Multi-hop scatternet formation and routing for large scale … · Multi-hop scatternet formation and routing 253 challenge to efﬁciently form a scatternet with basic requirements

Int. J. Ad Hoc and Ubiquitous Computing, Vol. 4, No. 5, 2009 251

Multi-hop scatternet formation and routing

for large scale Bluetooth networks

Wen-Zhan Song*

School of Engineering and Computer Science,Washington State University,Vancouver, WA 98686, USAE-mail: [email protected]*Corresponding author

Yu Wang

Department of Computer Science,University of North Carolina at Charlotte,Charlotte, NC 28223, USAE-mail: [email protected]

Chao Ren

School of Computer Science,Northwestern Polytechnical University,Xi’an, ShanXi 710072, ChinaE-mail: [email protected]

Changhua Wu

Department of Computer Science,Kettering University,Flint, MI 48504, USAE-mail: [email protected]

Xiang-Yang Li

Department of Computer Science,Illinois Institute of Technology,Chicago, IL 60616, USAE-mail: [email protected]

Abstract: This paper addresses the scatternet formation for large scale multi-hop Bluetoothnetworks. We first describe an efficient method to build a Connected Dominating Set (CDS)as the backbone of multi-hop Bluetooth network, then propose new algorithms to formthe dBBlue scatternets (Song et al., 2005) in each cluster. The final scatternet, M-dBBlue,guarantees the connectivity. Our experiment shows our scatternet seldom parks any node.We then propose a complete set of hierarchical routing methods for M-dBBlue which enablesthe self-routing inside each cluster. Moreover, our scatternet formation and routing algorithmdo not necessarily require position information of the node.

Keywords: Bluetooth; connected dominating set; dBBlue; multi-hop; scatternet formation.

Reference to this paper should be made as follows: Song, W-Z., Wang, Y., Ren, C., Wu, C.and Li, X-Y. (2009) ‘Multi-hop scatternet formation and routing for large scale Bluetoothnetworks’, Int. J. Ad Hoc and Ubiquitous Computing, Vol. 4, No. 5, pp.251–268.

Biographical notes: Wen-Zhan Song is an Assistant Professor in Computer Science fromWashington State University – Vancouver. His research interest spans sensor networks,peer-to-peer networks, distributed systems and algorithms. He received PhD from IllinoisInstitute of Technology in 2005, MS and BS Degree from Nanjing University of Science andTechnology in 1997 and 2000 respectively.

Copyright © 2009 Inderscience Enterprises Ltd.

252 W-Z. Song et al.

YuWang received the PhD Degree in Computer Science from Illinois Institute of Technologyin2004, theBEngDegree and theMEngDegree inComputerScience fromTsinghuaUniversity,China, in 1998 and 2000. He has been an Assistant Professor of Computer Science at theUniversity of North Carolina at Charlotte since 2004. His current research interests includewireless networks, ad hoc and sensor networks, mobile computing, and algorithm design.

Chao Ren is currently a PhD student at School of Computer Science in NorthwesternPolytechnical University, China. His research interests include data mining and networksecurity.

Changhua Wu obtained his PhD Degree in Computer Science from Illinois Institute ofTechnology in 2005, BA and MS Degrees from Hangzhou Dianzi University, formerlyHangzhou Institute of Electronic Engineering in 1998 and 2001 respectively. He is now anAssistant Professor at the Computer Science Department of Kettering University. His currentresearch areas include image analysis, computer vision, intelligent robot, andmobile networks.

Xiang-Yang Li is an Associate Professor of Computer Science at the Illinois Institute ofTechnology. He holdMS (2000) and PhD (2001) Degree at Computer Science fromUniversityof Illinois at Urbana-Champaign. He received his Bachelor Degree at Computer Science andBachelor Degree at Business Management from Tsinghua University, PR China in 1995.His research interests span wireless ad hoc and sensor networks, non-cooperative computing,computational geometry, optical networks, and cryptography.

1 Introduction

Wireless ad hoc networking has gathered significantresearch interests in the past years. Bluetooth (BluetoothSIG) is a promising low cost and low power wirelesstechnology, which enables portable devices to formshort-range wireless ad hoc networks based on a frequencyhopping physical layer. Bluetooth operates in theunlicensed 2.4GHz ISMband, with the frequency hoppingtechnique to alleviate the effects of the interference.The nominal bit rate of transmission is 1Mbps. It hasbeen widely predicted that Bluetooth will be the majortechnology for short range wireless networks and wirelesspersonal area networks. This paper deals with the problemof building multi-hop ad hoc networks using Bluetoothtechnology.

Bluetooth ad-hoc networking presents new technicalchallenges, such as scheduling, network forming androuting. According to the Bluetooth specification 2.0(Bluetooth SIG), when two Bluetooth devices come intoeach other’s communication range, one of them assumesthe role of master of the communication and the otherbecomes the slave. This simple one hop network is calleda piconet, and may include more slaves. The networktopology produced by the connection of piconets is calleda scatternet. There is no limit on the maximum number ofslaves connected to one master, although the number ofactive slaves at one time cannot exceed seven. If a masternode has more than seven slaves, some slaves must beparked. To communicate with a parked slave, a master hasto unpark it, thus possibly parking another active slaveinstead. The standard also allows multiple roles for thesame device: a node can be master in one piconet and aslave in one or more other piconets. However, one nodecan be active only in one piconet. To operate as a memberof another piconet, a node has to switch to the hopping

frequency sequence of the other piconet. Since each switchcauses delay (e.g., scheduling and synchronisation time),an efficient scatternet formation protocol can be one thatminimises the roles assigned to the nodes, without losingnetwork connectivity.

While several solutions and commercial products havebeen introduced for one-hop Bluetooth communication,the Bluetooth specification (Bluetooth SIG) does notindicate any method for multi-hop scatternet formation.The problem of scatternet formation has attracted manyattention with in past several years. Several criteria couldbe set as the objectives in forming scatternet. For example,the node degree of the scatternet should be small, if degreesof all nodes are smaller than eight, the scatternet can avoidparking any node. The formation and maintenance ofscatternet should have small communication overhead sothat it can be efficiently updated in dynamic networks.In this paper, we focus on scatternet formation for largemulti-hop ad hoc networks.

Previous literature on scatternet formation assumedthat devices are not able to communicate unless they havepreviously discovered each other by synchronising theirfrequency hopping patterns. Thus, even if all nodes arewithin direct communication range of each other, onlythose nodes, which are synchronised with the transmitter,can hear the transmission. Synchronising the frequencyhopping patterns is apparently a time consuming andpseudo-random process (Salonidis et al., 2001). In thispaper we assume that the problem of discovering allneighbours within transmission radius of a device isresolved by separate Bluetooth protocols (Salonidis et al.,2001; Basagni et al., 2002). These protocols are applicableas the pre-phase of our scheme.

This paper addresses the scatternet formation solutionsin large multi-hop Bluetooth networks. As shown inVergetis et al. (2005), it is a fundamental algorithmic

Multi-hop scatternet formation and routing 253

challenge to efficiently form a scatternet with basicrequirements for a large-scale Bluetooth network. Ourproposed method is based on a hierarchical structureto guarantee the network connectivity and provideefficient routing. We first propose a novel communicationefficient method to cluster the nodes, build a ConnectedDominating Set (CDS) as the backbone of multi-hopBluetooth network, and then construct the dBBluescatternets (Song et al., 2005) for each cluster, whichmakes self-routing within the cluster possible. A cluster isdefined by a dominator node and all its dominatee nodes.We propose several methods to construct the piconetson top of the backbone to provide high performance.The final scatternet, hereafter calledM-dBBlue, guaranteesthe connectivity, and its node degree is small in mostcases. Later our experiment showsmost nodes have degreesmaller or equal to seven.Moverover, all of ourmethodsdonot necessarily require position information of the nodes.

We then discuss how to route packets efficiently in ourscatternet using a hierarchical routing method. When twonodes want to communicate, the source node first sendsdata packets to the dominator node within its cluster usinga self-routing method supported by the dBBlue structure;the dominator node then forwards the data packets to thedominator node whose cluster contains the target node;the target dominator node then forwards the data to thetarget node using self-routing method provided by dBBluescatternet. Here the routing along the backbone could beany greedy methods (Karp and Kung, 2000; Kuhn et al.,2003) if position information is known, or on-demandrouting protocols such as AODV (Perkins and Royer,1999),DSR (Johnson et al., 2003), or table drivenprotocolssuch as DSDV (Perkins and Bhagwat, 1994).

In summary, the contributions of this paper are asfollows:

• We propose a new multihop scatternet formationalgorithm which first builds a CDS as the backbonethen forms the dBBlue scatternets inside eachcluster.

• For the CDS formation algorithm, we use themethod in Wan et al. (2002) to select the dominatorsand form clusters, and then propose a new efficientalgorithm (Algorithm 3) to find connectors to formthe connected backbone. Our new connectoralgorithm generates more connections than thespanning tree so that it enhances efficiency ofinter-cluster routing. Different from all previousmethods, we do not use local broadcast in ourmethod, since local broadcast cannot be performedefficiently in practice due to the constraint in MAClayer.

