FORMATION OF SCATTERNETS WITH HETEROGENEOUS BLUETOOTH DEVICES Paal Engelstad, Do Van Thanh, Tore E. Jonvik University of Oslo (UniK) / Telenor R&D, 1331 Fornebu, Norway {Paal.Engelstad, Thanh-van.Do, Tore-erling.Jonvik }@telenor.com ABSTRACT Bluetooth is an open specification for short-range wireless communication. It has gradually been extended to meet the demand for personal ad-hoc networking. A Bluetooth ad-hoc network consists of Bluetooth devices interconnected into piconets and piconets interconnected into scatternets. A number of scatternet formation algorithms have been proposed, but none has yet taken into account that Bluetooth devices may have very different capabilities. We have explored how three proposed scatternet formation algorithms perform when a share of the scatternet devices have low-capabilities and cannot forward packets. Simulations illustrated how the efficiency of the scatternet topology is deteriorated as the share of low-capability devices increases. We have also evaluated each algorithm with respect to its capability to accommodate low-capability devices in the scatternet. 1. INTRODUCTION Bluetooth is a technology for short-range wireless communication, and is designed for imposing low battery drain on the wireless terminal. Originally, it was intended for cable replacement between two devices, but the specification has gradually been extended to meet the demand for ad-hoc networking [1]. Due to low manufacturing costs (down to 5 USD/ Bluetooth chip) it is economically feasible to incorporate Bluetooth not only into computing devices, such as PDAs, mobile phones and portable PCs, but also into low-capability devices, such as head-sets, micro-displays and various sensors. A Bluetooth ad-hoc network consists of Bluetooth devices – often of widely different capabilities interconnected into piconets and piconets interconnected into scatternets. The Bluetooth specification outlines how two Bluetooth devices interconnect and form a piconet. When two devices come into each other’s communication range, they detect each other by first going through an INQUIRY phase. The following PAGE phase allows the devices to start establishing a connection. One of the devices obtains the role of a master, while the other becomes a slave. Each piconet consists of only one master, and maximum 255 slaves. However, the number of active slaves in a piconet cannot exceed seven. Although a device can only be master of maximum one piconet, it can be slave in a number of different piconets, or master in one piconet and slave in others. Hence, two piconets can be interconnected if one piconet master is also slave of the other piconet (i.e. as a ‘masterslave’ bridge), or if one device is slave of both piconets (i.e. as a ‘slave-slave’ bridge). It is anticipated that some low-capability devices might not participate in the scatternet at the same level as more advanced computing devices do. For example, due to low battery capacity, little memory or little processing power, some low-capability devices may not be capable or willing to forward packets on behalf of other nodes. They are bound to participate in scatternets as slaves while only more advanced computing devices may assume the role of a master, master-slave-bridge or slave-slave bridge. None of the existing algorithms proposed for Bluetooth scatternet formation takes the aspect of heterogeneous devices into consideration. We have therefore studied the performance of three proposed algorithms when low-capability devices are present. Amongst a wide number of proposed algorithms for scatternet formation, we chose three algorithms based on the following two basic criteria: 1. Weakly connected underlying topologies: A realistic scatternet formation algorithm should not assume a strongly connected underlying topology where each node may connect to any other node in the scatternet. Instead, the algorithm should work even when there might be at least two nodes in the same scatternet that are out of each other’s radio range. 2. Asynchronous scatternet formation: The algorithms should not mandate that all nodes start forming the scatternet at the same time (e.g. in terms of an algorithm comprising several distinct phases). Such synchronous formations would probably be difficult to co-ordinate in real usage scenarios. Instead, a Bluetooth device should be able to enter or leave the scatternet at any time. In these events, topology changes should be handled locally and not propagated throughout the scatternet. The resulting three algorithms that we have examined were Tree Scatternet Formation (TSF), Slave-slave-based Scatternet Formation (SSF) and Master-slave-based Scatternet Formation (MSF) [2] [3]. They are all asynchronous and designed for work well, even for weakly connected underlying topologies. The next section outlines relevant work on scatternet formation. Section 3 presents the framework that we http://folk.uio.no/paalee/ developed for the investigation of scatternets formation with heterogeneous devices. Based on this framework, we performed simulations. Simulation results are presented in Section 4. Finally, we present the conclusions of our work in Section 5. 