Cross-layer Optimized Routing for Bluetooth Personal Area Network

Transcription

1 Cross-layer Optimized Routing for Bluetooth Personal Area Network Leping Huang, Hongyuan Chen, T.V.L.N Sivakumar and Kaoru Sezaki Nokia Research Center Tokyo; Nagata-cho, Chiyoda-ku, Tokyo, Japan Institute of Industrial Science, University of Tokyo, Komaba, Meguro-ku, Tokyo, Japan, Abstract In this paper, we present some observations and analysis on Bluetooth PAN s performance degradation in multi hop network based on our experiments. We highlight a situation in which control packets can be received properly but not data packets, and its affects on Scatternet performance. Based on our analysis, we propose a Cross-layer Optimized Routing protocol for Bluetooth (CORB), which outperforms AODV based routing protocols. CORB is a QoS-extended AODV routing protocol that is optimized for Bluetooth MAC. It has two unique characteristics. First, it uses a new load metric (LM) in QoS routing protocol instead of number of hops as in conventional best effort routing. LM reflects nodes link bandwidth with respect to Bluetooth nodes role in the Scatternet. This helps CORB to bypass heavily loaded nodes, and find routes with larger bandwidth. Second, LM and some MAC layer parameters are dynamically adjusted in response to the changes in radio conditions. These two characteristics of CORB contribute to its improved stability, and rapid response to changing radio conditions. I. INTRODUCTION Bluetooth[1] is a promising wireless technology that enables portable devices to form short-range wireless network. It is assumed that many kinds of device such as home appliance, office electronics will be connected with each other through Bluetooth in the near future. The basic unit of Bluetooth network, Piconet, can only connect up to 8 nodes. With the proliferation of BT enabled devices, the demand to connect more than 8 devices increases. And also due to the limited capabilities of the devices, and the security concerns, it may not be possible to have Piconets of full capacity. Futher, it is possible that some nodes may not prefer to take the role of master or slave. Hence it is a natural requirement to develop technology that can connect multiple piconets to form a Scatternet while honoring the limitations of the nodes even if the most possible use case scenarios do not predict connectivity among more than 8 devices. Scatternet is an interconnected group of piconets. The node that connects multiple piconets is called PMP (Participant in Multiple Piconet) in Bluetooth specification. There are two types of PMP nodes. The node that attends multiple piconets simultaneously only as slave is called as S/S PMP, and the node that attends multiple piconets simultaneously, and has a master role in one of the piconets, is called M/S PMP in this paper. The methods regarding how to form a multi-hop network (which is called Network Formation), and how to route packet within scatternet (which is called PAN routing) are still under discussion at Bluetooth SIG[1]. The existing ad hoc routing proposals do not consider the specific features of Bluetooth scatternet, making the proposals inefficiently use the resources provided by Bluetooth MAC layer. In order to find routing protocols that are well suited for Bluetooth based PAN networks, at NOKIA Research Center, we have developed a test bed. From the experimental results obtained by running AODV on a Bluetooth Scatternet, we noticed that the system performance is greatly influenced by the topology of network, such as the number of slaves connected to a master, and the number of piconets connected to a PMP node. Furthermore, we also noticed that routing protocol become unstable when the quality of some links in the path deteriorates. These observations lead us to investigate a cross-layer optimization approach to increase the performance of QoS Extended AODV. End result of these investigations is the Cross-layer Optimized Routing protocol for Bluetooth. A word of caution is that we have not used the word optimization in its strict mathematical sense, but to denote the method by which the routing protocol utilizes additional information for improving its performance. In other words, the routing protocol is tuned (optimized) to work with BT specific Scattenets, and is not link agnostic as in the cases of traditional network routing protocols. The rest of the paper is organized as follows. Section 2 introduces the prior art of routing protocols