• When we form the dBBlue scatternets inside eachcluster, the method in Song et al. (2005) cannot beapplied directly since it only works for single-hopnetworks and the nodes in each cluster may not besingle-hop. We propose two novel methods to solvethe problems:

• decreasing the transmission radius to half sothat all dominatees of a dominator cancommunicate directly

• partition the dominattees into cliques such thateach cliques is single-hop.

In addition, we also propose two methods to buildthe one-hop scatternet dBBlue inside each clique. Theyapplies different role assignments for the dominatorsand connectors in the backbone to achieve some niceproperties. The final scatternet, M-dBBlue, guaranteesthe connectivity, and its node degree is small in mostcases.

We also describe how to perform IP-based routingin the M-dBBlue scatternet. The inter-cluster routing ishandled by a modified Bluetooth based RIP protocol onthe backbone, and the intra-cluster routing is derived fromthe self-routing mechanism of dBBlue.

The rest of the paper is organised as follows.In Section 2, we introduce our network model andreview two algorithms which will be used in theproposed scatternet formation method. In Section 3, wedescribe a novel multihop scatternet formation algorithm,which integrates the CDS-based backbone and dBBluestructure together seamlessly, and hence enjoys many niceproperties. In Section 4, we discuss in detail the IP-basedrouting solution forM-dBBlue scatternet.We evaluate ourstructures by simulation in Section 5. Some related worksare reviewed in Section 6. Finally, Section 7 concludes thepaper.

2 Preliminary

In Bluetooth specification 2.0 (Bluetooth SIG), piconetformation is performed in twophases: neighbour discoveryand link establishment. The neighbour discovery phaseis accomplished by the inquiry handshake procedures,implemented by the inquiry and inquiry scan command.Once two neighbouring devices complete an inquiryhandshake, only the node in inquiry mode knows the IDand weight information of the node in inquiry scanmode,not vice versa. To get the mutual information of eachother, two neighbouring nodes may set up a temporarypiconet that lasts only the time necessary to exchange theirID and other useful information (Petrioli et al., 2003).In the link establishment phase, each piconet is formedby one master and limited number of slaves, and eachnode decides its role locally based on the neighbouringinformation gathered in the neighbour discovery phase.The link establishment phase is achieved by the pagehandshake procedure, performed by the page and pagescan commands. Once two neighbouring nodes completea page handshake successfully, the node in page modeassumes master role, while the other node in page scanmode assumes slave role.

In this paper, for simplicity, we will not use theseengineering terms to describe our scatternet formation


algorithms in detail. Please refer to Petrioli et al. (2003) andBasagni et al. (2002) formoredetails of engineeringaspects.We assume that the problem of discovering all neighbourswithin transmission radius of a device has been resolved byseparate Bluetooth protocols, and any two neighbouringnodes can communicate directly to exchange information.Our goal is focusing on the topology formation with allpreferred properties in Bluetooth specification (BluetoothSIG), such as bounded node degree and zero role switch.In this section, we will discuss the network model andreview two algorithms, which form the ground for furtherdiscussion of the proposed approach.

2.1 Network model: a graph model

We assume that all devices have the same maximumtransmission range r. Thus, the set of Bluetooth devices Vdefine a Unit Disk Graph (UDG) G(V, r), in which twonodes are connected if and only if their distance is no morethan the maximum transmission range r. Notice that weonly assume that the underlying global communicationgraph is a UDG. The position information of the nodeis never used in our method. Given a subset U of V ,graph Gk(U) has an edge xy if and only if nodes x and yare at most k-hops away in the original communicationgraph G.

2.2 Connected Dominating Set: the virtualbackbone

CDS has been used as the virtual backbone for wirelessnetworks. In our proposed scatternet formation method,we will first form the clusters and build the connecteddominating set as the backbone of M-dBBlue scatternet.Assume V is the set of all nodes (Bluetooth devices).A subset S of V is a dominating set if any node u in Vis either in S or is adjacent to some node in S. Nodes inS are called dominators, while nodes in V − S are calleddominatees. A cluster is defined by a dominator node andall its dominatee nodes. Once some nodes, hereafter calledconnectors, are selected to formaconnectedgraph togetherwith dominators, the final structure is called CDS. Twodominators are said to be adjacent if they are within twohops of each other.

Several efficient methods (Alzoubi et al., 2002;Chlamtac and Farago, 1999; Wan et al., 2002; Wu andLi, 1999) have been proposed to construct a connecteddominating set, i.e., the backbone, of the networkmodelledby UDG. In this paper, we first use the method given inWan et al. (2002) to find a dominating set, then we design anovel communication efficient method to find a connectorfor each pair of dominators within two hops based onunicast communication only (unlike the previous methodsthat use broadcast communication model). The generatedCDS backbone is guaranteed to be connected.

For completeness, we briefly review the method inWan et al. (2002) here with our own interpretation.Their method uses a carefully chosen rank definition. Theranking of nodes is induced by an arbitrary spanning

tree T rooted at a leader. The message complexity oftheir method is O(n) if a leader is already known andO(n log n) if leader election is needed. Given a rootedspanning tree, the level of a node is the number ofhops in T between itself and the root of T . The rankof a node is then given by the ordered pair (level,ID), and such ranking gives rise to a total ordering ofthe nodes in the lexicographic order. After each nodeknows the rank of its own and all its neighbours, thealgorithm (Wan et al., 2002) first finds dominators by acolour-marking process. All nodes are initially markedwith white colour and will be marked with either grey orblack eventually. A node marked with black will becomedominator eventually. Each node maintains two localvariables: pendingChildrenNum and pendingLowNbNum.Variable pendingChildrenNum counts the number ofchildren who have not reported the completion and is thusinitialised to the number of children in the tree. VariablependingLowNbNum stores the number of lower-rankedneighbours who have not reported the status. Eachnode also maintains a blackList that records the IDs ofits black neighbours. The detailed method is given inAlgorithm 1.

LetU be the set of dominators constructed byAlgorithm1.It was shown that G2(U) is a connected graph (Wan et al.,2002; Alzoubi et al., 2002). Based on this, they gave acommunication efficient method to build a tree spanningall dominators as the final connected dominating set. Theyalso showed that the number of dominators within twohops of a dominator is at most 24. Since the graph G2(U)is connected, we only need to select one dominatee nodeas connector (or gateway) to connect two dominatorsseparated by two hops. Consequently, every connectornode is connected only to dominators, which implies thateach connector node has degree at most five. To minimisethe number of connectors, they build a spanning tree ofdominators as the connected dominating set. Thus, intheir method, two adjacent dominators are not necessarily


connected by a connector. In the worst case, two adjacentdominators may be connected by a path with O(k) hops,where k is the number of dominators found. However, inour multihop scatternet formation method, we prefer thatthe backbone keeps more connections than the spanningtree so that this can enhance the efficiency of inter-clusterrouting. Thus, we design a new communication efficientmethod (in Section 3.1) to find a connector for each pair ofdominators within two hops, instead of using the spanningtree. Figure 1 illustrates a backbone topology formed byour algorithm, in which each adjacent dominator pair isconnected by exactly one connector.

Figure 1 The disks represent the clusters. The solid linesconnecting dominators, connectors; the dashed linesconnect the dominatees and their dominators. Thedominators (denoted by black nodes) and connectors(denoted by grey nodes) form the network backbone(see online version for colours)

2.3 dBBlue scatternet for single-hop networks

Single-hop Bluetooth scatternet formation has been wellstudied inBarriere et al. (2003), Song et al. (2005), Salonidiset al. (2001) andLawet al. (2001). In a single-hopBluetoothscatternet, all wireless devices are within the radio vicinityof each other. In Song et al. (2005), Song et al. adoptthe well-known de Bruijn graph to build a self-routingscatternet, called dBBlue, with low-diameter O(log n) andbounded node-degrees. Each master has at most sevenslaves and each slave node exists in at most two piconets,and no node assumes both master and slave roles. Theyalso presented a scalable MAC assignment mechanismand a vigorous method to locally update the dBBluestructure using at most O(log n) communications when anode joins or leaves the network. The computation costis O(n) for static construction. Moreover, the congestionof every node is at most O(log n/n), assuming that a unitof total traffic demand is equally distributed among allpair of nodes. Since we will use dBBlue as the intra-clusterstructure in the proposed mutli-hop scatternet formationmethod, for completeness of presentation, we now brieflyreview the dBBlue scatternet formation algorithm fromSong et al. (2005).

The dBBlue scatternet construction is based on thewell-known de Bruijn graph (de Bruijn, 1946). Thede Bruijn graph, denoted by B(d, k), is a directed graphwith dk nodes. Assume that each node is assigned a

unique label of length k on the alphabet {0, . . ., d − 1}.There is an edge in B(d, k) from a node with labelx1x2 · · ·xk to any node with label x2 · · ·xky, wherey ∈ {0, . . ., d − 1}. It is well-known that the de Bruijngraph enables self-routing intrinsically. The self-routingpath from the source with label x1x2 · · ·xk to thetarget with label y1y2 · · · yk is x1x2 · · ·xk → x2 · · ·xky1 →x3 · · ·xky1y2 → · · · → xky1 · · · yk−1 → y1 · · · yk. Observethat, we could find a shorter route by looking for thelongest sequence that is both a suffix of x1x2 · · ·xk and aprefix of y1y2 · · · yk. Suppose that xi · · ·xk = y1 · · · yk−i+1is such a longest sequence. The shortest path between thesource and the target is x1 · · ·xk → x2 · · ·xkyk−i+2 → · · ·→xi−1 · · ·xkyk−i+2 · · · yk−1 → y1 · · · yk. Clearly, the routebetween any two nodes is at most k hops, i.e., B(d, k) hasdiameter k = logd n, where n = dk is the number of nodesof the graph.

