Bluetooth scatternet formation: A surveycs752/papers/bluetooth-002.pdfcomparisons are technically correct, Bluetooth has a much more powerful business model associ-atedwithit,basedontwokeypoints.Firstly,Blue-tooth

Ad Hoc Networks 3 (2005) 403–450

www.elsevier.com/locate/adhoc

Bluetooth scatternet formation: A survey

Roger M. Whitaker a,*, Leigh Hodge a,1, Imrich Chlamtac b

a Centre for Mobile Communications, School of Computer Science, Cardiff University, Queen’s Buildings,

5 The Parade, P.O. Box 916, Cardiff CF24 3XF, UKb Erik Jonsson School of Engineering and Computer Science, The University of Texas at Dallas,

P.O. Box 830688, Richardson, TX 75083-0688, USA

Received 1 October 2003; accepted 1 February 2004Available online 3 March 2004

Abstract

This paper describes the issue of piconet interconnection for Bluetooth technology. These larger networks, known asscatternets, have the potential to increase networking flexibility and facilitate new applications. While the Bluetoothspecification permits piconet interconnection, the creation, operation and maintenance of scatternets remains open.In this paper, the research contributions in this arena are brought together, to give an overview of the state-

of-the-art. First, operation of the Bluetooth system is explained, followed by the mechanism for link formation. Then,the issue of piconet interconnection is considered in detail. Processes for network formation, routing and intra- andinter-piconet scheduling, are explained and classified. Finally, the research issues arising are outlined.� 2004 Elsevier B.V. All rights reserved.

Keywords: Bluetooth; Scatternets; MAC; Routing; Scheduling; Connectivity

1. Introduction

The Bluetooth (BT) technology, as specified inthe Bluetooth System Version 1.1 [29], is the firstflexible, mass market protocol for wireless ad hoc

1570-8705/$ - see front matter � 2004 Elsevier B.V. All rights reservdoi:10.1016/j.adhoc.2004.02.002

* Corresponding author. Tel.: +44 29 2087 4812; fax: +44 292087 4598.

E-mail addresses: [email protected] (R.M. Whi-taker), [email protected] (L. Hodge), [email protected] (I. Chlamtac).

1 Supported by the EPSRC (GR/S23155/01).

operation. This means that global control of thenetwork is relinquished, permitting spontaneousdeployment without dependence on fixed infra-structure. However, decentralised network organi-sation is required to enable self-planning andmanagement, which is a key challenge in develop-ing fourth generation technology.

The BT specification has been converted intoIEEE standard 802.15 [16,88]. From the outset,the specification has been mass-market oriented,driven by a special interest group (SIG) of majormanufacturers, keen to develop a robust, flexible,

ed.

mailto:[email protected]

mailto:[email protected]

mailto:chlamtac@utdallas. edu

mailto:chlamtac@utdallas. edu

404 R.M. Whitaker et al. / Ad Hoc Networks 3 (2005) 403–450

and economically viable technology that can beeasily incorporated into a multitude of devices toenhance their functionality via short connectivity,using low power radio links (10–100 m) in the unli-censed 2.4 GHz industrial, scientific and medical(ISM) band. Most current Bluetooth devices offerthe modest range of 10 m using the (low) powerclass 2. This is a deliberate action, as Bluetoothis currently primarily a technology for personalarea networks (PAN), where device battery poweris limited. However, in the long term, higher powerand extended range could lead to a range of fur-ther applications, beyond personal area networks.Additionally the technology may well prove a use-ful starting point for the development of otherapplications, and systems, such as those in the mil-itary arena.

In its current form, Bluetooth is well on the wayto reaching impressive levels of penetration, helpedby its royalty free status. It is estimated that 35million chip-sets were produced in 2002, with anestimated rise to 510 million by 2006 [35]. Thecharacteristics of the Bluetooth chip [96] havefacilitated this, based on an occupancy of 90mm2, power consumption of 25 lA when idle witha peak of 25 mA, and a cost of less than 4 USAdollars per terminal for high volume production.

Since the conception of the Special InterestGroup (SIG) in 1998 and the release of specifica-tion (v1.1 in February 2001), Bluetooth has re-ceived a considerable amount of attention, beingvigorously marketed by the SIG who promise atechnology to ‘‘seamlessly connect all your mobiledevices’’. Commercial interest frequently centresaround its application (potential future applica-tion is given in [20] for example). An importantlimiting factor in developing future applicationsare the networking possibilities that the technol-ogy can facilitate. As currently specified, Bluetoothself-organises networks of up to eight active de-vices in piconets, but the specification permitsmuch more. Piconet interconnection is possible,permitting multi-hop links between out-of-rangedevices. However, mechanisms for establishingand maintaining such scatternets remain open,challenged by the need to provide totally decentra-lised, high performance solutions for potentiallydynamic environments.

Consequently, the extent to which the Blue-tooth networking capabilities can be advancedare receiving particular attention, and this is thefocus of our survey on the state-of-the-art. Therapid pace of research in this area has led to alarge, dispersed and fragmented literature, whichbroadly seeks to further investigate and extendthe operation of Bluetooth at the equivalent ofthe network layer in the traditional OSI model.This constitutes development for network forma-tion, maintenance, scheduling and routing in thecontext of Bluetooth.

We begin by introducing the Bluetooth specifi-cation and the operation of the system for a singlepiconet. To support scatternets, we then considerthe extension of functionality in the BT specifica-tion. The broad issues involved in defining a proto-col to support scatternet operation are described inSection 2. In Section 3, we describe the general is-sues which influence the design of protocols in thecontext of Bluetooth. In Section 4, we consider thedifferent possible topologies that have been pro-posed for scatternet applications. In Section 5 weconsider the problem of forming and maintaininga scatternet, specifically the process of establishingdevice role. In Section 6, we describe the differenttechniques that have been applied to assess net-work performance. In Section 7, we consider theissue of routing. Finally, in Section 8 we addressthe problem of scheduling inter- and intra-piconetcommunication.

1.1. Moving from piconets to scatternets

The original and first use of Bluetooth technol-ogy has been for small clusters of devices, whichoperate using the concept of a piconet. This is acollection of devices sharing the same communica-tion channel. Devices such as laptops, mobilephones, PDAs are currently the main proponentsof the technology, and it is likely to remain thisway in the short term. However, there are a num-ber of arguments which support the developmentof the technology for inter-connecting piconetsto form scatternets. Like many technologicaldevelopments, scatternets may precede concreteapplications. However, high levels of market pen-etration will increase the density of Bluetooth en-

R.M. Whitaker et al. / Ad Hoc Networks 3 (2005) 403–450 405

abled devices, thereby increasing the potential forscatternet based applications.

Scenarios requiring a greater amount of connec-tivity have received limited attention. In [38],the authors argue that scatternet functionality isimportant to allow flexible formation of Bluetoothpersonal area networks. Additionally, it is arguedthat scatternet functionality may also be used toimprove the performance of a group of nodes thatare either already part of a scatternet, or part ofseparate piconets. The roles of devices in suchnodes may be rearranged to adapt to a new trafficdistribution.

Bluetooth also can be used in a cellular fashionvia access points for wireless LAN applications,which creates further opportunities for scatternetapplications. Bluetooth WLANs are already in ser-vice for large scale commercial IP applications,such as conference scenarios like ‘‘Cebit 2003’’,where 150 Bluetooth access points were providedfor piconet formation [47]. Critics may be quickto dismiss the use of Bluetooth in this context,due to superior performance of the dominant wire-less LAN technology, WiFi (IEEE 802.11b). Thisoffers a higher data rate over a much greater dis-tance (50 m versus 10 m for Class 2 Bluetooth de-vices). In [20], it is pointed out that while thesecomparisons are technically correct, Bluetoothhas a much more powerful business model associ-ated with it, based on two key points. Firstly, Blue-tooth is designed for a myriad of wirelessPAN applications. Secondly, the cost and size ofthe technology means that it is being included inequipment by default. These points mean thatopportunities will exist to further network devices,including access points, for Bluetooth applications.

Finally, the low cost, power efficient specifi-cation of the technology means that there arepotential Bluetooth applications in other ad hocscenarios such as sensor networks [1]. These net-works involve a large number of densely deployedwireless sensors, which need to be low power andlow cost devices. The current literature acknowl-edges Bluetooth as being too expensive and toopower consuming for sensor network application.However, there are a number of additional pointswhich need to also be considered. The cost of asensor network node should be less than 1 dollar

to make the network feasible [74]. Clearly, the cur-rent Bluetooth production cost of 4 dollars [96]exceeds this, but it is feasible that this cost willfall further. In [27], it is argued that Bluetooth isnot currently efficient for sensor networks becauseturning them on and off adds to the energy con-sumption. However, this has to be contrastedagainst the additional features that a Bluetoothsensor network could provide, including higherdata rates than many sensor network solutions[1], off-the-shelf specification and compatibilitywith any other Bluetooth enabled device. Thiswould enable direct access to other wired and wire-less services. There may be scenarios where theseadvantages out weigh the increase in cost andadditional power consumption. If Bluetooth wereadopted in this context, scatternet formation forlarge networks would be essential.

2. The Bluetooth specification

Although it is not the purpose of our paper todescribe the BT specification in detail, we do pro-vide an overview to set the context of researchactivities. The Bluetooth radio system is definedto resolve the following fundamental issues in anad hoc communication scenario, concerning appli-cation of the radio spectrum, discovery of devices,connection establishment, channel allocation,medium access control, interference and powerconsumption. These aspects are the focus of ourdescription of BT. In Fig. 1, we present the organi-sation of the Bluetooth stack. There are now manydetailed expositions of various aspects of the BTsystem including books [17,62,70] and general arti-cles [11,31,32,38], from which we provide a briefsummary.

2.1. Radio layer

At the lowest level, the radio layer of the corespecification defines the wireless interface. Spreadspectrum frequency-hopping occurs in the 2.4GHz ISM band using either 79 or 23 radio frequen-cies in countries with restrictions in the ISM band.A fast hopping rate of 1600 hops per second occurs,using pseudo-random hopping sequences, which

RF

Baseband

Link Manager Protocol

L2CAP

Data

RFCOMM

HC

I Control

Service Discovery protocol

Audio

PPP

Applications

Fig. 1. Organisation of the BT stack.


provide an automatic method for controlling inter-ference from other sources in the unregulated ISMband. These frequencies are located at (2402 + a)MHz, for a = 0,1, . . . ,78. The modulation tech-nique is Gaussian Frequency Shift Keying (GFSK)which enables a simple low cost radio chip designand a transmission speed of up to 1 Mb/s. Thebaud rate is 1 Msymbols per second. Three powerclasses are defined for transmission, 20 dBm (100mW), 4 dBm (2.5 mW) and 0 dBm (1 mW). A ser-vice threshold of �70 dBm or better, is required.For further details of the radio layer, the reader isreferred to [32], which provides a description ofthe system including link budget characteristicsand signal-to-interference ratio information.

2.1.1. Hopping sequences

Frequency hopping provides interference diver-sity gain and frequency diversity gain, which arerequired for multiple channels to co-exist indepen-dently. The frequency hopping mechanism used byeach device operates in a ‘‘black box’’ fashion,with a device identity and clock as input, and a fre-quency as output. More specifically, the sequenceis selected by the unit identity, and the phase bythe clock. A huge number of sequences are avail-able to each unit, with each covering about 23 h.This prevents repetition of an interference patternwhen hopping sequences support different commu-nication channels in the same proximity. Thirty

two hops are designed to span about 64 MHz ofspectrum. This is designed to maximise interfer-ence immunity by spreading as much as possibleover a short time interval. Over a larger time inter-val, each frequency in the sequence has equalchance of occurrence, and perturbation of theclock and/or identity leads to an instantaneouschange in the selected hop. This permits devicesto switch between different hop sequences and per-mits devices to time share between two or morecommunication channels. This requirement meansthat storage of sequences in memory is infeasible.Sequences must be generated on the fly using log-ical circuitry.

2.2. Baseband

The baseband specifies how the radio layershould be employed to facilitate communicationbetween Bluetooth devices. This layer defines theconcept of a piconet, which is BT�s fundamentallogical topology for organising group-wise com-munication, under decentralisation. At this point,homogeneous devices are distinguished by theirBluetooth device address which is a unique 48-bitaddress, hard-coded into the Bluetooth chip.Additionally, each device holds a free-runningclock which ticks every half time-slot for a hoppingrate of 1600 hops per second. Exchange of clocksand BT addresses is fundamental to the formationof a piconet. This is defined as a group of devicesthat share the same communication channel. Thekey innovation is that piconets are formed indecentralised ad hoc manner, without interventionor assistance. Homogeneous devices take on oneof two different roles: master or slave. Each piconetconsists of exactly one device whose role is themaster, and at most seven other active deviceswhose roles are slaves. The role of master andslave are relative to one piconet, at one point intime. It is the master who defines the communica-tion channel used by all members of the piconet.The first device to initiate the formation of a pic-onet becomes the master. Every other device inrange is assigned a locally unique active memberaddress. These take up the role of the slave withinthe master�s piconet. At most seven active slavesparticipate in each piconet, but additional slaves


can be registered with the master and sustain theparked mode. Devices outside of any piconet sus-tain the stand-by mode.

The master functions to co-ordinate the intra-piconet communication. The master may commu-nicate with any member of the piconet, but slavesin the piconet may only communicate with the mas-ter. This means that the communication betweenslaves may only occur via the master. The principlerole of the master is maintenance of synchronisa-tion. This is organised by referencing slots in pico-nets as odd or even, according to the second leastsignificant bit of the clock of the master. To deter-mine the frequency hopping sequence in a piconet,slaves maintain the offset time between their clockand that of the master, using a slot dwell time of625 ls, and apply pseudo-random sequencing offrequencies. The master uses each time slot in thesequence to communicate with a slave, before mov-ing onto the next slave in the sequence. However,multi-slot packets requiring three or five slots arepermitted. When these are transmitted, the trans-mit frequency remains constant. The master thenresumes transmission with the slaves whose turnit would have been under single slot transmission.

Uplink and downlink between master and slaveoccurs using time division duplex, with the masteronly communicating in even numbered slots, andslaves in odd numbered slots. Furthermore, slavesmay only transmit to the master if the master hasjust transmitted to the slave, giving the master con-trol over the slave for purposes of controlling mul-tiple slot transmissions. Note that the piconetstransmission channel is fundamentally dependenton the clock of the master. Consequently a devicecannot be a master in two piconets simultaneously,since this would result in two piconets having thesame communication channel, resulting in signifi-cant co-channel interference between piconets.

2.2.1. Device admission

Prior to commencing communication in a pico-net, the slaves need to identify the address andclock of the master, and similarly the master needsto know the identity of slaves. This is organised byinquiry and paging phases defined at the basebandlayer. A scan procedure is carried out by listeningdevices who are idle.

The device carrying out the inquiry phase be-comes the master in a future piconet. The inquiryprocess is used to discover other devices andpaging is used to establish connection with them,via invitation. These are uni-directional processeswhich require the participation of both the in-quired and inquirer. Each needs to commencefrom a different initial state to ensure discovery.The principle problem that inquiry involves iscollision avoidance between receiver and sender.Two devices cannot exchange messages until theysynchronise the channel (i.e., frequency hoppingsequencing) for communication. Bluetooth usesthe simple approach of defining a globally adoptedhopping sequence known to all devices. Howeverthe inquirer hops at a greater speed than the in-quired, and the inquirer listens in between trans-missions for a response.

A further complication exists concerning multi-ple listeners. To avoid collision between responsesfrom different inquired devices, a randomisedback-off time is set by each device to stagger replytimes. Eventually the sender device collects thebasic information (such as device address andclock) from the listeners, which is then used inthe paging process. This procedure is performedby the inquirer, and is an invitation to join the pic-onet. With the information sent by the paging de-vice, the enquired device can join the piconet as aslave. After joining a piconet, the slave may nego-tiate role reversal, where the original master be-comes the new slave and vice versa.

The scan process is used by idle devices whichperiodically become active and listen for devicestrying to make a connection. Every time a devicewakes up, a different hop carrier is scanned for aperiod of approximately 10 ms. The hop carrierscanned is defined by the wake-up sequence. Thisis a cyclic sequence of 32 hops, in a pseudo-ran-dom order defined uniquely for each device. Herethere is a trade-off between power consumptionand response time.

More complex steps exist for admitting a newslave into an existing piconet. There are two pro-cesses by which this may occur, either active orpassive. The master may inquire on new nodes inits transmission range and invite them to join thepiconet. Alternatively the master may wait to be


discovered, by performing the scan phase. Forboth options, communication in the piconet mustbe suspended. Therefore a trade-off exists betweenpiconet capacity and latency of device admission.

2.3. Medium access control

The BT specification has been developed to per-mit co-existence of a large number of unco-ordi-nated communications. This is enabled by a largenumber of independent channels, each of whichis used by a restricted number of devices. Eachchannel is defined by a master device. For the mod-ulation technique adopted, a single frequency hop-ping channel in the ISM supports 1 Mb/s, andtheoretically, the channel with 79 carriers can sup-port 79 Mb/s. However, the non-orthogonality ofthe hop sequences means that this limit is unlikelyto be reached. Channel capacity must be shared be-tweenmembers of the piconet, and so the number ofactive slaves is deliberately limited to seven. Whenthe number of piconets is large relative to the num-ber of devices, channel capacity will be exploited.

The master/slave role within a piconet only ex-ists for the duration of the piconet. Once disbanded,each device could potentially resume as a master orslave. As well as defining a piconet, the device withmaster status controls all traffic across the piconetand admission control. Slots are used alternativelyfor master–slave and slave–master communication.To avoid collision of multiple slave transmissions,the master adopts a polling technique, whichinvolves the master defining which slave is to trans-mit in each slave-to-master slot. Only the slave ad-dressed in a master-to-slave slot may communicatein proceeding slot. If the master has information tosend a specific slave, the slave is polled implicitlyand can return information. If the master has noinformation to send, it polls the slave with a shortpolling packet. As the master schedules bi-direc-tional traffic, scheduling algorithms need to be ap-plied to efficiently use channel capacity. Althoughmasters operating in the same area use differentchannels, there is the probability of loosing a packetdue to another transmission using the same carrierhop. Information is transmitted without listeningfor a clear carrier. If data is received incorrectly, itis resent at the next opportunity. This does not

occur for voice calls. The simplicity of BT and thefast hopping speed makes collision avoidanceschemes inappropriate.

As portable low power devices are envisaged forBT operation, control of power consumption isimportant. A number of modes of operation havebeen defined to minimise power consumption. Inidle mode, the device performs the scan operationfor less than 1% of the time. The park mode re-duces device activity further, but can only be ap-plied to slave devices after a piconet has beenformed. In park mode, devices remain synchron-ised with the master but do not return a packetwith a payload. Park mode permits more thanseven slaves to participate in a piconet. A furtherlow power mode is the sniff mode, where the slavedoes not scan in every master–slave time slot, buthas a low duty cycle. Effectively the device onlywakes up periodically to communicate with themaster. Finally, the hold mode is used to put de-vices to sleep. This is used to suspend intra-piconetcommunication while the master searches for newdevices to admit to the piconet.

2.3.1. Link types and packet format

Two types of link can be supported. A singleasynchronous connectionless link (ACL) is sup-ported between a master and a slave, for servicessuch as bursty data traffic. Additionally, up tothree synchronous connection-oriented (SCO) linksmay be supported in a piconet, for services suchas voice traffic. The SCO link is a point-to-pointlink supported by reservation of duplex slots(odd and even). The ACL link is a point-to-multi-point link from the master to all slaves. AnACL link can use all of the remaining slots onthe channel not used for SCO links. The slottedstructure of the piconet channel means that ACLand SCO links can be mixed.

BT streams the information into packets, with asingle packet being sent in each slot. Each packetconsists of an access code (72 bits), packet header(54 bits) and payload (0–2745 bits). The access codeis used to ensure to identify the piconet to which thepacket belongs. Only if the access code matchesthat of the piconet master will the packet be ac-cepted by the recipient. This prevents packets fromone piconet being falsely accepted in another


piconet. The packet header contains link controlinformation, a 3 bit slave address to distinguishslaves in the piconet, a 1 bit acknowledgement forrepeat request, a 4 bit packet type code to define16 different payload types and an 8 bit header errorcheck code used to detect errors during transmis-sion. The header has maximum size of 18 informa-tion bits which are protected by further coding.

Four control packets are defined. These are theidentification packet, consisting of only the accesscode and used for signalling; the null packet whichcontains the access code and a packet header,being sent only to convey link control information;the poll packet, sent by the master to force a re-sponse from the slaves; and the synchronisation

packet, used to exchange real-time clock and iden-tity information between devices.

Both forward error correcting and packetretransmission schemes are included in BT. ForFEC, a 1/3-rate code and a 2/3-rate code are used.The 1/3-rate code is a simplistic approach where a3-bit repeat is used with a majority decision at therecipient. This is used on the packet header andcan additionally be used on the payload for SCOlinks. The 2/3 rate FEC code consists of a short-ened Hamming code which can be applied to thepayloads of either SCO or ACL links.

2.3.2. Link manager and host controller interfaceThe baseband state machine is principally con-

trolled by the BT link manager. This firmware isprovided with the link control hardware, and han-dles link setup, security and control. Its remit in-cludes control of paging, changing slave modes,and handling required changes in master/slaveroles. It also supervises the link and controlshandling of multi-slot packets. Link managerscommunicate with each other using the link man-agement protocol (LMP). This is organised byLMP packets which are sent in the payload ofpackets on asynchronous connectionless linksand are flagged by a bit in the ACL header.