over Bluetooth scatternet. In section 3, we present the phenomena that can directly influence the performance of routing protocol over Bluetooth PAN. In section 4, we analyze the reasons for the performance degradation. In section 5, we first present two cross-layer optimization methods that alleviate the problem of network performance degradation, and then our new routing protocol, Cross layer Optimized Routing for Bluetooth (CORB). Finally, we give evaluation result in section 6 and conclusions in section 7. II. ROUTING OVER BLUETOOTH SCATTERNET The routing protocols in the context of ad hoc networks have received considerable attention from the researchers because the routing protocols are vital to the successful operation of ad hoc networks, and are fundamentally different from their counter parts for planned networks. Some of the work that is submitted as experimental RFCs in IETF are: the dynamic source routing protocol (DSR) and Ad Hoc On-Demand Distance-Vector (AODV) routing [3]. The research emphasis has been on providing the shortest path routing and achieving a

2 high degree of availability in a dynamic network environment where the network topology changes frequently. Recently, Quality of Service (QoS) routing in MANET is also attracting attention. For example, Campell et. al proposed a hard-qos architecture INSIGNIA [5], which is similar to Integrated Service approach in the Internet, and a soft-qos architecture SWAN [4], which is similar to Differentiated Service. Charles Perkins also proposed a QoS extension for AODV [6], which mainly replaces the link metric with delay and bandwidth instead of number of hops. Most of the best effort routing protocols search the best route based on the number of hops from source to destination. Most of the QoS routing algorithms assume that nodes know the link bandwidth/delay beforehand, and use this information as metric in route search process. But this assumption is not true for the current Bluetooth links. There are some attempts at either developing new algorithms, or at customizing the existing algorithms for Bluetooth based ad hoc networks. Kargl et al proposed Bluetooth Scatternet Routing (BSR)[8], which is a reactive routing protocol similar to AODV or DSR but keeps additional information on the state of links. The authors proposed a cross layer optimization (between link, and network layer) to shorten the connection set-up delay. Liu et al proposed scatternet-route structure [9] to combine the scatternet formation with on-demand routing, thus avoiding connecting all nodes at the initial network startup stage and maintaining all Bluetooth links thereafter. This protocol results in large delays for setting up of routes due to the large delays involved in forming the links. III. PROBLEMS OBSERVED FROM EXPERIMENTS We have developed a test bed to evaluate the performance of routing protocol over Bluetooth PAN. We implemented AODV routing protocol to realize multi-hop networking. Since there were no Bluetooth chips that can support Scatternet functionality at the time of development, we have used multiple Bluetooth interfaces to implement the Scatternet functionality. We first conducted several experiments to evaluate the role of network topology on the system throughput. The details of the experiment are omitted here because of page limitation. We only give our result here. First, the throughput between a master and any of its slaves decreases with the increase in number of background slaves; (background slaves here mean the other slaves connected to the same master, but not actively transferring data.) it is more noticeable when a background slave is first introduced into the piconet. Second, throughput between a master and its directly connected PMP node decreases with the increase in number of piconets the PMP node serves. Similar results were also reported in [2]. In addition to the system performance evaluation studies, we also conducted experiments to evaluate the protocol stability as shown in Fig.1 and Fig.2. In a room, eight nodes were arranged on a conference table of about 2 meters by 2 meters. A Bluetooth PAN between these eight Bluetooth nodes was formed, in which AODV was used as routing protocol. Regarding the configuration of nodes position, position A was on the conference table, C was about 10 meters far from the E 7 D A B C D E Fig. 2. C 7 Location B Fig. 1. Unstable link A Experiment Layout Unbreakable link Delay(ms) location delay Time Throughput