CogMesh: A Cluster-based Cognitive Radio Network

Transcription

1 ogmesh: luster-based ognitive Radio Network Tao hen, Honggang Zhang, Gian Mario Maggio, and Imrich hlamtac RETE-NET Via Solteri 38, Trento, 38100, Italy bstract s the radio spectrum usage paradigm shifting from the traditional command and control allocation scheme to the open spectrum allocation scheme, wireless ad-hoc networks meet new opportunities and challenges. The open spectrum allocation scheme has potential to provide those networks more capacity, and make them more flexible and reliable. However, the freedom brought by the new spectrum allocation scheme introduces spectrum management and network coordination challenges. Moreover, wireless ad-hoc networks usually rely on a common control channel for operation. Such a control channel may, however, not always available in an open spectrum allocation scheme due to the interference and the need for coexistence with primary users of the spectrum. Instead, common channels most likely exist in a local area.in this paper, we propose a cluster-based framework to form a wireless mesh network in the context of open spectrum sharing. lusters are constructed by neighbor nodes sharing local common channels, and the network is formed by interconnecting the clusters gradually. We identify issues in such a network and provide mechanisms for neighbor discovery, cluster formation, network formation, and network topology management. The unique feature of this network is its ability to intelligently adapt to the network and radio environment change. I. INTRODUTION The radio spectrum usage is undergoing a paradigm shift from the traditional command and control allocation scheme to the open spectrum allocation scheme. The report published by the Spectrum-policy Task Force of Federal ommunications ommission (F) in 2002 [1], which aims at improving the way of utilizing the spectrum resource, catalyzed intensive research activities in this new field of open spectrum sharing. The cognitive radio (R), which was first coined by Mitola in 1999 [2], is a promising approach to achieve open spectrum sharing flexibly and efficiently [3] [4]. It is an intelligent wireless communication system that is aware of its radio environment and is capable of adapting its operation to statistical variations of the incoming radio frequency (RF) stimuli [3]. The research on the R has already penetrated into a variety of wireless networks and each layer of the network protocol stack [4]. IEEE [5] is the first standard based on the R. IEEE h [6] is going to bring R functions into Worldwide Interoperability for Microwave ccess (WiMX) networks for homogenous and heterogenous network coexistence. number of cognitive radio testbeds have been developed based on different architectures and radio technologies[7][8] [4]. The research on the R covers a wide range of areas, including spectrum analysis, channel estimation, spectrum sharing, medium access control (M), and routing. In this paper, we focus our study on the network formation issue of a decentralized R based mesh network, which is formed by secondary users of spectrums opportunistically utilizing the detected spectrum holes in an ad-hoc way. We name this kind of cognitive radio based mesh network as ogmesh. Basically, the ogmesh can be regarded as a multichannel multi-access network, in which the available channels of a node undergoes dynamic changes during the life time of the node. It is different from the traditional mesh or ad hoc networks in the sense that it can opportunistically utilize various spectral holes for smooth peer-to-peer communications by virtue of the unique R functionalities. orrespondingly, the topology management of the ogmesh is affected by two main factors: first, a common control channel may not always available for the whole network; and second, the topology of the network changes over time according to the presence of primary users and secondary users. Therefore, a distributed control plane becomes necessary for the network. However, until now most proposed spectrum control protocols for ad hoc open spectrum sharing networks assume the availability of a common control channel [4]. For instance, an open spectrum sharing protocol proposed in [9] extended the ideas of Request To Send/lear To Send (RTS/TS) exchange and Network llocation /07/$ IEEE

2 Vector (NV) in the IEEE M protocol for the open spectrum access, where the access operation relies on a common control channel. Ma proposed a dynamic open spectrum sharing M (DOSS-M) protocol for similar networks [10], in which a common channel is used for signaling. Zhao et al observed that although a very limited number of global common channels exist in a network, neighbor nodes may locally share numerous channels with others [11]. distributed grouping scheme was proposed in [11] to solve the common control channel problem. However, an efficient neighbor discovery process, which is important for an open spectrum access network, is absence. The challenges remaining for an ad-hoc open spectrum sharing network include the neighbor discovery, distributed control, and multi-hop communication. onsidering the nature of the ogmesh, we propose a cluster based approach to solve those problems. The contributions of this paper are the following: It investigates the issues to setup an ad-hoc open spectrum sharing network which coexists with primary users of the spectrums It proposes a decentralized cluster-based architecture to form a large scale network. It develops a distributed control scheme at the link layer. The neighbor information is used wisely to setup and maintain the network topology It provides mechanisms to adapt the network topology