The classical de Bruijn graph is balanced in the sensethat the labels of all nodes have the same length. Thede Bruijn graph can be generalised to any set of verticeswhose labels form a universal prefix set. In FraigniaudandGauron (2003), proposed a novel method to constructan efficient topology for P2P network based on thegeneralised de Bruijn graph defined on a universalprefix set.

“A universal prefix set is a set S of labels on an alphabetΣsuch that, for any infinite word w ∈ Σ�, there is a uniqueword in S, which is a prefix of w. The empty set is also auniversal prefix set.” (Fraigniaud and Gauron, 2003)

For instance, {00, 01, 100, 101, 110, 111} is a universalprefix set on alphabet Σ = {0, 1}, but {00, 01, 10} and{00, 01, 100, 1000, 101, 110, 111} are not. There is adirected edge from node u = x1x2 · · ·xk to another nodev in the generalised de Bruijn graph if x2 · · ·xk is theprefix of the label of node v. A generalised de Bruijngraph is pseudo-balanced if the lengths of the labels aredifferent by at most one. For simplicity, we still denotea pseudo-balanced de Bruijn graph on alphabet {0, 1} byB(2, k) if the node labels have length at least k bits and atmost k + 1 bits. We also say that a node from B(2, k) is atlevel k if its label has k bits. In this paper, we only considerthe balanced or pseudo-balanced binary de Bruin graphB(2, m).

The method Song et al. (2005) constructs a balancedde Bruijn graph B(2, m) as the initial backbone of thenetwork. It chooses integer m such that 2m−1 <

⌈n6

⌉ ≤2m. The choosing of m guarantees that there are enoughbridge slave nodes, which implies that no master nodeserves as bridge slave. The detailed method is given inAlgorithm 2. Once the initial topology construction isfinished, the token node t will be responsible for followingnode joining and leaving issues. Master nodes and bridgesform the backbone of Bluetooth scatternet, and a piconetworks like a node in de Bruijn graph. Figure 2 illustratesa dBBlue scatternet containing 48 nodes based on B(2, 3)graph. Notice that dBBlue scatternet does not work formultihop networks, since it requires every node can beconnected to other nodes.


Figure 2 dBBlue scatternet with 48 nodes based on B(2, 3)

3 M-dBBlue scatternet formation

Our M-dBBlue scatternet formation algorithms formultihop Bluetooth networks first builds a CDS asthe backbone of multi-hop Bluetooth networks, thenconstructs the dBBlue scatternets in each cluster.

3.1 Backbone construction for M-dBBluescatternet

In the proposed method, we first use the method given inWan et al. (2002) to find a dominating set, then adopts anew communication efficient method to find a connectorfor each pair of dominators within two hops based onunicast communication only (unlike the previous methodsthat use broadcast communication model). The method inWan et al. (2002) has been briefly introduced in Section 2.

The CDS backbone is guaranteed to be connected. Similarto Alzoubi et al. (2002) andWan et al. (2002), we can showthat the node degree of the connected dominating sets isbounded by a constant, and the hops and length stretchfactor are bounded by small constants. In addition, thenumber of dominators is at most a small constant factorof the minimum number of dominators.

In Wan et al. (2002), they used the spanning tree ofdominators to select the connector and form the connecteddominating set. However, in our M-dBBlue network,we prefer that the backbone keeps more connectionsthan the spanning tree to enhance the efficiency ofinter-cluster routing. Notice that Algorithm 1 finds a setof dominators with the following property: the backboneby connecting each pair of dominators separated by twohops is connected. Thus, we only need to find connectorto connect any pair of 2-hop adjacent dominators to formthe connected backbone. We try to minimise the numberof connectors usedwhile keeping at least one connector foreach pair of adjacent dominators. Therefore, a connectorcould be used to connect many pairs of dominators in ourmethod. To find a connector for each pair of dominatorsis not our innovation, several methods (Alzoubi et al.,2002, 2003) have been proposed. However, in all previousmethods (Wan et al., 2002; Alzoubi et al., 2002, 2003),they adopt the broadcast communication model to buildCDS graph. It is well known that local broadcast cannotbe performed efficiently in practice, due to the constraintin MAC layer that simultaneous broadcast by dominateescould cause massive signal interference which causeslarge latency. In Algorithm 3, we actually reduce thecommunication cost significantly by using unicast insteadof broadcast.

In our algorithm, each dominator maintains twolists: adjacentDominatorList and dominateeList, whichare initially empty. Here adjacentDominatorList recordsall adjacent dominators of this node, in addition,the connection flag is set for each pair of adjacentdominators acknowledged by a connector; dominateeListrecords all dominatees dominated by this node, whichis reserved for dBBlue scatternet construction as willbe seen later. Notice that, our method not onlygenerates a CDS-based backbone, but also splits all nodesinto separated clusters. Each dominatee also maintainstwo lists: blackList and neighbourDominatorList. TheblackList is generated in Algorithm 1, which recordsthe known dominator neighbours of the dominatee. TheneighbourDominatorList stores the dominator neighbourswhich need be connected by itself, if this node is aconnector. The detailed algorithm is given byAlgorithm 3.

It is easy to show that each 2-hop adjacentdominator pair is connected, since the connection willbe acknowledged by some dominatee for sure in thealgorithm.Figure 1 illustrates abackbone topology formedby our algorithm, in which each adjacent dominator pair isconnected by exactly one connector. One connector couldbe used to connect several dominators. For instance, nodeu serves connector role among three dominators A, Band C, hence node v will not be selected as dominators.


Figure 3 illustrates the procedure of finding connectors.Node A, B and C are three dominators, we assumethat ID(A) < ID(B) < ID(C). Both node u and vare dominatees which have blackListu = {A, B, C} andblackListv = {A, B} respectively. Since node A has thesmallest ID in their blackList, both u and v send aTRYCONNECTORmessage only to dominatorA. Thereare two cases here:

1 In Figure 3(a), the message from u first reaches A.Dominator A will select u as connector to connect Band C, then the message from v will be discardedsince it is redundant.

2 In Figure 3(b), the message from v first reaches A.Dominator A will select v as connector to connect B.When the message from u arrives, node u will beasked to connect with node C, in addition, it will alsobe asked to connect B and C because theirconnection flag has not been set.

Figure 3 Finding connectors with minimal communication:(a) u first reaches A and (b) v first reaches A(see online version for colours)

The following lemma which bounds the number ofdominators within k units from a node v is proved in Wanet al. (2002) and Alzoubi et al. (2002, 2003) by using asimple area argument.

Lemma 1: For every node v, the number of dominatornodes inside the disk centred at v with radius k-unitsis bounded by a constant �k = (2k + 1)2 for k > 1 and�1 = 5.

The bounds on �k can be improved by a tighter analysis.Therefore, as Wan et al. (2002) and Alzoubi et al. (2002,2003), we can prove the following theorem.

Theorem 2: In the CDS built by Algorithm 1 andAlgorithm 3, node degree is bounded by a constant.

Proof: For a dominator u, the number of neighbouringconnectors in the CDS is bounded by 24, since thenumber of dominators within two hops of u is boundedby �2 − 1 = 24 (remember that u is a dominator). For aconnector v, the number of its neighbouring dominator isat most five. �

However, the node degrees of dominatees who are not inthe CDS are not bounded. In other words, in a cluster thenumber of dominatees dominated by a dominator couldbe very large. Thus in the next section, we will use dBBluestructure to further form scatternet inside each cluster.

3.2 dBBlue scatternet formation in clusters

Forming the CDS and clusters provides the backbone andthe globe topology for our M-dBBlue scatternet, in thissection,wewill studyhow to form the scatternet inside eachcluster, i.e., how to assign the roles for each node and formthe piconets. After building the CDS, a cluster may havemany nodes (dominatees), and it is inefficient to connect alldominatees to the dominator. The node degree is preferredto be bounded by a constant number, where seven is thebest match to Bluetooth specifications. In this section,we will continue to describe our scatternet formationapproach in clusters based on the dBBlue protocol. ThedBBlue protocol only works for single-hop Bluetoothnetwork, where each device can directly communicate toall other devices. Unfortunately, each cluster (composedof a dominator node and all its dominatees) may not be asingle-hop Bluetooth network, i.e., some dominatee pairscould not communicate directly at all. Thus, the single-hopscatternet formation algorithm can not be applied heredirectly. There are two possible solutions here:

• make sure that all dominatees of a dominator nodecan communicate directly

• partition the dominatee nodes into groups (cliques)such that the dominatee nodes of each group cancommunicate directly.

For all nodes in a cluster, we propose two methods tobuild a dBBlue scatternet, which will be discussed in detail


later. Figure 4 illustrates the idea of applying the dBBluescatternet formation protocol directly to each cluster.