Variation due to Node Novement conference table, and E was about 20 meters far from the table. Initially, an active session was initiated through node 7 (the gray color at A). Then, node 7 was moved from A to C slowly, and then was kept at C for a while. After that, node 7 was moved from C to E slowly. The movement pattern is shown in the left part of Fig.2. In the experiment, the ping (ICMP echo) packet was sent periodically between node 1 and node 5 of this active session to monitor the change of Round Trip Time (RTT). The experimental results are presented in the right part of Fig.2. From this figure, we notice that the measured delay keeps stable for a while when the node begins to move from A, and then, it suddenly increases sharply to a large value (above 1500ms) when the node approaches the position C. When the node stays at C, the delay is kept above 1500ms with a large variation. When the node moves form C to E, delay first increases sharply, and then ping packet stops to arrive at the sender, but the link is still active for about 20 seconds after ping packet has stopped to come back. After that, RTT returns to normal range after the route search is triggered by route error, and an alternative path ( ) is selected. The major conclusion of this experiment is that the link availability indicated by PHY layer signal may not denote the existence of link useful for data communication in Bluetooth. Sometimes, although the PHY link is active, the link quality is too poor for any meaningful communication. In such condition, conventional best effort routing, which uses number of hops as routing metric, is not enough to guarantee any meaningful communication. IV. ANALYSIS From the experimental results we noticed that when the number of slaves connecting to a master increases, the link

3 bandwidth between master and each slave link decreases proportionally. When PMP node attends more piconets, its link bandwidth to each of those piconets decreases proportionally. These phenomena are caused by the characteristics of Bluetooth s MAC layer. Bluetooth radio uses a Master-Slave slotted TDD (Time Division Duplex) MAC layer protocol. In TDD scheme, because slave can t send packet unless it receives a POLL packet from its master, slaves needs to be polled by master periodically to avoid too much buffering. As a result of this scheme, master needs to waste some bandwidth on polling its slaves even when there is no data flow. This scheme causes the decrease of bandwidth between master and slaves when the number of slaves increases. Since each Bluetooth unit uses only one radio interface and can only participate in one piconet at a given time instant, a PMP node has to schedule its time for participating in more than one piconet. Furthermore, since masters determine timing and FH sequence in their piconets which are independent of each other; a PMP node also has to re-synchronize to the new master when switching to a new piconet, which results in switching overhead. These problems cause the degradation of bandwidth in the PMP node. As a result, nodes with different roles (master, salve or PMP) may have different freedom and capacity for packet forwarding. Additionally, the phenomenon shown in Fig.2 tells us that sometimes although the PHY layer indicates the link is active, the link quality is too poor for any meaningful communication. We call a link when its delay is extremely large as unstable link, and a broken link as imaginary link if it is not detected as broken by either master or slave. Phenomenon of imaginary link happens when nodes are already out of the communication range, but the MAC layer still assumes that the link is active. A master usually does not monitor slaves continuously but polls its slave once within a fixed period of time determined by LinkSupervisionTimeout. If for any reason, baseband packets are not received from a link for duration longer than that time, the link is deemed to be broken. The value of LinkSupervisionTimeout can be adjusted through Bluetooth link layer parameter LinkSupervisionTimeout, and the default value is 20 seconds, which causes the problem of imaginary link. The unstable link is caused by the mismatch of covering ability between control and data packet in addition to the unpredictable radio conditions. We define the mismatched area as gray zone, where control packet can be received properly, but data packet can t be received reliably. Similar problem in ad hoc network is first reported in [10]. When the receiver is within the gray zone of transmitter, although the control packets indicate the link s activeness, the packet error rate is very high. This results in a low-bandwidth, high-delay and large-jitter link, and sometimes the link become useless. We analyze this problem in detail here. Bluetooth specification supports six types of data packets (DM1, DM3, DM5, DH1, DH3, DH5) and 3 types of control packets in ACL mode[1]. The control packets are encoded with 1/3 FEC with maximum gross payload length of 240 bits. The DMx types of packet are encoded with 2/3 FEC with the x indicates the number of slots the packet occupies. The maximum payload TABLE I COMMUNICATION RANGE OF BLUETOOTH PACKET Type Range (m) Type Range (m) DH DM DH Control DH length of DM5 is 224 bytes. The DHx types of packet are encoded without any FEC. The maximum payload length of DH5 is 339 bytes. Small packets are less prone to packet errors since there are fewer bits to transfer than in large packets. Packets with stronger FEC are more robust to packet errors since FEC offers error correction. The differences in both FEC, and packet lengths contribute to the occurrence of gray zone. We calculate the gray zone range based on the following radio assumption: modulation (FSK), background noise (3.4e-11), TxPower (1mW), effective communication range (error ratio < 0.1%). The result is given at Table I. The result shows that there is about 10 meters gray zone between control packet and longest data packet DH5. Because we use FSK instead of Gaussian-FSK in calculation, the result may be different from the real gray zone range, but it does indicate the existence of gray zone problem, which influences the communication reliability when both receivers are close to the boundary of each other s radio coverage. V. PROPOSAL In this section, we first present two cross-layer optimization methods that alleviate the performance degradation problem, and then we propose our new routing protocol CORB, a variant of AODV routing protocol with cross-layer optimization. A. Load Metric Here we propose new link metric to reflect the bandwidth degradation, especially for heavily loaded PMP and master nodes. We assume a simplified round-robin fashion MAC model for both inter-piconet and intra-piconet scheduling to calculate the load metric. The major assumptions are as follows. 1) Master communicates with its slaves in a round robin fashion 2) S/S PMP node accesses masters in different piconets with same amount of time without employing any specific scheduling policy 3) M/S PMP node first decides the time slots for the piconet in which it works as slave, then uses the rest of time slots to control its own piconet. The round robin channel access is a very simple MAC algorithm that offers fair channel access to multiple nodes. Most of Bluetooth inter-piconet/intra-piconet MAC methods are based on it. The estimated bandwidth based on this model can be used as the proxy for link bandwidth and it is a good indicator of node s real link bandwidth and load level. In Bluetooth based networks, the presence of different nodes (master, slave, or PMP) on a path greatly affects the path

4 3 31 TABLE II SYMBOLS AND NOTATIONS UED IN LOAD METRIC Fig Example of Scatternet Topology Used to Explain Terminologies bandwidth, and delay characteristics. Conventional best effort routing protocols use number of hops to approximate the goodness of a route, which simply ignores this BT specific knowledge. We propose a load metric (LM) that gives different weights to different links according to the roles of nodes forming the link. LM is defined as the ratio of maximum link bandwidth to the estimated link bandwidth, where maximum bandwidth is the maximum bandwidth between master and a slave if there were only one slave in its piconet. The LM reflects the link quality in terms of the bandwidth. The notation used to calculate the load metric is listed in Table II. For calculations, the following four types of nodes are considered: Master (M), Slave (S), Slave-only PMP (SP) and Master-Slave PMP(MP). The combination of these four types can form six types of links, M,S, M, SP, M,MP, S,MP, S,SP and SP,MP, but two link types, S,SP and S, MP are not allowed by Bluetooth specification. If there are two SP nodes between two piconets, we do not differentiate between them because they have same value of load metric. We omit the situation that PMP node attends multiple piconets as slaves and has its own