to network and radio environment changes. The remaining of this paper is organized as follows: The network model and assumptions are discussed in section II. The network architecture are described in section III. The motivation to use the cluster-based structure is explained here. Next, based on the proposed network architecture, the cluster formation, network formation, and topology management issues are identified and discussed individually: Section IV introduces the M protocol; Section V describes the spectrum hole detection procedure; Section VI discusses the neighbor discovery and cluster formation; Section VII describes the intercluster connection; Section VIII addresses topology management issues and provides a distributed solution. The performance of the distributed algorithm for topology management is studied in section IX. The conclusion is drawninsectionx. II. NETWORK MODEL In a typical cognitive radio scenario, users of a given frequency band are classified into primary users and secondary users [4]. Primary users are licensed users of that frequency band. Secondary users, on the other hand, are unlicensed users that opportunistically access the spectrum when no primary users operating on that frequency band. The ogmesh network is formed by secondary users which utilize the spectrum holes for communications. The spectrum holes are white or gray spaces which are free of primary users or partially occupied by low-power interferers [7]. The nodes in the ogmesh are equipped with cognitive radio modules, which are capable of detecting and utilizing spectrum holes efficiently in a distributed way. given number of spectrum holes are available for the whole network and are identified by their unique channel IDs after effective spectrum sensing and channel state estimation. For each node its available spectrum holes depend on its location. For simplicity, we use the terms channel and spectrum hole interchangeably. We assume the spectrum holes detected by a node change in time but at a relatively slow rate, and the R nodes move only at a slow speed. The network topology, therefore, is relatively dynamic with stable status. Symmetric links are assumed in the network. simplified interference avoidance model, i.e., the overlay spectrum sharing model, is used in our network [4], where secondary users only use the spectrums that have not been occupied by primary users. Once detecting the presence of a primary user on a given frequency band, the secondary user simply vacates that band. It is the case in IEEE network [5]. nother interference avoidance model is the interference temperature model, which allows the coexistence of primary and secondary users [3]. However, it complicates the studied problem here and is thus not considered in this paper. In summary, the network we study in this paper has following characteristics: a) ll secondary users are cognitive radio enabled. b) Spectrum holes are location dependent and time varying, which means secondary users may own different channel set. c) There is no global common channel for the network. However, local common channels exist among adjacent users. d) Secondary users form an ad-hoc mesh network, which means distributed control scheme is deployed. e) In a same area, parallel transmissions on different channels are permitted. III. NETWORK RHITETURE Traditional multi-channel wireless systems usually use a global control channel for neighbor discovery and access control [12],[13]. However, it is not the case in the ogmesh. The ogmesh uses local control channels 169

3 for channel access control, where a distributed control scheme is applied. However, the dynamic changing of spectrum holes makes the channel control extremely complex. pure distributed control scheme like arrier Sensing Multiple ccess/ollision voidance (SM/) may not work. ccordingly, we introduce the concept of cluster into the network. node forms a cluster on a channel and invites adjacent nodes sharing the same channel to join its cluster. For convenience, the control channel of a cluster is called the master channel of that cluster. The node forming the cluster becomes the clusterhead, which is responsible for intra-cluster channel access control and inter-cluster communications. The channel access scheme is described in section IV, and the cluster formation process is detailed in section VI. By negotiating gateway nodes between clusters, clusters are interconnected into a large network. gateway node is a member of one cluster that is able to reach the member of other cluster. The cluster interconnection is illustrated in Fig. 1, and described in section VII. From Fig. 1, we can see clusters are interconnected in two cases: two clusterheads are connected by one gateway node, or connected by two gateway nodes when no node is 1-hop neighbor of two clusterheads. There are three types of members in a cluster: the clusterhead, ordinary node, and gateway node. For a network to be properly constructed, protocols hannel 1 Fig. 1. luster B luster lusterhead Gateway node hannel 2 luster Ordinary node lusterhead Gateway node Ordinary node lusters interconnected by gateway nodes and mechanisms are necessary to specify the behaviors of nodes under different network conditions. We divide these protocols and mechanisms into five parts and discuss them in the following sections: a) M protocol to support cognitive radio based multiple channel access b) Spectrum hole detection c) Neighbor discovery and cluster formation. d) Inter-cluster connection. e) Topology management. IV. M PROTOOL The cluster formation and inter-cluster connection are performed distributively based on the neighbor information of nodes. We provide mechanisms to enable