Some link controller hardware may include ahost controller interface (HCI) layer above the linkmanager. This is a firmware layer which is used toseparate the BT baseband and link manager from atransport protocol. The HCI defines a standardinterface independent method of communicating

with the firmware. Three standard transport mech-anisms are supported in BT: USB, RS-232 andUART. The HCI allows a Bluetooth applicationto access BT hardware in the absence of the trans-port layer or other hardware implementation de-tails. The HCI layer forms a component in theBT stack, but it does not constitute a peer-to-peercommunication layer since the HCI commandand response messages do not flow across the phys-ical link.

2.3.3. Link layer and L2CAP

All BT protocols above the link manager andHCI are software based. The lowest level is thelogical link control and adaptation protocol(L2CAP) specification which is effectively BTs linklayer. The L2CAP delivers packets received athigher layers to the other end of the link. It isrequired because baseband packet size is too smallfor transporting higher layer packets. The L2CAPresolves this, and operates over an ACL link pro-vided by the baseband. It performs segmentation,re-assembly and multiplexing of high level applica-tions above the HCI. The L2CAP supports themultiplexing of several logical channels over thedevices ACL links. A single ACL link is alwaysavailable between the master and any active slave.This provides a point-to-multi-point link support-ing data transfer. A detailed description of thechannels, channel state machine, connection, con-figuration, disconnection and L2CAP packets isgiven in [78].

2.3.4. The service discovery protocol and profile

specification

The service discovery protocol (SDP) is used todetermine which BT services are available on aparticular device. Each Bluetooth device can actas a client or server in discovering services withSDP. The SDP only provides information aboutthe different services which are available. Furtherprotocols (either from BT or elsewhere) must beused in order to use a service. In order to preservethe efficiency of the process and minimise theamount of information which needs to be carriedacross the BT link between devices, universally un-ique identifiers (UUIDs) are used. The SDP oper-ates with each server cataloguing the all available


services provided by the device, and each serveroperates search and browse facilities.

To support interoperability for a range of appli-cations, profile specifications define how a funda-mental range of operations should be organised.These define usage for generic access, service deliv-ery, mobile telephony interconnection, intercom,serial ports, headsets, dial-up networking, fax ser-vices, LAN access, and usage models for genericobject exchange, object push, file transfer andsynchronisation. Of particular interest are thearrangements for serial port and LAN access.The former case is facilitated by adopting theRFCOMM protocol. This emulates the signalson an RS-232 interconnection cable and is basedon the ESTI 07.10 standard. This permits the emu-lation and multiplexing of several serial ports overa single channel. The key benefit from adoptingthe RFCOMM is that it enables legacy applica-tions that have been written to operate using serialcables, to run over a BT link.

2.4. Inter-piconet communication

As well as permitting piconets to co-exist inde-pendently, the BT specification makes provisionfor connectivity between piconets. A collection ofconnected piconets is called a scatternet. Inter-pic-onet communication is facilitated due to packetbased communication over slotted links. Thismeans that particular Bluetooth devices can timeslice communication between the channels of mul-tiple piconets. Recall that a device can instanta-neously change hopping scheme with knowledgeof master identity and master clock. In order toallow jumps to be feasible, a guard time occursin the traffic scheduling to allow for slot misalign-ment between different piconets. Accordingly ahold mode is used to allow a unit to temporarilyleave one piconet and visit another piconet. Thehold mode may also be used as an additionalpower control mechanism. The role of the devicein each piconet is flexible, with the constraint thata device cannot be a master in more than one pic-onet, by definition. However, it is acceptable for adevice to be a master in at most one piconet and aslave in multiple other piconets. Beyond the func-tionality for devices to time-slice between multiple

piconets, support for the creation, operation andmaintenance of scatternets remains open.

3. Factors influencing scatternet protocol design

Bluetooth networks are distinguished fromother networks currently associated with ubiqui-tous and pervasive computing in a range of ways:spontaneous network formation, isolation frominfrastructure, simple low cost devices with powerconstraints and links with states that permit lowpower. The decentralised formation, maintenanceand operation of fully connected networks is spec-ified for small numbers of devices, organised viathe concept of a piconet. The pre-requisite inestablishing a piconet is that all devices are inrange at least one device, which is required tooperate as the master. This precludes obtainingfully connected networks in scenarios where

• a larger number of devices (i.e., greater than 8)require connectivity;

• not all devices requiring connectivity are withinrange of at least one member.

In both these scenarios, scatternets are required:that is multiple interconnected piconets. Animportant paper fully exposing the possibilities ofinter-piconect communication is [38]. Here it isstressed that in the context of personal area net-works (PANs), although a single piconet may besufficient, the possibility for a device in the PANto be present in multiple piconets is essential toallow the combination of various Bluetooth usagecases, examples of which are provided.

Although scatternets are supported in the BTspecification, protocols are not explicitly definedfor creation, maintenance and operation. Thecomplexity of these tasks significantly increases,when moving from single piconets to multiple con-nected piconets. The fundamental issues to beaddressed concern:

• Device statusEach device needs to determine its role withrespect to (possibly) multiple piconets, specifi-cally is role(s) as master and/or slave. This


defines the links between devices and thereforethe topology of the network, which influencesthe performance of the network. There is a largedegree of freedom in the number of feasi-ble alternative scatternets (see [12]), whichdefines a significant combinatorial optimisationproblem. This is made more difficult by thedecentralised nature of the problem, character-ised by a lack of a centralised entity with globalknowledge.

• RoutingIn a single piconet, any pair of devices are atmost two hops apart. The master has totalknowledge of all slaves, and controls their abil-ity to send and receive messages. Therefore in asingle piconet, packet delivery between any pairof devices becomes trivial. However, complexityis significantly increased in scatternets, since apacket may have to transverse multiple piconetsto reach its destination. Additionally theremay be a number of alternative routes, theknowledge of which must be organised anddistributed.

• Intra- and inter-piconet communication schedulesThe master in a piconet is required to share timebetween slaves. Some slaves will become signif-icantly more important than others, particularlyif they perform a relay function between distinctpiconets. Consequently schemes for schedulingbecome increasingly important. Additionally,devices involved in multiple piconets can onlyparticipate in one at a time. Consequentlyresource sharing schemes need to minimisepotential conflicts.

Mechanisms to determine device status, sched-uling and routing are influenced by many factors.These include distribution of devices, scalability,device differentiation, environmental dynamism,integration between co-ordination issues and level

of centralisation. These factors are important be-cause they serve as a guide in protocol design.

3.1. Distribution of devices

A commonly applied assumption in the designof protocols is that all devices are in mutual range(e.g., [4,49–51,53,56,76,80]). Under this assump-

tion, the primary aim is to increase the numberof devices which can connect and form a network.This assumption simplifies scatternet formation,because local (i.e., 1 hop) neighbourhood knowl-edge at each device constitutes global networkknowledge. This assumption is applied, based onthe observation that the current expectations forBluetooth are focussed on personal area network-ing, where a maximum network diameter of 10–15hops is envisaged.

3.2. Scalability

In the role of WPAN, network scalability is alimited concern. However, network protocolswhich scale will be of use if applications are devel-oped involving a large number of nodes, such as insensor networks, as surveyed in [1]. The scenarioswhich are considered range from personal area net-works (such as in [80] with up to 36 devices) to thelargest scenarios involving 2000 devices (e.g.,[102]). The issue of scalability is best addressed byprotocols which do not assume that all devicesare in range. For some protocols, analytical resultscan be derived on time complexity and messagingcomplexity, relative to number of devices. In theseapproaches, some properties of scaling are effec-tively guaranteed (e.g., [49]). As noted in [30], sim-ple protocols may have good scaling properties butat the expense of overall network quality.

3.3. Device differentiation

It is likely that the most appropriate scatternetfor a given collection of devices will depend oncharacteristics of the device such as battery life,computational abilities and likely traffic load. In[22] for example, these factors have been incor-porated into the protocol to create the network.However the predominant stance involves assum-ing devices are homogeneous. This follows fromthe specification, and the fact that any Bluetoothdevice can take a role as master or slave. Howevera number of authors make further logical hierar-chies between devices, particularly for the creationof scatternets. These devices take leading roles, fre-quently through acquiring and using greaterknowledge of other devices (e.g., [80,102]).


3.4. Environmental dynamism

As each device is autonomous in an ad hoc sce-nario, synchronisation between devices, in terms ofdevice start-up and activity time, cannot be as-sumed. This means that scatternet membership islikely to change and the network will need to re-pair or heal under conditions of failure. The casesunder which failure occurs, for example, are ad-dressed in [56]. Mechanisms will need to rapidly re-act in particularly dynamic environments, such asthose considered in [95]. On-demand techniquesare one solution to environmental change. This ap-proach is used in [57,58], for example, but operatesat the cost of increased overhead. Additionally,creation of networks with high levels of fault toler-ance are desirable, as sought in [95].

3.5. Integration between co-ordination issues

Each of the co-ordination issues can be ad-dressed in isolation, or with integration. However,intra- and inter-piconet scheduling is dependenton routing, which is dependent on the networktopology (i.e., collective device status), and thiscould potentially be exploited in protocol designand operation. Consequently, some authors haveargued for significant integration via cross layeroptimisation. A range of solutions and approachescan be undertaken. For example, [2] seeks to deter-mine network topologies which maximise thenumber of device pairs which can operate simulta-neously. The integration between scatternet struc-ture and inter-piconet communication is addressedin [43]. The development of protocols which formparticular topologies can aid routing. For exam-ple, [24] adopts ring structures only, while in [90],a tree structure is adopted with Bluetooth deviceaddresses carefully used to aid packet navigation.In [57,58], on demand network formation is devel-oped where routing is implicit.

3.6. Level of centralisation

A key issue in protocol design is the reliance oncentralisation, where one entity (e.g., single deviceor subset of devices) gains total network knowl-edge. The benefit is that such knowledge makes

higher performance solutions achievable. How-ever, this induces potentially greater overhead,against which increases in network quality mustbe assessed. Greater overhead is more likely tobe permissible in environments with little dyna-mism. In fully decentralised protocols (e.g., [98])no device has greater knowledge than any other.In partially decentralised protocols, hierarchiesexist based on the level of knowledge devices attain(e.g., [80]). The extreme case is when one nodegains total total knowledge to create the network.Independent of use as a protocol, entirely centra-lised problem solutions have been considered(e.g., [60]) and are challenging computationalproblems in their own right.

4. Link formation and network topology

The fundamental step in setting up a scatternetis local device discovery and the formation ofpoint-to-point links, where pairs of devices learnof each other�s identities and synchronise hoppingsequences to form piconets. In scatternet scenar-ios, a number of devices are required to participatein more than one piconet to facilitate network-wide connectivity. We describe the basic linkestablishment process provided in the Bluetoothspecification.

4.1. Bluetooth link establishment

Link establishment in Bluetooth is controlledby four processes:

• InquiryThis process consists of broadcasting inquirypackets, which do not contain the senders iden-tity or other any information.

• Inquiry scanIn this state, devices listen for inquiry packets,and upon detection of an inquiry packet, thedevice broadcasts an inquiry response packet.This contains the devices identity and clock ofthe device in inquiry scan mode.

• PageUnder this process, a device tries to establish aconnection with a device whose identity and


clock are known. Page packets are sent, whichcontain the sending devices address and clockfor synchronisation. The packets sent can onlybe received by those devices with particularidentities.

• Page scanIn this state, a device listens for a page packet.Receipt is acknowledged and synchronisationbetween the page and page scan devices isestablished.

These four processes need to be co-ordinatedand controlled in order to establish and maintainpiconets and scatternets. From the baseband spec-ification (see Section 2.2), Bluetooth defines a sim-ple process for link formation between a pair ofdevices. This process is asymmetric since the pro-tocol distinguished the devices as a sender and re-ceiver. The sender starts in the inquiry state whilethe receiver is in the inquiry scan state. Initiallythere is a frequency synchronisation delay untilthe sender transmits on the frequency on whichthe receiver is currently listening (recall that thesenders and receivers are hopping at differentrates). When the inquiry packet is received, the re-ceiver backs off for a random time period (uni-formly distributed in the range 0–639.375 ms).Hopping then resumes at the hop it was listeningto, immediately prior to back-off. A second fre-quency synchronisation delay then occurs, and asecond inquiry packet is received from the sender.

The receiver replies with a frequency hoppingsynchronisation (FHS) packet containing thereceivers address and clock value and the receiverenters the page-scan state. Upon receipt of theFHS packet, the sender enters the page state. Inthe page state, the sender transmits a Device Ac-cess Code (DAC) packet, that can be uniquelyheard by the receiver. The senders address con-tained in the FHS packet, is used to estimatethe phase of the receiver, in order to eliminatethe frequency synchronisation delay. The receiverresponds with a DAC packet, and the sender rep-lys with a FHS packet. The receiver then uses thisinformation to determine the channel hopping se-quence and the phase of the sender. At this pointthe receiver becomes the slave in the communica-tion channel and replys with another DAC packet.

As soon as the sender receives this DAC packet,the sender becomes the master in the connection,and both devices are now synchronised acrossthe same hopping sequence.

The delay components of link formation arerandom back-off delay and the frequency synchro-nisation delay, which occurs twice. In [80], this isshown to be at most 659.375 ms when 32-hopsare available for the universal frequency hoppingset used in the inquiry process, and 649.375 mswhen 16-hops are used. This connection establish-ment delay, however, depends on the pre-assign-ment of roles for the devices. In an ad hocscenario, no pre-assignment of roles will be avail-able, making this link establishment processproblematic.

4.1.1. Symmetric link formation protocol

In [79,80], a protocol for symmetric link forma-tion is developed, which is analysed in [81]. Theapproach is symmetric in the sense that no senderor receiver allocation is required. This is achievedby ensuring that the two nodes (nominally the sen-der and receiver) alternate node status between in-quiry and inquiry scan states, in an unco-ordinatedmanner. When the devices meet for a sufficientlylong time interval in opposite states, they try toconnect according to the asymmetric link forma-tion protocol.

Clearly the schedules for state alternation arecrucial in determining the link formation delay.In [81], it is shown that a random schedule withstate residence times following a common randomdistribution are preferable to deterministic solu-tions, following intuitive logic. The mean and var-iance of link delay formation for random statealternation are derived in [81].

Salonidis et al. [81] propose the scheme as a toolwhich can be generally used by protocols for scat-ternet formation ‘‘on the fly’’, which are builtabove this. Such an example is [102], where theprotocols are developed on the assumption thatdevices know the identifiers of their one-hopneighbours. The approach introduced in [80] iswell used, with this and similar mechanisms beingadopted by a range of authors including[7,9,10,53,72,76,83,94]. The symmetric protocolfor link formation is a suitable mechanism for


device discovery of single-hop neighbours, when atemporary piconet is formed specifically for thispurpose. The general issue of device discovery ina multi-hop rather than single-hop scenario hasbeen addressed in [6].

4.2. Network topology

Scatternet formation protocols are mainly re-quired to control the states of inquiry, inquiryscan, page and page scan at each device, so thatgroups of devices synchronise on common hop-ping sequences, to form interconnected piconets.This determines the device status as a master orslave in possibly multiple piconets. Additionally,the protocol is required to account for the dynamicnature of the ad hoc environment, where scatternetmembership is variable, and device residence in thescatternet is unpredictable.

The role that devices take in the scatternetdetermines the network topology and conse-quently the performance of the network. Forexample, the number of piconets a device belongsto, particularly in situations where are many po-tential master nodes, has significant implicationson characteristics of the network. Of crucialimportance are bridges, which are devices that re-side in more than one piconet. There are two typesof bridge. If the bridge is a master in one piconet,the bridge is of a master–slave type. Otherwise the

Master inone piconet

Master in one piconet and slave in another

Disjoint Piconets

Fig. 2. Examples of piconet a

bridge is a slave in all piconets to which it belongsand is a slave–slave type.

Bridge location and distribution is importantsince participating multiple piconets leads to in-creases in processing and communication over-head. Note that bridges can only be active is onepiconet at a time. Switching between piconetscan reduce throughput, and increases the demandson inter-piconet scheduling. However, tradedagainst this, bridge devices improve connectivityand thereby reduce path length between deviceswhich can improve throughput. An example of apossible scatternet topology with a variety of pos-sible bridges is displayed in Fig. 2. The protocolsintroduced for scatternet formation frequentlyuse heuristics to control possible configurationswhich can be formed, including:

• Bridge devices must never be masters. Thisreduces the scheduling burden on masters,who then only need to consider intra-piconetcommunication.

• The number of bridges per piconet is restricted.This restricts the number of potential inter-pico-net conflicts for bridging devices, but at theexpense of limited potential alternative routes.

• The number of piconets must be kept as smallas possible. This reduces the number of commu-nication channels used, and the potential forinterference.

Slave in one piconet

Slave in two piconets

Connected Scatternet

nd scatternet topologies.

Table 1A summary of topological features imposed in scatternet formation

Paper Bridge devicesmust never bemasters

Number ofbridges perpiconetrestricted

Number ofpiconets mustbe kept as smallas possible

Piconets mustnot be linkedby more thanone bridge

A device mustparticipate in alimited numberof piconets

Barriere et al. [5] • •Law et al. [51] • •Lin et al. [56] • • • •Petrioli et al. [71] •Ramachandran et al. [76] •Sato et al. [83] • •Salonidis [80] • •Stojmenovic et al. [89] •Wange et al. [98] •Zaruba et al. [102] •Zhang et al. [103] • • •Zhen et al. [104] •


• Piconets must not be linked by more than onebridge. This minimises the co-ordination issueswhich need to be addressed by scheduling.

• A device must participate in a limited numberof piconets. This controls the amount of inter-piconet scheduling at the device level.

The way in which these assumptions have beenapplied are summarised in Table 1.

4.3. Modeling

In order to represent the Bluetooth system forscatternet formation, a model is required. This ab-stracts the problem to some extent, and frequentlyinvolves graph theory, where devices represent thevertices and links between devices represent edges.A range of graph theoretic structures can be used inthe modeling process so that the resultant scatter-nets have topologies with particular characteristics.

Nodes may be labelled according to their partic-ular role in the scatternet. Indeed, this is a centraltask in formation of a scatternet, and is a pre-req-uisite for routing and scheduling. In addition tobeing labelled as a master or slave with respectto a single piconet, a node may be labelled as abridge (master–slave or slave–slave). For slavenodes, the degree of a vertex determines the num-ber of piconets to which it belongs. The degree of amaster determines the size of the piconet it defines.

Control of degree has been a central issue in manyscatternet formation models and algorithms, par-ticularly in [7,25,30,53,56,60,71,80,89].

Protocols for scatternet formation frequentlyuse the concept of a visibility graph. This repre-sents the graph of all possible links which could

be formed between devices in range. In this con-text, the protocols for scatternet formation mustlabel nodes and select edges from the visibilitygraph.

4.4. Classes of topology

Beyond applying particular characteristics inTable 2, protocols can be designed to create partic-ular network topologies, based on a graph theo-retic structure. In this section we review thefrequently used structures.

4.4.1. Trees

A tree is a connected graph with no cycles. Thismeans that a Bluetooth network which forms a treehas a minimal number of edges for connectivity.However this also means that there is no choiceof route for traffic between nodes: only a singlepath exists. Additionally this topology is suscepti-ble to partitions when links and devices fail.

In [43], a two level hierarchical tree structure isformed, composed of a root, followed by leaves.The root is the initiating piconet, and the slaves

Tab

le2

Asu

mm

ary

ofsc

attern

etfo

rmat

ion

and

maintenan

cepro

toco

ls

Pro

pose

rCen

tra-

lise

d/

global

coord

i-

nat

ion

Distrib

-

uted/

loca

l

coord

i-

nat

ion

Rad

io

range

Fixed

positions

Two

hop/

gateway

Multi-hop/

interm

e-

diate

gateway

Geo

metric

appro

ach

Gra

ph-

bas

ed

appro

ach

Rin

g

stru

c-

ture

Tre

e

stru

c-

ture

Hiera

r-

chical

stru

cture

Mes

h

stru

c-

ture

Lim

ited

num

ber

ofnodes

Min

imise/

lim

it

num

ber

ofpiconet

Power

man

age-

men

t

Weigh

ting

ofnodes

Optim

ise

capac

ity

Optim

ise

thro

ugh

-

put

Dyn

amic/

mainte-

nan

ce

Max

–

min

optim

i-

sation

Bar

rier

e

etal.[5]

••

••

Baa

tz

etal.[2]

••

••

••

Bas

agnian

d

Petrioli

[9,72]

••

•

Chiasser

ini

etal.[22]

••

••

••

••

Chun

etal.[24]

••

•

Cuom

o

etal.[25]

••

••

Law et

al.[51]

••

••

•

Liet

al.[54]

••

••

••

Liet

al.[53]

••

••

Lilak

iatsak

un

etal.[55]

••

•

Liu

etal.[57]

••

Lin

etal.[56]

••

•M

arsa

n

etal.[60]

••

••

••

Petrioli

etal.[71]

••

••

•

Sat

oet

al.[83]

••

Salonid

is

etal.[80]

••

••

•

Sto

jmen

ovic

etal.[89]

••

•

Sun

etal.[90]

••

•Tan

etal.

[92,94

,95]

••

••

Wan

get

al.

