IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 4, NO. 6, NOVEMBER 2005

2649

A Distributed Self-Healing Approach to Bluetooth Scatternet Formation Carla-Fabiana Chiasserini, Member, IEEE, and Marco Ajmone Marsan, Fellow, IEEE

Abstract—This paper proposes a distributed self-healing technique for topology formation in dynamic Bluetooth wireless personal area networks (BT-WPANs) and analyzes three new algorithms for scatternet formation. The three algorithms employ distributed procedures for the insertion of one or more nodes in a BT-WPAN, and are able to effectively compromise between the need for system efficiency and the desire to promptly adapt to topology changes. Depending on which algorithm is employed, the proposed approach generates BT-WPANs with different connectivity properties as well as topology structures. Index Terms—Bluetooth, topology formation, wireless personal area networks (WPANs).

I. I NTRODUCTION

W

IRELESS personal area networks (WPANs) are based on a new networking approach that enables short-range connections between various wireless devices. One of the most promising technologies for the deployment of WPANs is Bluetooth [1], which has been recently standardized as IEEE 802.15.1 WPAN. In the following, we assume the reader to be familiar with Bluetooth specifications and terminology. One of the most challenging issues in Bluetooth WPANs (BT-WPANs) concerns the formation of the network topology and its changes as nodes arrive or move out. Distributed algorithms for the creation of BT-WPANs have been proposed in [2]–[9]. In this paper, we propose a scatternet formation approach, which has the following nice properties: 1) it is completely distributed; 2) it applies to the general case of multihop topologies and does not require the use of extra hardware at the Bluetooth devices; 3) it fully meets the Bluetooth specifications; 4) it does not require the election of any node as responsible of special tasks; and 5) it is self-healing in that it is able to handle the topology changes that may occur in a BT-WPAN during its operation. Based on this approach, we present three scatternet formation algorithms that differ in the level of information about the BT-WPAN topology that is assumed during the insertion of a node. Depending on which algorithm is employed, our approach gives BT-WPANs with different connectivity properties as well as topology structures. All algorithms aim at satisfying the Bluetooth technology constraints while providing full network connectivity, high throughput, and reduced overhead.

Manuscript received November 10, 2003; revised December 27, 2004; accepted December 28, 2004. The editor coordinating the review of this paper and approving it for publication is A. Boucouvalas. This work was supported by the Italian Ministry for University and Research through the Virtual Immersive Communications (VICOM) and the Piattaforme Riconfigurabili per Interoperabilità in Mobilità (PRIMO) projects. The authors are with the Dipartimento di Elettronica, Politecnico di Torino, Torino 10129, Italy (e-mail: [email protected]; [email protected]). Digital Object Identifier 10.1109/TWC.2005.858297

II. P RELIMINARIES According to Bluetooth specifications [1], the establishment of a connection between any pair of nodes is performed through inquiry and paging procedures. During the inquiry procedure, a node may enter either the INQUIRY state or the INQUIRY_SCAN state. When a node in the INQUIRY_SCAN state receives an ID packet, it responds with an Frequency Hop Synchronization (FHS) packet if it wishes to be discovered. An FHS packet includes the identity of the responding node, an indication of its native clock and device class, plus 5 bits that are not used for the inquiry response. We propose to employ these 5 bits to convey some additional information from the responding node to the inquiring device. In particular, we assume that — 2 bits represent the node energy status (e.g., battery charge level below 25%, between 25% and 50%, between 50% and 75%, higher than 75%); — 2 bits represent the congestion level of the node (e.g., low, medium, high traffic load, or congested); — 1 bit gives an indication on the node connectivity status. This bit specifies whether the node belongs to an isolated piconet. We define a piconet to be isolated if either the piconet is not connected to any other piconet or all neighboring piconets are connected with this piconet only. When more than one node responds to an inquiry, the device that started the procedure has to decide to which node it should connect. The decision can be made based on the responses that the inquiring device receives and on the information it can acquire on the nodes in its proximity. By employing the proper inquiry access code (IAC), an inquiring node can search for a particular type of device, and a node responding to a generic inquiry can include its device class in the FHS packet. Next, the node that performed the inquiry selects the neighbor to connect with and starts paging it. In our work, we do not tackle the issue of neighborhood discovery. By relying on the work in [8], we assume that a node that is not connected to any other node alternates between the INQUIRY state and the INQUIRY_SCAN state till it gets in contact with other devices and starts a paging procedure. Instead, a node that is already connected to other devices only periodically enters the INQUIRY_SCAN state. Such an approach significantly speeds up the process of neighborhood discovery; indeed, as the scatternet gets formed, more nodes listen to the inquiry messages coming from unconnected devices. III. D ISTRIBUTED T OPOLOGY F ORMATION A LGORITHMS The three topology formation algorithms that we propose are based on distributed procedures to handle topology changes; in particular, we refer to the insertion of a node, while the removal of a node is omitted for the lack of space.