nodes exchange their 1-hop and 2-hop neighbors information, which includes neighbors s identity and their channel list. In the ogmesh, a node may only know partial of its neighbors at the initial stage. The clusters are formed based on the partial neighbor information. s nodes gradually collect more neighbor information based on the proposed neighbor discovery algorithm, clusters are reconstructed and interconnected to a more reliable network structure. ll these are done by the M protocol. The M protocol proposed here is a hybrid M protocol that consists of guaranteed access and random access periods, where the guaranteed access period is used for data transmission in and between clusters, and the random access period is used for control message exchange. For each cluster, channel access time is divided into a sequence of superframes. Each superframe consists of five main periods as shown in Fig. 2. The beacon period is issued by the clusterhead. It contains the time synchronization, control and resource allocation information of the cluster. The following period is the neighborhood broadcasting period (NBP). It is divided into a number of fixed length mini-slots. Each member of a cluster occupies one mini-slot and uses it to broadcast its identity and 1-hop neighbor list. n entry in the neighbor list includes the identity of the neighbor and its channel list. The master channel of a neighbor is indicated in the cluster list, through which a node knows how to reach the neighbor cluster. preamble is put at the beginning of each mini slot for other nodes identifying the broadcasting message if they miss the beacon. Moreover, the time and duration of random access periods in this superframe is broadcast in the Frame Map period of its mini-slot. neighbor of this node, once receiving its neighborhood broadcasting message, has the chance to exchange its neighbor information with the node in the following random access period. The location of a member s mini-slot is announced by the clusterhead in the beacon period. The number of minislots in a superframe is limited by a system parameter in order to avoid too many nodes crowding in one cluster. Next comes the data period. Parallel transmissions are permitted in this period if the transmission sessions use different channels. Time division multiple access 170

4 (TDM) is used in each channel. Following the data period, an intra-cluster random access period (RP) is used for cluster members exchanging control messages. The superframe is ended by a public RP. This period has multiple purposes. It uses for a node joining the cluster, nodes exchanging neighbor information, or clusters exchanging control information. The slotted LOH is used to resolve collisions in the RPs. Besides five main periods, there are one or several Superframe Beacon Neighbor Broadcast Period Fig. 2. Spectrum D etection Period Data Period Superframe structure Private Random ccess Public Random ccess spectrum detection periods scheduled in a superframe. During these periods, all members of a cluster keep silence and detect spectrum holes. It is desirable to synchronize the spectrum detection periods of adjacent clusters so as to reduce the false alarm of primary users. The false alarm is an event that a secondary user incorrectly reports the presence of primary users due to the interference from other sources. Since the superframes of different clusters are not required to be synchronized, the location of the spectrum detection periods varies from cluster to cluster. Even in a cluster, their location varies from superframe to superframe. distributed algorithm is demanded to synchronize the spectrum detection periods of adjacent clusters, and determine the location of those periods in superframes on a superframe by superframe basis. Note that the design of the M protocol can benefit from the spread spectrum technique for the coexistence of multiple clusters, in which two kinds of spreading codes are used in the periods of a superframe. One is the public spreading code, which is globally known to every node and used by the beacon, NBP, and public RP for broadcasting control messages. The other is the private spreading code used for data transmission. clusterhead chooses the private spreading code in a way so that adjacent clusters use different codes. The code distribution algorithm is required here. V. SPETRUM HOLE DETETION The spectrum detection is performed by each node periodically. The techniques used for spectrum hole detection include matched filter detection, energy detection, cyclostationary feature detection [4], and so on. In [3], multitaper spectral estimation plus singular value decomposition (MTM-SVD) is suggested for spectrum hole detection. We assume nodes in the network are equipped with one or combination of those detection techniques, and are able to detect all spectrum holes accurately and efficiently. Synchronization in an ad-hoc network is a challenge due to the absence of centralized coordination in nature. However, there are solutions to partially solve this problem [14]. ordeiro proposed an algorithm in networks to synchronize the superframe of different base stations [5]. The network scenario of is similar to our network in the sense that they both use the spectrum in an opportunistic way. However, uses fixed base stations for the channel access control, while in the ogmesh, the clusterhead of a cluster is not fixed. Therefore a more sophisticated mechanisms is demanded in the ogmesh. The initial idea is that we only synchronize the detection periods instead of superframes. time stamp algorithm can be employed to solve this problem. Each new cluster is stamped by its created time. If it detects a