communications, IEEE Journal on, vol. 15,no. 7, pp. 1265–1275, 1997.

[7] A. Amis, R. Prakash, T. Vuong, and D. Huynh, “Max-Min d-ClusterFormation In Wireless Ad Hoc Networks,” INFOCOM 2000, IEEE,vol. 1, 2000.

[8] T. Weiss and F. Jondral, “Spectrum Pooling: An Innovative Strategy ForThe Enhancement Of Spectrum Efficiency,” Communications Magazine,IEEE, vol. 42, no. 3, pp. S8–14, 2004.

[9] J. So and N. Vaidya, “Multi-Channel MAC For Ad Hoc Networks:Handling Multi-Channel Hidden Terminals Using A Single Transceiver,”Proceedings of the 5th ACM international symposium on Mobile ad hocnetworking and computing, pp. 222–233, 2004.

[10] C. Cordeiro, K. Challapali, and M. Ghosh, “Cognitive PHY and MAClayers for Dynamic Spectrum Access and Sharing of TV Bands,”Wireless Internet Conference (WICON 2006), Boston, Aug, 2006.

[11] L. Bao and J. Garcia-Luna-Aceves, “Topology Management In Ad HocNetworks,” Proceedings of the 4th ACM international symposium onMobile ad hoc networking & computing, pp. 129–140, 2003.

[12] V. Chvatal, “A Greedy Heuristic For The Set-Covering Problem,”Mathematics of Operations Research, vol. 4, no. 3, pp. 233–235, 1979.

[13] I. Chlamtac and A. Farago, “A New Approach To The Design AndAnalysis Of Peer-To-Peer Mobile Networks,” Wireless Networks, vol. 5,no. 3, pp. 149–156, 1999.