[98]

••

Zar

uba

etal.[102

]

••

••

•

Zhen

etal.

[104

]

••

••



in this piconet then form bridges (and masters) inthe new piconet. The authors find that this ap-proach has lower throughput than a non-tree alter-native topology where bridges are only masternodes, but performs better in terms of average sys-tem delays.

Routing considerations for tree based structureshave been considered in [90]. Routing in trees issimplified because there is a unique path betweensource and destination devices. This means thatthe routing considerations need only address dis-semination of this particular route. Sun et al. [90]propose a protocol to create and maintain routingtable information, through extending the idea of abinary tree search to sub-trees in the network. Pro-tocols are also introduced to create such treestructures.

In [94] the use of a tree topology is justified bysimplified routing, and the authors explicitly men-tion the tension between connectivity and latency.It is argued that minimising the total number oflinks in the topology reduces the contention fortransmission slots in the time division duplexscheme. The protocol proposed achieves a mini-mum number of average piconets per bridge byensuring that bridges participate in exactly twonodes. In [95], the tree formation algorithm is eval-uated in dynamic conditions.

A protocol to form so-called Bluetrees is pro-posed in [102]. This operates using two phases, inthe first of which sub-trees are formed using parti-cular nodes. The second phase concerns spanningthese sub-trees. The paper makes a number ofinteresting observations concerning tree structuresand Bluetooth. In particular, in an open, interfer-ence and obstacle-free environment, if a node nhas more than five neighbours then there are atleast two nodes among the neighbours that arethemselves neighbours. Additionally, it is notedthat most distributed algorithms (and conse-quently protocols) for finding a spanning tree ina network operate by creating subtrees and thenexpanding them to form a single spanning tree.

The advantages of the tree structure are basedin simplicity. A minimal number of links leads tosimple routing issues. However, this structure canonly be applied in dynamic environments if theprotocol can quickly respond to link failures,

which in the case of a tree, are certain to partitionthe network.

4.4.2. Gabriel and Yao graphs

These structures are closely related to the con-cept of a unit disk graph (UDG), in which nodesin the Euclidean plane are considered to define adisk of unit radius. In a UDG, an edge is definedbetween a pair of vertices if their Euclidean dis-tance is at most one. A constrained Gabriel graph

over a graph G contains an edge uv if and only if:

• uv 2 G;• the open disk when using uv as diameter does

not contain any node w from G such that bothuw and wv are in G.

Consequently a Gabriel graph produced from agraph G has vertices with a sparse geographicaldispersion. The Yao graph from a graph G selectsedges based on the proximity of associated nodes.A Yao graph with an positive integer parameter kis defined as follows. At each node u, any k equallyseparated rays originate and define k separatecones. In each cone, a shortest directed edge isselected.

The Yao and Gabriel graphs are used in [54] toconstruct and analyse general network topologiesfor wireless communication. One useful point isthat these graphs can easily be constructed in adecentralised manner. The investigation considersthe power efficiency of the shortest paths betweensource and destination, and the tolerance of nodemovement using a power stretch factor. In [53],the approach is extended specifically for scatternetformation, and the authors propose a new sparsegeometric structure for this purpose.

In [89], Gabriel and Yao graph structures alsofeature in a dominating set based approach forformation and maintenance of scatternets. A dom-inating set in a graph G is a subset of nodes who,collectively, are adjacent with all other nodes inthe graph. Consequently these sets are useful forpurposes of connectivity.

4.4.3. 1-Factors

In [2], the authors seek to construct scatternettopologies to ease the problem of inter-piconet


scheduling. A matching in a graph G is a subset ofedges such that no two are incident. A perfectmatching has n/2 edges, where n is the number ofvertices. When n is odd, a near perfect matchingdescribes a matching with (l � 1)/2 edges. Notethat (near-) perfect matchings are best possible.A graph which can be partitioned into (near-) per-fect matchings is 1-factorizable, as these matchingsare also known as (near-) 1-factors.

It is pointed out that a matching determines apossible communication pattern in the scatternet.This is because each edge represents a communi-cating master–slave pair, so the size of a match-ing (i.e., number of edges) determines thenumber of simultaneously active piconets. Conse-quently (near-) 1-factors are desirable. Baatzet al. [2] briefly describe how to build scatternetsfrom 1-factors of complete graphs, which occurwhen all devices are in transmission range.

4.4.4. Ring structuresIn contrast to purposely avoiding topologies

containing cycles (i.e., Section 4.4.1) some scatter-net topologies advocate their use, principally dueto increasing reliability. In [24], ring structure scat-ternets called Bluerings are proposed, consisting ofn devices where each device belongs to two pico-nets and has two links in total. Each device is botha master and a slave, and a BlueRing supports amaximum of bn

2c active links. This leads to a struc-

ture where there are exactly two mutually disjointpaths between source and destination for all pairsof devices. Since there is a minimal number of de-vices for a master to switch between, the authorspoint out that switching overheads will be limited.Additionally, routing is very simple since incomingpackets are merely forwarded. However, theauthors also point out two disadvantages of thering structure. Firstly large diameter rings couldhave high packet delay, and secondly, not all visi-bility graphs will have the necessary 2-connectivityto create a BlueRing.

In [56], the use of more general ring structures isexplored, via the BlueRing protocol (unrelated tothat in [24]). In this approach, relatively smallnumbers of piconets are sought, which are inter-connected to form a ring. A set of principles guidethe protocol: specifically, nodes participate in no

more than two piconets and two piconets areinter-connected by at most one bridge. Addition-ally, bridges are never masters, and piconets are re-quired to contain at least two slaves. This isbecause the protocol requires that distinct slavesare used to communicate with the adjacent pico-nets in the ring.

5. Scatternet formation and maintenance protocols

Protocols for scatternet formation are regardedby many authors as finite state machines, control-ling states including inquiry, inquiry scan, pageand page scan. The protocol ideally needs to func-tion independently at each device, without anysynchronisation between devices (e.g., [30]). Butthis makes it difficult to control the network pro-duced, or guarantee particular properties of thenetwork, such as end-to-end connectivity. A com-promise to this approach are staged or phased pro-tocols, where all devices perform activities with aloose degree of synchronisation. For example,three common phases are device discovery, piconetformation, and subsequent interconnection ofpiconets. Time intervals are defined in which eachdevice must perform each phase (such as neigh-bourhood discovery), but the devices are unsyn-chronised within phases. This loose degree ofsynchronisation makes it possible to prove proper-ties of the resultant network (e.g., [72]). Phasedprotocols for formation frequently assume en-

masse device start-up to establish these properties,since it is difficult to establish global results on net-work characteristics otherwise.

In this section, we consider the principle contri-butions concerning protocol development for scat-ternet formation. We describe how the protocolsoperate, and the rationale behind them.

5.1. BlueStars and BlueMesh

In [72] the concept of BlueStars is introduced.This is a three-phase protocol which operates bydiscovering topology, creating piconets and thenconfiguring these to form a scatternet. Piconetsare referred to as BlueStars and the connectedscatternet is termed a BlueConstellation. Topology


discovery is dependent on the symmetric knowl-edge of neighbouring devices. This is explicitlyengineered using temporary piconets. Weights ofneighbours are built up when device identities areexchanged.

In the second phase, individual piconets areformed using a dynamically computed weightwhich expresses the suitability of a node for themaster role. This leads to the selection of the mas-ter and slaves in disjoint piconets. Nodes are se-lected when they have the biggest weight in theirown neighbourhood. Once a node has decided onthe role of master, it pages neighbours to becomeslaves. The receiving node will join its first biggestneighbour who provides an invitation.

Finally, in the last phase, gateway devices (i.e.,bridges) are selected for interconnection of pico-nets. This uses information gathered during theformation of piconets. Each master determinesits neighbouring (i.e., 2-hop or 3-hop) master de-vices. A master is designated as an init master ifit has the highest weight among all its neighbour-ing masters. The init masters control the scatternetformation, selecting bridges with the largest weightamong alternatives. In the case of three hop neigh-bours, pairs of intermediate bridges with thehighest weight sum are chosen. Proof that theproperties of the protocol are fully specified, andanalysis of this approach is given in [7].

The approach is novel in its criteria for deviceself-selection: a device selects itself given knowl-edge of its own weight, and that of its neighbours.Weights are built up at the device discovery stageand represent the degree of a vertex in the visibilitygraph. The resulting scatternet has a mesh struc-ture in which there are multiple paths betweensource and destination pairs. Additionally the pro-tocol does not require devices to be in each othersrange. It is assumed that the location of devices re-mains static throughout, and maintenance of thescatternet is not considered. An important pointis made concerning the considerable device discov-ery time, which is caused by three factors: (1) theneed to adopt stochastic mechanisms so thatneighbouring pairs of devices can each discovereach other using uni-directional inquiry functions;(2) the impossibility of identifying the inquirerleads to the construction of temporary piconets

between neighbours which have already discoveredeach other; (3) lengthy back-off times as stipulatedin the BT specification. Simulation results in [7]show that significant improvements are possibleafter reducing the back-off time.

In BlueStars, the number of slaves is notbounded and the consequence of this is that someslaves may need to be parked, resulting in delays.This is resolved by a further protocol [71] calledBlueMesh, which produces scatternets with a num-ber of interesting properties. The protocol pro-vides multiple paths, the lengths of which arereported to perform well in comparison with thatbest possible in the visibility graph. Additionallymasters have at most seven slaves, and on average,each node does not assume more than 2.3 roles.

BlueMesh operates in two phases, the first ofwhich involves discovery of one and two hopneighbours. The second phase is an iterative proce-dure which consists of role selection followed bygateway selection. As with BlueStars, weights areapplied to determine which nodes are to be mas-ters, and the slaves in a piconet are selected as adominating set over the masters two hop neigh-bours. Numerous properties of the BlueMesh pro-tocol are proven in [71].

5.2. Scatternets via node insertion and removal

In [22], formation of scatternets is considered bydefining procedures to handle topology changes ina Bluetooth network. These are based on the inser-tion and removal of nodes, and the cases whenthese operations need to be performed. The aimof these procedures is to provide feasible scatter-nets which have desirable properties such as fullconnectivity, high throughput, and in particular,reduced overheads due to control messages. Themethod can also be used both to form and main-tain scatternets and has a high level of flexibilityin this regard.

For node insertion, a node wishing to joinstarts the process by broadcasting identity packets.Nodes in proximity, which are listening, respondwith a FHS (frequency hop synchronisation)packet if they are willing to accept the new neigh-bour. Possibly more than one neighbour may re-ply, in which case a decision may be required for


connection choice. The authors of [22] propose cri-teria to make this decision, using classification ofdevices and their status. A generic node class is de-fined by a triple (x,y,z) where x is an identifier forcomputational capabilities and battery capacity, yrecords the devices role (e.g., unspecified, slave,master etc.) and z indicates traffic load of the node.These triples are passed between devices usingshort encoding. The inquiring device then selectsits neighbour(s) for connection. The inquiring de-vice is given an ordered range of alternative nodesto page, choosing the highest ranking alternativewhich is permissible. The criteria for permissionare defined using the triple (x,y,z).

For node removal, the changes in networktopology caused by the nodes absence depend onthe role it played. Four different possibilities areconsidered based on the role of the device. Themost challenging scenarios are when bridges and/or masters stop participating. The removal proce-dures in these cases are structured so that alter-natives are sought using variations on insertionprocedures. The rules provided by the protocol en-sures that no negotiation or collective decisionmaking between subsets of nodes, is required.

5.3. BlueRings

The BlueRing protocol designed in [24] is de-signed to create a circular chain of devices, eachwith bridge status. Two alternative algorithms,NODE-ID and HEAD-SEEK-SCAN, are intro-duced to build up the required topology, princi-pally controlled by identifying and updating thehead and tail of chains of nodes, as they areformed. The principal difficulty in forming the re-quired ring structure is premature formation of aring consisting of only some of the nodes in thenetwork. The algorithms proposed overcome thisby first forming a linear structure consisting ofall nodes in the network, and then closing the ringby forming the last link only after a time outperiod.

Both of the proposed algorithms operate itera-tively for a fixed time in each round, using twoprocedures, SEEK and SCAN. SEEK is theauthors notation for the inquiry procedure. Theformer process denotes a device actively attempting

to discover other neighbouring devices. A seekersends out Bluetooth inquiry messages in the hopeof getting a response from other devices. SCANis the process of listening, and replies with an in-quiry message when it detects a seeker. The SCANand SEEK mechanisms emulate the processes pre-sented in [49].

The NODE-ID algorithm operates with iso-lated nodes having equal chance of performingSEEK or SCAN. At the same time, the head ofany line of nodes performs SCAN and the tail per-forms SEEK. All intermediate nodes in the chainremain idle. Criteria are defined such that the BTidentity of the head of any chain is always the larg-est identifier in that component. This is enforcedby ensuring a seeker component connects to ascanner component if and only if the identity ofthe seeker or head of the seeker component is lar-ger than the identifier of the scanner. The algo-rithm terminates when the seeker tail of aconnected component gets the first response fromthe scanner head for a consecutive number ofiterations.

In the HEAD-SEEK-SCAN algorithm, onlythe head of a line and isolated nodes performSEEK or SCAN operations. This means that onlyone device in each connected component performsscatternet formation in each round. This preventspremature cycle formation in each iteration ofthe algorithm. Unlike the NODE-ID algorithm,device identifiers do not play a role in operationof the algorithm. At each iteration, the isolatednodes and the head of chains perform SEEK orSCAN operations, with equal probability. Thesedevices maintain two variables, componentLengthand tailInfo which respectively detail the numberof devices in the component and the identifier ofthe tail of the component. When a seeker obtainsits first response from a scanner, the seeker pagesthe scanner to form a temporary connection, toenable the seeker and the scanner to exchangecomponentLength and tailInfo variables. The tem-porary connection becomes permanent if one ormore of the seeker or scanner are isolated nodes.Otherwise the tail of the shorter component is bro-ken and the tail of the shorter component forms alink with the head of the other component. Thealgorithm terminates when the head of a line


receives no response after a set number ofiterations.

For both algorithms, it is assumed that all de-vices are in mutual range, and that the processesare synchronised. It is shown that the HEAD-SEEK-SCAN algorithm outperforms the NODE-ID algorithm, and is comparable to the approachproposed in [49]. The main advantage of the Blue-Ring approach is that it provides a degree of rout-ing reliability (i.e., an alternative path) if a node orlink fails. Routing is simplified by packets beingforwarded around the ring. However there are dis-advantages. Firstly, although it may be possible tohave a single operational piconet, the location ofdevices can prohibit the formation of a ring. Onlyin cases where the visibility graph has 2-connectiv-ity can BlueRing formation take place. Secondly,network diameter will be necessarily high, leadingpotential packet delays. A further observationconcerns the number of piconets used. Contradic-tory to many other approaches which seek to min-imise the number of piconets (and consequentlythe number of communication channels andbridges), the ring approach uses a maximal num-ber of piconets, assuming devices are restricted topiconets only.

5.4. Distributed scatternet formation procedure

In [25], a distributed approach is defined in thedistributed scatternet formation procedure (DSFP),which forms a scatternet incrementally andsequentially. Devices are �inspected� by the DSFP,in sequence. Each device takes the decision onwhether to join the current network, based onthe decisions made by devices which have alreadyjoined. For a node to make a decision, at leastone other node in its neighbourhood must alreadyhave joined the DSFP, and is therefore part of thescatternet. The main steps in the procedure are: (1)neighbourhood discovery; (2) organisation of de-vices into sequential order; (3) inclusion of devicesin the scatternet, based on this order, decidingwhich connections to establish with those alreadyactive, and which role to assume, with the aim ofoptimising a target metric.

The approach uses adjacency matrices for check-ing the performance of network configurations, and

a range of different performance measures areconsidered. The approach is inspired by centra-lised approaches for network configuration, andspecific details of the distributed algorithm arenot fully described due to the short paper lengthof [25].

5.5. Simple and lightweight scatternet formation

In [30], a process is introduced based on ran-domisation. The approach is intentionally simpleand lightweight, with each device using only neigh-bourhood information for decision making. Thisminimises overhead, increasing the protocols abil-ity to rapidly adapt to changes. The approach isused to gain a better understanding of when moresophisticated approaches are required.

The protocol operates using a range of cases.Synchronisation is defined as a cycle of inquiryand scan between a pair of nodes. All nodes haveone of four states (unassigned, master, slave,bridge) and initially all are unassigned.

When two nodes synchronise for the first time,and both are unassigned, the one with the highestdegree becomes master, and the other becomes aslave in the piconet defined by the highest degreenode. When two nodes synchronise and one isunassigned and one is a master, the unassignednode joins the piconet of the master if it has lessthan seven slaves. When two nodes synchroniseand one is unassigned and the other is a slave,the unassigned node becomes the master of anew piconet, and the other node joins the piconetas a slave unless it is already a bridge in b piconets.Finally, if two nodes discover each other and nei-ther is unassigned, then a range of cases are con-sidered. If the both are masters then neitherchanges state. If one is a master and the other isa slave in a different piconet, then the slave be-comes a bridge unless it is already a bridge in b pic-onets. Optionally the master may refuse the newslave if it already has a bridge to the slave�spiconet.

This approach represents a simple way of creat-ing and interconnecting piconets. No phased ap-proach is required and all devices need not be inrange. The protocol can also be easily extendedto permit device removal. However properties of


the scatternet (such as connectivity) cannot beguaranteed and are explored empirically in [30].

5.6. Scatternets via merging, moving and migration

In [49–51] various aspects of a new scatternetformation protocol are introduced. The protocolis designed to operate on isolated but in range de-vices, and operates under synchronisation. Thealgorithm operates with devices being partitionedinto components, with a component constitutinga set of interconnected devices. A component canbe a single device, a piconet or scatternet. In eachcomponent there is one device which is leader.There is no freedom in choice of leader for the sin-gle device case. For a piconet, a master is leader.For a scatternet, one master is leader. Leaders per-form a centralised role over the partition, and maycome in and out of retirement as required.

All leaders iteratively perform a MAIN proce-dure in synchronisation, and initially all devicesare leaders. In the MAIN procedure, a leadercalls SEEK (i.e., inquiry) on a probabilistic basis(1/3 < p < 2/3). Otherwise the leader activatesSCAN. This means that each leader has morechance of running SCAN than SEEK. Also duringeach round, the leader structure means that onlyone device in each component is running SEEKor SCAN.

When a leader executes SEEK, it tries to ac-quire a new slave which is running SCAN. How-ever, the probabilistic nature of the protocolmeans that this is not always successful. Therefore,if a leader is unable to contact a slave after a cer-tain time, the leader gives up and tries again withthe MAIN procedure in the next round. In eachround, a graph theoretic matching may be foundbetween the SEEK devices and SCAN devices.The time required in SEEK and SCAN modes isinvestigated in [49].

When a leader u running SEEK connects toa slave v running SCAN, a procedure CON-NECTED is invoked. The new link leads to a lar-ger connected component, and reorganisation ofthe piconets and leader occurs via three opera-tions: MOVE, MERGE and MIGRATE. Theseoperations involve ensuring that new, larger con-nected components have only one leader, and

involve the redistribution of slaves to masterswithout leader status, as far as possible. This re-quires the assumption that every pair of devicesare within range. However in [50], extensions areproposed which do not require this assumption,and operations are defined which permit the proto-col to be applied in dynamic environments. Theprotocol is shown to have O(log n) time complex-ity and O(n) message complexity.

5.7. Partial Delaunay triangulation scatternet

formation

In [53], a decentralised geometric algorithm isintroduced which creates a degree-limited scatter-net. The approach seeks to consider the geograph-ical dispersion of devices and select links so thatthe resultant topology has a planar property.Unusually, the protocol requires knowledge ofthe geographic location of devices, and GPS isposed as a solution to this problem. Other authors(e.g., [7]) have criticised this assumption as beingunrealistic. The protocol is synchronous in itsoperation and ensures that no piconet has morethan seven slaves. The approach is conceptuallycomplex, involving a series of phases in whichthe topology is built up by selection of links sub-ject to graph topology. We briefly describe these.

Initially the neighbourhood discovery phaserenders each node with knowledge of potentialneighbours within transmission range. The nextoptional phase involves partial Delaunay triangu-lation. This is a new class of graph topology, whichis planar and more dense than other common sub-graphs. In the next phase, the degree of each nodeis limited to seven, by applying the Yao structure,and master–slave relationships are formed in sub-graphs. Subsequently a clustering based approachfollows, consisting of several iterations. In eachiteration, nodes with unallocated role havinghigher keys than any of their unallocated-roleneighbours apply the Yao structure to bound thedegree, decide master–slave relations and informall neighbours about either deleting an edge ortheir master–slave decision. Two ways of definingthe master–slave relation are considered: nodewith initially higher degree becomes master or acluster based approach. In the cluster-based


approach, a dominating set of masters in the de-gree limited subgraph is implicitly constructed,and some gateway piconets are added to preserveconnectivity. It is mentioned that creation andmaintenance require small overhead.

In [89], further geometric approaches are pro-posed for formation and maintenance based onthe concept of the dominating set. Geometric solu-tions to scatternet formation can potentially leadto higher quality scatternets because they use moreinformation on the distribution of the devices.However, gaining location information is poten-tially expensive and out of context with the mis-sion of Bluetooth as a low cost technology.Realistically, the cost of location informationestablishment will need to fall substantially beforegeometric approaches could be applied to scatter-net formation.