neighbor cluster has a time stamp earlier than its own, the cluster synchronizes its detection period to that neighbor cluster and replaces its time stamp with the neighbor cluster s one. t the end, all clusters use the detection period of the earliest formed cluster. For a large scale network, it may take long time to achieve the synchronization. VI. NEIGHBOR DISOVERY ND LUSTER FORMTION The neighbor discovery and cluster formation process are introduced together since they are highly related. For convenience, we give the following definitions: the host cluster of a node is the cluster that the node belongs to; the neighbor cluster of a node is the cluster that the node does not belong to but has 1-hop neighbors as its members; the total neighbor clusters of all members of a cluster are called the cluster s neighbor clusters. The neighbor discovery is performed during clusters NBPs. When a node wants to join the network, it first detects the available channels. Then it scans one of its channels for a given period of time, waiting for beacons on that channel. The node starts the scanning process from the lowest frequency band channel, which is called the lowest channel. The scanning time on a channel is chosen so that it exceeds the period of the longest superframe. We call a scanning period as scanning interval, and the first scanning interval a new node starts as the first scanning interval. If there is a 171

5 H I {3} {1,3} B B E F D I {3} () H {1,3} () () E F D {1,3} G {1,3} () (E) (E) G (E) H B E F D I {3} () {1,3} H B () () E F D (I) I {3} () {1,3} (I) () () 1. onnect Graph 2. Stage I, luster formed 3. Stage II, luster E formed 4. Stage III, luster I formed {1,3} () G G {1,3} () (E) (E) (E) Fig. 3. luster formation neighbor cluster on the frequency band a node listens on, it is able to capture its beacon during a scanning interval. We divide the first scanning interval into three cases: no message comes; a beacon comes; or neighbor messages come but no beacon comes. In the first case, the node forms a cluster on the scanning channel and becomes the clusterhead. In the second case, the node requests to join the cluster through the public RP of the cluster. If the clusterhead accepts the request, it assigns a mini-slot to the requesting node. Starting from next superframe, the new joining node broadcasts its neighbor list in that mini-slot. However, if there is no empty minislot in a cluster, the clusterhead will reject the request. The requesting node then chooses the second lowest channel to start a new scanning process, or form its own cluster if finding the detected clusters are all full after iterating all channels. The third case means the node has neighbor clusters but it is 2-hop away from clusterheads. The node then records neighbor information, and tries to exchange neighbor information with that neighbor through the public RP of the corresponding neighbor cluster. fter that, it continues its scanning process on the next available channel. If the node can not find a channel satisfying the case one and two after iterating all channels, it starts its own cluster on a randomly chosen channel. fter a node joining a cluster, it periodically chooses from its channel list a non-master channel to scan so as to discover other neighbor nodes. n algorithm can be developed to intelligently choose the non-master channel according to the neighbor information the node detects. For instance, if it discovers new 2-hop neighbors on a non-master channel, it listens on that channel first. Let us explain the neighbor discovery and cluster formation by an example illustrated in Fig. 3. The numbers in the bracket close to each node are available channels of that node. The smaller number represents the lower spectrum hole. We assume the spectrum holes do not change during the cluster formation procedure. The edge between two nodes indicates they can hear each other. ssume the node is the first node forming the cluster on the channel 1. The cluster is labeled as the cluster. Its 1-hop neighbors B,, D listen on their lowest frequency band, i.e. the channel 1, detect the beacon issued by the cluster. They join the cluster through corresponding association processes. From the neighbor discovery process, the node B knows the node is its 1-hop neighbor, and the node D is its 2-hop neighbor. Next, the node E, F, G form a cluster on the channel 2. ssume the node E forms the cluster, labeled as the cluster E. The node F, G join the cluster E right after. The 172

6 node B listens on the non-master channel 2. It discovers E, F as its 1-hop neighbors, and G as its 2-hop neighbor. The cluster and E therefore are interconnected by the node B. Then, assume the node I forms the cluster I on the channel 3. The node H receives B s broadcasting message and detects B as its 1-hop neighbor. However, H can not receive beacons from the cluster. It starts a new listen process on the channel 3 and finally joins the cluster I. The node H informs B that its new neighbor list through the public RP of the cluster. The node B knows from H there is a cluster on the channel 3. It knows the neighbors H, I on the channel 3 through a scanning process on that channel. Furthermore, the node will know B has new neighbors H and I from the NBP of the cluster and finally know its neighbor I on the channel 3. t this stage, three clusters are formed, and the clusterheads has enough neighbor information for inter-cluster connection. The clusters then negotiate with each other to form a large network through the public RP of each other. VII. INTER-LUSTER ONNETION The network is formed by interconnecting clusters through gateway nodes. gateway node of two clusters is intermediate