piconet M,S,S, because many result shows that it is very inefficient for multi-hop communication[7]. Besides, the PMP degree (D ) of both master and slave is one, while that of SP and MP is the number of piconets in which they are slaves. The formulas to calculate LM for four different types of links are listed in (1) to (4). Comparing with two adjacent links, we can see that the link with larger LM indicates smaller link bandwidth. The link with largest value of LM indicates the bottleneck along the path from source to destination. As a result, the path metric is defined as the maximum LM among the links forming the path. The path with smallest metric indicates the path with largest bandwidth. LM Mi S ij = D i (1) LM Mi SP ij = D SP ij D i (2) LM Mi MP ji = max D j +1,D i (3) LM MPji SP jk = max D j +1,D i (4) For a path p = L 1,L 2,...,L i,...,l n, the LM is the maximum of each link s LM is given below. LM(p) =max{lm(l 1 ),LM(L 2 ),...,LM(L n )} (5) 23 Symbol Name Explanation M i Master Master node of piconet i S ij Slave j-th slave of picoent i SP ij Slave PMP PMP that attends both piconet i and piconet j as slave MP ij Master PMP Master of piconet i and slave of piconet j D i Slave degree number of slaves in piconet i D i PMP degree Number of piconets that a PMP node attends as slave B 0 Maximum bandwidth Maximum bandwidth between master ans slave in the case of only one slave in a piconet Estimated bandwidth between node B i j Estimated bandwidth i and node j b i j Normalized ratio of estimated bandwidth B i j bandwidth to maximum bandwidth B 0 LM i j k Load Metric Load metric of path(i-j-k), inverse of normalized bandwidth Proof: Load Metric We first calculate the normalized bandwidth b, then calculate LM from b s reciprocal. If two nodes of a link are within one piconet, the bandwidth is determined by the scheduling strategy of master node. If one end of the link is PMP node, the bandwidth depends on the two adjacent piconets that the PMP node serves. As a result, (1) and (2) can be easily deduced from MAC protocol assumptions. The proof of (3) and (4) is explained by a sample of link M 1 MP 21 S 21 link shown in Fig.3. The normalized bandwidth b M1 MP 21 S 21 is the smaller value of b M1 MP 21 and b MP21 S 21. According to the assumption 3, b M1 MP 21 varies between 0 and 1/D 1, while b MP21 S 21 is determined by the duration left in MP 21 after it reserves part of its time to attend piconet 1. Let α denotes the variable bandwidth reserved for MP 21 to attend piconet 1, then the relationship between those variables is given in (6). From (6), we find that when α is larger than 1/(D 2 + 1), b M1 MP 21 is larger than b MP21 S 21. Consequently, the normalized bandwidth of path M 1 MP 21 S 21 is 1/(1+D 2 ). Otherwise, the normalized bandwidth is 1/D 1. The result is presented in (6), by which we can prove the the correctness of (3) and (4). b M1 MP 21 =1/D 1 = α b MP21 S 21 =(1 α) 1/D 2 where 0 α 1/D 1 b M1 MP 21 S 21 = min b M1 MP 21,b MP21 S 21 1 LM M1 MP 21 SP 21 = b M1 MP 21 S 21 = max D 1,D 2 +1 (6) B. Optimization Based on Radio Layer Information To solve unstable link and imaginary link problems, we integrated link layer information into routing protocol. The LM we proposed in previous sub-section is an indicator of link bandwidth in normal radio condition. When radio environment

5 degrades, the link bandwidth also decreases sharply. It is necessary to reflect this change into link metric at each node in the path. In our proposal, every node periodically measures the RSSI(Radio Signal Strength Indicator). When the RSSI is smaller than a predefined threshold value, the link metric in the link table is increased to a large value. This change is immediately updated to the routing table entries corresponding to this node and other related nodes by some route maintenance method. For example, if there is an active path (a b c d e), the RSSI of link c d is below the threshold. First, the link entry of c d in node c is updated to a larger value, then, the metric for the route entries that use the affected link are also modified. The cost metric of routes c d, c e in node c is updated to a larger value based on the new LM of link c d. After that, this information is propogated to the source node of the flow. All the nodes on the route to source node would update their routing table to reflect the bad link quality on the route. E.g. routing table in node b is also updated after receiving update message. At node b, routing table