Figure 4 Two approaches to partition the network to clusters,such that any two nodes in each cluster can talk witheach other directly: (a) all nodes use half-radius tobuild CDS and (b) divide the dominatee nodes intocliques (see online version for colours)

The first approach (Figure 4(a)) is based on the followingobservation: if all dominatee nodes lie inside the diskcentred at the dominator with radius equal to half ofthe transmission range, then all dominatee nodes areguaranteed to be able to communicate directly, i.e., beingone-hop. Then, in the backbone construction phase, wemay set everyBluetoothnode inpower-savingmode so thateach node decreases its transmission radius to half of themaximum. To make sure that the backbone constructedthis way is still connected, we need that the communicationgraphG(V, r/2) be connected. In other words, if each nodein the wireless ad hoc network decreases its transmissionradius to r/2, the network is still connected. Once themodified backbone is built, every node returns to normaloperation mode with transmission radius equal to r.Consequently, each cluster is a complete graph.

This approach is straightforward but need strongrequirement of the network to ensure the connectivity ofthe backbone. Practically, the condition that G(V, r/2) isconnected could be satisfied frequently. Nonetheless, thesecond solution can be applied in general. In the remainingof the paper, we always assume the scatternet formationand routing is based on the second approach. All followingmechanisms can be easily transformed to support the firstapproach.

The second approach (Figure 4(b)) is using somealgorithms such as those in Ishii and Kakugawa (2002),Balas and Yu (1986), Wang et al. (2003), and Carter andPark (1993) to partition the dominatee nodes into cliques(groups such that the dominatee nodes of each groupcan communicate directly). If the position information isavailable for each node, a simple and efficient methodto divide into cliques is as follows. We can dividethe transmission region of the dominator node intosix equal-sized cones (with degree π/3) and notice that allthe dominatee nodes from a cone are guaranteed to forma clique. By this way, we may divide a cluster into at mostsix cliques.

We now continue the one-hop scatternet formation ineach clique and present two algorithms to build a dBBlue

scatternet for each clique. In both algorithms, we will letthe dominator nodes be slaves of piconets. Notice thatthis approach is different from all previous scatternetformationmethodsbasedon the connecteddominating set,in which the dominator is naturally assigned master roleinstead. We will show that assigning dominator slave rolesactually produces scatternet with several nice properties.

In our first method, we build a dBBlue scatternetusing all nodes in a clique. Notice that a connector nodeappears in several cliques (at least two since it is connector).If the role assignment is not treated carefully, a connectornode (as connector u in Figure 5(a) in two cliques: onefor dominator v2, the other for dominator v1) may beassignedmaster role inonedBBlue scatternet and slave rolein another dBBlue scatternet. This multi-role assignmentcould cause some delay in scheduling packets. In an idealsituation, we would like the connector to be only slaveof a couple of piconets without being master node in anypiconet.

Figure 5 (a) All nodes in each cone of unit-radius cluster forma dBBlue scatternet and (b) all dominatees and thedominator in each clique of unit-radius cluster form adBBlue scatternet (see online version for colours)

In our second method, we will first build a piconet foreach connector: the connector being themaster node of thepiconet and all its dominator nodes (at most five) being theslavenodes of thepiconet. For each cliquepartitioned fromthe dominatee nodes of a dominator, we build a dBBluescatternet using all dominatee nodes and the dominatornode, excluding the connectors. In other words, we buildtwo level piconets: the top level piconets are built for thebackbone nodes, and the low level piconets are built forall dominatees and dominators. Here the dominator nodeis used to connect each low level dBBlue scatternet to thetop level backbone, serving as the gateway (as node v1 inFigure 5(b)which connectsu’s piconet and t’s piconet fromtwo levels).

Figure 5(a) illustrates the first method which describedin detail in Algorithm 4. Notice that we made some specialtreatments with the dominator and connectors in thealgorithm, because we need diminish the probability toassign them master roles in the dBBlue structure. Thoughwe can not avoid the role switch between bridge andmaster absolutely, it is not difficult to show that, at mostone connector need assume dual roles, even in the worstsituation that there are no dominatees in the clique, whichrarely happens in practice as we will see in our simulationexperiments.

Notice that there are three kinds of nodes in theM-dBBlue scatternet built by Algorithm 4: dominator,


dominatee and connector. We can prove the follwingtheorem about their nodes degree in the M-dBBluescatternet.

Theorem 3: The node degree of a node u in theM-dBBluescatternet built by Algorithm 4 is bounded by:

• 7, if u is a dominatee;

• c, if u is a dominator, where c is the number ofcliques that u has inside its cluster. If the positioninformation is available, c ≤ 6.

• 35, if u is a connector; If there is at least onedominatee node in a clique, the bound reduces to 5.

Proof: We prove the node degree bounds for three kindsof nodes one by one. For a dominatee node, its nodedegree is obviously bounded by seven according to theproperty of dBBlue scatternet since it only appears in onedBBlue scatternet. For a dominator node, according toAlgorithm 4, the dominator assumes pure slave in theone-hop dBBlue scatternet. So the degree of a dominatornode is bounded by the number of cliques it has insideits cluster. If we use the position-based method, there areat most six dBBlue structures in a cluster, so its degree isbounded by six. However, if we do not use position-basedmethod, the number of cliques may be very large. For aconnector node, it could exist in at most five clusters, ineach cluster it can at most have seven neighbours in theworse case. This gives a bound of 35. However, as long asthere is at least one dominatee node in a clique, we can letthat dominatee assume master role and the connector beits pure slave. Hence, the connector’s degree is boundedby five. Therefore, in the M-dBBlue scatternet built byAlgorithm 4, node degree is less or equal to sevenwith highprobability.Aswill see later in simulation results (Figures 8and 9), the average node degree of dominators is almostalways smaller than seven. �

Another approach is to exclude the connectors fromparticipating in the one-hop dBBlue scatternet formationin each clique, so that no nodes need assume both masterand slave roles. Figure 5(b) illustrates the second methodwhich is described in detail in Algorithm 5. Notice that theconnector u will be a master node who forms a piconetwith dominators v1 and v2 and assigns pure slaves to them.

The leader t will become a master node and v1 become itspure slave.

Similarly, we can analyse the node degree of theM-dBBluescatternet built by Algorithm 5 and get the followingtheorem.

Theorem 4: The node degree of a node u in theM-dBBluescatternet built by Algorithm 5 is bounded by:

• 5, if u is a connector;

• 7, if u is a dominatee;

• �2 + c − 1, if u is a dominator, where c is the numberof cliques that u has inside its cluster. If theposition information is available, c ≤ 6.

Proof: For a connector node (as node u in Figure 5(b)),it does not participate in one-hop dBBlue construction, soall its neighbours must be dominators, which is at mostfive. For a dominatee node, it only participates in one-hopdBBlue formation, so its degree is always bounded byseven. For a dominator node (as node v1 in Figure 5(b)),it assumes the bridge slave role for the backbone. It is inat most �2 − 1 piconets for the backbone since it has atmost �2 − 1 = 24 neighbouring connectors. Additionally,it could be in several cliques, assume c cliques. Then thedegree of dominator is atmost �2 + c − 1 under pessimisticestimation. If the position-based method is used, there areat most six cliques, thus the degree is at most �2 + 6 − 1 ≤30. As will see later in simulation results (Figures 8 and 9),the average node degree of dominators is much lower thanthese bounds. �

In summary, Algorithm 4 could build a degree-7Bluetoothscatternet as long as there is one dominatee in eachclique and position information is available. To evaluatethe performance of the two algorithms, simulation isconducted in Section 5.

Dynamic network. Notice that so far, we concentrate onthe static networks. As Vergetis et al. (2005) pointed out,the time required to generate a stable topology for a largenumber of nodes may be very large, thus it is hard toefficiently update the scatternet in a dynamic environment.However, our scatternet method has the potential ofdynamic updating with node joining or leaving, since boththe backbone and the dBBlue structure of each clustercan be maintained efficiently. For the backbone built bythe method in Wan et al. (2002) or (Algorithms 1 and 3,


it may need many massages to update or rebuild since aspanning tree is involved, but inAlzoubi et al. (2002, 2003),they gave a more message efficient backbone formationmethod. We can direct adopt their method to form theCDS forM-dBBlue. InAlzoubi et al. (2002), they describedthe detail about how to maintain their backbone in mobileenvironment. For the dBBlue structure, in Song et al.(2005), they described a vigorousmethod to locally updatethe structure dBBlueusing atmostO(log n)messageswhena node joins or leaves the network. In most cases, the costof updating the dBBlue is actually O(1) since a node canjoin or leave without affecting the remaining scatternet.The number of affected nodes is always bounded fromabove by a constant when a node joins or leaves thenetwork. We leave the formal overhead analysis andsimulation study of our algorithms in mobile networks asour future work.

4 Routing in M-dBBlue scatternet

Since Bluetooth networks are usually regarded as theextension of internet, in this section,we propose a completeIP-based routing mechanism to integrate M-dBBluescatternet with internet seamlessly.