node, through which two clusters can exchange control messages and data. They are chosen by clusters through a certain algorithm. There are two cases for the inter-cluster connection: two clusters are overlapping, or non-overlapping. In the first case at least one node belonging to any of the clusters is 1-hop neighbor of two corresponding clusterheads. This node is called the intermediate node. One of the intermediate nodes is chosen by its clusterhead as the gateway node to reach the peer cluster. The gateway node in this case is called 1-hop gateway node. In the second case, if two clusters are non-overlapping but there are nodes belonging to two clusters can hear each other, they are chosen to interconnect two clusters. We call those nodes 2-hop gateway nodes. In the following, we describe the procedure of the cluster interconnection in detail. The 1-hop gateway interconnection is illustrated in Fig. 4. The node and are clusterheads of two clusters, named the cluster and, respectively. The node B1 and B2 are a member of the cluster and, respectively. The clusterhead knows it can reach the cluster through the node B1. It chooses the node B1 as the gateway node to the cluster and commands the node B1 to inform the clusterhead its choice. The node B1 listens on the beacon of the cluster and sends the gateway choice message to the clusterhead through the private random access period of the cluster. When the clusterhead has control messages send to the clusterhead, it firstly sends to the node B1. Then the node B1 sends to the clusterhead through s private random access period. In the reverse path, the clusterhead sends control messages to the node B1 in the cluster s private random access period. The node B1 relays the messages to the clusterhead. The clusterhead can select the node B2 as its gateway node to the cluster, and uses the path to B2 to when it has control messages send to the clusterhead. Note that it does not matter whether the cluster or operates on the same frequency band since the gateway node works in a store and forward mode. The 2-hop gateway interconnection is more complex since it involves the coordination of two gateway nodes. The 2-hop gateway interconnection is illustrated in Fig. 4. Similar to the 1-hop gateway scenario, the node, are clusterheads of cluster and, and the node B1 and B2 belongs to the cluster and, respectively. The cluster and are not overlapping. But the node B1 and B2 can hear each other. The clusterhead knows the cluster can be reached through the node B1. It initializes the gateway setup process by sending to the node B1 a request message. The node B1 relays the request to the node B2 in the cluster s public RP. The node B2 receives the request and relays it to the clusterhead. The acknowledgement flows from the clusterhead to through the node B2 and B1 following the similar procedure. s nodes discovering more neighbors through the NBPs, Ordinary node hannel 2 hannel 1 luster B1 luster B2 1-Hop Gateway Fig. 4. lusterhead Gateway node lusterhead Gateway node Ordinary node luster B2 lusterhead Gateway node B1 luster 2-Hop Gateways Inter-luster connection hannel 1 Ordinary node the network connectivity can be gradually improved. Based on link layer connections, the upper layer protocols are able to deliver end-to-end services. 173

7 VIII. TOPOLOGY MNGEMENT There are several motivations for efficient topology management in the ogmesh. First of all, the random nature of the I phase makes the formed clusters hardly being optimized results in line with the physical topology. The number of clusters can be reduced while maintaining the network connectivity. s a result, the control overhead is reduced, and the spectrum efficiency is improved. Secondly, in the cognitive radio scenario, the available channels for each node fluctuate with regard to the radio environment. onsequently, the topologies of clusters are not static, and the reconfigurations are required over time. For this matter, the topologies of clusters must be optimized time by time so as to adapt to the radio environment. In the following section, we examine network maintenance issues at the link layer, and provide corresponding solutions. The issues include: 1) Nodes join the network; 2) Nodes leave the network; 3) Spectrum holes of a node change; 4) lusterhead shifts the master channel; 5) luster merging.. Nodes Join Network When a new node joins the network, it first detects the spectrum holes. Then it starts the aforementioned process in section VI to join a cluster. The problem here is how the node knows its neighbors and how its neighbors on different master channels know this node. The node and its neighbors in the same cluster know each other through the neighbor discovery process. From the neighbor information it obtains, the node can identify potential neighbor clusters and their master channels. It can, thereafter, actively shift to the master channels of other clusters for the neighbor discovery. On the other hand, when its 1-hop neighbors detect a new neighbor, they will broadcast their updated neighbor list through neighbor discovery process. Other nodes in neighbor clusters, when discovering this node as their 2-hop neighbor, may schedule scanning processes to check if it is their 1-hop neighbor. fter several rounds of neighbor information exchange, the new node and its neighbors will obtain accurate neighbor information accordingly. B. Nodes Leave Network When a node leaves the network, it is important to inform its neighbors the leaving event timely. Note that a node may have different roles in a cluster. The cluster and neighbors of the leaving node should detect and handle the leaving event properly according to the node