A Two-hase Scatternet Formation rotocol for luetooth Wireless ersonal Area Networks Yoji Kawamoto 1, Vincent W.S. Wong 2, and Victor C.. Leung 2 1 Network & Software Technology Center, Sony Corporation, Tokyo, Japan 2 Department of Electrical and Computer Engineering, The University of ritish Columbia, Vancouver, C, Canada E-mail: kawamoto@sm.sony.co.jp, vincentw@ece.ubc.ca, vleung@ece.ubc.ca Abstract luetooth is a promising short-range radio technology for wireless personal area networks. There is an interest to expand the coverage of such networks by interconnecting them to form scatternets. ost of the scatternet formation algorithms recently proposed in the literature do not support dynamic topology changes, and the master or bridge nodes in the resulting scatternet may become the traffic bottleneck and reduce network throughput. In this paper, we propose a Two-hase Scatternet Formation (TSF) protocol to support dynamic topology changes while maintaining a high aggregate throughput. In the first phase, a Control Scatternet is constructed to support topology changes and route determination. The second phase creates a separate On-demand Scatternet whenever a node wants to initiate data communications with another node. The On-demand Scatternet is torn down when the data transmissions are finished. Since all the time slots in an On-demand Scatternet are dedicated to a single communication session, a high aggregate throughput is achieved at the expense of a slightly higher connection setup delay. I. INTRODUCTION luetooth [1] is a short-range radio technology that uses the Frequency Hopping Spread Spectrum (FHSS) technique with Time-Division Duplex (TDD) in the license-free 2.4 GHz Industrial, Scientific and edical (IS) band. Due to its low cost and low power, luetooth is considered a promising technology for wireless personal area networks. Various consumer electronics devices (e.g., laptop computers, DAs, digital cameras) have begun to provide luetooth wireless communication capability. luetooth devices are required to form a piconet before exchanging data. Each piconet has a master unit that controls the channel access and frequency hopping sequence. Other nodes in the piconet are referred to as the slave units. In a luetooth piconet, the master node can control up to seven active slaves simultaneously. Several piconets can be interconnected via bridge nodes to create a scatternet. ridge nodes are capable of timesharing between multiple piconets, receiving packets from one piconet and forwarding them to another. Over the past few years, a number of scatternet formation algorithms have been proposed in the literature. The luetooth Topology Construction rotocol (TC) [2] consists of three phases: coordinator election, role determination, and actual connection establishment. The protocol assumes that all the devices are within the same radio range. In the luetree protocol [3], a piconet is first constructed by a coordinator called the lueroot. It is followed by piconet creation with other slaves. A rooted spanning tree called the luetree is finally constructed. The luenet protocol [4] is a distributed protocol that does not require any coordinator, and has a better performance when compared with luetree. Note that none of these protocols [2]-[4] supports dynamic topology changes (i.e., dynamic join/leave by any device after the scatternet has been constructed). The protocol proposed by Law and Siu [5] supports dynamic join by introducing the SEEK, SCAN, and CONNECTED procedures. In the current luetooth specification version 1.1 [1], all the packet transmissions need to be routed via the master or bridge nodes. When the traffic load is high, these master or bridge nodes may become the network bottleneck. Furthermore, in order to support dynamic join/leave of mobile devices within a scatternet, a number of time slots in the master nodes may need to be allocated for the dynamic topology updates. This further reduces the time slots available at the master nodes for data transmissions. In this paper, we propose a Two-hase Scatternet Formation (TSF) protocol with the aim of supporting dynamic topology changes while maintaining a high aggregate throughput. In the first phase, a Control Scatternet is constructed for control purposes (i.e., to support dynamic join/leave, route discovery, etc). The second phase is invoked whenever a node needs to initiate data communications with another node. A dedicated piconet/scatternet is constructed on-demand between the communicating nodes. Since the On-demand Scatternet can dedicate all the time slots to a single communication session, it has the capability to provide a high throughput and a small end-to-end data transfer delay. The On-demand Scatternet is torn down when the data transmissions are finished. This paper is organized as follows. Section II presents the Control Scatternet formation and scheduling algorithms in phase 1. Section III describes the phase 2 procedures to construct the On-demand Scatternet. erformance analysis and simulation results are presented in Section IV. Conclusions are given in Section V. II. HASE 1: CONTROL SCATTERNET FORATION A. Scatternet Formation Algorithm One of our design goals is to minimize the number of piconets in the Control Scatternet. Since the Control 0-7803-7700-1/03/$17.00 (C) 2003 IEEE 1453