5.8. BlueRing of trees

Although of the same name as the protocolintroduced in [24], the BlueRing protocol intro-duced in [56] has an entirely different format. In[24], all devices in the scatternet were containedin the ring topology. In [56], the protocol involvesa ring of masters and slaves, and can be viewed asa ring of piconets. The structure of this topologymakes possible easy schemes for uni-cast andbroadcast communication. These schemes arestateless in the sense that no routing informationneeds to be defined and recorded. Single pointand multi-point failures are tolerated by the proto-col. It is assumed that all devices are in range ofeach other for start up of the protocol.

A centralised formation mechanism, similar tothat in [80], is proposed, which involves the identi-fication of leaders, which control the setup of thescatternet. A binary parameter, RING-MEM, ismaintained by each device, to indicate whether ithas become a member of the BlueRing. Construc-tion involves two phases. In stage one, each devicechooses to operate inquiry with a probability p andperform inquiry scan with probability 1 � p. Wheninquiry meets inquiry scan, the two devices set up atemporary piconet to exchange RING-MEM anddevice address information. A so-called winning

device is established at this stage. The device with

RING-MEM=1 wins when the other device hasRING-MEM=0. In the case of a tie, the devicewhich holds the most information on other deviceidentities wins. The loser also provides the win-ner with all its information on other BT deviceidentities. Subsequently the temporary piconet isremoved. The winner becomes a leader if no fur-ther inquiry or inquiry scan message is receivedwithin an inquiry timeout period. The leaders entera page stage, trying to collect other non-leaders,which forms stage two of the procedure.

In stage two, the leader designates several de-vices as masters by paging them and setting up atemporary piconet. The process for achieving thisis not fully detailed in [56]. For each designatedmaster, the leader provides its information ofits slaves, including downstream and upstreambridges. The mechanism for selecting the bridgesis not described. Criteria for maintaining a formedscatternet are detailed, based on two scenarios: sin-gle-point failure and multi-point failure. The lattercase is more demanding as it could require refor-mation of the ring structure if bridges are missingor masters are missing. Case-by-case analysis is gi-ven on how to resolve all possible failures. Thisprotocol is specifically designed to integrate the is-sues of formation, maintenance and routing, butmay be limited by the distribution of the devices.

5.9. Scatternet formation via clustering

In [76], scatternet formation is approached fromthe problem of clustering in ad hoc networks. Twoapproaches to clustering are proposed: a random-ised distributed linear complexity algorithm whichconstructs a minimal set of star-shaped clustersand a deterministic algorithm. The algorithmsare developed generally and then applied to Blue-tooth. The assumption is that all devices are withinmutual transmission range. The goal of the algo-rithms is to maximise the number of nodes in eachcluster so that the number of piconets isminimised.

The randomised algorithm operates in twophases, the first of which results in devices beingprovisionally designated as master or slave. Thesecond phase corrects the effect of randomnessintroduced previously, by using a deterministic


algorithm to decide on the final set of masters andslaves. A super-master is elected which collectsinformation about all the nodes. The super-mastercan then run a centralised algorithm to form a net-work of desired topology between piconets result-ing from the clusters. Each node conducts t roundsof Bernoulli trials with a defined probability ofsuccess p. A node which is successful at least oncebecomes a master designate, otherwise the node isa slave designate. In the second phase, the desig-nated masters select designated slaves which par-ticipate in electing the super-master node. Alarge amount of information is passed between de-vices during the election process which implicitlybuilds up the knowledge of the network.

In the deterministic algorithm, the initial phaseuses the idea that nodes discovering each otherform a tree of responses, the root of the tree beinga master, who collects information about other de-vices which connect to the component. Criteria areformulated which ensure that the component nec-essarily has a tree structure. Subsequently the sec-ond phase of the randomised algorithm is repeatedto elect the super-master, who again is given thislabel due to having centralised knowledge of thetopology. The authors do not propose particularmethods for inter-connection of the piconets giventhe knowledge of the master. Throughout, thealgorithms operate using careful control of theinquiry, inquiry-scan, page and page-scan modes.Both algorithms are synchronous in the sense thatdevice failure is not permitted, and new deviceswill not be connected if they become active afterthe initial phase. The authors conclude that therandomised algorithm outperforms the determinis-tic approach.

5.10. Scatternet formation and extension for

maintenance

A protocol is introduced in [83] which performsthree separate processes: neighbour discovery, for-mation and adjustment for changing location of afixed number of devices. Neighbourhood discoveryis performed by setting up temporary links whichare removed once identities are exchanged. Thisprocess continues for a fixed amount of time andthen all nodes start the formation phase. Each

device sets a random timer to start this phase, toprevent competition by all devices trying to con-struct piconets simultaneously. A device remainsin page scan state until the timer is expired. Whenexpired, the device enters the page state, and triesto construct a piconet as master. The device pagesthe devices previously discovered, and contactsthem one-by-one. If the paging is successful, themaster–slave relationship is established, and theslave cancels its formation timer and does not tryto construct a piconet. A master continues pagingneighbours until it has seven slaves, or there are noneighbours to page. In the process of piconet for-mation, it is ensured that no piconets share morethan one device and no device belongs to morethan two piconets.

In subsequent phases, the aim is to increase theconnectivity by making connections between dis-connected but local piconets, using neighbourhoodinformation held by slaves from the neighbourhooddiscovery phase. Again, a random back-off is usedto stagger the each piconets activities. When acti-vated, the master identifies slaves that belong toan other piconet, and the master obtains the ad-dresses of devices in the adjoining piconet(s). Thenthe master identifies all neighbours of slaves whichare not connected to the adjoining piconet(s). Themaster selects the minimum number of slaves cov-ering the disconnected devices, and instructs thosedevices to go into piconet formation phase. This re-sults in a bridge between disconnected components.

Finally, a scatternet adjustment protocol is de-fined to deal with device movement within thescatternet. This involves the slaves belonging to ex-actly one piconet periodically going into inquiry-scan state, and disconnected nodes which wereformally masters maintaining inquiry state to forma piconet, and former slaves perform the inquiryscan state. This means that the network will main-tain a fixed number of masters. It is not assumedthat all devices are in mutual range of each other.The scatternet adjustment protocol is shown tomaintain a degree of network connectivity.

5.11. Bluetooth topology construction protocol

In [80], a symmetric protocol is defined, withoutthe need for sender or receiver role to be prede-


fined. This is achieved by forcing the two nodes toalternate independently between the inquiry stateand the inquiry-scan state, and when there is suffi-cient time overlap between opposing states, theconnection will be established between devices.The schedules for alternation are explored in[81], where analytical calculation of the meanand variance for link formation delay are pre-sented when state residence times follow a com-mon random distribution.

This symmetric link establishment procedure isextended to a distributed scatternet formation pro-tocol. The protocol seeks to collect informationabout all nodes in the network prior to formingthe scatternet. The aim is to ensure that scatternetsare formed which are appropriate to the deviceswhich participate. The Bluetooth topology con-

struction protocol (BTCP) consists of three phases:initially a leadership election process which is cho-sen to permit asynchronous operation of devices inthe network; role determination is then performedand finally actual connection establishment isperformed.

In the co-ordinator election, an asynchronous,distributed election of a co-ordinator node occurs.This node will ultimately hold centralised informa-tion on the count, identities and clocks of all net-work devices. The process requires all nodes tohave a variable recording the number of votes, ini-tialised as 1. Each node then alternates betweeninquiry and inquiry-scan states. When a node pairdiscover each other, they compare votes. The de-vice with the largest number of votes in the pairis classed as the winner. In the case of a tie thedevice with the largest address is the winner. Theloser device sends to the winner, all the deviceFHS packets of the nodes it has won so far. Theloser then terminates the connection to the winnerin the pair and enters the page-scan state. Thismeans that it will not be able to hear inquiry mes-sages and will only page messages from nodes thatpage it in the future, thereby removing the loosingdevice from the election process, and preparing itfor subsequent protocol phases. At this point thewinning device has increased its votes by the num-ber of votes carried by the looser, and it continuesthe election process by resuming the alternationbetween inquiry and inquiry scan. The process

continues until one device remains as winner,and all others are in page-scan state, awaiting tobe paged by a node with information about them.

The second phase concerns role determination.The unique remaining winner acts as co-ordinatorand has the FHS packets of all nodes. Initially theco-ordinator checks if more than one piconet is re-quired. If not the case, the co-ordinator becomesmaster and connections with all other nodes aremade, who become slaves. If more than onepiconet is required, the co-ordinator must decideon the role that each device must perform in thefinal scatternet. In [81], for n devices, it is shownthat for the scatternet to be fully connected, theminimum number of masters p satisfies

p ¼ 17 �ffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffiffi289 � 8n

p

2

& ’; 1 6 n 6 36:

Note that number of devices is fixed at a maximumof 36. The co-ordinator selects itself and p � 1other devices as masters. [p(p � 1)]/2 other nodesare defined as the scatternet bridges, and theremaining devices are distributed equally amongthe masters for slave-only designation. A tempo-rary piconet is created between all masters andthe co-ordinator to distribute the allocation ofslaves per piconet. Finally in the third phase, ac-tual connection establishment occurs.

The election mechanism is a novel approachto creating and identifying devices with superiornetwork topology knowledge. Throughout, theunderlying assumption is that all devices arewithin transmission range of each other. Also, set-ting up the temporary piconet between masters inthe second phase caps the number of piconets whichare permissible in the final scatternet. The forma-tion algorithm assumes en-masse device start up.

5.12. Self-routing Bluetree scatternet formation

In [90], a tree-shaped scatternet formation pro-tocol is introduced which seeks to minimise thenumber of piconets involved, and facilitate simplerouting. The protocol operates using only theordering of device addresses and a number of oper-ations are defined based on this information. Animportant concept in the protocol is the definition


of range. For a given device i in a tree, this is a paircontaining the smallest and largest device addressesrooted by i. The tree is structured so that a devicei has range (xi,yi), then any device with identitygreater than xi and less than yi must be routed ati. This permits easy navigation around the tree.

Under this protocol, all devices, independent oftheir status or connectivity, infinitely cycle througha range of modes which perform inquiry, inquiryscan, page, page scan and connection. These areincorporated with a range of bespoke operationsto join, update range and prune connected compo-nents as they grow. However in order to maintainthe range property, a lock mechanism is usedwhich permits only one connected component ata time, to be admitted to any particular tree. A fea-ture of the protocol is that when two connectedcomponents join, disconnection (i.e., pruning) ofnodes may be required to ensure that the resultanttree has the required range property. It is the roleof the root in the inquired tree to find the correctposition to incorporate another connected compo-nent and update the range values in the tree. Theprotocol does not specify the point at which devicerole (master/slave) is established in creating thepiconets. The protocol operates in an asynchro-nous manner and does not require all devices tobe in mutual range. The operation of the protocolis not always guaranteed under changing member-ship (node failure), and the particular circum-stances under which problems occur aredocumented in [90]. The protocol seeks to inte-grate network formation with routing issues, sothat routing induces minimal overhead.

5.13. Tree scatternet formation

In [94,95], a protocol is presented for formingscatternets with a tree structure. The authors focuson designing the protocol to deal with frequentchanges in scatternet membership. In particular,the protocol seeks to converge to a steady state,in which the scatternet is fully connected. The pro-tocol handles both nodes arriving incrementallyand en-masse, and similarly for departure. Ran-domisation is used to ensure a balance betweeneach nodes data communication and communica-tion to maintain the scatternet. At any point in

time, the tree scatternet formation (TSF)-generatedscatternet is a forest of connected tree compo-nents. A connected component with only a singledevice is called a free-node. Each node operatesautonomously with only local communication,and alternates between two states: FORM andCOMM. In the COMM state, the node is involvedin data communication with other nodes in its con-nected component.

The FORM state has two sub-states, composedof inquiry and inquiry scan, as advocated in [80]for symmetric link formation. The protocol per-mits either free nodes to connect to free nodes, freenodes to connect with non-root nodes, root nodesto connect only to root nodes, and non-root nodesto connect only to free nodes. These properties en-sures that TSF produces loop-free topologies. Thejoining node takes its role (master/slave) to pre-serve the current structure of device roles in the in-quired component. Time division between FORMand COMM states is important. If the connectedcomponent large, more time is spent in theCOMM state, and conversely so. The amount oftime in the COMM state is set proportionally tothe degree of the node, dependent on the nodesage, which is checked via a threshold. Consider-ation is given to the optimal value for the randominterval in which devices should stay in the FORMstate.

The protocol also gives consideration to healingproperties and link loss. When a master losses con-nection to a slave, it only need check whether it is afree node. When a slave loses connection to a mas-ter, it updates its node type and sets its age to zero.A leaf node in this situation becomes a free nodeand an internal node becomes a root node. Theauthors note that any root is only permitted tohave seven slaves and consequently this restrictsthe total size of the scatternet, but it is reportedthat this does not impede the performance of theprotocol for tests involving up to 100 nodes.

5.14. Bluenet scatternet formation scheme

The scatternet formation scheme proposed in[98] is designed to produce a flat network structurewithout any hierarchy. In comparison with a treestructure, it is shown that although the protocol


spends more resources on maintaining scatternetlinks, more communications traffic can be carried.The protocol does not require all devices to be inrange, and is designed for the setup of the network.

The phased procedure begins by each node lo-cally building up the visibility graph. Then thenodes follow by randomly entering the page state,to try and invite a fixed number of neighbours tojoint its future piconet. The invited nodes becomeslaves and stop paging. Devices which remain iso-lated follow in the next phase by paging all neigh-bours, to try and gain connection in at least onepiconet. If more than one piconet is found inwhich it can participate, then the device becomesa bridge. In the final phase, the formed piconets,which at this stage are largely disjoint, seek inter-connection. The master of each piconet instructsslaves to set up outgoing links. The mechanismfor achieving this is not provided. The resultantscatternet is not guaranteed in the sense that notall BlueNets are connected even when the initialtopologies are. However the probability of connec-tivity occurring increases with device density.

5.15. Blueroot and distributed Bluetrees

In [102], two scatternet formation protocols areintroduced for the creation of tree topologies. Thefirst protocol, Blueroot, creates a tree structureusing a designated node, called the Blueroot. It isassumed that each node knows whether or not itis the Blueroot, the identifiers of the one-hopneighbours, and whether they are currently partof a piconet. The Blueroot node pages its neigh-bours, one by one, implying that the Blueroot willbe master. If a node is paged and it is not currentlypart of any piconet, it accepts the page, thusbecoming a slave to the paging node. Once a nodehas been assigned the role of slave, it initiates pag-ing all of its neighbours one by one, and so on.

When building a Bluetree as described, it is pos-sible that a master is assigned too many slaves. Ageometric observation is used to reconfigure theBluetree, to ensure that no master has more thanfive slaves. It is observed that in an interference-and obstacle-free environment, if a node has morethan five neighbours, then there are at least twonodes among the neighbours which are themselves

neighbours. Based on this observation, an algo-rithm is executed at each master node. If a masterm has more than five slaves, the master polls theslaves to find the identifiers of neighbours, andto find out how many slaves they handle them-selves. Using this information, the master can findtwo slaves, say s1 and s2, which could possibly con-nect. Node s1 establishes a connection with s2,where s2 becomes a master, while s2 is instructedto disconnect from m. This operation retains thetree property and ensures than piconets have a lim-ited, feasible number of active slaves.

In an extension to Blueroot, a protocol calleddistributed Bluetrees, is introduced. This is a treeformation protocol with further distribution, sav-ing on setup time. As with the protocol describedabove, connectivity is established while maintain-ing a limited number of roles per device. Initiallya collection of nodes, called init nodes, are selectedin a distributed manner. These are selected as nodeswith the highest identifier in its neighbourhood.These nodes initiate the Blueroot protocol in paral-lel, with two modifications. Firstly, when a piconetconnection is established, the slave will be informedabout the identifier of the root of the tree. Sec-ondly, when paging neighbourhood nodes whichare already part of a Bluetree, information onrespective roots is exchanged. The information col-lected via these modifications is used in the secondphase of the protocol. This connects sub-tree scat-ternets into one spanning the entire visibility graph.This is performed using a virtual graph, whose vir-tual nodes are the Bluetrees formed in the firstphase of the protocol. An edge exists between vir-tual nodes in the virtual graph if the respectiveBluetrees can be connected via an edge from thevisibility graph. On dedicating one of these virtualnodes the Blueroot, the Blueroot algorithm can berun on the virtual graph.

The protocols are principally designed for en-masse formation of the scatternet, but they areunusual in being performed on a large number(up to 2000) of devices.

5.16. Blue-star island scatternets

The Blue-star island protocol [104] operateswith each device continuously cycling through


inquiry, inquiry-scan states, on a randomised basis.When a handshake takes place between totally iso-lated nodes, a link is established and the master/slave status is allocated. After joining a piconet,slave nodes periodically enter the inquiry-scan stateto facilitate discovery by other masters. Bridge sta-tus is permitted for slaves up to a predefined limiton the number of interconnecting piconets. Nodeswith master status periodically enter the inquiry-scan state, until a limit on the number of slavesper piconet is reached. If a master node scans an-other master node, it breaks all current links andjoins the piconet as a slave. All slaves in the brokenpiconet become totally isolated devices withouteither master or slave status, and start the cycleof inquiry, inquiry scan to find new links.

This simple protocol does not guarantee con-nectivity, but is highly flexible, with no device syn-chronisation required, and is fully decentralised. Itis also sufficient to deal with changes in scatternetmembership and dynamic scenarios. The approachgives some consideration to scheduling (via mini-mising the number of piconets) and also mentionsthe issue of route discovery in this context.

5.17. Centralised scatternet formation

A number of authors have dealt with the prob-lem of scatternet formation with centralisedknowledge. This is of limited use for the applica-tion of a protocol, but is useful for two purposes:

1. finding the best possible performance for agiven visibility graph;

2. gaining knowledge on scatternet configurationsand performance.

The considerable size of the search space (see [12]),in terms of number of feasible scatternets, pre-cludes analytical or exhaustive treatment of thisproblem. An alternative is to take a distributionof devices, and consider the best possible perfor-mance which can be achieved, using the globalknowledge (i.e., visibility graph). On applicationof optimisation techniques, the resultant scatternetcan be directly compared with the scatternetformed by the decentralised protocol. Thisapproach was first advocated in [61], using sim-

ple heuristic algorithms to generate feasible scat-ternets, which were then analysed in detail.Subsequently, more sophisticated optimisationapproaches have been introduced.

In [22,60], a mathematical programming formu-lation is used. This involves minimisation of thetraffic load at the most congested node, subjectto constraints concerning full network connectiv-ity, piconet size and number of piconets, specifiedtraffic requirements (desired source–destinationflows), and the number of roles devices perform.The approach assumes that routes are specified be-tween the source and destinations, and the amountof traffic is also specified. This input reduces thesearch space (i.e., number of possible solutions)because edges along the routes must necessarilybe included in the scatternets produced. Thereforethe approach focuses on specifying the roles ofdevices.

In other papers considering centralised scatter-net formation [53,54,102], the thrust has been toadapt centralised approaches for decentralisedapplication.

5.18. Summary

In Table 2, we present a classification of ap-proaches for scatternet formation and mainte-nance. Characteristics of each approach arecharacterised in terms of control mechanisms,topology and optimisation objectives. Below wedefine the terms used in Table 2.

• Control mechanism� Global control––control is centralised at a sin-gle device. This node gains full knowledge ofall other devices in the area and performs scat-ternet formation, informing each node of itsrole and the links between them.

� Local control––scatternet formation is decen-tralised. Devices determine their roles in theabsence of a fully centralised co-ordinator.

• Topology� Radio range––all devices are in radio range ofeach other.

� Fixed positions––the positions of devices arefixed.


� Gateway––allows piconets to be intercon-nected only by common slaves.

� Intermediate gateway––bridges are allowed tobe masters in one piconet.

� Geometric approach––uses location informa-tion (theoretically via GPS) to determinetopology.

� Graph based approach––uses a specific graphto determine the topology, other than ringor tree.

� Ring structure––devices are formed into a ringgraph based structure.

� Tree structure––devices are formed into a treegraph based structure.

� Hierarchical structure––nodes form a logicalhierarchy (other than slave/master).

�Mesh structure––nodes form a graph with theexplicit requirement of at least two pathsbetween device pairs.

� Limited number of nodes––the approach isapplicable to a limited numbers of devicesonly.

• Optimisation and control of the topology�Minimise/limit the number of piconets––tocontrol interference and inter-piconetcommunication.

� Power management––to minimise consump-tion in the network formation process.

�Weighting of nodes––to reflect the suitabilityfor use as masters.

� Optimise capacity––the approach directlyconsiders this aspect.

� Optimise throughput––the approach directlyconsiders this aspect.

� Dynamic/maintenance––a repair response tochanges in scatternet membership.

�Max–min optimisation––explicit optimisationof quantifiable multiple objectives.

6. Topology performance measurement

Evaluating the communication performance ofan ad hoc wireless network is a challenging prob-lem in its own right [97]. A single, overall best pro-tocol for scatternet formation and maintenance

does not exist. Firstly, different scenarios requiredifferent performance properties. Secondly, thereare various different ways to assess the perfor-mance characteristics of both the protocol andthe topology it creates and maintains. We considerthe alternative approaches for assessing scatternetperformance.

6.1. Properties of the protocol

Properties of the protocol relate to its ability tocreate and maintain scatternets. Low time com-plexity is particularly important in situations ofen-masse start-up. Low message complexity is par-ticularly important in highly dynamic environ-ments were scatternet membership is frequentlychanging, and generally important for minimisingpower consumption.