Internet is comprised of many separate administrativedomains or Autonomous Systems (AS). There are twolevels routing protocol, intra-domain and inter-domain,in internet. Intradomain routing protocols, such as RIPand OSPF, route packets within a AS; while interdomainrouting, currently handled by the Border GatewayProtocol (BGP), routes packets betweenASes.Weproposea similar hierarchical way to implement the IP-basedrouting in M-dBBlue scatternet. The inter-cluster routingis handled by a modified Bluetooth based RIP protocolon the M-dBBlue backbone. The intra-cluster routingis derived from the self-routing mechanism describedin Song et al. (2005).

Inter-cluster routing: The backbone of M-dBBluescatternet works in a similar way as Internet, so wemay apply any IP-based routing protocol here directlywithout much modification, such as the modified RIPprotocol. In M-dBBlue scatternet, each cluster is assigneda network number, and every node in the cluster isdynamically assigned an IP address with same networknumber. See Figure 6 for an illustration of possiblenetwork number assignment of the clusters derived fromFigure 1. The routing along the backbone could also beany greedy method (Karp and Kung, 2000; Kuhn et al.,2003) if geometry information is known,or source-initiatedon-demand routing protocols such as AODV (Perkins andRoyer, 1999), DSR (Johnson et al., 2003), or table drivenprotocols such as DSDV (Perkins and Bhagwat, 1994).Since such kind of routings are well-studied, we omit thedetails of routing along the backbone here.

Intra-cluster routing: In each clique of a cluster, dBBluestructure (Song et al., 2005) intrinsically provides theself-routing mechanism based on the labels derived from a

pseudo-balanced de Bruijn graph. To enable the IP-basedrouting in a clique, we need map the IP address to thecorresponding label. Given the IP address of the targetnode, the source node need know the label of the targetnode if self-routing based on de Bruijn graph is used.One possible approach to solve this is to store all pairsof (IP, label) for all nodes in a node, e.g., the dominatornode of the cluster. The source node always queries thedominator node for the label of the target node.Notice thatsuch queries can be conducted using self-routing since thelabel of the dominator node is always fixed in our dBBluestructure. This centralised approach is simple, however, itsuffers several disadvantages: the traffic storm problem tothe dominator node, the single failure of the dominatornode breaks the network, and so on.

Figure 6 Intercluster routing in the backbone of M-dBBluescatternet

We propose to use a distributed storage of the (IP, label)pairs. Eachmaster node u in the dBBlue structuremanagesa lookup table, which stores the (IP, label) pairs of thosenodes whose key has u’s piconet ID as prefix. Noticethat the label is generated when we construct the dBBluescatternet for each clique. The key of a node is somevalue computed from its IP, e.g., the hash value of its IP.Table 1 illustrates apossible lookup table stored in amasternode with label 1011. Notice that, we represent the labelof a master node by its piconet ID. It is assumed thatthe key value of any pair stored in this lookup table hasprefix 1011.

Table 1 The lookup table in the master node u = 1011

IP-Address Piconet ID MAC

216.47.152.22 00100 000216.47.152.44 01001 100...

......

216.47.152.90 1100 010

Assume that the length of every piconet ID in the dBBluescatternet is between m and m + 1. Each node first mapsits host address, the suffix of its IP address, to a binary keywith length m + 1. The mapping technique could adoptany hashing function or simply translate its host addressto binary format which is then abbreviated or extendedto (m + 1)-bits key. The node then forwards its key and(IP, label) pair to the target master node through thelabel-based routing in dBBlue scatternet. Notice that thetarget master node in which the pair will be stored has alabel being a prefix of the key. Since the labels of the nodesare universal prefix free, the target master node is unique.


Notice the simple abbreviation/extension is efficientbut could cause unbalanced overload between masternodes with small and large labels. In contrast, thehash based mapping is believed to achieve uniformlydistribution with high probability. In this paper, forsimplification of our presentation, we adopt the latter tomap host IP address to a key. For example, in a cone,suppose m = 4 and the network number of its clusteris 216.47.152.∗ with mask 255.255.255.0. The node withIP address 216.47.152.90 first translates its host address90 to the binary format 1011010, then abbreviates itssuffix to a (m + 1)-bits key 10110. The node 216.47.152.90forwards its (IP, label) pair to the master node with label1011 since the key 10110 has 1011 as prefix. Anothernode with IP address 216.47.152.12 first translates its hostaddress 12 to the binary format 1100, then extends it to a(m + 1)-bits key 11000 by simply appending 0. Similarly,the node 216.47.152.12 forwards its (IP, label) pair to themaster node 1100 or 11000, whichever exists.

Consider the case that a node u wants to send packetsto target node v in the same dBBlue scatternet while onlyIP address of node v is known.W.l.o.g., suppose themasternode w holds the (IP, label) pair of node v. Node u firstmaps the IP address of node v to a key, say k. Two options,which are illustrated by Figure 7, could be used to send outthe packet:

• Search and forwarding. Node u queries the dBBluebackbone based on key k and gets the label of node vfrom the master node w. Node u then forwards thepacket targeting v through dBBlue routing protocol.Figure 7(a) illustrates the mechanism.

• Packet-in-tunnel forwarding. Node u adds anadditional header with the key k to the packet thensends the packet out. The routing of the packet isbased on the label of k. Once the master node w getspacket, it strips out the header and relays the packetto node v according to node v’s label in its lookuptable. Figure 7(b) illustrates the mechanism.

Figure 7 Label based intradomain routing: (a) search andforwarding and (b) packet-in-tunnel forwarding(see online version for colours)

Remember the path between any two nodes in dBBluestructure is atmost 2m + 2 hops. The former approach canreduce the overall workload of dBBlue structure since thedata packet travels through the network at most 2m + 2hops, while the data packet travels at most 4m + 4 hops inthe latter case. But the latter approach does not need keepthe packet before getting the target label as in the formerapproach, and the packet can reach the target faster if the

time difference between transmitting different size packetsis negligible, because the total communication path is atmost 6m + 6 hops in the former while atmost 4m + 4 hopsin the latter. On the other hand, the latter approach cankeep the anonymity of node w hence increase security.

We continue to describe the IP-based routing forthe node pairs within different cliques of one cluster.Suppose that the dominator keeps an IP address rangetable for each cone. For the packets targeting a nodein other cones in the cluster, the dBBlue protocol willfirst forward them to the cluster dominator, which thenforwards the packets to the target clique. Eventually thepacketwill reach the target through the intra-clique routingas described above.

5 Performance evaluation

We have conducted extensive simulation to study theperformance (topology properties) of different multi-hopscatternet structures proposed in this paper. In ourexperiment, we generate n wireless nodes, with uniformtransmission range, randomly distributed in a square areawith side length 40 units. To get a stable result, werandomly generate 100 samples of the network for eachcase and construct the scatternet using different methodsfor each sample. When running Algorithms 4 and 5,we divide each cluster into six equal-size cones, i.e., sixcliques. The average communication hops between allpairs of nodes is obtained by actually running messagepassing mechanism in the scatternet. For inter-clusterrouting, we first find the shortest path between eachpair of dominators/connectors. To follow the shortestpath during routing, every node only needs to record thenext hop neighbour to any other nodes. For intra-clusterrouting, we suppose that the label of the target node isknown. In other words we assume that the IP-address hasalready been mapped to a label in a distributed way asdescribed in Section 4.

All experiment results are shown in Figures 8 and 9. Inthe figures, we useCDS1 to denote the CDS constructed byAlgorithm1 (findingdominators) andAlgorithm3 (findingconnectors), and CDS2 to denote the CDS constructedby the algorithm from Wan et al. (2002). To distinguishthe two proposed different scatternet formation methodsfor clusters, we use Exclude to denote Algorithm 5, andInclude to denote Algorithm 4.

As we expect, M-dBBlue scatternet formed based onthe backbone CDS1 does provide smaller average hopsbetween any pair of nodes than CDS2 structure. Thus,more energy is saved. The tradeoff is that the averagedegree of dominators and connectors in CDS1 is a little bithigher than the scatternet based onCDS2. However, this istolerable since the average degree of connectors is still lessthan four, and the average degree of dominator is less thaneight if CDS1 is used. Notice that only connector could beassigned master role. Thus, every master on the backbonewill have at most seven slaves with high probability, whilethe master in each cluster is guaranteed to have at mostseven slaves.


Figure 8 Performance evaluation of our different scatternet construction methods when node transmission range varies in the range[5, 38] and the number of nodes is fixed at 200: (a) average degree of dominators; (b) average degree of connectors;(c) proportion of nodes with dual-roles; (d) proportion of nodes with degree > 7 and (e) average communication hops

On theotherhand, theM-dBBlue scatternet,whenExcludemethod is used to form scatternet for each cluster, haslower diameter than that produced by Include method.In addition, no nodes assume dual role in the scatternetformed by the Exclude method. The Include method hasits own advantage: the average degree of dominators andconnectors is smaller, and almost no node has degree morethan seven (See Figures 8(d) and 9(d)).