Table 1 Extension of ID packet for node i Elements Symbol bits Number of G i 6 0 63 Neighbours Role R i 3 {UNDEFINED, ASTER, SLAVE, RIDGE2,, RIDGE6} etric value V i 12 Node density at the vicinity of a node ode D i 1 {DISCONNECTED, CONNECTED} Unique ID U i 48 luetooth Device (D) Address Scatternet is used for the exchange of signaling and control information, the number of piconets can be minimized by removing the limitation of the number of slaves in a piconet. This can be achieved by putting the slave nodes into the park mode [1]. In this section, we begin by presenting the distributed Control Scatternet formation algorithm. It is followed by introducing a new scheduling algorithm to support dynamic topology changes and arbitrary number of slaves in the piconet. The formation of the Control Scatternet is divided into three periods during which (1) the nodes discover their neighbours with whom they exchange status information, (2) the nodes congregate into piconets by electing a master among the nodes within the radio range of each others, and (3) the piconets are connected to form the Control Scatternet. During periods 1 and 2, which durations are T 0 and T 1, respectively, each node senses for adjacent nodes by alternating between INQUIRY and INQUIRY SCAN states randomly. We assume that the values of T 0 and T 1 are greater than 5 sec. Nodes exchange neighbour information via the ID packets [1] in the inquiry messages. We propose an extension of the ID packet to include the following information (see Table 1). Node i has (G i, R i, V i, D i, U i ) as its neighbour information where G i denotes the number of neighbours, R i denotes the role, V i denotes the metric value, D i denotes the mode, and U i denotes the unique ID. To initialize the algorithm, node i sets G i = V i = 0, R i =UNDEFINED, D i = DISCONNECTED, and U i = D (luetooth Device) address. Each node maintains a database, called the neighbour information base (NI), of the information of each of its neighbours. The NI is initially empty. Note that we limit the number of neighbours to 63 nodes as shown in Table 1 in order to reduce the size of an inquiry message. The distributed Control Scatternet formation algorithm described below provides the procedures to exchange neighbour information and update the NI. eriod 1. 0 t < T 0, Sensing Neighbours During this period, each node alternates between INQUIRY and INQUIRY SCAN states randomly. Each node updates its NI based on neighbour information received during the INQUIRY SCAN state by creating a new entry for the neighbour if one does not exist in the NI, or updating the existing entry in the NI with the new neighbour information. To enter the INQUIRY state, each node calculates its own neighbour information based on the entries in its NI. G i is set to the number of entries in the NI. We define the metric value V i as the sum of the neighbours G values. Let F i denote the set of currently known neighbours of node i for which entries exist in the NI of node i. Then the current value of metric V i is: V =, 1 i N (1) i G j j F i where N denotes the number of nodes in the network. For instance, if a node has three neighbours (G i = 3) which G j are 5, 6, and 6, respectively, then V i = 17 from (1). Since V i /G i gives the average connectivity of the neighbours of node i, the metric value can indicate whether the node s neighbourhood is sparse or dense. To maximize the piconet coverage in the Control Scatternet, this metric is used during the master node selection. After receiving an ID packet, the node starts the Random ack-off Timer (RT) and sends its FHS packet [1] to the source node after the expiration of the RT. Since a node may receive several ID packets, it prepares several RTs simultaneously to distinguish between the source nodes. eriod 2. T 0 t < T 1, Election of aster Nodes After the expiration of T 0, each node i, whose R i = UNDEFINED, tries to decide its own role initially based on the following rules: Rule R0: Node i keeps R i = UNDEFINED if there exists a node j F i such that D j = CONNECTED. Otherwise, go to rule R1. This is the case when a Control Scatternet has already been constructed and node i tries to join this Control Scatternet. Node i will switch to the AGE SCAN state and wait for the AGE message from the master node. Rule R1: Node i sets R i = SLAVE if there exists one node j F i such that R j = ASTER; or R i = RIDGEn if there exists n nodes j F i such that R j = ASTER; Otherwise, go to rule R2. For instance, if node i receives neighbour information indicating that three neighbours have assumed ASTER role, it sets R i = RIDGE3. The value of n represents the node degree of the bridge node, i.e., the number of masters that the bridge node may be connected to. Rule R2: Node i sets R i = ASTER if, for all j F i, R j = UNDEFINED, and one of the conditions is true: (a) G i > G j, (b) V i < V k for all k F i and G i = G k, (c) U i < U k for all k F i and G i = G k and V i = V k. 1454