• Time complexityTime complexity describes the protocols abilityto scale in terms of time taken to create a fullyconnected scatternet. This is particularly rele-vant for en-masse device start up, where big�O� notation can be used to describe the worse-case effects of doubling the number of deviceson the proportion of time taken. In many simpleprotocols, it is not possible to derive this analyt-ically (e.g., [30]), since network connectivity isnot guaranteed. In these situations, empiricalobservation is necessary. However, someauthors have developed formation protocolsfrom which time complexity can be analyticallyderived. For example, [49] determines networkformation in O(log n), (assuming n nodes). In[76], O(n) time complexity is established. Arange of authors [7,8,14,49,80,83,95] makeempirical measurements on this aspect.

• Message complexityMessaging complexity describes the protocolsability to scale in terms of number of messagesrequired for purposes of creating a fully con-nected scatternet. As messaging for networkformation precludes use of the network forcommunication purposes, it is important tominimise messaging, particularly in dynamicenvironments where only a small time windowexists for exchanging set up messages. In [50],


messaging is considered explicitly, and it isestablished that for n devices, the protocol hasmessaging complexity of O(n). The same resultis established for the protocol introduced in[76]. For other protocols, empirical observationof the total number of messages is not consid-ered, because this is better suited to analyticaltreatment. In preference, time complexity isnormally empically observed.

• Environmental dynamismThe performance of scatternet formation andmaintenance protocols in dynamic environ-ments is important for ad hoc operation. Proto-col efficiency in dynamic scenarios has receivedlittle attention other than in [95], where a rangeof critical cases are considered. These focus onhealing partitions and dealing with en-masseand incremental departures and arrivals.

6.2. Properties of the scatternet

The performance of a scatternet topology canbe evaluated using measures which are either traf-fic dependent or traffic independent. Traffic depen-dent measures require the specification of atraffic profile, defining source and destination ofpackets, and packet flows are considered. Thisneeds specification of routing techniques (see Sec-tion 7) and scheduling techniques (see Section 8).If studying Bluetooth topology in isolation, aswith many scatternet formation protocols, thismakes evaluation difficult. Consequently, trafficindependent performance measures are frequentlyused. We consider the range of traffic independentmeasures used in the literature.

• Average shortest path lengthIn the absence of a pre-defined routing protocoland without definition of traffic flow, the poten-tial for high quality routes can be assessed byshortest path length between source and destina-tion devices. This can be benchmarked againstthe best-possible shortest path between corre-sponding devices in the visibility graph. The dis-parity between shortest path length in thescatternet and that in the visibility graph canbe taken as a performance measure. A range of

authors (e.g., [7,8,14,25,56,60,71,72,95,98,102,104]) use this approach. However, the measurehas to be considered in a wider context, by not-ing that shortest paths may not necessarily leadto high levels of throughput, particularly if mul-tiple paths use common devices which effectivelyform a bottleneck. The authors [14,49,104] con-trast the shortest path approach by consideringmaximum network diameter. Shortest pathassessment is likely to be a realistic measure inthe case of light traffic loads.

• Traffic flow across a bottleneck

A couple of authors [22,60] have directlyaddressed the potential drawback of shortestpath length by considering the performanceof traffic flow at bottlenecks identified in thenetwork. This is potentially useful but in prac-tice, the identification of such bottlenecks islikely to depend on the flow of traffic acrossthe network. Consequently assumptions mustbe introduced concerning routing across thenetwork.

• Average number of slaves per piconetThis is an important metric because the num-ber of slaves per piconet dictates the quantityof piconets required across the network. In aconnected scatternet of n devices, with eachpiconet containing k devices, there must be atleast d(n � 1)/ke piconets (see [51]). As eachpiconet has its own channel, the number ofpiconets also dictates the number of channelsper scatternet. When the number of piconetsis high, the potential for collisions and channeldegradation is increased. However, within apiconet, a potentially higher link capacity isexperienced. Consequently protocols generallyseek to maximise piconet size (up to sevendevices), with the exception of [24], whichmaximises the number of piconets by forminga ring structure in which every piconet con-tains exactly two devices. A range of authors[7,8,22,49,53,71,72,104] monitor the perfor-mance of protocols taking into account theaverage number of devices.

• Average number of roles per deviceA range of authors [7,8,49,53,71,72,102] notethat this factor impacts on the performanceof the network by the switching overhead that


it induces. Consequently a number of authorsconstrain this, while others choose to measurethis factor as a minimisation objective.

• Capacity and throughputScatternet capacity broadly relates to the net-works ability to carry traffic, and this is inter-preted in various ways by different authors. In[24], this is approximated by measuring themaximum number of successful simultaneousdevice-pair communications possible. In [25],the effect of piconet polling and number ofslaves per piconet are used to estimate per-linkcapacity. Traffic dependent metrics based onthe residual capacity are introduced which aredependent on assumptions for routing. Thepaper [84] derives an approximate function forcapacity based on a range of characteristics thatrelate to the network. Throughput describes themaximum flow which can be achieved betweensource and destination devices and has beenadopted for analysing scatternet performancein [56,57,60,73,94,98]. Well known algorithms(such as the Ford–Fulkerson approach asdescribed in [98]) can be readily applied. Thepaper [94] is unusual in deriving approximatefunctions for assessing throughput in the con-text of a scatternet.

• Average path latencyThis metric seeks to approximate the communi-cation latency between device pairs in the scat-ternet. Tan et al. [94] attribute this to threefactors: hop count, intra-piconet schedulingdelay and inter-piconet bridging delay. Thesefactors are dependent on the policy adoptedfor scheduling and routing, which poses prob-lems given the current lack of consensus onthese issues, and lack of realistic scatternet traf-fic patterns. Consequently in [94], a metric isintroduced which is independent of schedulingand traffic, and approximates the average pathlatency between all pairs of source and destina-tions for a given scatternet, by observing thelatency contributions from inter- and intra-piconet scheduling. This approach is alsoadopted in [24]. Alternative to this is direct sim-ulation of delay (e.g., [43]). However thisrequires the assumptions on traffic, schedulingand routing. These are present in some

approaches to scatternet formation, particularlywhen there is a well-defined underlying topol-ogy structure such as a tree or ring. In suchcases (e.g., [57]) fewer approximate metrics needto be imposed.

• InterferenceIt is widely acknowledged that the fast fre-quency hopping scheme adopted by Bluetoothis resilient to interference. Consequently mostauthors choose to neglect interference whenassessing scatternet performance. The onlyauthors to include this factor for neighbouringpiconets are Lin et al. [56]. As 79 frequenciesare used for hopping, each of which is equallylikely, the probability that a time slot suffersinterference is approximated as (78/79)R�1

where R is the number of piconets in transmis-sion range of each other. Similar analysis isapplied in [32].

• Network lifetimeAs mobile devices are likely to be powerrestricted, energy efficiency is important toenhance network lifetime. Few authors directlyaddress this issue. The only paper to considerthis directly for BT is [73], which defines thenetwork lifetime as the time until at leastone device exhausts all battery power. Twopower saving techniques are introduced (oneusing battery power based master/slave switchand the other using distance based powercontrol) to manage traffic flow across the scat-ternet.

• Reliability of the networkIf a protocol can rapidly respond to link anddevice failure, then the protocol will be able toensure reliability, via the provision of potentialroutes across the network. However, the extentto which outage causes critical failures willdepend on the provision of alternative routesin the scatternet. Flat network structures (e.g.,[72]) have been advocated for increased reliabil-ity, and the provision of alternative paths hasbeen considered in [24,56]. Specific networktopologies, or networks with certain character-istics, are frequency sought since propertiessuch as reliability may be guaranteed toa certain extent, such as in a mesh configu-ration.


6.3. Performance simulation

To assess performance in the presence of traf-fic, simulation is required. The IBM extension tothe network-simulator �ns� [69] is termed BlueHoc[15,48], and is open-source, written in C++. Thisemulates the Bluetooth stack and the operationswhich are available in the Bluetooth specifica-tion, and has been adopted in [50,57,94,95,104].BlueHoc is a useful starting point for protocolperformance measurement and has been subse-quently enhanced in [7] for example, by addi-tional mechanisms to handle collisions andenable devices to alternate between inquiry andinquiry scan. In [34], further development is re-ported, to consider node movement. A Javaalternative to BlueHoc is Simjava [85], which isadopted in [68] for purposes of simulating servicediscovery.

7. Routing

To facilitate multi-hop communication, apacket will need to be relayed via a number ofmasters and bridges before it reaches its intendeddestination. A number of authors (e.g., [13,99])point out that routing in the context of Bluetoothdiffers to that for general ad hoc network scenar-ios. Consequently, a different set of design com-promises are required for routing techniques, ascompared to those being developed for generalad hoc applications (see [77] for an overview) bythe Internet Engineering Task Force (IETF) mo-bile ad hoc network (MANET) working group.For example, in [13], it is pointed out that due topacket size limitation at the baseband layer,MANET style solutions will require fragmentationand packets at each relay device. Consequently,there will be an increased buffering requirementat each device leading to a higher delay at eachhop. This means that it is advantageous to supportforwarding across the network at the slot level, asit will reduce the buffering that is required.

It is likely that the Internet protocol (IP) will becommonplace in the context of scatternets, andtherefore it is possible that this layer could be usedfor routing. However, a number of arguments are

proposed in [38] against sole reliance on this layer,for scatternet routing. In particular, a routingfunction on the IP layer would need to be adapted,so that routing function can be integrated into thescatternet formation function. This violates theprinciple of keeping the IP layer independent fromthe link layer. Additionally, the IP routing ap-proach would preclude other non-IP based appli-cations. Bluetooth devices are not compelled tohost an IP layer to be part of an operationalpiconet. It is concluded in [38] that the best wayto manifest routing functionality is at the layer be-neath IP.

The context in which Bluetooth scatternets arelikely to operate, also affects the design require-ments. For personal area network applications,traffic characteristics, mobility patterns and scalingrequirements will also differ from classic ad hocnetworks. It is likely that in many instances, scat-ternets will be quasi-static, short lived and small.Therefore, scalability and adaptivity features maynot always be essential. Instead, protocol simplic-ity, power and bandwidth conservation are partic-ularly important. Additionally, for efficiencypurposes, integration with scatternet formationprocesses is desirable.

As for other scenarios, choices for routing strat-egies fall into the broad categories of pro-activeand re-active approaches. Traditional routing pro-tocols, used for Internet routing for example (andalso in ad hoc networks), are pro-active in thatthey maintain routes to all nodes. This type of ap-proach requires storage for routing tables, andprocedures are required to generate the routing ta-bles and ensure that they are up to date, as the adhoc nature of scatternets means that links will betemporary. In a reactive protocol, routes are deter-mined dynamically. We describe the ways in whichthese approaches have been proposed in the con-text of Bluetooth scatternets. It is notable that verylimited development has occurred in this area.

7.1. A reactive protocol

The protocol proposed in [13] is a source-ori-ented approach, which creates routes only whendesired by the source device. The commonly usedmethod of representing routes in source routing


headers is to include a list of the node identities.This would be a wasteful approach in Bluetoothscatternets. To avoid this, Bhagwat et al. [13] seekto reduce the overhead by introducing an efficientroute representation for uni-cast and multi-casttraffic. This approach is called the routing vectormethod (RVM), and relies on representing piconetsby a local identification number (LocID), selectedlocally by bridges. Within a piconet, a 3-bitMAC address (MacAddr) is used to identify theslaves. The sequence of LocID and MacAddr val-ues is used to identify routes, as opposed to usinga sequence of Bluetooth identities. This leads to asubstantial reduction in the number of bits perhop, induced as an overhead.

When a packet is sent, the sequence of ad-dresses corresponding to the route to be followedis written to the route vector field of the header.A source routing algorithm is employed to deter-mine the path between the source and destinationnodes. Since scatternets are mobile units, a proto-col is proposed where two or more maximally dis-joint routes are determined, and each packet issent via all routes. Discovery of the first route oc-curs using a flooding approach across the entirescatternet, using search packets. Each bridge thatpropagates a search packet appends the piconetidentities from which it has received the packetand also appends the packets next destination.When the final destination first receives a searchpacket, it returns a reply packet along the reversepath. The second route is built similarly to thefirst. Once routes have been defined, a packet issent from the source to the destination via theRVM route. When a relay receives a packet, itsends the packet to the master of the piconet cor-responding to the first LocID, to be forwarded tothe unit whose MAC address is given by the firstMacAddr. Before sending, the first pair (LocID,MacAddr) are removed from the list.

On demand route discovery is recommended bythe authors on the assumption that only a small setof devices will need to communicate at once. Theroute is kept in the packets, not routing tables, tominimise storage overheads. The authors indicatethere will be a limit beyond which such networkscannot be scaled, through they expect this ap-proach to suffice in the majority of applications.

7.2. Hybrid protocols

Hybrid protocols combine the characteristicsof reactive and pro-active protocols. In [23], anapproach is presented, derived from a DynamicSource Routing (DSR) approach [40] designedfor ad hoc networks, which has been adjusted tofit Bluetooth characteristics. When forming aroute, this approach considers the delay that maybe expected at masters due to connections to manyslaves. A Blueroute layer manages a cache at eachBluetooth unit, which may contain routes fromthis unit to other units. When performing routediscovery, the source checks its cache to see if ithas a route to the destination. A process is re-peated where a unit floods the packet to its neigh-bours when it is not the packets destination anddoes not have a cached route to the destination.If the unit has a cache of the destination route, thisis used to determine the route. No detail is givenregarding the size of cache at each Bluetooth node.

In [44], a hybrid reactive-pro-active ZoneRouting Protocol approach that employs a combi-nation of pro-active and reactive schemes is pro-posed. Since Bluetooth nodes are expected tohave low resources, schemes that require storageof large routing tables, may not be viable. Usinga hybrid approach allows Bluetooth masters tostore smaller routing tables that provide routinginformation for nodes within a maximum numberof hops (denoted MAXHOPS). Similarly to [23], ifa node does not have a full path to the destination,reactive path discovery is employed by floodingthe scatternet with a route request. The first nodethat has a path to the requested node, sends aroute reply packet back to the source. This packetfollows the route taken to reach this node and thisinformation along with the nodes path knowledgeis used to define the route.

The parameter MAXHOPS determines the per-formance of the scheme. Simulations show that thepro-active part of the scheme was able to commu-nicate large amounts of routing information at lowoverheads. A high value of MAXHOPS leads tosmall route acquisition latencies, but at the ex-pense of higher routing overhead and higher infor-mation storage costs. It is also shown that innetworks with large numbers of idle nodes, the


pro-active part may cause unnecessary overhead,particularly for large values of MAXHOPS.

7.3. Power conserving routing

Many wireless network scenarios require energyefficient network protocols. A general survey ofthis area is given in [41]. Recently, several Blue-tooth specific routing methods have been proposedthat aim to conserve power.

7.3.1. Power dependent routing

In [73], a reactive routing approach is proposedwhich is aimed at providing energy efficient tech-niques by utilising knowledge of the available bat-tery power of the Bluetooth devices as a costmetric in choosing the routes. It is assumed thatduring scatternet formation, each piconet is givena unique piconet ID (PID). When a device wantsto discover a route to a destination, a flooding ap-proach is employed in which it sends a route re-quest packet to its master. The master appendsits PID and its available battery power to the re-quest packet and forwards it to all associatedbridge points in the piconet. Each bridge appendsits Bluetooth address and adds its available powerlevel to the cost field and forwards the packet to allother piconets it is associated with. This processcontinues until the request packet reaches the des-tination device. The destination may get multiplecopies of the route request packet through differ-ent routing paths each having different costs. Thedestination waits for a specified amount of timeto get multiple copies of the route request packetthrough different paths. It then selects the paththat has the maximum cost field (i.e., maximumcumulative battery power in the path). The desti-nation then sends a route reply packet to thesource device. The route reply packet containsthe selected route vector.

7.3.2. On demand routing

In [58], an on demand approach to the building,routing and scheduling of a Bluetooth scatternet ispresented. The most common approach for scat-ternet formation protocols is to interconnect allBluetooth devices at the initial network startupstage and maintain all Bluetooth links thereafter.

In reality, the physical and data link only needto be created and supported when two Bluetoothdevices wish to communicate. If no communica-tion is required between Bluetooth devices, theycan go into a low power standby state. This powersaving principle has been ignored by most scatter-net formation algorithms, which tend to intercon-nect all Bluetooth devices as a complete scatternetat the initial startup stage, and maintain the fullconnectivity of the network at the data link level.Instead of forming a complete scatternet, theauthors propose a scatternet-route structure thatenables the dynamic establishment of Bluetoothlinks only along the traffic routes. The differencesbetween this approach and the complete scatternetapproach are similar to those between table drivenrouting protocols and on-demand routing proto-cols. The former maintains network-wide routeinformation for fast setup, but incurs substantialsignalling traffic and power consumption. The lat-ter suffers long route setup delays, but is morepower efficient in a mobile ad hoc environment.Where power is a concern, as in a Bluetoothscenario, the on demand approach may befavourable.

As with reactive routing approaches, a routediscovery packet (RDP) is flooded to the wholenetwork to find a route to the destination. Uponreceiving the first RDP packet, the destinationnode sends a route reply packet (RRP) back tothe source along the discovered route. When trans-ferring the RRP, point-to-point Bluetooth linksare created to connect the members of the newroute and at the same time, the routing tables ofthese devices are filled with the new route discov-ery information. When the RRP arrives at thesource device, the scatternet route is ready forthe first data packet to be sent from the source.Analysis and simulation show that whilst this ap-proach has non-negligible setup delay, it achieveshigh link utility, stable network throughput, andfast packet transmission.

7.4. Self-routing structures

One approach to simplifying the potentiallycomplex issue of routing is to create a topologyin which there are unique paths between source


and destination. Such structures necessarily haveto be trees. In [90], this approach has been ex-tended further by structuring the hierarchy in thetree so that messages are self-routing in the sensethat they need carry no routing information, onlysimple rules to guide their way through the topol-ogy, which can be regarded as a search tree. Fur-ther details of this approach are given in Section5.12. Slightly more complex than the tree topologyis the cycle or ring structure, which also leads tovery simple routing. This has been tackled in[24,56], as described in Sections 5.3 and 5.8.

8. Scheduling

Scheduling is required to transfer packetsbetween devices which may reside in the same ordifferent piconets. The Bluetooth specification con-tains only limited information on how schedulingis to be performed. The nature of Bluetooth scat-ternets means that whilst existing scheduling ap-proaches used in wireless networks are applicablefor intra-piconet communications, more consider-ation is required to effectively apply them in topol-ogies involving interconnected piconets.

Within a piconet, scheduling is undertaken bymaster devices, and controls the devices choicefor communication with (and therefore from) aslave in the time division duplex. The master deci-des which slave is the one to next access the chan-nel. A slave is authorised to deliver a packet to themaster only if it has received a polling packet fromthe master. The polling cycle determines the orderin which the slave are polled, and the number ofslots for data transmission. To manage messageforwarding between piconets, consideration mustbe given to the effects of scheduling on bridgedevices, which reside in multiple piconets. Anamicable relationship must be established betweenschedules in overlapped piconets to avoid conflicts:a master should only poll a bridge when it is notengaged in communication in another piconet.Furthermore, inter-piconet scheduling must satisfyinter-piconet traffic demand, so that bottlenecksare minimised. This requires consideration ofhow best to divide a bridges time between multiplepiconets. The queue at a device refers to the num-

ber of packets awaiting transmission. A schedule isexhaustive if the queue at each device is clearedbefore the next device is polled. We consider theadvances which have been made concerning intra-and inter-piconet scheduling for Bluetooth.

8.1. Intra-piconet scheduling

Intra-piconet scheduling only considers thescheduling problem for a single piconet. Three is-sues need to be addressed in the design of anintra-piconet scheduler:

1. slave activity: avoidance of polling slaves whichhave no data to transmit;

2. fairness: give preference to slaves based on theirrelative importance;

3. efficiency: minimise the delay for slaves who arewaiting to transmit data.

This section begins with an examination ofintra-piconet schedulers that build upon an exist-ing scheduling approaches.

8.1.1. Round Robin based scheduling

Round Robin (RR) is one of the simplest andmost widely used scheduling algorithms, mostcommonly used to schedule CPU time. Applyingthis to intra-piconet scheduling means that slavesare polled in a cyclic fashion. A variety of RRbased approaches have been proposed and evalu-ated. These are distinguished by the criteria fordefining the cycle, and the length of time (slots)each master–slave pair is permitted for communi-cation. In [21], Capone et al. consider the problemof designing an efficient and simple polling andscheduling scheme for Bluetooth. Three RR sched-uling schemes are proposed:

• Pure Round Robin (PRR): A fixed cyclic orderis defined and a single chance to transmit is givento each master–slave queue pair according to thecyclic order. This scheme is not exhaustive andeach slave is granted a fixed number of slots.

• Exhaustive Round Robin (ERR):As with PRR afixed order is defined, but the scheme is exhaus-tive and it does not switch to the next slave untilboth the master and the slave queues are empty.


• Exhaustive Pseudo-cyclic Master queue length(EPM): A dynamic cycle order is defined atthe beginning of each cycle (each slave is visitedexactly once per cycle) according to a decreas-ing master to slave queue length order.