1536-1276/$20.00 © 2005 IEEE

2650

IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 4, NO. 6, NOVEMBER 2005

In several cases, insertion procedures must select to which node the incoming device should connect, choosing among nodes with similar characteristics. In all such cases, the preferred node is chosen considering sequentially: 1) the node energy level: the highest energy node(s) is (are) preferred; 2) among maximum energy nodes, the node traffic load is considered: the lowest traffic node(s) is (are) preferred; 3) among maximum energy and minimum traffic nodes, the minimum distance node is preferred.1 A. First Node Insertion Procedure The first distributed procedure for the insertion of a new node into the BT-WPAN topology (IN1) assumes that only a minimal amount of information about the preexisting BT-WPAN topology is acquired by the new node and used to decide how to best enter the network. As described in Section II, a node wishing to join the BT-WPAN starts an inquiry procedure by broadcasting ID packets [1]. The nodes in its proximity, which are listening on the predefined channel for inquiry messages, respond with an FHS packet if they are willing to accept a new neighbor. When more than one node replies, the new node has to decide which node it should connect to. The decision can be made based on the responses that the new node receives and on the information it can acquire about the nodes in its proximity. In the case of this algorithm, only the bit of connectivity status is used in the inquiry response that is carried by the FHS packet. This bit specifies whether the node belongs to an isolated piconet. We define a piconet to be isolated if either the piconet is not connected to any other piconet or all neighboring piconets are connected with this piconet only. Let Nnew be the node starting the inquiry. Depending on the responses received from the neighbors, Nnew must decide which node it will page. A node replying to Nnew may belong to either an isolated or a non-isolated piconet. In addition, it may be: 1) a master with less than seven connected active slaves; 2) a master with already seven connected active slaves; 3) a master and bridge; 4) a slave and bridge; or 5) a slave. The algorithm followed by Nnew to select the node(s) to page is influenced by the fact that some nodes belonging to an isolated piconet (call it an isolated node) reply to the inquiry procedure. • If at least one isolated node replies to the inquiry procedure, then Nnew first tries to establish a new piconet to interconnect those nodes. In order to achieve this goal, Nnew becomes a master of a new piconet where all isolated nodes are slaves (if the number of isolated nodes that replied to the inquiry procedure is less than seven; otherwise, only six isolated nodes, chosen according to their class in the following order: master, master and bridge, slave, slave and bridge, and resolving ties through the preference rules described before) and act as bridges toward their former piconet.