After the above role decision procedure, each node continues to exchange neighbour information with its neighbours and updates its own NI as in period 1. ased on the received neighbour information, node i modifies its role as follows. Rule R3: If R i = ASTER, then set R i = SLAVE if there exists node j F i such that R j = ASTER and U j < U i. iconet#1 iconet#2 S R SLEE S S R SLEE S 1 roadcast Fixed Interval m oll1,2 Ack1,2 olln-1,n AckN-1,N Access Window This election procedure is important in period 2 as it avoids the presence of several masters within radio range of each other due to miscalculation in the algorithm. iscalculations can occur due to nodes not starting their algorithms at the same time or the loss of neighbour information due to transmission errors. Rule R4: If R i ASTER and R j ASTER for all nodes j F i over some time in period 2, then repeat master election procedure using rule R2 for role determination. eriod 3. t > T 1, Connection of iconets into Scatternet In this period, nodes that have been elected ASTER start connecting to the neighbouring nodes using the page procedure [1]. Other nodes change to the AGE SCAN state to connect to a master. A node changes its mode to CONNECTED when it is connected to another node. efore the connection procedure, each master node sorts its neighbours into those assigned SLAVE role and those assigned RIDGE role, and further sorts the bridge nodes according to their node degree n and then according to their unique IDs. The master starts connecting the slave nodes first, and then the bridge nodes from the highest node degree to the smallest, and then from the smallest unique ID among the bridge nodes with the same degree n. A bridge node with the highest degree and the smallest unique ID may be connected to several masters, and it is considered by surrounding masters as the preferred bridge node for the Control Scatternet. If a master detects duplicated bridging with adjacent piconets, it will detach the duplicated bridge node(s) using the L command [1]. This ensures that within a Control Scatternet, only one bridge node exists between two piconets. Recall that in order to maximize the coverage and minimize the number of the piconets, we allow each master to connect to more than seven slaves. To enable this condition, a master allocates active mode (am) addresses only to bridge nodes. Other slaves, called pure slaves, return the am address after connecting to the master, and then switch into park mode [1]. Consequently, we limit the number of bridge nodes in a piconet to 6 to leave at least one am-address for assignment to pure slaves on demand. After the establishment of a piconet, the master sends all of its slave and bridge nodes neighbour information to adjacent piconets, and broadcasts neighbour information received from adjacent piconets to all nodes in its own piconet, in order to update the neighbour information in each node. This information is used for the On-demand Scatternet formation in phase 2. Fig. 1. Scheduling in a piconet. Scheduling in the Control Scatternet In association with our Control Scatternet formation algorithm, we propose a new time slot scheduling mechanism as shown in Fig. 1 to accommodate an arbitrary number of slaves in one piconet and to enable the support of dynamic changes of the network topology. The time slots within a given fixed interval are divided into three periods, namely pure slaves (S), bridge nodes (R), and sleep (SLEE) periods. The R periods are synchronized between piconets to make the Control Scatternet more efficient. Synchronization of the R period among all piconets is achieved by exchanging scheduling information between masters when a bridge node is connected. In the S periods, pure slaves are controlled by their masters in the same manner as park mode terminals [1]. The length of the S period in each piconet depends on the number of slaves in the piconet. ure slaves listen for the broadcast messages by which the master assigns at least one access window for them to access the piconet; however, they do not need to wake up periodically as described in [1]. If a pure slave needs to communicate with another node, it will make a request to the master using the access window. Then the master allocates an am-address to the pure slave so that it can send control messages to setup an On-demand Scatternet. The duration from the end of a S to the beginning of the next R period is used for such control communications. In the luetooth specification [1], bridge nodes use the hold request to inform the master about its timing to switch over to another piconet. This method increases the bandwidth and power usage because the bridge node has to inform the master every time it switches over to another piconet. y synchronizing the R period between piconets and assigning specific time slots to the bridge nodes in each R period, the bridge nodes can switch between piconets at pre-defined time slots without informing the masters about the timing of each switch over. C. Support of Topology Changes aster nodes can use the SLEE period either to accept new nodes, or to allocate additional time slots for communications, or to sense for adjacent nodes. ridge nodes can also use the SLEE period to sense for adjacent nodes. ure slaves, on the other hand, continue to sense for adjacent nodes as long as possible even when they are within their respective S periods as mentioned earlier. 1455