2. RELATED WORK 2.1. Bluetooth link formation The link formation process is specified in the Bluetooth baseband specification [1]. The link formation comprises an Inquiry phase followed by a Page phase. In both phases the hopping pattern is limited to 32 frequencies (or only 16 frequencies in a few countries). The Inquiry process allows a master node to discover and collect clock and address information about neighboring devices. The page process uses this information to establish a bidirectional frequency hopping communication channel. master receives the response and establishes a connection with the slave. 2.2. Related work on scatternet formation Miklos et al. [4] undertook an early study on scatternet formation. They used heuristics to generate scatternets with some desirable properties, and evaluated them by means of simulations. Johansson et al. [5] performed link-layer simulations of piconets. Later Salonidis et al. [6] presented a symmetric link formation scheme where no configuration of potential master or slave roles is necessary. Instead, each node alternates between the INQUIRY and INQUIRY SCAN states (Figure 2). Figure 2. The symmetric link formation scheme proposed by Salonidis et al. [6] Figure 1. Connection State Diagram for Bluetooth link formation During the Inquiry process a potential master (i.e. the device to be the master of the connection to be formed) will first have to enter the INQUIRY state, while a potential slave will have to enter the INQUIRY SCAN state (Figure 1). A device in the former state alternates between transmitting short identity-packets containing an Inquiry Access Code (IAC) and listening for responses. A node in the latter state, on the other hand, listens for packets from devices in the INQUIRY state, performs an IAC check and responds when appropriate with an Inquiry Response Message containing the device address and other device specific information. Once the Inquiry is successful, the potential master enters the PAGING state and tries to establish a connection with the specified potential slave. Meanwhile, the potential slave has shifted to the PAGE SCAN state. It receives a Paging packet from the potential master and processes the Device Access Code. If appropriate, it returns a response to the master and finally enters the connection state. The http://www.unik.no/personer/paalee Using this scheme, they also proposed the Bluetooth Topology Construction Protocol (BTCP) [6]. BTCP comprise an initial procedure in which a scatternet leader is first elected. The election phase means that BTCP is a synchronous protocol (although devices are allowed to act in a distributed fashion). The election algorithm also requires a strongly connected underlying topology in which each node is in direct communication distance with every other node in the scatternet. After election, the elected leader determines the scatternet topology and instructs other devices on how to form the scatternet. As the resulting topology is not detailed by the protocol, BTCP represent an election protocol more than a complete scatternet formation algorithm. On the contrary, the Bluetree algorithm, proposed by Zaruba et al. [7], assumes that a node knows a-priori whether it is a scatternet leader (or ‘root’) or not. It then uses some distributed spanning tree algorithms to form a hierarchical scatternet tree of master-slave bridges rooted at the scatternet leader. Since the root initiates the scatternet formation while other devices must wait to be paged, the formation algorithm is synchronous. However, unlike BTCP, it does not require a strongly connected underlying topology. Zaruba et al. also proposed to use slave-slave bridges to interconnect different Bluetrees. Wang et al. [8] go one step further and propose a non-hierarchical (however, still synchronous) variation of Bluetrees. Unlike the previous proposals, the LMS algorithm, proposed by Law et al. [9] is an asynchronous approach in which a scatternet is built as nodes join the scatternet tree, and nodes can join anytime. The scatternet formation overhead is kept to a minimum since one node is designated the task to perform INQUIRY or INQUIRY SCAN on behalf of the whole scatternet. However, this scheme requires a strongly connected underlying topology. Tan et al. [2] proposed the Tree Scatternet Formation (TSF) algorithm. A hierarchical scatternet is formed asynchronously as nodes join the tree. Unlike LMS, the algorithm allows for a weakly connected underlying topology of devices. Two trees may combine into one tree as the roots of the trees get within communication range. Neither LMS nor TSF require the aforementioned symmetric link formation proposed by Salonidis et al. [6]. Finally, Engelstad et al. [3] proposed SSF and MSF as non-hierarchical counterparts to TSF. In addition to being asynchronous and accommodating weakly connected underlying topology of devices, these