entry to c, d and e is adjusted to a large value. There are several methodsto propogate the change of link metric in intermediate nodes. One is to measure the route quality by periodical probe frame, another is to send report to source node by an intermedaite node after it detects link degration. Source and destination node decides maximum tolerable value of route metric in route establishment phase. When updated route metric is larger than this value, source or destination node will trigger route search. However, when the link degradation happens between two intermediate nodes, the change may not be propagated to the end host on time because of unstable link status. Sometimes, unstable link status may even result in the failure of propagation of control messages like route maintenance packet. Thus, we propose the use of dynamic LinkSupervisionTimeout adjustment to trigger link break faster when the link degradation is so serious that control message cannot be propagated. It means whenever the RSSI is smaller than threshold, the LinkSupervisionTimeout is also reduced. If the control message cannot be propagated on time, link break is triggered. Such link break is immediately reported to end host by current AODV route maintenance mechanism and finally leads to the route research at the end host. The relationship among route maintenance, RSSI and LinkSupervisionTimeout is shown in Fig.4. When the link instability is detected by reading RSSI, local link and routing table information is updated. When one end of the link is end host, it directly trigger route search. If both ends of the link are intermediate node, this information may not be propagated to the end host on time. When route maintenance information is propagated properly, end host will start route search after receiving this information. If the link is too unstable to send control message like route maintenance, link break is detected by the LinkSupervionTimeout (which depends on link quality.) This would finally result in route search at the end host. The integration of these methods guarantees the stability of routing protocol. route metric > threshold Reroute is triggered at the end host YES Link status become unstable RSSI < threshold Local link/route metric is updated End host? Route metric at end host > threshold YES Link break -> RRER packet is propagated to end host NO Generate route maintenance packet Propagated properly NO Adjust LinkSupervisionTimeout Fig. 4. Block Diagram of Route Maintenance under Different Radio Condition C. Proposed routing protocol: CORB Based on the proposed LM and inter-layer optimization, we propose our cross-layer optimized routing for Bluetooth (CORB), which is a QoS extended AODV with optimization for Bluetooth radio. As discussed above, current Bluetooth specification does not provide delay or bandwidth information directly. To find a route with maximum bandwidth, we use LM as the link metric in route search process. It means that the path with smallest LM is used for packet forwarding. When an originator needs a route to a destination, it broadcasts a Route Request (RREQ) over the Scatternet. The destination node, when it receives RREQ, responds by unicasting a Route Reply (RREP) to the originator node. LM is piggybacked to the originator by RREP message. Intermediate node updates LM in RREP message with the larger value between LM in RREP and its link LM. After the originator receives RREPs, it selects a route with smaller LM, or in other words, larger bandwidth, to forward data packets. Another feature of CORB is its route maintainance. In besteffort routing protocol, route maintainance is triggered only when intermediate node detects link break and sends out Route Error (RERR) message. In CORB, route maintainance decision also depends on the stability and quality of links on the route. It can detect link degradation and select alternative path before link breaks. CORB uses the optimization based on radio layer information discussed above. In our implementation of CORB, source node is responsible for sending probe packet to monitor route quality. When it detects that link degradation causes violation the negotiated threshold(qos violation), source node initiate route search to find an alternative route. Additionally, when source node receives link break notification from intermediate nodes due to imaginary and unstable links, source node also starts route search process. The summary of the CORB algorithm is given in Table III.