The difference in performance between ERRand EPM is shown to be negligible, and this sug-gests that the use of a pseudo-cyclic dynamicscheme based on partial information of the queuestatus is not necessary since the fixed cycle schemegives good performance. The PRR scheme per-forms the poorest, in general. While exhaustiveschemes such as ERR are shown to be effectivein symmetric scenarios, they may be less appropri-ate where traffic demand is asymmetric. An addi-tional drawback of exhaustive schemes is thatslaves with large amounts of data to transmitmay capture the channel, leading to unfair relativedistribution of bandwidth. Whilst it clearly makessense to offer more bandwidth to slaves with highloads, some attempt must be made to do thisfairly. The simplest way is to modify the ERRscheme, limiting the number of slots that can beperformed each pair cycle. This functionality isemployed in the Limited Round Robin (LRR) ap-proach. At low loads, the delays of the LRRscheme are higher than that of ERR. At higherloads, LRR can achieve lower delay than ERRdue to the wastage of slots by the ERR scheme.

A further proposed enhancement to the LRRscheme, defined in [21], extends it to consider theactiveness of slaves. A performance gain is madeby reducing the rate of visit to queues which havebeen empty in the last visits and hence, shouldhave a lower probability of being the longerqueues. This should also reduce the time wastedpolling idle slaves. This leads to extension of theLRR scheme to the Limited Weight Round Robin(LWRR) scheme [21]. LWRR adopts a weightedround robin algorithm with weights dynamicallychanged according to the observed queue status.At the beginning, each slave is assigned a weightequal to maximum priority (MP), which consti-tutes a variable which is set within the scheduler.

The weight is used to calculate how many cyclesa slave must wait before it is next polled, where thenumber of cycles to wait is given by the MP slave

weight. Each time a slave is polled and no data isexchanged between master and slave, the weight ofthe slave is reduced by 1, until a minimum weightof 1 is reached. In this case, the slave has to wait amaximum of MP � 1 cycles before it gets a chanceto send data. When an exchange occurs between aslave and its master, the weight of the slave is in-creased to the MP value. Evaluation of LWRRshows that it always performs better than the otherschemes, and is usually very close to that of theERR.

8.1.2. Extended Round Robin schedulingThe work of Lee et al. [52] extends that in [21].

A new polling scheme is proposed, called Pseudo-Random Cyclic Limited slot-Weighted RoundRobin (PLsWRR). This has two important prop-erties. Firstly, as in LWRR, it attempts to distin-guish between slaves on the basis of their‘‘activeness’’, which constitutes their traffic his-tory. Secondly, the polling order for each cycle isdetermined in a pseudo-random manner. LWRRhas two disadvantages due to the number of slotsin a polling cycle being variable. Firstly, an inac-tive slave needs to wait for a long time to get achance to exchange data packets, if the previouspolling cycles have large numbers of slots. Thiscan lead to high delay for an idle slave. Secondly,an idle slave is polled frequently if the previouspolling cycles have a small number of slots. Thismay reduce the efficiency of the system. ThePLsWRR scheme extends the LWRR scheme byemploying a RR algorithm in which the weightsof slots are changed according to the activenessof the slave in the previous cycle.

Initially, each slave is assigned a slot-weightequal to Max-Slot Priority (MSP). This value isa variable which can be set in the system, takento be 160 slots in the simulation, which equatesto a maximum waiting time of 100 ms. When aslave is polled and no data is exchanged, the slotweight of the slave is reduced by the number ofslots in the previous cycle. As with LWRR, thelowest weight value is 1 and if data is exchangedbetween the slave and the master, the slot-weightof the slave is increased to the MSP. PLsWRRuses the number of slots as opposed to numberof cycles (as in LWRR) to reduce the polling to


less active slaves. PLsWRR guarantees that slaveswait a maximum of MSP slots to get a chance tobe polled. This makes the behaviour more reli-able than LWRR, in which the slave waits for abounded number of cycles, but the length of thesecycles may be variable. As a result, unlike LWRR,PLsWRR works effectively regardless of the lengthof the previous polling cycles, avoiding the twodisadvantages of LWRR.

Evaluations show that PLsWRR outperformsLWRR. It is also shown that the pseudo-randomcyclic order polling in PLsWRR provides fairnessto the flows, with all flows in the piconet havingequal shares of the total capacity.

8.1.3. Efficient double cycle scheduling

The efficient double cycle (EDC) scheduling ap-proach introduced in [18,19] seeks to preserve thefairness of a typical round robin technique but in-crease efficiency by avoiding wasted polling of de-vices with no data to send. Polling only the activedevices leads to a polling sub-cycle, which was ini-tially addressed in [37]. The EDC approach ex-tends the application of a polling sub-cycle to aseparation of up-link and down-link scheduling.In the downlink direction, the master has knowl-edge of packet queues for slaves. In the otherdirection, it can be assumed that the master hasa probabilistic knowledge based on the feedbackthe master gets when polling the slaves. Thismeans that the master can only estimate the prob-ability of an inactive slave by exploiting knowledgeof slave transmissions in previous cycles.

A partial decoupling in the scheduling of down-link and uplink transmissions is achieved by defin-ing a double polling cycle, which consists of anuplink polling sub-cycle (cycleup) and a downlinkpolling cycle (cycledw). During both these cyclesthe master updates a subset of slaves that are eligi-ble for polling. For the downlink cycle, these aredefined as E(DW), and E(UP) denotes the mem-bers eligible for the uplink cycle. Different rulesare applied for the selection of slaves in E(UP)and E(DW). These rules unsure that there is fair-ness separation for the different directions: speci-fically cycleup ensures fairness in the uplinkdirection, and similarly for cycledw in the downlinkdirection. The sets E(UP) and E(DW) are simulta-

neously updated at the beginning of each cycle,with knowledge of traffic loads gained from theprevious cycle. The set E(DW) is determined bythe master, using knowledge of queue lengths.For the uplink polling cycle, if a slave has sent apacket with a null payload in the previous cycle,then its polling window is doubled (until a maxi-mum value is reached). Otherwise a polling win-dow of 1 is used, and therefore the device ispolled in the next cycle. Simulation results pre-sented in [18,19] show that EDC significantly im-proves the throughput of TCP connections whencompared to a round robin scheduling approach.

8.1.4. Adapting to flow and utilising multiple slots

In addition to reducing the polling frequency toslaves with no data to transmit, improvements canbe gained from reducing the polling frequency forslaves with low volumes of data to transmit. Daset al. [26] describe three algorithms that each ad-just the polling rate of slaves, based on the amountof data they have to transmit.

In the first approach, called Adaptive Flow basedPolling (AFP), the polling rate is increased forslaves that have more than a threshold value (buf-thresh) of data in the buffer, and decrease the poll-ing rate for those slaves that have no data to send.This is controlled by updating the polling intervalfor each device. Initially this is a uniform (default)value, but subsequently changes according to thenumber of packets in the slaves buffer. If thereare more than buf-thresh packets in the buffer, thepacket is sent and the polling interval is set to thedefault value. In this case, there is a high flow ratefor this slave, and hence its polling interval is re-duced so that it can be served more frequently. Ifthere are less than buf-thresh packets in the buffer,the packet is transmitted and the polling interval isunchanged. If a poll packet is transmitted and anull packet is received, the current polling intervalat the device is doubled (up to a defined limit).

In the second algorithm, the Sticky algorithm,the rate of polling is not affected, but slaves withhigh levels of data to transmit are allowed to trans-mit multiple packets. Each slave is serviced in acyclic fashion and if the slave has more than buf-thresh packets in its buffer, a maximum level ofnum-sticky packets are transmitted. If the number


of packets in the buffer is less than buf-thresh, asingle packet is transmitted as in RR scheduling.

The final algorithm, Sticky Adaptive Flow basedPolling (Sticky AFP) combines aspects of AFP andSticky, such that when the slave has more thanbuf-thresh packets to send, a maximum of num-sticky packets are transmitted.

A simulation is used in [26] to assess the theTCP throughput for different values of num-sticky,and a comparison is performed between the threealgorithms. Both AFP and Sticky algorithms givesignificantly better performance than RR. TheSticky algorithm is found to have the lowest end-to-end delay, whilst Sticky AFP has the highest.Sticky reduces queue occupancy by transmittingmultiple packets from queues with high backlog,thereby preventing queue overflow and reducingend-to-end delay. On the other hand, Sticky AFPcauses a marked increase in end-to-end delay ofintermittent constant bit rate traffic, because flowis set infrequently for such bursty sources. It isconcluded in [26] that either AFP or Sticky (witha high value of num-sticky) result in the best over-all performance.

8.1.5. Predictive fair polling

The predictive fair polling (PFP) approach isbased on consideration of best effort traffic [100].This means that the slaves get the same fractionof their fair share resources, where a fair share isequal to the share that the slaves would have beengiven when a generalised processor sharing systemwas used. This is opposed to the case where fair-ness determines that slaves get the fraction of theirnegotiated quality of service requirements.

The scheduling operates without informationabout the offered load. This means that neitherpacket size nor arrival times are known for sched-uling purposes. Consequently a range of estima-tors are used, the output from which is fed into adecision making process. The estimators usedapproximate traffic demand and data availability.The traffic demand estimator calculates a movingaverage of inter-arrival times of packets. The dataavailability predictor calculates the probability Piof each slave i having a baseband packet waitingfor transmission to the master. This leads to a fairshare measure fsi for each slave i, defined as

fsi ¼P iPnk¼1 Pk

;

where n is the number of slaves. The fair sharemeasure is compared with the number of pollssince the last poll to slave i (denoted si). For eachslave i, this is called the fraction of fair share de-noted Ffsi. If si P fsi, this is defined as 1.Otherwise

Ffsi ¼sifsi

:

The decision on whom to poll next is determinedusing fraction of fair share and Pi. Clearly, thepolling of a slave i with Pi = 1 and Ffsi = 0 is ur-gent, while polling a slave with Pi = 0 and Ffsi = 1isn�t urgent. In between these extremes, the pollingdecision is made using a measure of urgency, de-noted Ui, for each slave i. This is defined as

Ui ¼ aP i þ ð1 � aÞð1 � FfsiÞ; 0 6 a 6 1:

Tuning the setting of parameter a affects the rela-tive impact of the fraction of share compared toPi, in the measure of urgency. The decision onwhom to poll next is determined by the slave withthe highest Ui value. The authors find that tuningthe value of a to 0.1 leads to the low response times(sum of waiting times and service times of packets)when such packets are generated under Poissonprocesses. The authors also propose extensionsto deal with quality of service traffic handlingand duplex traffic handling.

8.1.6. Prioritising links where master and slave

have data to transmit

In [42], the proposed approach to polling ismotivated by three issues: (1) state of the queuesat the master and slaves; (2) the traffic arrival pro-cess at these queues; and (3) the packet length dis-tribution at the master and slave. Two schedulingpolicies are introduced, Priority and K Fairness,which are master–slave queue state dependent. Inthese approaches, master–slave pairs are distin-guished based on the state of the queues at themaster and slaves, the master having a separatequeue for each slave. If a node (master or slave)has data to send, the node is in mode 1, otherwisemode is 0. This leads to four distinct states for amaster–slave pair. It is assumed that binary infor-


mation regarding the status of the queue at a slaveis known at the master, with transfer using free bitsin the payload packet header.

In the Priority Policy (PP), higher priority isgiven to master–slave connections in the 1–1 stateover master–slave pairs in the 0–1 or 0–1 states(which have equal priority), with exclusion of the0–0 state. The PP policy achieves higher through-put than a pure round robin policy since connec-tions in the 1–1 state are given a proportionallyhigher number of time slots.

The K Fairness policy imposes a fairness bound,where waiting is limited by the factor K slots.Round robin scheduling is performed among allmaster–slave connection pairs that are in 1–1, 0–1 or 1–0 states. When the scheduler is at a 0–1 or1–0 pair, its service is sacrificed to a 1–1 connec-tion provided that the difference between theservice provided by any two back-logged connec-tions does not exceed K slots. Counters are main-tained to track the excess or deficit servicereceived for the queue across each master–slavepair. The master–slave pair that has received themaximum excess service (service sacrificed to it)from other pairs is denoted qmax and the master–slave pair that has sacrificed the maximum serviceto other master–slave pairs is denoted qmin. Bothqmin and qmax are defined with respect to each ofthe back-logged queue pairs (i.e., 1–0, 0–1, or 1–1). When scheduling a 0–1 or 1–0 type connection,if the difference received by qmax and qmin exceedsK, no further service is sacrificed by the pair beingscheduled. This ensures that the max–min fairnessbound [46] is maintained. This policy achieveshigher throughput than RR and also maintainsfairness. KFP gives better throughput than PPwith more fairness. In PP the unfairness rises qua-dratically with an increase in p. In KFP, theunfairness is bounded and increases linearly. Thefairness bound is potentially useful in defining aQoS guarantee for KFP.

8.1.7. Queue management based on capacity and

multi-slot framing

The performance of a scheduling approach isreflected in the profile of queues, since this is indic-ative of speed and flow. It is essential to control thelength of the queue and also the amount of time

any packet remains delayed in the queue. Ideallythe scheduler should smooth traffic fluctuationand avoid long delays. The use of multi-slot fram-ing can be advantageous for a number of reasons.Firstly, multi-slot framing benefits effective pay-load transmission as a 1-slot frame has associatedoverheads besides the payload. The subsequentslots in a multi-slot frame can avoid overheadssuch as the frame header. Secondly, differentspeeds can be used for different link directions.

Luo et al. [59] present a scheduling algorithmwhich seeks to address these issues. They showthat in a scenario with exact knowledge of queuestatus for each slave at the master, high perfor-mance can be achieved. However, this is not prac-tical in an applied scenario, due to the overheadsrequired to obtain this information at the master.Consequently, the authors introduce the notionof predictive capacity assignment as a replacementto the real time queue analysis. The result is Adap-tive Capacity Ratio with Intelligent Frame Schedul-ing (ACRIFS). In ACRIFS, the master is assumedto have a queue for each slave. The capacitythreshold of these queues is calculated periodicallybased on equations that consider the traffic and thepropagation delay of the link. The scheduler usesthe maximum queue length at the master to selectthe queue for transmission, and in addition, seesif the transmitted slot is within the capacity thresh-old. A predictive capacity ratio is used to encour-age the master to determine multi-slot framingwithout requiring knowledge about queue statusfrom the slaves, making ACRIFS a practical solu-tion. Simulation and analysis of this approachshows that it significantly outperforms the bench-mark RR approach, when considering packetdelay.

8.1.8. Packet-type priority approach

The scheduling approach in [82] is unique inconsidering a mix of voice and data packet types,each of which may have different levels of priority.A mixed data and voice traffic profile is simulated,modelled by an interrupted Bernoulli process. Per-formance is considered based on mean packetdelay and probability of packet loss for data traf-fic. A Bluetooth frame is used in which eight SCOslots are reserved for voice traffic and 16 ACL slots


are reserved for data traffic. For the ACL slots,scheduling of packets of two possible different datatypes is considered (denoted D1 and D2).

Three schemes are considered. New schemescalled priority (PR), alternating priority (AP) arecompared against a round robin scheme. In thePR scheme, voice has the highest priority, D1has the second priority and D2 has the lowest pri-ority. In the AP scheme, slots are reserved forboth data traffic components in a round robinmanner, but the unused slots can be used byother data traffic. The round robin approach uses3-slot cycles.

The effectiveness of these scheduling methodsare studied by varying the traffic load and bursti-ness of the data streams and varying queue capac-ity. The results show, that with an infinite queue,the length of the queue grows dramatically whenoffered load and burstiness are high. In all threescheduling schemes, queue capacity has an essentialimpact on loss and mean delay for all studied trafficcases. Empirical results show that PR and AP haveapproximately equal probability of packet loss, sig-nificantly lower than that of the RR approach. TheAP scheme also has the advantage of fairness andefficiency in comparison to the other schemes.

8.1.9. Bin packing

Yang et al. [101] apply the concept of bin pack-ing to the scheduling problem. The aim of the gen-eral bin packing problem is to pack items ofvarious sizes into a set of bins of fixed capacitysuch that the number of bins required is mini-mised. In off-line bin packing, the size of all itemsis known and (near) optimum solutions can befound through the use of exhaustive or meta heu-ristic approaches. In online bin packing, the size ofitems is not known and they must be packed asthey arrive (items already packed cannot be re-moved or rearranged). Scheduling in Bluetoothcan be viewed as an online bin packing problemwith a limited amount of look ahead, since thescheduler can potentially determine the size ofqueued packets at the head of the line (HOL) foreach slave, and decide which to pack (i.e., includein the current frame).

Two scheduling strategies are considered––Look Ahead (LA) and Look Ahead Limited Round

Robin (LARR). In LA, a Next Fit (NF) approachis used in which bins are filled in sequence (analo-gous to filling frames in sequence). The NFDextension to NF involves arranging HOL packetsin non-increasing order of size before they arepacked in frames. Problems with this approachare two-fold. Firstly, every time a packet is sched-uled, the new largest HOL packet must be deter-mined. Secondly, since the algorithm alwaysattempts to schedule the largest packet, it may besome time before packets from some slaves arescheduled (starvation).

Look Ahead Round Robin attempts to over-come these problems. It combines the simplicityof RR with the LA approach to avoid starvationand reduce the computational complexity by nolonger attempting to find the largest HOL packet.Instead, slaves are serviced in a RR fashion, and ifthe packet fits the frame, it is scheduled. This ap-proach differs from round robin because whenthe packet does not fit, instead of waiting for thenext frame, the algorithm looks ahead and at-tempts to schedule a packet from the next slavein the RR sequence. Analysis and simulation showthat LA and LARR perform significantly betterthan RR. Since LARR is less complex and avoidsstarvation whilst maintaining a similar level of per-formance, the authors advocate this scheduling ap-proach for Bluetooth.

8.1.10. Interference aware scheduling

The only approach to scheduling which tries tominimise the effects of interference is that in [28].In this paper, an IEEE 802.11 system operatingin direct spread spectrum mode is considered asthe source of interference, though the approachmay be adapted to any source of interference.The approach exploits the fact that devices in thesame piconet will not be subject to the same levelof interference on all channels. The approach dis-tributes channels to devices to maximise through-put whilst maintaining fairness of access. Thealgorithm, called Bluetooth Interference Aware

Scheduling (BIAS), has three components, namely,a channel estimation procedure, a procedure thatweights devices based on channel access priority,and a credit function that controls fair access tobandwidth.


The channel estimation procedure is used to de-tect the presence of interference in the frequencyband. Each slave maintains a table that indicatesif a frequency is used or clear, depending onwhether the bit error rate exceeds a threshold.Since the master controls all transmissions, theslaves send their frequency usage table to the mas-ter which can then make use of this data in thechannel estimation phase to optimise thefrequency allocation on each time slot, to avoidpacket transmission on a channel with high inter-ference. A credit system is used to control thebandwidth allocated to each device, in order to en-sure that no device gets more than its fair share ofthe available bandwidth. Priority is given to de-vices with few good channels over those withhigher channel availability. Simulation resultsshow that BIAS can reduce the probability ofpacket loss where interference exists, but with aslight increase in the mean access delay for theworst-case scenario.

8.2. Inter-piconet scheduling

Inter-piconet scheduling is required to co-ordi-nate the activities of bridge devices which residein multiple piconets. Such bridge devices mustswitch between piconets to enable inter-piconetcommunication. The switching process inducesan overhead due to guard time: each switch maycost up to two slots, which are not available forcommunication. This occurs as a consequence ofthe clock, the speed of which may differ slightly be-tween masters. Bluetooth permits a clock drift of20 parts per million against the ideal timing duringactivity. In low power modes this is extended to250 parts per million.

Within a piconet, the master always expects aslave to be available. This means that conflictscan arise, in which a master polls a bridge whenthe bridge is engaged in another piconet, causingslot wastage. Therefore the scheduling algorithmmust ensure that slaves are available to a masterwhen it wishes to communicate with them. A sim-ple solution to this problem uses slot reservation,in which specific slots used in the piconet are re-served for particular master–slave pairs. Two mas-ters sharing a common slave cannot reserve the

same slot to poll the same bridge. These slots arecalled rendezvous points. As with intra-piconetscheduling, inter-piconet scheduling must beresponsive to the network capacity considerations.In particular, it is essential that bridge device areable to focus capacity on the piconet with thegreatest load at any moment in time.

All approaches to inter-piconet scheduling arerequired to tackle common issues. The rendezvouspoints must be determined, and the rendezvouswindow must be controlled. Also, this needs tobe integrated with intra-piconet scheduling, thedynamic nature of traffic, quality of servicerequirements, and interference considerations (ifapplied). There are many different ways ofapproaching these issues, from the creation ofdeterministic processes based on simple heuristics,to controlled, randomised processes.

The Bluetooth sniff state is commonly employedto determine rendezvous points. This is a powersaving mode that allows the slave device to reducethe duty cycle. Instead of listening to the master inevery downlink slot, the sniff mode allows a slaveto listen to a master in a reduced number of timeslots. Thus the sniff mode relieves a slave from itsfull duty cycle, permitting the bridge to switch be-tween piconets. Additionally, hold mode ap-proaches have been proposed, along with a callfor a new Bluetooth mode called jump. We beginby looking at approaches which are non-loadadaptive. Then we consider how these have beenextended to adaptive scheduling which supportsQoS based approaches. Finally we conclude withan overview of an interference aware approach.