1 A node can estimate its distance from another Bluetooth device by using the received signal strength indicator (RSSI).

Then, Nnew tries to connect the newly created piconet to other non-isolated piconets by considering the nonisolated nodes that replied to the inquiry procedure. If no such nodes exist, no further action is taken. If some non-isolated nodes replied to the inquiry procedure, then Nnew selects one of them according to their role in the following order: slave, slave and bridge, master, master and bridge, master with seven slaves, and resolving ties through the preference rules described before. The selected node becomes a slave in the new piconet created by Nnew (actually, if the node was a slave, it becomes a slave and bridge; while if it was a master, it becomes a master and bridge). At this point, Nnew tries to merge its newly created piconet with a preexisting piconet. Indeed, if Nnew during the inquiry procedure received the replies of all slaves of a master or master and bridge that joined the new piconet of Nnew , and if space is available in the new piconet, then Nnew adds to its new piconet all slaves of the master of the preexisting piconet and the master itself becomes only a slave of Nnew . • If no isolated node replies to the inquiry procedure, then Nnew selects the node to page according to the following order of preference. 1) Master with less than seven connected active slaves. Nnew selects the preferred master node. Denote the chosen master with µ. Nnew pages µ and creates a new piconet, where Nnew is the master and µ is a slave. Afterwards, the two nodes switch their roles, so that Nnew becomes a slave in the piconet controlled by µ. 2) Slave. Two possible cases are considered. a) Nnew has enough processing and energy capabilities to be a master. Then, Nnew forms a new piconet by paging one or more of the slaves that have responded to its inquiry (the number of slaves is a parameter of the algorithm). Such slave nodes become bridges between the new piconet and their former piconet. b) The characteristics of Nnew force it to act as a slave. Nnew chooses the preferred node to page among those that replied to the inquiry and forms a new piconet with the selected device. Afterwards, the two nodes exchange roles so that in the new piconet Nnew becomes a slave and the selected device becomes the master and a bridge toward its former piconet. 3) Slave and bridge. As before, two possible cases are considered. a) Nnew has enough processing and energy capabilities to be a master. Node Nnew selects the slave and bridge node to page on the basis of preference. A new piconet is created, where Nnew acts as the master and the selected node as a slave. b) The characteristics of Nnew force it to act as a slave. Nnew chooses the preferred node to page among those that replied to the inquiry and forms a new piconet with the selected device. Afterwards, the two nodes exchange roles, so that in the new

IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 4, NO. 6, NOVEMBER 2005

piconet Nnew becomes a slave and the selected device becomes the master and a bridge toward its former piconet. 4) Master and bridge. Nnew chooses the preferred node to page among those that replied to the inquiry and forms a new piconet with the selected device. Afterwards, the two nodes exchange roles, so that in the new piconet Nnew becomes a slave and the selected device becomes the master and a bridge toward its former piconet. 5) Master with already seven connected active slaves. Nnew chooses the preferred node to page among those that replied to the inquiry and forms a new piconet with the selected device. A new piconet is created, whose master is Nnew . Then, there are two possible ways to proceed. a) In the new piconet, Nnew remains the master and the selected node acts as a slave and bridge. b) The two nodes exchange roles. To do so, the selected node puts one of its slaves in park mode and takes Nnew as an additional slave. The slave in park mode must then search for a new piconet by performing an insertion procedure on its turn. B. Second Node Insertion Procedure The second procedure (IN2) assumes that the new node can acquire a rather detailed information about the preexisting BT-WPAN topology by contacting the masters of the preexisting piconets. The masters have the information about the piconet connectivity, i.e., they know which piconets are reachable from any piconet. While it is natural to assume that a master knows which piconets are directly reachable through the bridges in its piconet, the global connectivity information can be obtained through an exchange of information among masters. Like in the previous procedure, a node wishing to join the BT-WPAN starts an inquiry procedure and the nodes in its proximity willing to accept a new neighbor respond with an FHS packet. The new node selects which of the responding nodes to page and forms a new piconet in which it acts as a master. Also in this case, let Nnew be the node starting the inquiry procedure. Depending on the responses received from neighboring nodes, Nnew must decide which node it will page. If a master or master and bridge replies to the inquiry, Nnew pages up to seven such nodes and includes them as slaves in the newly created piconet. Those nodes become master and bridge in their original piconets. Using the data channels in the new piconet, Nnew can obtain detailed information from such master and bridge nodes about the piconet connectivity and can expel from the new piconet the masters of those piconets that already communicate. If some slave or slave and bridge replies to the inquiry, after having paged the master or master and bridge nodes as just explained, if less than seven slaves are left in the piconet, Nnew pages such slave or slave and bridge nodes and includes them in the new piconet so that all become slave and bridge nodes. Using the data channels in the new piconet, Nnew can obtain