6 TABLE III SUMMARY OF CORB ALGORITHM Types Route search Route maintenances Source Broadcast RREQ (1) Send probe packet periodically, (2) start route search again, when metric at returned probe is larger than threshold Destination Reply the RREQ with Send back probe packet Relay RREP Update metric according to LM algorithm Update metric according to LM algorithm and current RSSI VI. EVALUATION Test bed was built using laptops and commercially available Bluetooth devices, and we implemented necessary protocols. Our CORB is developed based on Bluetooth BNEP/PAN API. We tested the performance of our routing protocol on testbed in the same environment as described in section 2. For CORB, the imaginary link problems has almost disappeared, and we have not noticed any packet losses caused by imaginary link. The unstable link problem is also alleviated. In most of the cases, we can only notice one frame with abnormally large delay. To numerically measure the performance of our protocol, we conducted following experiment. The experiment environment was same as given in SectionIII. When the mobile node is moved from location B (within 10 meters range of node A) to location E, we measured the time taken by the routing module to trigger the new route search due to link loss. These measurements were done for both AODV, and CORB. Regarding to the parameter of CORB, nodes measure the quality of links with its neighbor every 5 seconds. The experiments were done 20 times for each routing protocol. Results are graphically shown in Fig.5. The results shows that the route search delay for CORB is about 39 seconds compared with 234 seconds for AODV. The variation of delay for CORB is also much smaller than that for AODV. Large deviation of search delay for AODV shows that it is not suitable to depend only on default configuration of the link supervision mechanism to detect link break. The results for CORB protocol show that by incorporating link layer information into routing protocol, the route stability can be greatly improved, and the routing protocol can respond quickly to the adverse changes in link conditions. The LM used in CORB is an estimated value, and not a measured value. By substituting the measured LM in place of estimations, the performance of CORB can be further improved. Duration ( in seconds) CORB Fig. 5. Route Switch Delay Route switch time AODV quality information. We also reported the relative performance of CORB, and AODV when they were implemented over actual HW/SW test bed and tested in indoor radio conditions. Performance of CORB needs to be compared with other routing protocols either by simulation or by implementation. We also observed that the increase in number of Piconets in a small geographical area could cause serious interference problem., which in turn, can also influence scatternet s system performance. These are some of the issues that need to be studied in the future. REFERENCES [1] Bluetooth SIG, Bluetooth Specification 1.2, Dec [2] Leping Huang, et.al, Proposal for QoS ad hoc routing over Bluetooth PAN, IEICE General conference 2003, Sendai Japan, March [3] Charles E. Perkins, et al. Ad Hoc On Demand Dis-tance Vector (AODV) Routing IETF RFC 3561, July [4] A. Veres, A.T. Campbell, M. Barry and L-H. Sun, Supporting Service Differentiation in Wireless Packet Net-works Using Distributed Control, JASC, Vol. 19, No. 10, pp , October [5] S.-B. Lee, G.-S. Ahn, X. Zhang, and A.T. Campbell, IN-SIGNIA: An IP-Based Quality of Service Framework for Mobile Ad Hoc Networks, J. Parallel and Distributed Computing, vol. 60, no. 4, pp , Apr [6] Charles E. Perkins, Elizabeth M. Royer, and Samir R. Das. Quality of Service in Ad hoc On-Demand Distance Vector Routing. IETF Internet Draft, draft-ietf-manet-qos-00.txt, July [7] Gy. Miklos, A. Racz, Z. Turanyi, A. Valko, P. Johansson, Performance Aspects of Bluetooth Scatternet Formation, poster at MobiHOC 2000, [8] Frank Kargl, Stefan Ribhegge, Stefan Schlott, Michael Weber. Bluetooth-based Ad-Hoc Networks for Voice Transmission, Hawaiian International Conference on System Sciences 36, Hawaii, USA, January [9] Yong Liu, Myung J. Lee, Tarek N. Saadawi, A Bluetooth scatternetroute structure for multihop ad hoc networks, JSAC, 21( ): [10] H. Lundgren, E. Nordstrom, and C. Tschudin, Coping with Communication Gray Zones in IEEE b based Ad hoc Networks, The Fifth International Workshop on Wireless Mobile Multimedia,WoWMoM 2002,Atlanta, Georgia,U.S VII. CONCLUSIONS AND FUTURE WORKS In this paper, we reported, and analyzed the performance anomalies of BT PAN network. Based on the analysis we proposed a routing protocol that is optimized for Bluetooth networks (CORB). New link metric LM is introduced into routing protocol to bypass heavily loaded nodes. Bluetooth MAC layer information is filtered into routing protocol to improve the stability and reaction time to changes in link