algorithms also obey two design principles targeted at enhancing the efficiency of the resulting scatternet [3, 6, 8]: a. b. Minimal Piconet Overlap: Two piconets should be connected by at most one bridge node. Minimal Bridge Degree: A bridge node should participate in at most two piconets, i.e. each slaveslave bridge is slave of exactly two masters, while each master-slave bridge is slave of exactly one other master or master-slave. Engelstad et al. [3] introduced a role-transition matrix to describe state transitions of asynchronous scatternet formation algorithms that are obeying the two aforementioned design principles. At some point in time every device starts out as a free node. A regular free node F may then encounter and connect to another device. Whether it becomes master (i.e. in role-state M) or slave (i.e. in role-state S) depends on the role of the node it connects to and can be read from the role-transition matrix of the scatternet formation algorithm. When a slave or master connects to a node of another piconet, it may obtain an additional slave-role or master-role and become a bridge between the two piconets. Whether it becomes a masterslave bridge (i.e. in role-state ms) or a slave-slave bridge (i.e. in role-state ss) can also be found from the roletransition matrix. The ‘Maximum Bridge Degree’ design principle prohibits that a node obtains more than two roles. Thus, the upper part of the role-state diagram depicted in Figure 6 describes how regular nodes participate in scatternet formation [3]. An example of a role-transition matrix is shown in Figure 3. It describes the transitions of the hierarchical TSF algorithm. 'Before' refers to a node's state before interconnection, while 'after' refers to the resulting state. Note that if a node disconnects, on the other hand, the state transition will go in the opposite direction. Figure 3: Role-transition matrix for TSF. A more detailed matrix is provided in [3]. Since MSF, unlike TSF, is not bound to the hierarchical approach, a master is allowed to connect to free nodes, slaves or master-slave bridges, resulting in highly meshed scatternets. The latter connections always produce master-slave bridges, and no slave-slave bridges. The role-transition matrix of MSF is shown in Figure 4. It differs from that of TSF in role-transitions 3, 8, 11, 12, 13 and 23. Figure 4. Role-transition matrix for the MSF algorithm, which generates only ms-bridges. The matrix is simplified by omitting transitions with an ss-bridge in the initial state, since MSF results in scatternets without ssbridges. SSF - in contrast to TSF and MSF - never produces master-slave bridges in role-transitions 2, 6, 8 and 12. This restriction means that unlike in MSF, masters in SSF have no way of connecting to other masters or master-slaves, but may connect to slaves. In this case, the connected master remains a master, while the slave connected to becomes a slave-slave bridge. The role-transition matrix of SSF is shown in Figure 5. Figure 5. Role-transition matrix for the SSF algorithm, which generates only ss-bridges. The matrix has been simplified by omitting transitions with an ms-bridge or an ss-bridge in the initial state. How MSF and SSF are realized by using INQUIRY/INQUIRY-SCAN and PAGE/PAGE-SCAN is outlined in [3]. How TSF is realized is detailed in [2]. 3. A FRAMEWORK FOR HETEROGENEOUS DEVICES To explore the ability of a scatternet formation algorithm to accommodate heterogeneous devices, we assumed that a predefined share of the devices were of low capabilities, and would only attach to scatternets as slaves. Compared to the framework in [3], we introduced two new roles; Low-capability Free node (lf) and Low-capability Slave (ls). A low-capability device starts out in the “Lowcapability Free” role-state (lf), while a high-capability computing device, starts out in the regular free role state F. When a Low-capability Free node lf encounters and connects to another node, it can only assume the role of a Low-capability Slave ls. A Low-capability Slave cannot connect to other devices. The lower part of the role-state diagram in Figure 6 illustrates how low-capability devices participate in scatternet formation. Figure 6: Role-state diagram for scatternet formation. (x represents the role of the node connected to, i.e. x∈{F, M, M', S, ms, ms', ss, lf, ls}. With two new roles, the role transition matrices in Figure 1, 2 and 3 must be expanded with additional transitions. These additional transitions are shown in Figure 7. Figure 7. Expanding the role-transition matrices by adding additional transitions. For simplicity transitions with a Low-capability Slave (ls) or a slave-slave (ss) in the initial state have been omitted. Transition a of Figure 7 means that two Lowcapability Free nodes cannot interconnect because both devices require to become slave of a connection. A Lowcapability Free node lf may, however, connect to a regular high-capability node, and become a slave. Transitions b, c, d and e show that the regular node connected to must become master of the new connection. SSF does not generate master-slave connections, and for this algorithm lf-nodes are thus not allowed to connect to regular slaves (transition c). For MSF and TSF, on the other hand, lfnodes can connect to regular slaves, which become masterslaves after the connection is set up. Hence, with the link formation scheme proposed for SSF and MSF in [3], lf-nodes will always go in INQUIRYSCAN, and will not alternate between INQUIRY and INQUIRY-SCAN like regular free nodes. 