When a pure slave senses a new node, it forwards the new neighbour information including the new node s clock and D address to its master using the access window. A new node within the coverage area of the Control Scatternet will learn that there are one or more adjacent nodes already in the CONNECTED mode when it goes through the Control Scatternet formation periods 1 and 2. In this case the new node changes its state to AGE SCAN during the formation period 2 according to rule R0 above. The master node pages the new node during the SLEE period to connect the new node to its piconet. If the new node fails to connect to a master after the expiration of T 1, it repeats period 2 and rules (R1-4) without changing its state to AGE SCAN. This occurs if the new node is out of range of surrounding masters and needs to form a new piconet. Whenever a pure slave finds a new ASTER among its neighbours, it goes to AGE SCAN state to connect with the new master. ure slaves that become connected to two masters change their role to RIDGE2 accordingly. When a master node leaves the Control Scatternet, it will choose a new master node from those in its NI. When a bridge node leaves the Control Scatternet, it will inform its master nodes, which will then choose another bridge node from those in their NIs. III. HASE 2: ON-DEAND SCATTERNET FORATION The second phase is invoked whenever a node wants to initiate data communications with another node. A new scatternet or piconet is constructed on-demand between the communicating nodes and torn down when the data transmissions are finished. In this section, we describe the procedures to create the On-demand Scatternet. Recall that after the Control Scatternet formation in the first phase, each master node maintains all the information of its slaves and bridges nodes within its piconet and adjacent piconets. This information is exploited during the second phase. The route selection mechanism for the On-demand Scatternet can be based on any on-demand source routing protocols proposed for wireless mobile ad hoc networks. In this paper, we use the dynamic source routing (DSR) protocol [6] for route selection. An example is used to illustrate the procedures (see Fig. 2). In Fig. 2(a), there are eleven nodes distributed within an area. After the first phase, a Control Scatternet is formed which consists of two piconets connected by a bridge node (see Fig. 2(b)). Now, we assume that a source node S wants to initiate data communications with destination node D (see Fig. 2(c)). The On-demand Scatternet procedures include the following steps: Step 1: Route Selection based on DSR [6] (i) The source node S sends a Route Request (RREQ) message to its master node. (ii) If master node has the destination s route in its routing cache, it sends a Route Reply (RRE) message to the a S c d a. distributed nodes without connections b. The Control Scatternet is constructed c. RREQ/RRE/REQ/RE are exchanged d. On-demand scatternet is constructed Fig. 2. D m source node S. Otherwise, the master node sends a RREQ to all of its bridge nodes by adding its D address in the RREQ message. (iii) When bridge node receives a RREQ, it forwards the message to other master nodes. Note that whenever a bridge or master node receives a RREQ, it checks the request ID and discards the message if it indicates a duplicate [6]. In Fig. 2(c), bridge node forwards the RREQ message to master node m. Note that bridge node does not need to add its D address into the RREQ since each master knows neighbour piconets connection as a result of the construction of the Control Scatternet. (iv) If a master node has the destination node information, it will send a RRE on the reverse path instead of forwarding the RREQ to the destination. The RRE includes piconet route information which is represented by masters D address, and it is preserved in each master s route cache. In Fig. 2(c), master node m sends the RRE via bridge node to master node. (v) Finally, the master node (to which the source node is connected) receives the RRE. This RRE includes the piconet route. The master node forwards this RRE message to the source node. In Fig. 2(c), the solid arrow lines denote the route of this source/destination pair. Step 2: articipating Nodes Selection (i) The master node, which the source node communicates with, begins to select the participating nodes for the On-demand Scatternet. The master node uses its own slaves and adjacent piconets neighbour information to choose the participating nodes by following chosen piconet route. The master node sends a ath Request (REQ) message to the bridge node via the route determined in Step 1. The REQ includes the same b S D m, m: master node : bridge node : pure slave nodes S: source node D: destination node A Two-hase Scatternet Formation rotocol 1456