8.2.1. Non-load-adaptive scheduling

Johansson et al. [39] propose a periodic rendez-vous approach, called theMaximum Distance Ren-

dezvous Point (MDRP) algorithm. The basis of theMDRP algorithm is that redenzvous points (RPs)are as far away from each other as possible. TheMDRP algorithm uses a periodic super frame, thatis common to all nodes in the scatternet, and de-notes the time period between two RPs for anymaster–bridge pair. For a pair of masters con-nected via a bridge, the distance between the RPsindicates the amount of time the bridge spends ineach piconet, referred to as the rendezvous window.


The MDRP algorithm attempts to distributethe time spent in each piconet by a bridge equally,by maximising the distance between each RP of abridge. A master node assigns RPs to bridges asfollows. Suppose the bridge belongs to i other pico-nets, and has RPs r1, r2, . . . , ri with these piconets.The master node has j other bridges to which it hasassigned RPs ri+1, ri+2, . . . , ri+j. These RPs defineslot numbers in the super-frame and are in everysuper frame period. Based on the list of i + j RPsalready defined, new RPs are defined at the middleslot in the largest interval between successive exist-ing RPs. This is a simple mechanism for allocationof time between piconets. However, the authorspoint out that such simplicity and robustnesscomes at a price, since the algorithm is not loadadaptive to temporary changes in network traffic:simply the RV allocation is dependent on the num-ber of piconets a bridge is associated with. Otherauthors have tried to define adaptive methodscapable of changing with the dynamic nature ofthe traffic. We consider these in detail.

8.2.2. Adaptive scheduling using a RV max–min

optimisation approach

In [45], Kazantzidis et al. seek to extend the ren-dezvous point approach introduced in [39] by notonly selecting non-conflicting rendezvous points,but seeking optimal choices of RV point. This isaddressed in the context of mobility, where newpiconets join existing piconets, at which pointRV point selection needs to be considered. Theaim is to select RV points so that the minimumforwarding throughput between piconets is closeto optimal. The underlying rationale is to avoidestablishing RV points that take away forwardingthroughput from other piconet pairs. In this way,bottleneck links are avoided and overall forward-ing throughput across the scatternet is maintained.

Note that the establishment of a new RV pointfor a visiting piconet not only affects the through-put between the home-visiting piconet pair, but italso affects the throughput of all connections ofthe visiting piconet. The proposed approach in-volves optimising forward throughput equationsat bridge devices. This has a potentially expensiveoverhead, involving the optimisation of localforwarding throughput equations. However the

overall aim in this instance is determination ofhigh quality solutions.

8.2.3. Adapting to wasted polls and traffic load

Zhang et al. [103] propose a Flexible Scatternet-wide Scheduling (FSS) scheme that can adaptscheduling based on wasted polls and traffic load.The scheme is based on a switch table concept,which is formed when the scatternet is created.This table directs bridge devices to switch betweenthe multiple piconets to which it belongs. Conflictsare avoided by the master polling slaves only atslots when the bridge device is synchronised withthe master.

The switch table is maintained and updated, toadjust for load and improve system performance.When the scatternet is initiated, a switch table isgenerated for each bridge, where master nodesare considered in round robin order and a switchoccurs every frame. The FSS scheme acts uponthe scheme using two algorithms: a flexible(intra-piconet) scheduling algorithm and a switchtable modification algorithm.

In the scheduling algorithm, the master pollsthe slaves in a weighted round robin manner. Ini-tially, the weight of each slave is computed basedon the estimated traffic load in each master–slavelink. This means that in a cycle, only a subset ofits slaves may be polled. In order to decide the fre-quency of polling of slaves for a master, each slavehas a polling weight tuple (P,R), where P indicatesthat the slave should be polled every P cycles, andR represents the maximum frequency that a slavecan be polled in a cycle. A master can dynamicallychange the polling weights based on estimation oftraffic load, based on previously wasted polls. If apoll is wasted (both slots allocated for polling arenot used), the offending slave has P increased (upto a certain threshold), otherwise P is decreaseduntil it reaches 1. For slaves where P = 1, if a pollis wasted, R is decreased (until it reaches 1) and Pis increased. If a poll is not wasted, R is increaseduntil it reaches an upper bound.

While the scheduling algorithm is able to flexi-bly schedule slots used to poll slaves and bridgenodes, it cannot change the quantity or positionsof slots that a bridge node can use. To address thisproblem, the switch table modification algorithm


permits a bridge device to dynamically update itsswitch table, based on the traffic load. This allowsa device to change its switching pattern betweenpiconets. The switch table update can be initiatedby the master or slave. For the slave, the bridgemonitors the outgoing queue length and incomingqueue length for each of its master–slave links. Themaster with the longer queue length should getmore time slots compared to a master with a rela-tively low queue length. Based on this idea, themaster with the larger queue length borrows slotsfrom the master with the relatively low queuelength. The authors define detailed data structuresand processes by which the borrowing process isimplemented.

8.2.4. Adaptive scheduling using a credit scheme

and adaptive presence point density

Baatz et al. [3,4] present an adaptive schedulingwhich approach utilises the concept presence

points. These are defined points where communica-tion between a master and slave may commence.Presence points are used as an alternative to defin-ing local schedules, and enable each device toquickly determine whether its peer is in the samepiconet. If so, communication may begin betweenthe devices, otherwise another presence point maybe tried, without loosing a significant amount ofbandwidth. Additionally, the length of a particularcommunication period is not pre-determined, butdepends on current link utilisation and the amountof data ready to exchange. As the communicationschedule is determined online rather than a priori,simple calculations are required to determine howlong to stay in specific piconets. Additionally, pres-ence points must be relatively dense so that thereare still some available, even in the presence ofinterference.

The authors integrate presence points with slotsused by the sniff mode of operation. Sniff slots areregarded as possible presence points at which peersmay start communicating. In order to decide whento abort an ongoing sniff event in order to use anupcoming sniff slot, a priority scheme is intro-duced, where each device associates a particularpriority for each of its links. These priorities aredevice centric, and different devices may give dif-ferent priorities to the same link. A link has higher

priority relative to another link if it has previouslybeen untreated unfairly, relative to other links sup-ported by the device. This is quantified by trackingthe number of slots used by each device.

A potential problem that the authors addresswith this approach is the frequent switching, andtherefore bandwidth wastage, that it might induce.Consequently in [4], the density of sniff slots is re-duced exponentially over time, by deliberate skip-ping of presence points. This scheme is called theAdaptive Presence Point Density (APPD) scheme,where each device manages an internal parameter,which determines the time between use of sniffslots. Each device may increase this parameter if,for example, sniff slots are not taken or no datatransmission occurs.

8.2.5. Inter- and intra-piconet scheduling using

pseudo-random sequences of RV points

Racz et al. [75] propose a lightweight approachto scheduling known as the Pseudo-Random Coor-dinated Scatternet Scheduling (PCSS) algorithm.The approach involves bridge devices assigningmeeting points with their peers such that the se-quence of meeting points follows a pseudo-ran-dom sequence which is different for each pair ofdevices. Consequently uniqueness of the pseudo-random sequence guarantees that meeting pointswith different peers of the same node will collideonly occasionally. The key advantage is that thisremoves the need for explicit information ex-change between peer devices, which significantlyreduces complexity.

The algorithm uses checkpoints, equivalent toRV points. Checkpoints are assigned for a pairof devices, based on the Bluetooth clock of themaster and the MAC address of the slave. Thisscheme guarantees that the checkpoint sequencegenerated by the master and the slave is the same(without need for communicating the sequence),while also ensuring the sequences belonging to dif-ferent node pairs will be different. The probabilityof a collision, where a device can attend only oneof the checkpoints, is dependent on the fre-quency of checkpoints. This is chosen dependenton the free capacities of the node or on the amountof data to transmit. Dynamic adjustment is per-mitted, with an increase or decrease the intensity


of checkpoints depending on the amount of userdata to be transmitted and on the capacity of thedevice. In [75], PCSS was verified in having higherthroughput than the equivalent approach withoutthe protection using pseudo-random sequencing.

8.2.6. Dynamically scheduled meetings

The advantage of PCSS as proposed in [75] isthat it achieves scheduling with little overhead.However, because it is based on a randomisedscheme, as the density of nodes grows, schedulingconflicts will arise where one node is actively wait-ing for another node that is busy communicationwith some other node. Tan et al. [93] present a Lo-cally Coordinated Scheduling (LCS) approach thatis able to co-ordinate nodes in a manner that elim-inates all scheduling conflicts, but with associatedadditional overhead.

The LCS approach is designed to dynamicallyadjust the schedule based on workload conditions.It is based on the concept of scheduled meetings, atwhich devices meet to exchange data. At the end ofeach meeting, the nodes negotiate the start timeand minimum duration of the next meeting. Ateach occurrence, a parent device sends a list of pos-sible future meeting start and finish times to a childdevice, and the child device replies with desiredstart and finished times of one future meeting,which fall within one of the meeting periods sug-gested by the parent.

The role of LCS is to monitor traffic character-istics associated with each link and arrive at an effi-cient scatternet-wide schedule based on them.Computation of start time is based on whetherthe data rate is increasing, decreasing or stable.This enables LCS to respond to varying trafficconditions quickly, without wasting resources.The duration of the next meeting is based on queuesize, and the past history of transmissions, in orderto set a duration that is just large enough to ex-change all back-logged data. Similarly, the meetingrecess interval is adjusted according to the datarate and the nature of traffic flows.

LCS is comprehensive in the sense that it utilisesa range of further techniques to achieve efficientscheduling. For example, meetings with similarcharacteristics are grouped to reduce bandwidthwastage and end-to-end latency. Additionally,

meetings at various parts of the scatternet arescheduled in a hierarchical fashion to exploit theuse of parallel communications, to improvethroughput. LCS also tries to tolerate disruptionin connectivity, by providing a fall back communi-cation mechanism where nodes are not able tocommunicate during agreed meeting periods.

8.2.7. QoS

Son et al. [86,87] develop two inter-piconetscheduling approaches, based on satisfying mea-sures for quality of service (QoS). In [86], theaim of inter-piconet scheduling is to provide justenough capacity for quality of service, via thebridge devices time sharing activities. Quality ofservice in this context is defined generally, andcould be included as measurements on buffer size,packet delays, and numbers of packets queueing atdevices in the network. Such traffic information isperiodically sent via the link manager to masterswhich take part in inter-piconet communication.The offered traffic is estimated and predicted froma record of the collected information. This is com-pared with the allocated capacity, and the inter-piconet scheduler decides whether the QoS canbe satisfied. If not, the QoS is changed accord-ingly. This process is performed at regular inter-vals. The QoS calculation is quite complex andconsequently look-up tables are proposed for realtime application.

In the second approach [87], resource allocationfor inter-piconet scheduling is formulated as a con-vex optimisation problem, with multiple objectivesthat characterise the maximisation of total net-work flows and the minimisation of the total costof flows. A distributed iterative capacity allocationscheme is proposed to perform the optimisation.This seeks to divide network capacity to links insuch a way that maximises the network flowsand satisfies the QoS requirements (such as timedelays, queueing sizes, etc.) as formulated in thegoverning convex optimisation equations. Onceallocation of capacity has occurred at a device,allocations of link capacities are iterated alongeach link for routes across the network. Capacityto support the longest route paths is allocated asa priority. The authors indicate that there are someopen issues that need to be addressed, such as the


amount of messaging required by this approach,and issues that arise due to mobility of nodes inscatternets.

8.2.8. Load adaptive and hold mode inter-piconetscheduling for small scatternets

Har-Shai et al. [33] restrict their focus to smallscatternets using a load adaptive algorithm

(LAA). Under LAA, the algorithm determinesthe duration of the bridge activity in the differentpiconets such that the delay incurred by packetsrequiring inter-piconet connection is reduced.The algorithm adapts based on a range of factorsrelated to the network performance, includingdevice activity and queue size. Progression abovedefined thresholds for any of the factors promptsa switch between piconets (it is assumed that onlya pair of piconets are operating). The hold modeutilized is at the bridge to enable one Bluetoothdevice to leave a piconet. The main differencebetween commonly used sniff and hold modes isthat the duration of the hold period must be setevery time the slave is placed into hold mode,whereas the parameters of sniff mode are set onceand can be used repeatedly. Consequently theparameter setting for the hold mode determinesthe performance of the algorithm.

8.2.9. A new Bluetooth mode––jump modeJonasson et al. [36] propose an additional Blue-

tooth mode called jump. This mode includes a setof communication rules that enable efficient scat-ternet operation by offering flexibility for a deviceto adapt its activity in different piconets, under dif-ferent traffic conditions. The jump mode operatesin a distributed manner, allowing each device tojump in and out of the jump mode link. By default,a jumping node is absent from a link. When ajumping node wants to be present in a link, ithas to signal its presence to the peer node of thelink.

The basic premise is that each jumping nodedivides its time into time windows of pseudo-random length, denoted rendezvous (RV) windows,spending one or more RV windows in a piconetbefore jumping to another piconet, and signalsits decision to all concerned nodes. When a jump-ing node is present on a jump mode link during an

RV window, capacity is distributed by the intra-piconet scheduling mechanism. A jumping masterpolls its slaves as usual, and jumping slaves thathave signalled their presence are incorporated intothe masters polling scheme.

The essence of this scheme is that a jumpingnode uses the same sequence of RV windows inall of its piconets. Employing such sequences ofRV windows in conjunction with the simple signal-ling protocol makes it possible for all the jumpingnodes peers to know whether it will be present ontheir link during each RV window. One advantageis that under this approach, the dependencies be-tween the inter- and intra-piconet schedulers arebe kept to a minimum, enabling more or less inde-pendent evolution in both fields. The paper de-scribes only the mechanisms of the jump mode,and consequently algorithms can be further devel-oped to implement the mode in scheduling.

8.2.10. Interference aware schedulingSun et al. [91] point out that the in situa-

tions where inter-piconet scheduling is required,piconets are within range of each other, and there-fore interference considerations are important.Although the effects of interference are minimisedby fast frequency hopping, it cannot be completelyavoided. The simulation reported by Sun et al. [91]shows that interference can degrade system perfor-mance by 30%.

However, if the master of a piconet knows thehopping sequence of an adjacent piconet, it canavoid interference by not sending packets whenboth piconets hop to the same frequency. Anybridge has the information on the hopping se-quence of each master, and can distribute this be-tween adjacent piconets. Interference can then beavoided by careful scheduling as described in[91]. Assuming that a slaves reply is a single slot,when a master intends to send an n-slot packet,it can first check if any of the adjacent piconets willuse the frequency in the next n + 1 slots. If someadjacent piconets will use the frequency, the mas-ter checks to see if its device address is greater thanthose of masters in the other adjoining piconets. Ifit is (or no other adjacent piconets hop to the fre-quency) then the master can safely transmit.Otherwise, the master suppresses transmission,


waits for the next time slot and starts the wholeprocess again.

This scheme avoids interference between adja-cent piconets, and can potentially be implementedwith a range of scheduling approaches. A smallcommunication overhead is required when thegateway node sends the masters information tothe adjacent piconets masters. Additionally, somecomputational overhead is required where themaster pre-calculates the hopping sequence of theadjacent piconets along with its own. Also, whentraffic load is not heavy, this scheme may reducebandwidth utilisation since the master may sup-press its transmission, although other piconets onthe same frequency, do not transmit. However,when traffic load is medium to high, this schemeshould help alleviate the interference, thus increas-ing bandwidth utilisation. This scheme does notconsider interference from piconets that are phys-ically close, but topologically far away (i.e., thosethat are physically close but do not share a bridgewith the piconet).

8.2.11. Queue theoretic analysis

The performance of a range of bridging strate-gies have been assessed analytically in [63–67].Queueing theory has been applied to a range ofdifference scenarios, each scenario focusing on adifferent topology configuration at the point ofbridging. In [63,67], the scenario where bridgeshave a master and slave role has been considered.In [65,67], the scenario where bridges have onlyroles as a slave are considered. These contributionsfocus on modelling access delay, and consider theassociated probability distribution of end-to-enddelay times for both local (intra-piconet) andnon-local traffic. Sensitivity of various parameters,on piconet performance, is also considered. Anal-ysis suggests that the main criteria in minimisingend-to-end packet delay should be in minimisingthe end-to-end delay for inter-piconet rather thanintra-piconet traffic. This is suggested as a practi-cal alternative to minimising a weighted averageof local and non-local delays. Analysis also indi-cates that the optimum value for the time intervalbetween bridge exchanges is dependent on currentvalues of traffic parameters, such as burst arrivalrate and mean burst size.

9. Conclusions

Bluetooth is an interesting development in per-vasive and ubiquitous communication becauseit represents the first mass market, low costtechnology, with opportunities for high levels ofpenetration. Developing protocols for scatternetformation and maintenance is important, as thisopens up increased possibilities for flexible net-working and new applications. The state-of-the-art in this area has developed rapidly, and anumber of significant contributions have alreadybeen made. However, a number of challengingkey issues remain, and the following observationscan be made.

Firstly, simplifying assumptions, particularly inscatternet formation (e.g., en-masse device start-up) limit the applicability of numerous proposedprotocols. Further development is this area wouldbe beneficial. Secondly, it is largely the case that is-sues of network formation, scheduling and routinghave been studied in isolation. Given the depen-dencies between these issues, further integrationof the issues may lead to higher performance solu-tions. Thirdly, there has been limited systematicbenchmarking applied to the proposed techniques.This makes it difficult to draw conclusions on therelative effectiveness of alternative approaches. Fi-nally, the operation of the technology undermobility has received little attention. The extentto which the operation of interconnected piconetscan be sustained under mobility, is importantgiven that potential applications may occur forgeographically dynamic devices. Consideration oflimits of device mobility on network operation,including the limits of mobility speed that couldbe sustained from the networking (not just trans-mission) point of view, is an interesting open ques-tion. The ongoing development of methodologiesfor scatternet formation is being documented atwww.scatternet.org.

References

[1] I.F. Akyildiz, W. Su, Y. Sankarasubramaniam, E.Cayirci, Wireless sensor networks: a survey, ComputerNetworks 38 (2002) 393–422.

http://www.scatternet.org


[2] S. Baatz, C. Bieschke, M. Frank, P. Martini, C. Scholz, C.Khul, Building efficient bluetooth scatternet topologiesfrom 1-factors, in: Proceedings of the IASTED Interna-tional Conference on Wireless and Optical Communica-tions, WOC 2002, Banff, Alberta, Canada.

[3] S. Baatz, M. Frank, C Kuhl, P. Martini, C. Scholz,Adaptive scatternet support for bluetooth using sniffmode,in: Proceedings of the 26th Annual IEEE Conference onLocal Computer Networks, LCN 2001, pp. 112–120.

[4] S. Baatz, M. Frank, C. Kuhl, P. Martini, C. Scholz,Bluetooth scatternets: An enhanced adaptive schedulingscheme, in: Proceedings of the 21st Annual Joint Confer-ence of the IEEE Computer and Communications Soci-eties, INFOCOM 2002, pp. 782–790.

[5] L. Barriere, P. Fraigniaud, L. Narayanan, J. Opatrny,Dynamic construction of blue-tooth scatternets of foxeddegree and low diameter, in: Proceedings of the Four-teenth Annual ACM-SIAM Symposium on DiscreteAlgorithms, 2003.

[6] S. Basagni, R. Bruno, C. Petrioli, Device discovery inbluethooth networks: a scatternet perspective, LectureNotes in Computer Science, Vol. 2345, Springer, Berlin,2002, pp. 1087–1092.

[7] S. Basagni, R. Bruno, C. Petrioli, Performance evaluationof a new scatternet formation protocol for multi-hopbluetooth networks, in: The 5th International SymposiumWireless Personal Multimedia Communications, 2002,pp. 208–212.

[8] S. Basagni, R. Bruno, C. Petrioli, A performance com-parison of scatternet formation protocols for networks ofbluetooth devices, in: Proceedings of the First IEEEInternational Conference on Pervasive Computing andCommunications, 2003 (PerCom 2003), pp. 341–350.

[9] S. Basagni, C. Petrioli, Multihop scatternet formation forbluetooth networks, in: Proceedings of the 55th IEEEVehicular Technology Conference, VTC Spring 2002,pp. 424–428.

[10] S. Basagni, C. Petrioli, A scatternet formation protocolfor ad hoc networks of bluetooth devices, in: Proceedingsof the 55th IEEE Vehicular Technology Conference, VTCSpring 2002.

[11] P. Bhagwat, Bluetooth: technology for short-range wire-less apps, IEEE Internet Computing 5 (3) (2001) 96–103.

[12] P. Bhagwat, S.P.Rao,On the characterization of bluetoothscatternet topologies. Available from <http://www.winlab.rutgers.edu/~pravin/publications/publist.htm>.

[13] P. Bhagwat, A. Segall, A routing vector method (RVM)for routing in bluetooth scatternets, in: Proceedings of the1999 IEEE International Workshop on Mobile Multi-media Communications (MoMuC �99), pp. 375–379.

[14] V. Bhatnagar, G. Kesidis, Bluetooth scatternet formationusing proximity information of an election protocol, in:Proceedings of the Joint 2nd IEEE International Confer-ence on Net Working and IEEE International Conferenceon Wireless LANs and Home Networks (ICN�02 andICWLHN�02), Atlanta, August 26–29, 2002.

[15] BlueHoc. Available from <http://www.oss.software.ibm.com/developerworks/opensource/bluehoc/>.