s role. There are two kinds of leaving events: a node may leave the network following a disassociation process or disappear suddenly due to node malfunction. We call the former the disassociation process and the latter the absent process, which are described in the following. Nodes with different roles have different disassociation processes. For an ordinary node, it informs its clusterhead through a specific message sent in the private RP of the host cluster. The clusterhead announces the leaving of the node by broadcasting a special message in its mini-slot. The members of the host cluster detect the leaving event through the special message, updating their neighbor list accordingly. Members of the host cluster will inform their neighbor clusters the leaving event. Those neighbor clusters in turn inform their members who are neighbors of the leaving node accordingly. When a gateway node disassociates from the network, it informs the clusterhead its leaving. The clusterhead will start a process to negotiate a new gateway node. The remaining leaving procedure follows the ordinary node disassociation process. The disassociation process of a clusterhead is little more complex. For a cluster, a secondary clusterhead can be optionally chosen by the primary clusterhead during the lifetime of the cluster, which is the member of the cluster with maximum 1-hop neighborhood to other cluster members. When a cluster has a secondary clusterhead, the primary clusterhead hands over the clusterhead role to the secondary clusterhead before it leaves the network. The members of the old cluster who are 1-hop neighbors of the new clusterhead remain in the cluster. Those who are 2-hop neighbors of the new clusterhead leave the cluster and start an aforementioned node joining process. The procedure to inform the neighbors the leaving event of the clusterhead is similar to that in the ordinary node disassociation process. If a cluster has no secondary clusterhead, the clusterhead broadcasts a cluster dismissed message to its members. The members then start node joining processes accordingly. The members may inform the neighbor clusters the disassociation of the clusterhead through public RP of those clusters. The neighbor clusters dismiss their gateway nodes to this cluster and update the neighbor lists of their members accordingly. In an absent process, a node leaves the network without informing its cluster. So it is important to detect the leaving event timely. For an ordinary node, the clusterhead detects the leaving event through the neighbor discovery process. node has to broadcast its neighbor list in the assigned mini-slot at least once during a 174

8 given number of superframe periods. The clusterhead and its members infer the leaving of the node through the absence of corresponding neighborhood broadcasting messages during that period. Once the leaving event is detected, the remaining procedure is similar to that in the disassociation process. If a clusterhead disappears, the members of the cluster will fail to received beacon. They can therefore infer the leaving of the clusterhead. If there is a secondary clusterhead, it takes over the clusterhead role. Otherwise, the cluster is dismissed. In addition to actively detecting the leaving events, each node maintains an expired timer for each entry in its neighbor list. Without receiving a neighborhood broadcasting message from its neighbor in the given expired time, the node deletes the entry of that neighbor from its neighbor list. Different expired time are used for different types of neighbors. The 1-hop neighbors have shorter expired time than 2-hop neighbors. The neighbors belonging to the same cluster have shorter expired time than those belonging to different clusters. The expired timer of 1-hop neighbor entry is renewed if the node receives neighborhood broadcasting message from the corresponding neighbor during the expired time. For a 2-hop neighbors entry, its expired timer is updated if the node finds that neighbor listed in received neighborhood broadcasting messages. Once an 1-hop neighbor entry is expired, the node inform its host and neighbor clusters accordingly.. Spectrum Holes hange Each node senses spectrum holes periodically. If it detects new spectrum holes, it simply informs its cluster and neighbor clusters accordingly. The new spectrum holes will not immediately affect the operation of the host and neighbor clusters. However, if some spectrum holes of a node are not available any more, the host and neighbor clusters may malfunction if their control functions rely on those spectrum holes. The network needs rapidly adapt to spectrum hole occupied events. If a node detects its master channel is not available, it hand over to one of its neighbor cluster, and inform its old host cluster and other neighbor clusters the change. If no neighbor cluster is available, it forms a cluster by itself. If a gateway node detects the channel used to connect the peer cluster become unavailable, it informs its clusterhead to adapt to the change. If a gateway node shifts to another cluster due to the spectrum hole change event, its previous host cluster is able to detect the handover event through the neighbor discovery process in NBP and deal with this event accordingly. In case the master channel of a clusterhead is not available, the cluster will be dismissed. D. luster Shift Master hannel clusterhead has the need to change it master channel for several reasons: the current master channel is too crowd because several neighbor clusters share the same channel; the channel quality deteriorates due to the increasing interference; there are many 1-hops neighbors on other channels that can be