4. SIMULATIONS 4.1. Simulation setup By simulations we placed Bluetooth devices in a 30mby-30m square. For simplicity, a node was allowed to communicate with all other nodes located within 10m. This is represented by arcs in the visibility graph as defined in [4] It was not allowed to communicate with any nodes outside this range. Nodes were placed in an arbitrary position inside the square, and they arrived one by one to a maximum of 40 nodes. Each arriving node was categorized either as a lowcapability node by a pre-determined probability P or as a regular node (by probability 1-P). The arriving node interacted with other nodes and scatternets already formed within the square, while obeying the rules of the scatternet formation algorithm simulated. The resulting scatternet configuration was used as a starting point when the next node arrived in the square. This simulation run was repeated 100 times. Simulations were done using Matlab. Upon the arrival of a node in the square, the arriving node detected and started connecting to an arbitrary of its neighboring nodes in accordance with the scatternet formation algorithm. It continued detecting and connecting to neighbors until it had tried connecting to all of its neighbors. 4.2. Evaluation metrics A number of different evaluation metrics can be used to assess the capability of an algorithm to accommodate heterogeneous devices [3]. Since the purpose of scatternet formation is to ensure connectivity for communication between devices, an overall goal is to maximize connectivity. Connectivity is a particularly significant evaluation parameter with the presence of low-capability devices, because such devices may easily become disconnected (e.g. if there are no masters or master-slaves to connect to). To maximize connectivity, an algorithm should interconnect a given set of nodes into as few disconnected scatternets, N, as possible. It is evident that the theoretical lower limit, N0,, of the number of disconnected scatternets formed by any algorithm, is found by traversing the visibility graph, as defined in [4], instead of traversing the links formed by the scatternet algorithm. Hence, the connectivity-ratio, rconn, defined as rconn = N/N0 (1) gives a good measure of the overall connectivity provided by the scatternet formation algorithm. The same simulation configuration was run for all algorithms under study, and the average connectivity ratio was calculated for each algorithm and for each number of nodes in the square. A number of researchers emphasize that a low average shortest path (ASP) of a scatternet indicates an efficient scatternet topology ([3], [8]). The ASP-ratio is defined as the ASP of a scatternet relative to the minimal ASP, ASP0: rasp = ASP/ASP0 devices, a fully meshed slave-slave based algorithm (like SSF) is preferable. The connectivity performance of SSF is comparable with that of MSF for low node densities (i.e. few nodes in Figure 8), and better for higher node densities. (The advantages of a slave-slave-based approach as compared to master-slave based algorithms are explained further in [3].) TSF had considerably worse connectivity ratio than both SSF and MSF, due to the hierarchical nature of TSF. We continued simulating scenarios with non-zero probability P that an arriving node is a low-capability device. Figure 9 and 10 show the result for P = .50 and P = .75, respectively. (2) ASP is found by traversing the links formed by the scatternet algorithm between the nodes only within one disconnected scatternet. ASP0 is found by traversing the links of the visibility graph between the same set of nodes. 4.3. Simulation Results We performed a number of simulations scenarios, each scenario characterized with a given probability, P, that an arriving node would be a low-capability device. For each scenario, we plotted the connectivity ratio to see how the presence of low-capability devices influences scatternet connectivity. Figure 9. P = .50: With 50% low-capability devices, the slave-slave based approach (SSF) still provides better connectivity than the master-slave based approach (MSF). TSF is hardly capable to accommodate heterogeneous devices due to its hierarchical nature. Figure 8. P = 0: With only homogeneous devices, slave-slave based SSF has better connectivity performance as compared to master-slave-based MSF and TSF. (TSF is worse than MSF, mainly due to the hierarchical tree structure that TSF requires.) First we simulated a scenario with a probability of zero that an arriving node would be a low-capability