piconet route and the participating nodes D address and clock value. In Fig. 2(c), the REQ follows via the solid arrow line. (ii) All the intermediate master nodes along the path also select the participating nodes using the same method as (i) when they receive the REQ. Note that a master may be able to choose another participating node when it detects that the selected node is currently a member of another On-demand Scatternet. (iii) The master of the destination node sends a ath Reply (RE) message along the reverse path and signals the participating nodes it has chosen to start building the On-demand Scatternet. (iv) The destination node changes its state to AGE and acts as master to connect to its neighbour selected to participate in the current On-demand Scatternet. Each participating node starts in the AGE SCAN state to connect as a slave to the previous participating node along the reverse path. It then changes to the AGE state and acts as master to connect to the next node. (v) Each node in the On-demand Scatternet synchronizes its timing with the destination node. Finally, the On-demand Scatternet is created. The destination node becomes a master node and the source node becomes a slave node in the On-demand Scatternet. All other nodes in the On-demand Scatternet are master/slave nodes bridging two piconets. In Fig. 2(d), nodes S,, and D form an On-demand Scatternet (shaded). Note that each node in the On-demand Scatternet continues to maintain its connection to the Control Scatternet. Each master node in the Control Scatternet maintains information about which of its slave or bridge nodes are currently participating in an On-demand Scatternet. The On-demand Scatternet is torn down when the data transmissions are finished. Each participating node informs its corresponding master that it is available for future On-demand Scatternet formation. IV. ERFORANCE ANALYSIS In this section, we first present an estimation of the connection setup delay for the On-demand Scatternet. We then compare the performance between our proposed scheme and TC [2] in terms of the aggregate throughput. A. Connection Setup Delay A piconet/scatternet has to be created on-demand whenever two devices within a Control Scatternet want to exchange data. This means that a connection setup delay is always incurred whenever an On-demand Scatternet is built. Since the Control Scatternet knows each participant s clock and device address, it can start from the page procedure. In general, there are 32 frequencies in the page process [1], and the average time of matching the paging frequency between a paging and paged node is 10 ms if we assume this period is uniformly distributed between [0, 20 ms] [2]. Suppose there are K participants (including the source and destination nodes) in the On-demand Scatternet, the average time of the paging process is (K - 1) 10 ms. After matching the paging frequency, there are several negotiations such as synchronization and link channel setup between the master/slave nodes. We assume that this negotiation period is 20 slots (12.5 ms). Thus, the connection setup delay T C is: T C = 10 ( K 1) + 12.5 ( K 1) = 22.5 ( K 1) ms (2) For instance, suppose the number of nodes K is equal to 10 in the On-demand Scatternet. Then the connection setup delay is equal to 202.5 ms. In view of this delay, small data packets such as address resolution and the service discovery packets could be sent over the Control Scatternet rather than building an On-demand Scatternet for their transmissions. roadcast packets could also be sent in the Control Scatternet because it supports broadcasting inherently.. Comparison with TC We compare the performance between our proposed two-phase scatternet formation (TSF) protocol and the TC [2] via simulations. For TC proposed in [2], all the nodes have to be within the same radio range and the maximum number of nodes is equal to 36. For fair performance comparison, we assume that in both TSF and TC, the number of nodes in the network is equal to 36 and all the nodes are within the same coverage. With 36 nodes, there are 8 piconets and 28 bridge nodes within the scatternet created by TC. These piconets are connected to each other by bridge nodes. This gives a maximum hop count of 4 for any communication path. On the other hand, one piconet is created in the Control iconet after the first phase in our proposed TSF protocol. For both TSF and TC, we assume that asymmetric DH5 packets are used for data communication and DH1 packets are used for signaling/control messages. For TC, the theoretical maximum aggregate throughput is via a single hop communication, such that all connections are between a master and a slave node. Since there are eight piconets within the scatternet, the theoretical