merged into this cluster; from historical statistic information, the current master channel needs to be vacated for primary users in some time periods. It is the clusterhead s duty to determine the channel shift. It inform its members the shift operation. fter all member nodes confirm the change, the clusterhead starts the beacon in the new master channel from the next superframe. The neighbor clusters of this cluster is informed after the shifting. The gateway nodes to neighbor clusters are reselected. In case a cluster can benefit from the channel shift while few of its members have no spectrum hole on the new master channel, it requests those members to leave the cluster. lgorithms can be customized to meet different performance goals. E. Merge lusters The clusters of the network are formed spontaneously according to the initial network conditions. The resultant topology may not be optimized. Moreover, due to the nature of open spectrum sharing, the network topology may undergo frequent changes during the lifetime of the network. The cluster structures, therefore, need to be adjusted in accordance with the network conditions. The cluster merging process provides a method to adapt cluster structures to the radio environment automatically and distributively. The clusters that share some properties are merged in a way that the overall number of clusters in the network is reduced. onsequently, the communication overhead is reduced. ccording to different criteria, different merging algorithm can be developed. For instance, statistic learning algorithms can apply here to well adapt to radio environment changes. Lots of cluster formation algorithms have been proposed for ad hoc networks so far [15]-[16]. They are different on the criteria to select clusterheads. However, they have some critical problems to be utilized in the ogmesh: They usually assume a single channel in the network; 175

9 They are designed for the fixed network topology, and therefore lack the adequate capability to adapt to dynamic physical topology changes; Most of them only guarantee the network connectivity. The cluster configuration may not be optimized; Some approaches need the full topology knowledge of the network. s a result, a unique cluster merging approach is demanded for a ogmesh network. We use a cluster merging algorithm based on the minimal dominating set (MDS) in the graph theory. The MDS problem is proven to be NP-hard even when the complete network topology is available [16]. However, a suboptimum DS can be obtained through a local minimum election of the dominators by a heuristic algorithm. The algorithm is run periodically and distributively on each node and only relies on the neighbor information to determine the locally optimized cluster configuration. s a result, the collection of clusterheads is gradually converged to a sub-optimum DS. We call the algorithm the local minimal dominating set (LMDS) merging algorithm. The algorithm works as follows. From its neighbor list, a node, denoted as the node, obtains the node set that includes all members of its 1-hop neighbor clusters and its host cluster. It is the target node set to be optimized. The objective is to construct clusters based on an MDS of the graph formed by the node set so that the number of clusters in the graph can be minimized. The MDS is obtained by a heuristic algorithm, which takes the multiple channels of a node into account. First, a cluster is formed by taking the node as the clusterhead and the channel with maximal degree as the master channel. The 1-hop neighbors of the node in the node set are assigned to the cluster if they share the master channel with the node. The members of the formed cluster are eliminated from the node set. The remaining nodes is processed as the following: s in the Max Degree algorithm [17], a node with max degree on a channel is chosen to form a cluster with corresponding neighbors in order until all nodes join the network. Finally, the new cluster configuration comes out with selected clusterheads, master channels and members of each cluster. If the number of resultant clusters is smaller than the current one, the node starts a negotiation process to reconfigure its surrounding clusters. Due to the space limitation, we do not list the detailed algorithm here. The performance of the algorithm is shown in section IX. IX. SIMULTION STUDY The setup of the simulation is the following. Multiple nodes of the ogmesh are randomly placed in a 600m 600m 2-dimension square according to the Poisson distribution. The maximum transmission range of each node is set to 100m. The available channels for a nodes are determined by its location in the square. The square is divided into 16 equal size sub-squares. Secondary users in a same sub-square share identical available channels. The available channels for secondary users in the sub-square is randomly picked from a channel pool (P). We specify that each sub-square has at least one available channel. Two reference algorithms are developed for performance comparison. The first is the lowest ID algorithm (Lowest ID) [17], in which the node with the lowest ID among its neighbors has the highest priority to form a cluster. The second is the max degree algorithm (Max Degree) [17], in which the node with max degree among its neighbors forms a cluster first. Fig. 5 shows the number of clusters obtained by different Fig. 5. luster statistic in stationary channel condition algorithms under the stationary channel condition, in which the detected spectrum holes do not change during the lifetime of nodes. The number of channels in the P is set to two. It is seen from Fig. 5 that after the I phase, the number of clusters is high. However, after the optimization, the number of clusters is reduced significantly and kept below 30 even when the total number of nodes becomes 210. It shows the capability of the proposed algorithm to merge small clusters. Moreover, the proposed algorithm exhibits similar efficiency as Max Degree, and better performance than Lowest ID. 176