Figure 8 illustrates the performance variation whenthe uniform transmission range of nodes varies in [5, 38],while the total number of nodes is fixed at 200. Thefirst observation is that the average hops of shortestpath between all pair of nodes first drop then rise toa number around O(log n) when the transmission rangeincreases. This is because when the transmission rangeis small, the communication graph is sparse and the


Figure 9 Performance evaluation of our different scatternet construction methods when the number of nodes varies in the range from40 to 480 and the node transmission range is fixed at eight: (a) average degree of dominators; (b) average degree ofconnectors; (c) proportion of nodes with dual-roles; (d) proportion of nodes with degree > 7 and (e) averagecommunication hops

backbonehas largediameter,which consequently increasesthe communication hops for nodes from different clusters;when the transmission range becomes larger, the diameterof the backbone becomes smaller and the scatternetdegenerate to an one-hop dBBlue scatternet in the extremecase, which gives a O(log n) bound on communicationhops. Figure 8(a) and (b) show that the average degreein backbone reaches the peak when the transmissionrange is around half of the simulation field width. Itdrops down after the peak because the size of backboneshrinks.

Figure 9 illustrates the performance variation whenthe number of wireless nodes varies in [40, 480], while thetransmission range of each node is fixed at eight units.The average degree in backbone keeps almost constantafter the density reaches some extent. In both Figures 8and 9, the proportion of nodes with dual roles, and theproportion of nodes with degree exceeding seven dropwhen the network density increases. Notice we need nottest the degree of dominatees because Algorithm 2 canguarantee the perfect degree bound as discussed in Songet al. (2005).


6 Related work

In this section we review the solutions of scatternetformation for multi-hop networks by dividing them intofive categories. A survey of the current research in theformation of Bluetooth scatternet is presented inWhitakeret al. (2005).

Tree based methods. Zaruba et al. (2001) proposed twodistributed tree-based methods for forming connectedscatternet. In both methods, the resulting topology istermed as BlueTree. The first method is initiated by asingle node (blueroot). A rooted spanning tree is built fromthe blueroot. Each node is a slave of its parent and amaster of its children in the tree. In the second method(Zaruba et al., 2001), several roots are initially selected.Each of them then creates its own scatternet as in the firstmethod. After that, sub-tree scatternets are connected intoone scatternet spanning the entire network. Rememberthat the tree topology suffers from a major drawback: theroot is a communication bottleneck. In addition, dynamicupdating that preserves correct routing is not discussedin these protocols. There are several modified versions ofBlueTree, such as themethods byDong andWu (2003) andHuang et al. (2003, 2006), to reduce the communicationoverhead or increase the connectivity. Cuomo et al. (2003)also proposed a tree-based scatternet formation algorithmSHAPER for multi-hop networks, which focuses on theself-healing behaviour of the tree structure: i.e., it is ableto dynamically reconfigure the scatternet after topologicalvariations due to mobility or failure of nodes. Guerin et al.(2003) proposed depth/breath first search andMST-basedscatternet formation schemes for unit disk graphs in twoand three dimensions, but their schemes are not localised.

Cluster based methods. Petrioli et al. (2003) and Basagniand Petrioli (2002) described a multihop scatternetformation scheme based on clustering scheme (Linand Gerla, 1997). The constructed scatternet is calledBlueStars. The protocol proceeds in three phases: devicediscovery, partitioning of the network into piconets (stars)by clustering, and interconnection of the piconets to aconnected scatternet. All clusterhead nodes are declaredmaster nodes in a piconet, with all nodes belongingto their clusters as their slaves. In the third phase,BlueConstellation, some of the slaves become masters ofadditional piconets, i.e., become master-slave bridges, toassure the connectivity of the scatternet.However, piconetsin the scatternet may have more than seven slaves. Thismay result in performance degradation, as slaves need tobe parked and unparked in order to communicate withtheir master. A performance evaluation of the clustering-based scatternet formation scheme (Basagni and Petrioli,2002) is given by Basagni et al. (2002) and Petrioli et al.(2003). To fix the unbound slave number (Petrioli et al.,2004; Petrioli and Basagni, 2002) modified their protocol(Petrioli et al., 2003; Basagni and Petrioli, 2002) andproposed a new scatternet called BlueMesh. The idea ofbounding the slave number in BlueMesh is again based

on the observation that if a node in unit disk graphhas more than five neighbours then at least two of themmust be connected. Same with BlueStars, the selection ofthe masters is based on the node weights. However, theselection of slaves is performed in such a way that if amaster has more than seven neighbours, it only choosesseven slaves among themso that via them it can reachall theothers. Variants of clustering-based scatternet formationschemes (Wang et al., 2002; Guerin et al., 2002) werealso proposed. Both clustering processes (Wang et al.,2002; Guerin et al., 2002) follow a random fashion. Initialconnections are made by nodes entering scan or inquiryscanphases at random.Already existingmaster nodes havepriority in attracting more slaves up to the limit. Aftereach node is assigned master or slave role, or is unableto join any piconet or attract any neighbour as its slaveto create its own piconet, some bridge piconets are addedto connect the scatternet. However, both methods (Wanget al., 2002; Guerin et al., 2002) do not always lead toa connected structure. A two-phase bluetooth scatternetformation algorithm is introduced in Li and Yang (2005),in which the election of the Bluetooth masters or bridges isbased on device (the energy supply) and link (the ReceivedSignal Strength) characteristics. Initially, all randomdistributed nodes form a series of isolated piconets withbounded number of slaves within k; and then interconnectthese piconets into a connected scatternet-eBlueScatter,in which master and bridge nodes constitute a connecteddominating set. Recently, a Group-Scatternet FormationAlgorithm (GSFA) is proposed in Shih et al. (2007).

Position based methods. Li et al. (2004) proposed thefirst position-based schemes that construct degree limitedand connected piconets in multihop networks withoutparking any node. Notice that the schemes in Petrioliet al. (2004) and Petrioli and Basagni (2002) can alsoachieve bounded degree scatternet. Their neat schemedoes not require position information, but instead thelocal information is extended to two hop information,with a two round device discovery phase for obtainingnecessary information. In Li et al.’s solution, nodes knowtheir positions and are able to establish connectionswith other nodes within their transmission radius in theneighbour discovery phase. The degree of each node islimited to seven by applying Yao structure (Yao, 1982),and the master-slave relations are formed in createdsubgraphs. This phase follows clustering based approach,and consists of several iterations. In each iteration,undecided nodes with higher keys than any of theirundecided neighbours apply Yao structure to bound thedegree, decide master-slave relations on the remainingedges, and informall neighbours about either edge deletionor master-slave decision. The experiments confirmed goodfunctionality of created Bluetooth networks in additionto their fast creation and straightforward maintenance.Basagni et al. (2003, 2004) described the results of anns2-based comparative performance evaluation amongfour major solutions for forming multihop scatternet(Li et al., 2004; Petrioli et al., 2003; Zaruba et al., 2001;


Wang et al., 2002). They found that device discoveryis the most time-consuming operation, independentlyof the particular protocol to which it is applied. Thecomparative performance evaluation showed that due tothe simplicity of its operationsBlueStars is by far the fastestprotocol for scatternet formation. However, BlueStarsproduces scatternets with an unbounded, possibly largenumber of slaves per piconet, which imposes the useof potentially inefficient Bluetooth operations. Theyproposed a combined solution by applying aYao structureon each piconet, to limit the degree of each master nodeto seven. This is a variant of the clustering-based scheme(Li et al., 2004).

On-demand methods. Most above scatternet formationprotocols tend to interconnect all Bluetooth devices at theinitial network startup stage and maintain all Bluetoothlinks thereafter. The master or bridge nodes in theresulting scatternet may become the traffic bottleneckand reduce network throughput. To make the scatternetstructuremore suitable to serve inmobile ad hoc networks,several on-demand methods (Liu et al., 2003; Kawamotoet al., 2003; Pagani et al., 2004; Chou and Chang, 2004;Tekkalmaz et al., 2006; Zhang and Riley, 2005) (tobuild scatternets only along the multihop routes withtraffic demands and eliminate unnecessary link and routemaintenances) are proposed recently.

QoS based methods. Marsan et al. (2002) studied howto construct the optimal topology that provides fullnetwork connectivity, fulfills the traffic requirements andthe constraints posed by the system specification, andminimises the traffic load of the most congested node inthe network, or equivalently its energy consumption. Byusing a min-max formulation, they provided a centralisedsolution based on integer linear programming. Chiasseriniet al. (2003) extended the work of Marsan et al.and enhanced the optimisation problem by adding theconstraints on the network capacity. (Augel and Knorr,2004) proposed an approach of scatternet formationin which the formation is dependent on the QoSrequirements of the applications. In their solution, toavoid larger degree which may cause negative influenceon throughput, nodes with high degree stop paging andinstruct a neighbour with a low degree to start paginginstead. Each device may try to influence the topologydepending on the QoS requirements. They described ageneral scatternet formation design guidelines for QoSapplications, but did not present any particular scatternetformation protocol. Melodia and Cuomo (2004a, 2004b)and Cuomo and Melodia (2002) discussed the scatternetformation issue in Bluetooth by setting a frameworkfor scatternet analysis based on a matrix representation,which allows developing and applying different metrics.They identified several metrics (capacity, average load,or path length) both in a traffic independent and ina traffic dependent context, and showed the relevantnumerical results. Then, a distributed algorithm forscatternet topology optimisation was introduced, thatsupports the formation of a locally optimal scatternet

based on a selected metric. Numerical results obtainedby adopting this distributed approach to optimise thenetwork topology were shown to be close to the globaloptimum. Cuoma et al. (2004a, 2004b) extended theirwork (Melodia and Cuomo, 2004a, 2004b; Cuomo andMelodia, 2002) and provided an integrated approach forscatternet formation and quality-of-service support (calledSHAPER-OPT) by combining the tree-based scatternetformationalgorithmSHAPER (Cuomoet al., 2003) and theDistributed Scatternet Optimisation Algorithm (DSOA)(Melodia and Cuomo 2004a, 2004b; Cuomo andMelodia,2002). A combination of Dynamic Slot Assignment (DSA)and piconet partitioning is proposed in Cordeiro et al.(2006). With DSA, the piconet master dynamically assignsslots to slaves so as to allow them to communicatedirectly with each other without any intervention fromthe master, which make a multicast-like communicationfeasible within the piconet.