maximum throughput is equal to 723.2 kbps 8 = 5.7856 bps. For TSF protocol, the theoretical maximum throughput corresponds to each node being dedicated to its own communication session. Since there are one master node and 35 pure slaves in the Control Scatternet, there can be at most 35/2 = 17 On-demand piconets where each one is dedicated to only one communication session. The theoretical maximum aggregate throughput is equal to 723.2 kbps 17 = 12.2944 bps. In our simulation model, we randomly choose pairs of source and destination nodes and determine the average aggregate throughput. We allow a node to serve more than one On-demand iconet and assume that each On-demand piconet is synchronized with each other. This assumption is feasible by exploiting the master s clock value in the Control iconet 1457
Aggregate Throughput (b/s) 12 11 10 9 8 7 6 5 4 3 2 1 0 TSF without acket Collision TSF with acket Collision TC 0 10 20 30 40 50 60 Number of Connections Fig. 3. Throughput comparison between TC and TSF for each On-demand iconet. The simulation time is 10 5 time slots where each slot corresponds to 625 µs. Each point is averaged over 1000 simulation runs. ackets from different piconets will collide if they use the same frequency slot during their transmissions. To account for this fact, we also include the modeling of packet collisions in our simulations. Figure 3 shows the average aggregate throughput versus the number of data connections. In Figure 3, we use the phrase TSF with(out) packet collision to indicate whether or not the modeling of packet collisions is included in the simulation. Simulations results indicate that our proposed TSF protocol has a higher average aggregate throughput when compared with TC. The throughput increases when the number of connections increases. Note that the chances of packet collisions increase when the number of On-demand iconets increases. Since our proposed TSF protocol constructs an On-demand piconet for each data connection while TC uses the existing scatternet for data and control transmissions, we expect that the average end-to-end delay of our proposed TSF protocol is lower than TC. In both cases, the average end-to-end delay increases when the number of connections increases. For TC, the increase in delay is due to the sharing of time slots with other connections within the piconet. For TSF, the increase in delay is due to both the sharing of time slots and the increase in occurrence of packet collisions. Our current work includes the packet collision avoidance mechanism for the TSF protocol. Collision avoidance may be feasible by either modifying the hopping frequency or limiting the number of the On-demand piconets within the Control Scatternet. However, such modification may have significant impact on the design of the luetooth modules. Note that the co-existence mechanism between luetooth and WLAN [7] is currently under-study within the IEEE 802.15 working group [8]. Some of the proposals in [8] may be used to reduce the packet collisions in our proposed TSF protocol. V. CONCLUSIONS In this paper we have proposed a novel two-phase scatternet formation (TSF) protocol to improve the communication efficiency of luetooth wireless personal area networks while supporting dynamic changes in network topology. In the first phase, a Control Scatternet is constructed which is used for control and signaling purposes. The second phase creates an On-demand Scatternet whenever a node wants to exchange data with other nodes. We have presented simulation results to show that our proposed TSF protocol achieves a higher average aggregate throughput when compared with TC. ACKNOWLEDGEENTS The authors would like to thank Xiuying Nie for the comments on an earlier draft of this paper. Yoji Kawamoto would like to thank Dr. C. K. Toh and Sony Corporation for giving him an opportunity to spend one year at the University of ritish Columbia. REFERENCES [1] luetooth Special Interest Group, Specification of the luetooth System, Core, version1.1, http://www.bluetooth.org/specifications.htm [2] T. Salonidis,. hagwat, L. Tassiulas, and R. Laaire, Distributed Topology Construction of luetooth ersonal Area Networks, in roc. IEEE INFOCO 01, Anchorage, Alaska, April 2001. [3] G. V. Zaruba, S. asagni, and I. Chlamtac, luetrees Scatternet Formation to Enable luetooth-ased Ad Hoc Networks, in roc. IEEE ICC 01, Helsinki, Finland, June 2001. [4] Z. Wang, R. J. Thomas, Z. Haas, luenet a New Scatternet Formation Scheme, in roc. 35th Hawaii International Conference on System Science (HICSS-35), ig Island, Hawaii, January 2002. [5] C. Law and K.-Y. Siu, A luetooth Scatternet Formation Algorithm, in roc. IEEE Symposium on Ad Hoc Wireless Networks 2001, San Antonio, Texas, November 2001. [6] D.. Johnson, D. A. altz, Y.-C. Hu, and J. G. Jetcheva, Dynamic Source Routing rotocol for obile Ad hoc Networks, February 2002. IETF Internet Draft (work in progress). [7] The editors of IEEE 802.11, IEEE Std 802.11b-1999, 1999. [8] IEEE 802.15 WAN TASK GROU 2 (TG2), http://www.ieee802.org/15/pub/tg2.html 1458