2651

from such slave and bridge nodes the ID of the master of their piconets. Thus, Nnew that has obtained at the previous step the information about the piconet connectivity can label the slaves that are connected (directly or indirectly) to the masters that already belong to the new piconet. If no master was inserted in the new piconet at the previous step, Nnew does not have information about the connectivity among the piconets of its slaves, but it can at least know which nodes belong to the same piconet (through the master ID) and label all but one of those. Nnew can then expel from the new piconet all labeled slaves. As a last step, the algorithm tries to reduce the number of piconets by merging or eliminating useless piconets. If only masters replied to the inquiry, Nnew releases its role as a master, thus eliminating the new piconet, and becomes a slave and bridge in all piconets. If only slaves replied to the inquiry, no piconet reduction is possible. If both masters and slaves replied to the inquiry, Nnew incorporates the whole piconet of a master if it can reach all its slaves and if the number of nodes permits. C. Third Node Insertion Procedure The third procedure (IN3) assumes that the new node can acquire the piconet connectivity information from any node that replies to the inquiry procedure. This assumption is justified by the fact that, while the connectivity information is generated by an exchange of information among masters, it is possible to disseminate this information to all slaves over the Bluetooth data channels. Like in both previous procedures, a node wishing to join the BT-WPAN starts an inquiry procedure, and the nodes in its proximity respond with an FHS packet if they are willing to accept a new neighbor. The new node selects which of the responding nodes to page and forms a new piconet in which it acts as a master. Also in this case, let Nnew be the node starting the inquiry procedure. Depending on the responses received from neighboring nodes, Nnew must decide which node(s) it will page. Nnew pages up to seven nodes, preferring masters to slaves, and includes them as slaves in the newly created piconet. Those nodes become bridges toward their original piconets. Using the data channels in the new piconet, Nnew obtains from such master and bridge or slave and bridge nodes information about piconet connectivity and can expel from the new piconet the nodes of those piconets that already communicate. Finally, the algorithm tries to reduce the number of piconets by merging or eliminating useless piconets, as in procedure IN2. IV. N UMERICAL R ESULTS We evaluate the performance of the proposed node insertion algorithms via simulation. We consider a square service area of 1000 m2 and a piconet radius of 10 m. The topology is constructed incrementally, with new nodes arriving at random time instants and taking a random position within the service area. In all plots, the performance metrics of interest are shown as functions of the node density (in nodes per square meter) that is reached at the end of the scatternet formation process; we denote by N the total number of nodes in the network

2652

IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 4, NO. 6, NOVEMBER 2005

TABLE I PERCENTAGE GENERATED BY THE PROPOSED ALGORITHMS : CONNECTED TOPOLOGIES AND TREE TOPOLOGIES

Fig. 1. Average number of piconets versus node density for the topologies obtained through the insertion algorithms IN1, IN2, and IN3.

Fig. 2. Average number of bridge nodes per piconet in the topologies obtained through the insertion algorithms IN1, IN2, and IN3 as a function of node density.

at the end of the formation procedure. Plots are obtained by averaging the results of 5000 runs for each value of node density; each run corresponds to a different instance of the random variables of the system model. The results presented here are derived by assuming that all Bluetooth devices have the same computational capabilities and energy resources; also, since we do not tackle the issue of traffic routing and scheduling, traffic connections are not simulated. The values of the delays associated with the inquiry and paging procedures are as reported in [10]. We assume that six slots is the time needed for two nodes to exchange the connectivity matrix in the IN2 and IN3 algorithms.2 Table I presents some results concerning the connectivity properties and topology structure of the scatternets constructed through our approach for different values of node density. The table reports the percentage of connected networks that we obtain starting from topologies that are connected at the physical layer, and, among the connected BT-WPANs, the percentage of topologies with a tree structure. We notice that the IN1 algorithm is not able to guarantee connectivity, although the percentage of unconnected topologies is very low. On the contrary, the IN2 and IN3 algorithms always produce connected scatternets; the difference is that, except for very small values of node density, the network topology obtained through IN2 is a mesh, while the IN3 scheme always gives a tree structure. These results suggest that, depending on the algorithm that is applied (i.e., the amount of information on the BT-WPAN available during an insertion procedure), we obtain topologies with different connectivity properties and logical structure.