device. This scenario corresponds to scatternet formation with homogenous Bluetooth devices [3]. Indeed, Figure 8 shows that the connectivity ratio is farther from the theoretical optimal connectivity ratio of one. With no low-capability Figure 10. P = .75: In the extreme case with 75% low-capability devices, the slave-slave based approach (SSF) outperforms the master-slave based approaches (MSF and TSF). With an average 50% share of devices being of low capability (Figure 9), the connectivity of MSF deteriorates faster than that of SFF for low node densities (i.e. for few number of nodes in Figure 9). For higher node densities SSF deteriorates relatively faster than MSF, but not so much that MSF is preferable. In total, SSF performs better than MSF. We also observe that the strictly hierarchical approach, TSF, is hardly capable of accommodating heterogeneous devices, because a device that cannot transform from slave to master-slave will clog the hierarchical formation process. As an extreme scenario, we tested an average share of low-capability devices of 75% (Figure 10). We observed that the performance of the meshed master-slave-based MSF algorithm breaks down, while SSF on the contrary performs surprisingly well. We also found that with heterogeneous devices (P = 50%) there are no significant differences between the average shortest path performances of the three algorithms (Figure 11). formation algorithm must create topologies that easily accommodate connectivity for such devices. This is easily measured by the connectivity ratio [3]. We simulated three proposed algorithms for Bluetooth scatternet formation in different scenarios. Each scenario had a different average share of low-capability devices. TSF had little ability to accommodate low-capability devices, because a low-capability device will clog the hierarchical scatternet formation. The non-hierarchical algorithms, on the other hand, can be highly meshed, and rules to ensure a hierarchical structure are not present. Hence, the non-hierarchical algorithms showed considerably higher ability to accommodate low-capability devices. Of the non-hierarchal algorithms, the slave-slave based scatternet formation scheme, SSF, outperformed the scheme based on master-slave bridges, MSF. With no lowcapability devices, the SSF algorithm was preferable. When we then added low-capability devices to the system, the connectivity of MSF deteriorated faster than that of SFF, and MSF gradually broke down. SSF on the other hand, performed surprisingly well. Hence, out of the three algorithms we studied, the nonhierarchical, fully meshed slave-slave based approach to scatternet formation, SSF, were the only algorithm that was capable of accommodating scatternet formation of heterogeneous devices with acceptable performance. REFERENCES [1] [2] [3] [4] [5] Figure 11. P = .50: There are few significant differences between TSF, SSF and MSF with respect to the ASP-ratio. 5. SUMMARY AND CONCLUSIONS To our knowledge no work has yet been published on scatternet formation with respect to heterogeneous devices, although Bluetooth will be incorporated into devices of widely different capabilities. We have therefore developed a framework to investigate scatternet formation with heterogeneous devices, and performed simulations based on this framework. The presence of low-capability devices will influence the scatternet performance. When low-capability devices will only participate as slaves in scatternets, the scatternet [6] [7] [8] [9] “Specification of the Bluetooth System,” Bluetooth Special Interest Group document, http://www.bluetooth.com/, Dec. 1999. Tan et al., ”Forming Scatternets from Bluetooth Personal Area Networks”, MIT Technical Report, MIT-LCS-TR-826, http://nms.lcs.mit.edu/projects/blueware/body.htm, October 2001. Engelstad et al., ”Asynchronous formation of non-hierarchical Bluetooth scatternets“, to appear in proceedings of 3G Wireless Conference 2003, In Press. Miklos et al., ”Performance Aspects of Bluetooth Scatternet Formation”, MobiHoc 2000, pp147-148, June 2001. Johansson et al., ”Short-range radio based ad-hoc networking: performance and properties”. In Proceedings of the IEEE International Conference on Communications 1999, vol. 3, pp.1414-1420, 1999. Salonidis et al. ”Proximity Awareness and Ad Hoc Network Establishment in Bluetooth”, Technical Research Report, http://www.isr.umd.edu/CSHCN, 2001. Zaruba, G.V., Basagni, S., Chlamtac, I., “Bluetrees – Scatternet Formation to Enable Bluetooth-Based Ad Hoc Networks”, In Proceedings of the IEEE International Conference on Communications (ICC) 2001, Helsinki, Finland, June 2001. Wang, Z., Thomas, R.J., Haas, Z., “Bluenet – a New Scatternet Formation Scheme”, 35th Hawaii International Conference on System Science (HICSS-35), Big Island Hawaii, January 7-10, 2002. Law, C., Mehta, A.K., Sui, K. -Y., “Performance of a New Bluetooth Scatternet Formation Protocol”, In Proc. of the ACM Symposium on Mobile Ad Hoc Networking and Computing (MobiHoc) 2001, Long Beach, CA, USA, Oct. 2001.