10 The efficiency of the LMDS algorithms under different Fig. 8. algo luster statistic in dynamic channel condition, before LMDS Fig. 6. Number of clusters after I phase, in stationary channel condition, with various spectrum holes Fig. 9. algo luster statistic in dynamic channel condition, after LMDS Fig. 7. Number of clusters after LMDS algo, in stationary channel condition, with various spectrum holes Ps is shown in Fig. 7. From the figure, we can see the algorithm has the similar performance when the number of spectrum holes are greater than one. When the size of the P is one, the network is reduced to a single channel network. Fig. 6 shows that in I phase, the size of the P has small impact on the number of formed clusters. Fig. 8 and Fig. 9 present the number of clusters under the dynamic channel condition, in which the detected spectrum holes vary during the lifetime of the nodes. s expected, when the network has multiple channels, the channel changes significantly affects the cluster configu- ration. s seen from Fig. 8, the number of clusters dramatically increases after the radio environment change. However, after optimization, the number of clusters is reduced to the same level before the radio environment change. It verifies that the proposed algorithm is able to adapt to the radio environment change under different channel conditions. X. ONLUSION In this paper, we provide a framework to form a ogmesh network in the context of open spectrum sharing scheme. The network is constructed in a distributed way and provides coexistence with primary users of the spectrums. The basic unit of the network is the cluster, which is a sub-network formed by a group of neighbor 177

11 nodes sharing common channels, and coordinated by a selected node in the cluster called clusterhead. The network is constructed by interconnecting clusters after they learn each other through neighbor discovery process. This paper provides mechanisms for each node to efficiently exchange neighbor information over multiple channels. Moreover, issues in cluster formation, network formation, and topology management are addressed and corresponding solutions are provided. distributed topology management algorithm is proposed and its performance under various channel conditions is studied. It is the objective of this paper to bring traditional wireless ad hoc networks into the open spectrum sharing scenario. Note that it is a new area where lots of problems remains. For future work, we will develop efficient algorithms for neighbor discovery, channel allocation, and topology management. Moreover, we will evaluate the network performance under different network conditions. [12] J. So and N. Vaidya, Multi-hannel M For d Hoc Networks: Handling Multi-hannel Hidden Terminals Using Single Transceiver, Proceedings of the 5th M international symposium on Mobile ad hoc networking and computing, pp , [13] S. Wu,. Lin, Y. Tseng, and J. Sheu, New Multi- hannel M Protocol with On-Demand hannel ssignment for Multi-Hop Mobile d Hoc Networks, International Symposium on Parallel rchitectures, lgorithms, and Networks, I-SPN, pp , [14] Q. Li and D. Rus, Global lock Synchronization in Sensor Networks, omputers, IEEE Transactions on, vol. 55, no. 2, pp , [15]. Lin and M. Gerla, daptive lustering For Mobile Wireless Networks, Selected reas in ommunications, IEEE Journal on, vol. 15, no. 7, pp , [16]. mis, R. Prakash, T. Vuong, and D. Huynh, Max-Min d- luster Formation In Wireless d Hoc Networks, INFOOM 2000, IEEE, vol. 1, [17] L. Bao and J. Garcia-Luna-ceves, Topology Management In d Hoc Networks, Proceedings of the 4th M international symposium on Mobile ad hoc networking & computing, pp , REFERENES [1] Federal ommunications ommission, Spectrum Policy Taks Force, Rep. ET Docket No , Nov., [2] J. Mitola III and G. Maguire Jr, ognitive Radio: Making Software Radios More Personal, Personal ommunications, IEEE [see also IEEE Wireless ommunications], vol. 6, no. 4, pp , [3] S. Haykin, ognitive Radio: Brain-Empowered Wireless ommunications, Selected reas in ommunications, IEEE Journal on, vol. 23, no. 2, p. 201, [4] I. kyildiz, W. Lee, M. Vuran, and S. Mohanty, Next Generation/Dynamic Spectrum ccess/ognitive Radio Wireless Networks: Survey, omputer Networks, vol. 50, no. 13, pp , [5]. ordeiro, K. hallapali, and M. Ghosh, ognitive PHY and M layers for Dynamic Spectrum ccess and Sharing of TV Bands, Wireless Internet onference (WION), [6] J. Sydor, Messaging nd Spectrum Sharing Between d-hoc ognitive Radio Networks, IEEE International Symposium on ircuits and Systems, [7] T. Weiss and F. Jondral, Spectrum Pooling: n Innovative Strategy For The Enhancement Of Spectrum Efficiency, ommunications Magazine, IEEE, vol. 42, no. 3, pp. S8 14, [8] R. Brodersen,. Wolisz, D. abric, S. Mishra, and D. Willkomm, ORVUS: ognitive Radio pproach For Usage Of Virtual Unlicensed Spectrum, White paper, vailable for download from eecs. berkeley. edu/research/mm, [9] S. Sankaranarayanan, P. Papadimitratos,. Mishra, and S. Hershey, Bandwith Sharing pproach to Improve Licensed Spectrum Utilzation, Dyspan 2005, Baltimore, Nov, [10] L. Ma, X. Han, and. Shen, Dynamic Open Spectrum Sharing M Protocol for Wireless d Hoc Networks, Dyspan 2005, Baltimore, Nov, [11] J. Zhao, H. Zheng, and G. Yang, Distributed oordination In Dynamic Spectrum llocation Networks, Dyspan 2005, Baltimore, Nov,