[16] R.C. Braley, I.C. Gifford, R.F. Heile, Wireless personalarea networks: an overview of the ieee p802.15 group,Mobile Computing and Communications Review 4 (1)(2000) 26–33.

[17] J. Bray, C.F. Sturman, Bluetooth: Connect WithoutCables, Prentice-Hall, Englewood Cliffs, NJ, 2001.

[18] R. Bruno, M. Conti, E. Gregori, Wireless access toInternet via bluetooth: performance evaluation of the edcscheduling algorithm, in: Proceedings of the 1st Work-shop on Wireless Mobile Internet, 2001, pp. 43–49.

[19] R. Bruno, M. Conti, E. Gregori, Bluetooth: architecture,protocols and scheduling algorithms, Cluster Computing5 (2002) 117–131.

[20] S. Buttery, A. Sago, Future applications of bluetooth, BTTechnology Journal 21 (3) (2003) 48–55.

[21] A. Capone, M. Gerla, R. Kapoor, Efficient pollingschemes for bluetooth picocells, Proceedings of the IEEEInternational Conference on Communications (ICC2001), Vol. 7, 2001, pp. 1990–1994.

[22] C.F. Chisasserini, M.A. Marsan, E. Baralis, P. Garza,Towards feasible topology formation algorithms forbluetooth-based wpans, in: Proceedings of the 36thAnnual Hawaii International Conference on SystemsSciences, 2003, pp. 313–322.

[23] C.-S. Choi, H.-W. Choi, Dsr based bluetooth scatternet,in: ITC-CSCC, 2002.

[24] F. Chun-Choong, C. Kee-Chaing. Bluerings––bluetoothscatternets with ring structures, in: IASTED InternationalConference on Wireless and Optical Communication,2002.

[25] F. Cuomo, T. Melodia, A general methodology and keymetrics for scatternet formation in bluetooth, in: Pro-ceedings of the IEEE Global Telecommunications Con-ference, 2002, GLOBECOM �02, pp. 941–945.

[26] A. Das, A. Ghose, A. Razdan, H. Sarab, R. Shorey,Enhancing performance of asynchrounous data trafficover the bluetooth wireless ad hoc network, in: Proceed-ings of IEEE INFOCOM 2001.

[27] S. Cho, A. Chandrakasan, E. Shih, B.H. Calboun, Energyefficient link layer for wireless microsensor networks, in:Proceedings of the IEEE Computer Society AnnualWorkshop on VSLI (WVLSI�OO), Orlando, FL, 2001,pp. 16–21.

[28] N. Golmie, R.E. Van Dyck, A. Soltanian, Interference ofbluetooth and IEEE 802.11: simulation modeling andperformance evaluation, in: Proceedings of the 4th ACMinternational workshop on modeling, analysis and simu-lation of wireless and mobile systems, 2001, pp. 11–18.

[29] Bluetooth Special Interest Group. Available from<http://www.bluetooth.com/>.

[30] R. Guerin, E. Kim, S. Sakar, Bluetooth technology: keychallenges and initial research, in: Proceedings of theConference on Network and Distributed Simulations.

[31] J.C. Haartsen, The bluetooth radio system, IEEE Per-sonal Communications 7 (1) (2000) 28–36.

http://www.winlab.rutgers.edu/~pravin/publications/publist.htm

http://www.winlab.rutgers.edu/~pravin/publications/publist.htm

http://www.oss.software.ibm.com/developerworks/opensource/bluehoc/

http://www.oss.software.ibm.com/developerworks/opensource/bluehoc/

http://www.bluetooth.com/


[32] J.C. Haartsen, S. Mattisson, Bluetooth––a new low-power radio interference providing short-range connec-tivity, Proceedings of the IEEE 88 (10) (2000) 1651–1661.

[33] L. Har-Shai, R. Kofman, A. Segall, Inter-piconet sched-uling in bluetooth scatternets, in: Proceedings of theOPNETWORK Conference, 2002.

[34] C. Hsu, Y. Joung, An ns-based bluetooth toplogyconstruction simulation environment, in: Proceedings ofthe 36th Annual Simulation Symposium, 2003.

[35] In-Stat/MDR. Available from <http://standards.instat.com/>.

[36] N. Johansson, F. Alriksson, U. Jonsson, Jump mode––adynamic window-based scheduling framework for blue-tooth scatternets, in: Proceedings of MOBIHOC 2001, pp.204–211.

[37] N. Johansson, U. Korner, P. Johansson, Evaluation ofscheduling algorithm for bluetooth, in: Proceedings ofIFIP Broadband Communications, Hong Kong, 1999.

[38] P. Johansson, M. Kazantzidis, R. Kapoor, M. Gerla,Bluetooth and enabler for personal area networking,IEEE Network 15 (5) (2001) 28–37.

[39] R. Johansson, R. Kapoor, M. Kazantzidis, M. Gerla,Rendezvous sheduling in bluetooth scatternets, in: Pro-ceedings of the IEEE International Conference on Com-munications (ICC 2002), pp. 318–324.

[40] D.B. Johnson, D.A. Maltz, Dynamic source routing in adhoc networks. Available from <http://www.ietf.org/inter-net-drafts/draft-ietf-manet-dsr-05.txt>.

[41] C.E. Jones, K.M. Sivalingam, P. Agrawal, J.C. Chen, Asurvey of energy efficient network protocols for wirelessnetworks, Wireless Networks 7 (4) (2001) 343–358.

[42] M. Kalia, D. Bansal, R. Shorey, Data scheduling and sarfor bluetooth MAC, Proceedings of the 51st IEEEConference on Vehicular Technology, Vol. 2, 2000, pp.716–720.

[43] M. Kalia, S. Garg, R. Shorey, Scatternet structure andinter-piconet communication in the bluetooth system, in:Proceedings of IEEE National Conference on Communi-cations, 2000.

[44] R. Kapoor, M. Gerla, A zone routing protocol forbluetooth scatternets, in: IEEE Wireless Communicationsand Networking Conference, 2003, pp. 1459–1464.

[45] M. Kazantzidis, M. Gerla, On the impact of inter-piconetscheduling in bluetooth scatternets, in: Proceedings of theInternational Conference on Internet Computing, IC2002, pp. 37–43.

[46] S. Keshav, An Engineering Approach to ComputerNetworking: ATM Networks, the Internet, and theTelephone Network, Addison-Wesley, Reading, MA,1999.

[47] R. Kraemer, personal communication with ceo of less-wire. Available from <http://www.lesswire.com>, 2003.

[48] A. Kumar, BlueHoc manual. Available from <http://oss.software.ibm.com/bluehoc/tutorial/docpage.html>.

[49] C. Law, A.K. Mehta, K. Siu, Performance of a newbluetooth scatternet formation protocol, in: Proceedingsof the ACM Symposium on Mobile Ad Hoc Networking

and Computing 2001, Long Beach, CA, USA, October2001.

[50] C. Law, A.K. Mehta, K. Siu, A new bluetooth scatternetformation protocol, Journal of ACM Mobile Networksand Applications 8 (5) (2003) 485–498.

[51] C. Law, K. Siu, A bluetooth scatternet formationalgorithm, Proceedings of the Global Telecommunica-tions Conference, IEEE GLOBECOM �01, Vol. 5, 2001,pp. 2864–2869.

[52] Y.-Z. Lee, An efficient and fair polling scheme forbluetooth, Proceedings of MILCOM, Vol. 2, 2002, pp.1062–1068.

[53] X. Li, I. Stojmenovic, Y. Wang, Partial Delaunaytriangulation and degree limited localized bluetoothmultihop scatternet formation, Draft, 2003.

[54] X. Li, P. Wan, Y. Wang, O. Frieder, Sparse powerefficient topology for wireless networks, in: Proceedings ofthe 35th Annual Hawaii International Conference onSystem Sciences (HICSS �02), pp. 3871–3880.

[55] W. Lilakiatsakun, A. Seneviratne, Wireless home net-works based on a hierarchical bluetooth scattenet archi-tecture, in: Proceedings of the Ninth IEEE InternationalConference on Networks, pp. 481–485, 2001.

[56] T. Lin, Y. Tseng, K. Chang, C. Tu, Formation, routingand maintenance protocols for the bluering scatternet ofbluetooths, in: Proceedings of the Hawaii InternationalConference on System Science (HICSS-36), Big Island,Hawaii, January 6, 2003.

[57] Y. Liu, M. Lee, T. Saadawi, On demand formation ofbluetooth scatternet, Proceedings of MILCOM, Vol. 2,2002, pp. 1069–1074.

[58] Y. Liu, M.J. Lee, T.N. Saadawi, A bluetooth scatternet-route structure for multihop ad hoc networks,Proceedings of the IEEE Journal on Selected Areas inCommunications 21 (2) (2003) 229–239.

[59] R. Luo, R.M. Edwards, G.A. Manson, Enhancingbluetooth scheduler with predictive link capacity assign-ment plus multi-slot framing, Proceedings of the 10thInternational Conference on Telecommunications (ICT2003), Vol. 2, 2003, pp. 959–964.

[60] M.A. Marsan, C.F. Chiasserini, A. Nucci, G. Carello, L.De Giovani, Optimizing the topology of bluetoothpersonal area networks, Proceedings of the IEEE Info-com, 2002.

[61] G. Miklos, A. Racz, Z. Turanyi, A. Valko, P. Johansson,Performance aspects of bluetooth scatternet formation,in: Proceedings of the First Annual Workshop on Mobileand Ad Hoc Net working and Computing, MobiHOC2000, pp. 147–148.

[62] B. Miller, C. Bisdikian, Bluetooth Revealed: The Insider�sGuide to an Open Specification for Global WirelessCommunications, Prentice-Hall, Englewood Cliffs, NJ,2000.

[63] V.B. Misic, J. Misic, Bluetooth scatternet with a master/slave bridge: a queueing theoretic analysis, Proceedings ofthe IEEE Global Telecommunications Conference, GLO-BECOM �02, Vol. 1, 2002, pp. 207–211.

http://standards.instat.com/

http://standards.instat.com/

http://www.ietf.org/internet-drafts/draft-ietf-manet-dsr-05.txt

http://www.ietf.org/internet-drafts/draft-ietf-manet-dsr-05.txt

http://www.lesswire.com

http://oss.software.ibm.com/bluehoc/tutorial/docpage.html

http://oss.software.ibm.com/bluehoc/tutorial/docpage.html


[64] V.B. Misic, J. Misic, Bridges of bluetooth county:topologies, scheduling and performance, Proceedings ofthe IEEE Global Telecommunications Conference, GLO-BECOM �02, Vol. 1, 2002, pp. 207–211.

[65] V.B. Misic, J. Misic, Minimzing end-to-end delays inbluetooth scatternet with a slave/slave bridge, in: Pro-ceedings of the Eleventh International Conference onComputer Communications and Networks, 2002, pp.634–639.

[66] V.B. Misic, J. Misic, Modelling bluetooth piconet perfor-mance, IEEE Communications Letters 7 (1) (2003) 18–20.

[67] V.B. Misic, J. Misic, Performance of bluetooth bridges inscatternets with exhastive service scheduling, in: Proceed-ings of the 36th Annual Hawaii International Conferenceon System Sciences, January 6–9, 2003, pp. 311–320.

[68] N.A. Nordbotten, T. Sekie, N.D. Aavaag, Service dis-covery in bluetooth scatternets.

[69] NS: the network simulator. Available from <http://www.isi.edu/nsnam/ns/>.

[70] C.E. Perkins, Ad Hoc Networking, Addison-Wesley,Reading, MA, 2001.

[71] C. Petrioli, S. Basagni, Degree-constrained multihopscatternet formation for bluetooth networks, in: Proceed-ings of the IEEE Globecom 2002, Taipei, Taiwan,November 17–21, 2002.

[72] C. Petrioli, S. Basagni, I. Chlamtac, Bluemesh: Degree-constrained multihop scatternet formation for bluetoothnetworks, ACM/Kluwer/SPIE Mobile Networks andApplications 9 (1) (2004) 33–47 (Special Issue onAdvances in Research of Wireless Personal Area Net-working and Bluetooth Enabled Networks).

[73] B.J. Prabhu, A. Chocklingham, A routing protocol andenergy efficient techniques in bluetooth scatternets, in:Proceedings of the IEEE International Conference onCommunications, ICC 2002, Vol. 5, 2002, pp. 3336–3340.

[74] J. Rabaey, J. Ammer, J.L. da Silva Jr, D. Patel, Pico-radio: ad hoc wireless sensor networking of ubiquitouslow-energy sensor/monitor nodes, in: Proceedings of theIEEE Computer Society Annual Workshop on VSLI(WVLSI�00), Orlando, FL, 2000, pp. 9–12.

[75] A. Racz, G. Miklos, F. Bubinszky, A. Valko, A pseudorandom coordinated scheduling algorithm for bluetoothscatternets, in: Proceedings of the 2001 ACM Interna-tional Symposium on Mobile Ad Hoc Networking &Computing, 2001, pp. 193–203.

[76] L. Ramachandran, M. Kapoor, A. Sarkar, A. Arggwal,Clustering algorithms for ad hoc wireless networks, in:Proceedings of the DIAL-M Workshop, 2000, pp. 54–63.

[77] E.M. Royer, A review of current routing protocols forwireless networks, IEEE Personal Communications 6 (2)(1999) 46–55.

[78] K.V.S.S.S.S. Sairam, N. Gunasekaran, S.R. Redd, Blue-tooth in wireless communication, IEEE CommunicationsMagazine 40 (6) (2002) 90–96.

[79] T. Salonidis, P. Bhagwat, L. Tassiulas, Proximity aware-ness and fast connection establishment in bluetooth, in:Proceedings of the First Annual Workshop on Mobile

and Ad Hoc Networking and Computing, MobiHOC2000, pp. 141–142.

[80] T. Salonidis, P. Bhagwat, L. Tassiulas, R. LaMaire,Distributed topology construction of bluetooth personalarea networks, in: Proceedings of the Twentieth AnnualJoint Conference of the IEEE Computer and Communi-cations Societies, INFOCOM, Vol. 3, 2001, pp. 1577–1586.

[81] T. Salonidis, L. Tassiulas, Performance issues of blue-tooth scatternets and other asynchronous tdma ad hocnetworks, Technical Research Report CSHCN TR 2002-23, Centre for Satellite and Hybrid CommunicationNetworks, University of Maryland, USA, 2002.

[82] V. Sangvornvetphan, T. Erke, Traffic scheduling in blue-tooth network, in: Proceedings of the Ninth IEEE Inter-national Conference on Networks, 2001, pp. 355–359.

[83] T. Sato, K. Mase, A scatternet operation protocol forbluetooth ad hoc networks, Proceedings of the 5thInternational Symposium on Wireless Personal Multi-media Communications, Vol. 1, 2002, pp. 223–227.

[84] A. Seth, A. Kashyap, Capacity of bluetooth scatternets,2002.

[85] Simjava. Available from <http://www.dcs.ed.ac.uk/home/hase/simjava>.

[86] L.T. Son, H. Schioler, O.B. Madsen, Predictive schedul-ing approach in inter-piconet communications, in: Pro-ceedings of the 4th International Symposium on WirelessPersonal Multimedia Communications, 2001.

[87] L.T. Son, H. Schioler, O.B. Madsen, Hybrid distributediterative capacity allocation over bluetooth network, in:Proceedings of the 8th International Conference onCommunication Systems (ICCS 2002), pp. 583–588.

[88] IEEE standards. Available from <http://standards.ieee.org/catalog/>.

[89] I. Stojmenovic, Dominating set based bluetooth scatternetformation with localized maintenance, in: Proceedings ofthe International Parallel and Distributed ProcessingSymposium, IPDPS 2002, pp. 148–155.

[90] M. Sun, C. Chang, T. Lai, A self-routing topology forbluetooth scatternets, in: Proceedings of the InternationalSymposium on Parallel Architectures, Algorithms andNetworks, I-SPAN �02, pp. 13–18.

[91] M.-T. Sun, S. Wang, C.-K. Chang, T.-H. Lai, H.Sawatari, H. Okada, Interference aware mac schedulingand sar policies for bluetooth scatternets, Proceedings ofthe IEEE Global Telecommunications Conference (GLO-BECOM �02), Vol. 1, 2002, pp. 11–15.

[92] G. Tan, Interconnecting bluetooth-like personal areanetworks, in: Proceedings of the 1st Annual OxygenWorkshop, 2001.

[93] G. Tan, J. Guttag, A locally coordinated scatternetscheduling algorithm, in: Proceedings of the 27th AnnualIEEE Conference on Local Computer Networks,LCN �02, 2002.

[94] G. Tan, A. Miu, J. Guttah, H. Balakrishnan, Formingscatternets for bleutooth personal area networks, MITTechnical Report, MIT-LCS-TR-826, 2001.

http://www.isi.edu/nsnam/ns/

http://www.isi.edu/nsnam/ns/

http://www.dcs.ed.ac.uk/home/hase/simjava

http://www.dcs.ed.ac.uk/home/hase/simjava

http://standards.ieee.org/catalog/

http://standards.ieee.org/catalog/


[95] G. Tan, A. Miu, J. Guttah, H. Balakrishnan, An efficientscatternet formation algorithm for dynamic environ-ments, in: Proceedings of the IASTED Communicationsand Computer Networks (CNN), 2002.

[96] Texas Instruments. Available from <http://www.ti.com/rd/brf6100>.

[97] C.-K. Toh, M. Delwar, D. Allen, Evaluating the com-munication performance of an ad hoc wireless network,IEEE Transactions on Wireless Communications 1 (3)(2002) 402–414.

[98] Z. Wang, R.J. Thomas, Z. Haas, Bluenet––a newscatternet formation scheme, in: Proceedings of the 35thAnnual Hawaii International System Sciences, HICSS �02,pp. 779–787.

[99] J.P.F. Willekens, Ad hoc routing in bluetooth, LectureNotes in Computer Science, Vol. 2213, Springer, Berlin,2001, pp. 130–144.

[100] R. Ait Yaid, G. Heijenk, Polling best effort traffic inbluetooth, Wireless Personal Communications 23 (2002)195–206.

[101] D. Yang, G. Nair, B. Sivaramakrishnan, H. Jayakumar,A. Sen, Round robin with look ahead: a new schedulingalgorithm for bluetooth, in: Proceedings of the Interna-tional Conference on Parallel Processing Workshops,2002, pp. 45–50.

[102] G.V. Zaruba, S. Basagni, I. Chlamtac, Bluetrees––scat-ternet formation to enable bluetooth-based ad hocnetworks, Proceedings of the IEEE International Confer-ence on Communications, ICC 2001, Vol. 1, 2001, pp.273–277.

[103] W. Zhang, G. Cao, A flexible scatternet-wide schedulingalgorithm for bluetooth networks, in: Proceedings of the21st IEEE International Performance, Computing, andCommunications Conference, 2002, pp. 291–298.

[104] B. Zhen, J. Parj, Y. Kim, Scatternet formation ofbluetooth ad hoc networks, in: Proceedings of the 36thAnnual Hawaii International Conference on SystemSciences, HICSS �03, 2003, pp. 312–319.

Roger M. Whitaker holds a Ph.D.degree in Discrete Mathematics (1999)and a B.Sc. degree in Mathematics andManagement Science. He is a lecturerand a co-director of the Centre forMobile Communications, School ofComputer Science, Cardiff University,UK. Prior to this position, he carriedout research for the UK Radiocom-munications Agency into spectrumefficiency. His research addresses the

application of Computer Science to the design, co-ordination
and optimisation of wireless networks and systems. He is cur-rently leading a number of externally supported research pro-jects in this area.
Leigh Hodge holds a Ph.D. in Com-puter Science (2002) and a B.Sc. degreein Computer Science (1995), both fromthe School of Computer Science, Car-diff University, Wales, UK. He is cur-rently a research associate at theCentre for Mobile Communications atCardiff University. His currentresearch addresses the application ofmeta-heuristic and evolutionaryapproaches to the optimisation of

wireless network design for Bluetooth and ad hoc networks.

Imrich Chlamtac holds a Ph.D. degreein computer science from the Univer-sity of Minnesota. Since 1997 he hasbeen the Distinguished Chair in Tele-communications at the University ofTexas at Dallas and holds the titles ofSackler Professor at Tel Aviv Univer-sity, Israel, The Bruno Kessler Hon-orary Professor at the University ofTrento, Italy, and University Professorat the Technical University of Buda-

pest, Hungary. He is a Fellow of the IEEE and ACM societies,
a Fulbright Scholar and an IEEE Distinguished Lecturer. He isthe winner of the 2001 ACM Sigmobile annual award and theIEEE ComSoc TCPC 2002 award for contributions to wirelessand mobile networks, and of multiple best paper awards inwireless and optical networks. He has published over 300papers in refereed journals and conferences, and is the co-author of the first textbook on Local Area Networks (Lexing-ton Books, 1981, 1982, 1984) and of Mobile and WirelessNetworks Protocols and Services (John Wiley, 2000). He servesas the founding Editor-in-Chief of the ACM/URSI/KluwerWireless Networks (WINET), the ACM/Kluwer Mobile Net-works and Applications (MONET) journals and the SPIE/Kluwer Optical Networks (ONM) Magazine.
http://www.ti.com/rd/brf6100

http://www.ti.com/rd/brf6100

Documents

Bluetooth scatternet formation: A surveycs752/papers/bluetooth-002.pdfcomparisons are technically correct, Bluetooth has a much more powerful business model associ-atedwithit,basedontwokeypoints.Firstly,Blue-tooth