Fig. 1 presents the average number of piconets in BT-WPANs obtained by using our approach as a function of node density. Given the number of nodes, the number of piconets that is formed in the scatternet is a critical issue; however, as shown in Fig. 1, the value of this parameter necessarily increases with the increase of node number (hence of density). As expected, the IN3 scheme, which produces tree structures, always gives the smallest number of piconets; instead, the IN2 algorithm includes in any new piconet a large number of preexisting nodes, thus creating quite a large number of piconets. The IN1 scheme gives performance close to the IN3 algorithm, although IN1 is likely to form meshed topologies rather than trees. The reason of this behavior is that the IN1 algorithm often forms meshed topologies but with a lower connectivity degree than the IN2 scheme. Fig. 2 presents the average number of bridge nodes per piconet versus the node density for the IN1, IN2, and IN3 algorithms. As previously pointed out, the IN2 scheme typically includes in each piconet formed by a newly arrived node several preexisting devices, thus creating many bridge nodes. Such a behavior is evident from the plot in Fig. 2. The smallest number of bridges is obtained by using the IN3 algorithm, since it always forms tree topologies, while the IN1 scheme gives intermediate performance between the results obtained through IN2 and IN3. We highlight that, on one hand, a large number of bridges may be a desirable feature because it enables us to distribute the interpiconet traffic among several nodes. On the other hand, the presence of many bridge nodes implies that the overhead due to a node switching from one piconet to another is significant and increases the energy consumption of bridges. Which solution is to be preferred depends on the BT-WPAN application. We point out that, in the case of the IN3 scheme, the number of bridges per piconet is about equal to 1.7 because a single

2 Namely, one slot to send the information request, three slots to send the connectivity information carried on a DH3 packet [1].

IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 4, NO. 6, NOVEMBER 2005

Fig. 3. Average number of interconnected piconets per bridge node as a function of the node density and for the topologies obtained through the insertion algorithms IN1, IN2, and IN3.

Fig. 4. Average number of adjacent piconets for each piconet in the BT-WPAN as a function of node density. The results obtained through insertion algorithms IN1, IN2, and IN3 are presented.

bridge node may interconnect more than two piconets. This is why, as shown in Fig. 3, the average number of piconets interconnected by a single bridge node is slightly greater than two in the topologies formed by the IN3 algorithm. We would like the average number of piconets interconnected by a bridge node to be as close as possible to two so that the overhead due to bridge nodes switching from one piconet to another is low. We can see that for any value of node density, the average number of interconnected piconets per bridge is very close to the desired value of two for both IN1 and IN3 schemes. Instead, in the case of IN2, the average number of interconnected piconets per bridge, hence the bridge node overhead, becomes quite significant for node density values greater than 0.05. The average number of adjacent piconets for each piconet in the scatternet is presented in Fig. 4. In accordance with the results shown in the previous plots for the IN1 and IN3 schemes, this parameter is very close to two. Conversely, the IN2 algorithm generates highly connected meshed networks; thus, in this case, each piconet has several neighboring piconets, allowing for small traffic delivery delays. By comparing the plots referring to schemes IN1 and IN3 in Figs. 2 and 4, we can see that the average number of bridges per piconet is much larger for IN1 than for IN3, especially for high node densities, while the average number of adjacent piconets is similar for the two schemes. This means that IN1 generates several bridges between the same pair of piconets.

2653

Fig. 5. Average number of roles per node versus node density for the topologies produced by the insertion algorithms IN1, IN2, and IN3.