Recently, Vergetis et al. (2005) investigated issues thatBluetooth may face in a large-scale ad hoc network.They showed deciding whether there exists at least oneconnected topology that satisfies the Bluetooth constraintsis NP-hard. However, the NP-hardness is only true whenthe network is a general graph or in a three-dimensionalspace. If the network is a UDG (as we assumed in thispaper), the problem of deciding whether connectivity isfeasible and constructing a connected topology whichsatisfies the desired degree constraint become polynomialcomplexity. Vergetis et al. also showed by simulationsthat the time required to generate a stable topology for alarge number of nodes can be large (their method is basedon relative neighbourhood graph). They pessimisticallyconcluded that Bluetooth may not be widely used inbuilding large ad hoc network. Indeed, establishing alarge Bluetooth scatternet is a very challenging problem.However, we still believe that Bluetooth technology canbe used to form large-scale networks, especially when thenetwork is static. We hope that results from this papercan lead to design of more efficient hierarchical structure,which increases the network’s capability of handling largenumber of devices.

7 Conclusion

In this paper, we proposed a practical solution for largemulti-hop Bluetooth scatternet formation and IP-basedrouting mechanism according to Bluetooth specification.We proposed a novel communication efficient method tobuild a CDS as the backbone of multi-hop Bluetoothnetwork. Then dBBlue scatternet is formed for eachcluster. The final scatternet, M-dBBlue, guarantees theconnectivity and each cluster has self-routing property.Our experiment shows the majority of nodes have degreesmaller or equal to seven which means our scatternetseldom parks any node. Our scatternet also enjoysefficient updating, since both the backbone and thedBBlue structure of each cluster can be maintainedefficiently in a dynamic environment. Ourmethod does not


need any position information of Bluetooth devicesfor scatternet construction and packets routing. It isinteresting to notice that M-dBBlue Bluetooth networkintrinsically supports the future P2P applications, sinceeach cluster supports the content based routing throughpseudo-balanced de Bruijn structure.

Acknowledgement

The work ofWen-Zhan Song was supported in part by USNational Aeronautics and Space Administration (NASA)under Grant No. NNX06AE42G. The work of Yu Wangwas supported in part by the US National ScienceFoundation (NSF) under Grant No. CNS-0721666.The work of Xiang-Yang Li is partially supported by USNSF under Grant No. CNS-0832120, National NaturalScience Foundation of China under Grant No. 60828003,National Basic Research Program of China (973 Program)under Grant No. 2006CB30300, the National HighTechnology Research and Development Program ofChina (863 Program) under Grant No. 2007AA01Z180,Hong Kong RGC under Grant No. HKUST 6169/07and HKBU 2104/06E, and CERG under GrantPolyU-5232/07E.

References

Alzoubi, K., Wan, P-J. and Frieder, O. (2002) ‘Message-optimal connected-dominating-set construction for routingin mobile ad hoc networks’, Proc. ACM Int. Symposiumon Mobile Ad-Hoc Networking and Computing (MobiHoc),Lausanne, Switzerland, pp.157–164.

Alzoubi, K., Li, X-Y., Wang, Y., Wan, P-J. and Frieder, O.(2003) ‘Geometric spanners for wireless ad hoc networks’,IEEE Transactions on Parallel and Distributed Processing,Vol. 14, No. 4, pp.408–421.

Augel, M. and Knorr, R. (2004) ‘Bluetooth scatternet formation– state of the art and a new approach’, Proc. 17thInternational Conference on Architecture of ComputingSystems (ARCS), LNCS 2981, Augsburg, Germany.

Balas, E. and Yu, C.S. (1986) ‘Finding a maximum cliquein an arbitrary graph’, SIAM J. Computing, Vol. 15,pp.1054–1068.

Barriere, L., Fraigniaud, P., Narajanan, L. andOpatrny, J. (2003)‘Dynamic construction of bluetooth scatternets of fixeddegree and low diameter’, Proc. 14th ACM-SIAM Symp.on Discrete Algorithms (SODA), Baltimore, Maryland,pp.781–790.

Basagni, S. and Petrioli, C. (2002) ‘A scatternet formationprotocol for ad hoc networks of Bluetooth devices’,Proceedings of the IEEE Semiannual Vehicular TechnologyConference, VTC Spring 2002, May, Birmingham, AL,pp.6–9.

Basagni, S., Bruno, R. and Petrioli, C. (2002) ‘Performanceevaluation of a new scatternet formation protocol formulti-hop Bluetooth networks’, Proceedings of the 5thInternational Symposium on Personal Wireless MultimediaCommunications, WPMC 2002, 27–30 October, Honolulu,Hawaii, pp.208–212.

Basagni, S., Bruno, R. and Petrioli, C. (2002) ‘Device discoveryin bluetooth networks: a scatternet perspective’, Proc.IFIP-TC6 Networking Conference, Networking 2002, Pisa,Italy, pp.1087–1092.

Basagni, S., Bruno, R. and Petrioli, C. (2003) ‘A performancecomparison of scatternet formation protocols for networksof bluetooth devices’, Proc. IEEE International Conferenceon Pervasive Computing and Communications (PerCom),Dallas-Fort Worth, Texas, pp.341–350.

Basagni, S., Bruno, R., Mambrini, G. and Petrioli, C.(2004) ‘Comparative performance evaluation of scatternetformation protocols for networks of bluetooth devices’,Wirel. Netw., Vol. 10, No. 2, pp.197–213.

Carter, R. and Park, K. (1993) How Good are GeneticAlgorithms at Finding Large Cliques: An ExperimentalStudy, Technical Report BU-CS-93015, Computer ScienceDept., Boston University.

Chiasserini, C-F., Marsan, M.A., Baralis, E. and Garza, P.(2003) ‘Towards feasible distributed topology formationalgorithms for Bluetooth-based wpans’, Proc. 36th HawaiiInternational Conference on System Science (HICSS-36),Big Island, Hawaii.

Chlamtac, I. and Farago, A. (1999) ‘A new approach to designand analysis of peer to peer mobile networks’, WirelessNetworks, Vol. 5, pp.149–156.

Chou, M-T. and Chang, R-S. (2004) ‘Blueline: a distributedBluetooth scatternet formation and routing algorithm’,IASTED International Conference on Parallel andDistributed Computing and Networks (PDCN), Innsbruck,Austria, pp.153–158.

Cordeiro, C., Abhyankar, S. and Agrawal, D.P. (2006) ‘Scalableand QoS-aware dynamic slot assignment and piconetpartitioning to enhance the performance ofBluetooth adhocnetworks’, IEEETransactions onMobile Computing, Vol. 5,No. 10, October, pp.1313–1330.

Cuomo, F. and Melodia, T. (2002) ‘A general methodologyand key metrics for scatternet formation in bluetooth’,Proceedings of IEEE Globecom 2002, November, Taipei,Taiwan.

Cuomo, F., di Bacco, G. and Melodia, T. (2003) ‘Shaper:a self-healing algorithm producing multi-hop Bluetoothscatternets’, Proceedings of IEEE Globecom 2003,December, San Francisco, CA, USA.

Cuomo, F., Melodia, T. and Akyildiz, I.F. (2004a) ‘Distributedself-healing and variable topology optimization algorithmsfor qos provisioning in scatternets’, IEEE Journal onSelected Areas in Communications, Vol. 22, No. 7,pp.1220–1236.

Cuomo, F., di Bacco, G. and Melodia, T. (2004b) ‘Optimizedscatternet topologies for personal area networking indynamic environments’, Proceedings of IEEE IEEEInternational Conference on Communications (ICC 2004),June, Paris, France.

de Bruijn, N. (1946) ‘A combinatorial problem’, KoninklijkeNederlandse Academie van Wetenschappen, Vol. 49,pp.758–764.

Dong, Y. and Wu, J. (2003) ‘Three Bluetree formationsfor constructing efficient scatternets in Bluetooth’, Proc.7th Joint Conference on Information Sciences, Cary,North Carolina, pp.385–388.


Fraigniaud, P. and Gauron, P. (2003) The Content-AddressableNetwork d2b, Tech. Rep. Technical Report TR-LRI-1349(also appeared in 22nd ACM Symp. on Principles ofDistributed Computing (PODC)).