In Fig. 5, we show the average number of roles taken by each node in the BT-WPAN, which is a key scatternet parameter. Indeed, it represents the number of piconets in which a node participates, thus accounting for the number of bridges in the scatternet as well as for the average number of interconnected piconets per bridge. In the case of the IN1 and IN3 schemes, the average number of roles per node is less than 1.5 for any value of node density; this implies that the obtained topologies are very efficient and the overall overhead in communication is small. This result is especially surprising for the IN1 algorithm, which gives a meshed network in the majority of cases. In the case of the IN2 scheme, the average number of roles per node becomes significantly higher as the node density increases. This is again due to the high number of bridges generated by IN2 and by the fact that these bridges interconnect several piconets. Clearly, this may be a drawback of scheme IN2; however, it ensures that we always obtain a connected topology—a highly desirable property of IN2. V. C ONCLUSION In this paper, we described and evaluated a distributed self-healing approach for the distributed topology formation in dynamic Bluetooth wireless personal area networks (BT-WPANs). Simulation results indicate that two of the proposed algorithms, IN2 and IN3, are capable of generating connected topologies whenever the physical network characteristics allow, while IN1 falls short of this objective. While IN1 and IN2 generate meshed topologies, IN3 generates trees. Algorithm IN2 generates meshed topologies with higher interconnection degree than IN1. R EFERENCES [1] Bluetooth Core Specifications. (2004). Overland Park, KS: Bluetooth SIG, Inc. Version 2.0. [Online]. Available: http://www.bluetooth.org [2] C. Petrioli, S. Basagni, and I. Chlamtac, “BlueMesh: Degree-constrained multi-hop scatternet formation for Bluetooth networks,” MONET, Special Issue on Mobility of Systems, Users, Data and Computing, vol. 9, no. 1, pp. 33–48, Feb. 2004. [3] C. Law and K.-Y. Siu, “A Bluetooth scatternet formation algorithm,” in IEEE Symp. Ad Hoc Wireless Networks, Global Telecommunications (GLOBECOM), San Antonio, TX, Nov. 2001, pp. 2864–2869. [4] G. V. Zaruba, S. Basagni, and I. Chlamtac, “Bluetrees—Scatternet formation to enable Bluetooth-based ad hoc networks,” in Proc. IEEE Int. Conf. Communications (ICC), Helsinki, Finland, Jun. 2001, pp. 273–277.

2654

IEEE TRANSACTIONS ON WIRELESS COMMUNICATIONS, VOL. 4, NO. 6, NOVEMBER 2005

[5] S. Basagni and C. Petrioli, “A scatternet formation protocol for ad hoc networks of Bluetooth devices,” in Proc. IEEE Vehicular Technology Conf. (VTC) Spring, Birmingham, AL, May 2002, pp. 424–428. [6] I. Stojmenovic, “Dominating set based Bluetooth devices,” in Workshop Advances Parallel and Distributed Computational Models, Fort Lauderdale, FL, Apr. 2002, p. 148b. [7] Z. Wang, R. J. Thomas, and Z. J. Haas, “Bluenet—A new scatternet formation scheme,” in Proc. 35th Hawaii Int. Conf. System Science (HICSS35), Big Island, HI, Jan. 2002, p. 61. [8] G. Tan, A. Miu, J. Guttag, and H. Balakrishnan, “An efficient scatternet formation algorithm for dynamic environments,” in Proc. Int.

Association Science and Technology Development (IASTED) Int. Conf. Communications and Computer Networks (CCN), Cambridge, MA, Nov. 2002, pp. 1–7. [9] F. Cuomo, T. Melodia, and I. F. Akyildiz, “Distributed self-healing and variable topology optimization algorithms for QoS provisioning in scatternets,” IEEE J. Sel. Areas Commun., vol. 22, no. 7, pp. 1220–1236, Sep. 2004. [10] T. Salonidis, P. Bhagwat, L. Tassiulas, and R. LaMaire, “Distributed topology construction of Bluetooth personal area networks,” in Proc. IEEE Information Communications (INFOCOM), Anchorage, AK, Apr. 2001, pp. 1577–1586.