Guerin, R., Kim, E. and Sarkar, S. (2002) ‘Bluetoothtechnology key challenges and initial research’, Proc.SCS Communication Networks and Distributed SystemsModeling and Simulation CNDS, pp.157–163.

Guerin, R., Rank, J., Sarkar, S. and Vergetis, E. (2003)‘Forming connected topologies in bluetooth ad hocnetworks’, Proceedings of ITC’18, Berlin.

Huang, T-C., Yang, C-S., Huang, C-C. and Bai, S-W. (2003)‘Hierarchical grownBluetrees (HGB) – an effective topologyfor Bluetooth scatternets’, Proc. International Symposiumon Parallel and Distributed Processing and Applications(ISPA 2003), LNCS 2745, Aizu, Japan.

Huang, T.C., Huang, C.C. and Bai, S.W. (2006) ‘Hierarchicalgrown bluetrees (HGB): an effective topology forBluetooth scatternets’, International Journal ofComputational Science and Engineering, Vol. 2, Nos. 1–2,pp.23–31.

Ishii, H. and Kakugawa, H. (2002) ‘A self-stabilizing algorithmfor finding cliques in distributed systems’, Proc. 21st IEEESymposium on Reliable Distributed Systems (SRDS’02),Suita, Japan, pp.390–395.

Johnson, D., Maltz, D. and Hu, Y-C. (2003) The DynamicSource Routing Protocol for Mobile Ad Hoc Networks(DSR), Internet Draft, draft-ietf-manet-dsr-09.txt.

Kawamoto, Y., Wong, V.W.S. and Leung, V.C.M. (2003)‘A two-phase scatternet formation protocol for Bluetoothwireless personal area networks’, Proc. IEEE WirelessCommunications and Networking Conference (WCNC03),New Orleans, LA.

Karp, B. and Kung, H.T. (2000) ‘GPSR: Greedy PerimeterStateless Routing for wireless networks’, Proc. ACM/IEEEInternational Conference on Mobile Computing andNetworking (MobiCom), Boston, Massachusetts, USA,pp.243–254.

Kuhn, F., Wattenhofer, R. and Zollinger, A. (2003) ‘Worst-caseoptimal and average-case efficient geometric ad-hocrouting’, Proc. 4th ACM Int. Symposium on Mobile Ad-HocNetworking and Computing (MobiHoc), Annapolis,Maryland, USA, pp.267–278.

Law, C., Mehta, A.K. and Siu, K.Y. (2001) ‘Performance of anew bluetooth scatternet formation protocol’, Proc. ACMSymposium on Mobile Ad Hoc Networking and ComputingMobiHoc, Long Beach, California, USA, pp.183–192.

Li, X-Y., Stojmenovic, I. and Wang, Y. (2004) ‘Partial delaunaytriangulation and degree limited localised Bluetoothmultihop scatternet formation’, IEEE Transaction onParallel and Distributed Systems, Vol. 15, No. 4,pp.350–361.

Li, X. andYang, X.Z. (2005) ‘eBlueScatter: an energy-efficient adhoc network formation algorithm over Bluetooth’, WSEASTransactions on Information Science and Applications,Vol. 2, No. 8, August, pp.1034–1045.

Lin, C.R. and Gerla, M. (1997) ‘Adaptive clustering for mobilewireless networks’, IEEE Journal of Selected Areas inCommunications, Vol. 15, No. 7, pp.1265–1275.

Liu, Y., Lee, M.J. and Saadawi, T.N. (2003) ‘A Bluetoothscatternet-route structure for multihop ad hoc networks’,IEEE Journal on Selected Areas in Communications,Vol. 21, No. 2, pp.229–239.

Marsan, M.A., Chiasserini, C.F., Nucci, A., Carello, G. andde Giovanni, L. (2002) ‘Optimising the topology ofbluetooth wireless personal area networks’, Proc. IEEEINFOCOM, New York, NY, USA, pp.572–579.

Melodia, T. and Cuomo, F. (2004a) ‘Ad hoc networkingwith Bluetooth: key metrics and distributed protocols forscatternet formation’, Ad Hoc Networks (Elsevier), Vol. 2,No. 2, April, pp.185–202.

Melodia, T. and Cuomo, F. (2004b) ‘Locally optimal scatternettopologies for Bluetooth ad hoc networks’, Proceedings ofFirst Working Conference on Wireless On-demand NetworkSystems (WONS 2004), January, Madonna di Campiglio,Italy.

Pagani, E., Rossi, G.P. and Tebaldi, S. (2004) ‘An on-demandBluetooth scatternet formation algorithm’, Proceedings ofFirst Working Conference on Wireless On-demand NetworkSystems (WONS 2004), January, Madonna di Campiglio,Italy.

Perkins, C.E. and Royer, E.M. (1999) ‘Ad-hoc on demanddistance vector routing’, Proceedings of the 2nd IEEEWorkshop on Mobile Computing Systems and Applications,February, New Orleans, LA, pp.90–100.

Perkins, C.E. and Bhagwat, P. (1994) ‘Highly dynamicdestination-sequenced distance-vector routing (dsdv) formobile computers’, Computer Communications Review,October, pp.234–244.

Petrioli, C., Basagni, S. and Chlamtac, I. (2003) ‘Configuringbluestars: multihop scatternet formation for bluetoothnetworks’, IEEETransactions onComputers, Vol. 52, No. 6,pp.779–790.

Petrioli, C., Basagni, S. and Chlamtac, I. (2004) ‘Bluemesh:degree-constrained multi-hop scatternet formation forBluetooth networks’, Mobile Networks and Aplications,Vol. 9, No. 1, pp.33–47.

Petrioli, C. and Basagni, S. (2002) ‘Degree-constrained multihopscatternet formation for bluetooth networks’, Proc. IEEEGLOBECOM, Taipei, Taiwan, pp.33–47.

Salonidis, T., Bhagwat, P., Tassiulas, L. and LaMaire, R. (2001)‘Distributed topology construction of bluetooth personalarea networks’, Proc. IEEE INFOCOM, Anchorage, Alaska,pp.1577–1586.

Shih,K.P.,Wang, S.S. andSu, J.H. (2007) ‘A scatternet formationalgorithm for efficient routing on bluetooth networks’Journal of Information Science and Engineering, Vol. 23,No. 3, May, pp.819–836.

Song, W-Z., Li, X-Y., Wang, Y. and Wang, W. (2005) ‘dBBlue:low diameter and self-routing bluetooth scatternet’, Journalof Parallel and Distributed Computing, Vol. 65, No. 2,pp.178–190.

Tekkalmaz, M., Sozer, H. and Korpeoglu, I. (2006) ‘Distributedconstruction and maintenance of bandwidth and energyefficient bluetooth scatternets’, IEEE Transactions onParallel andDistributed Systems, Vol. 17,No. 9, September,pp.963–974.

Vergetis, E., Guerin, R., Sarkar, S. and Rank, J. (2005) ‘Canbluetooth succeed as a large-scale ad hoc networkingtechnology?’, IEEE Journal on Selected Areas inCommunications, Vol. 23, No. 3, pp.644–656.


Wan, P-J., Alzoubi, K. and Frieder, O. (2002) ‘Distributedconstruction of connected dominating set in wireless ad hocnetworks’, Proc. IEEE INFOCOM, New York, NY, USA,pp.1597–1604.

Wang, R.L., Tang, Z. and Cao, Q.P. (2003) ‘An efficientapproximation algorithm for finding a maximum cliqueusing hopfield network learning’, Neural Computation,Vol. 15, pp.1605–1619.

Wang, Z., Thomas, R.J. and Haas, Z. (2002) ‘Bluenet – anew scatternet formation scheme’, Proceedings of the35th Hawaii International Conference on System Science(HICSS-35), Big Island, Hawaii.

Whitaker, R.M., Hodge, L. and Chlamtac, I. (2005) ‘Bluetoothscatternet formation: a survey’, Ad Hoc Networks, Vol. 3,No. 4, July, pp.403–450.

Wu, J. and Li, H.L. (1999) ‘On calculating connected dominatingset for efficient routing in ad hoc wireless network’,Proc. 3rd ACM International Workshop on DiscreteAlgorithms and Methods for Mobile Computing andCommunications, Seattle, Washington, pp.7–14.

Yao, A.C-C. (1982) ‘On constructing minimum spanningtrees in k-dimensional spaces and related problems’,SIAM J. Computing, Vol. 11, pp.721–736.

Zaruba, G.V., Basagni, S. and Chlamtac, I. (2001) ‘Bluetrees– scatternet formation to enable bluetooth based adhoc networks’, Proc. IEEE ICC, St. Petersburg, Russia,pp.273–277.

Zhang, X. and Riley, G.F. (2005) ‘Energy-aware on-demandscatternet formation and routing for bluetooth-basedwireless sensor networks’, IEEE CommunicationsMagazine, Vol. 43, No. 7, July, pp.126–133.

Website

Bluetooth SIG, Specification 2.0 of the Bluetooth System,http://www.bluetooth.com/

Documents

Multi-hop scatternet formation and routing for large scale … · Multi-hop scatternet formation and routing 253 challenge to efﬁciently form a scatternet with basic requirements