Class-based Packet Scheduling Policies for Bluetooth

Class-based Packet Scheduling Policies for Bluetooth Vishwanath Sinha, D. Raveendra Babu Department of Electrical Engineering Indian Institute of Technology, Kanpur - 08 06, INDIA vsinha@iitk.ernet.in, d raveendra@yahoo.com Abstract Bluetooth is a wireless ad-hoc network concept primarily intended to eliminate the cables between portable and/or fixed electronic devices. In this paper, a non-real-time communication service in Bluetooth is considered. Depending on its QoS (Qualityof-Service) requirements, the non-real-time traffic is divided in two classes: a) class: delay-tolerant traffic like email and paging; and b) class: delaysensitive traffic like ftp and remote log-in. class is given priority over class. We propose two new scheduling policies with an attempt to provide high performance service in terms of delay and throughput to higher priority class while fairly distributing the link resources to lower priority class.. Introduction The Bluetooth technology was primarily developed to provide wireless interconnection between small mobile devices. In any communication system, it is important to support various traffic with QoS gaurantees. In this paper, we examine in detail the performance of non-real-time communication service in Bluetooth. Depending on its distinct characteristics and QoS requirements, the non-real-time traffic can be divided into two classes: a) class: delaytolerant traffic like paging and email; and b) class: delay-sensitive traffic like ftp and remote log-in. The main distinguishing factor between these classes is how delay sensitive the traffic is. To satisfy the QoS needs of different connections, nodes present in the network need to have priority and scheduling mechanisms. We propose two new scheduling policies that provide QoS gaurantess to class traffic. In our proposed methods, class traffic is given priority over class traffic.. Bluetooth Technology Bluetooth is a universal short range wireless communication system operating in the.45 GHz ISM band. The technology is based on FH-CDMA using 79 carriers MHz spaced []. Bluetooth channels use FH/TDD (Frequency Hop/Time Division Duplexing) scheme. Two or more units sharing the same channel form a piconet, where one unit acts as master and the other units act as slaves. Up to seven slaves can be active in a piconet. Bluetooth uses centralized TDD scheduling as the MAC (Media Access Control) protocol. There is a strict alternation of slots between the master and the slaves. The master packet transmission starts in even slots, while the slave transmission starts in odd slots. A Bluetooth packet can be,3 or 5 slot length. A slave may transmit the packet in the slave-to-master slot only if it is polled by the master in the preceding slot. Between any two devices, which are forming a master-slave pair, two types of links can be established. These are the (a) Synchronous Connection Oriented (SCO) link used for time bounded data, such as voice, and the (b) Asynchronous Connection Less (ACL) links for data. In practice, an SCO link is implemented by reserving duplex time slots at regular intervals. The ACL link works as traditional packet-switched network. In this paper, only the ACL type of link will be considered. Multiple piconets overlapping in the same area will form a scatternet. 3. Round-Robin Scheduling One of the conventional scheduling algorithms, Round-Robin (RR) scheduling can be used to schedule the data in a piconet. In RR scheduling amongst slaves, each master-slave (m-s) connection is alloted a pair of slots for the transmission of packets. Using RR scheduling it is possible to provide each m-s

pair, generating either class or class packets, a fair access to the channel. But a simple RR scheduling is not suitable for Bluetooth as it is unable to minimize dealy for interactive sessions like class. 4. Priority Queueing at master A PQ at master model is developed when the Bluetooth system is operating as infrastructure network. Here, in a piconet, master could be the Fixed Access point or the Base station. We consider downstream traffic in a piconet consisting of one master and seven slaves. In this PQ at master model, master maintains one priority queue, for all the slaves, for a class traffic and an individual output class queue for each slave, as shown in Figure. class ( λ ) traffic classifier packets class from ( λ ) higher layers master priority queue ms0 ms ms ms3 ms4 ms5 ms6 Base band queues at master scheduler MAC TX RX Figure : Illustration of downlink with a Priority Queueing (PQ) at master model. With this priority queueing model, the scheduling of packets is as follows: The master always gives priority to a packet from class queue which is nonempty in its transmission. Master keeps on serving the class packets in a FIFO (First-In-First-Out) manner until the priority queue becomes empty. So the service process at the priority queue is an exhaustive service process. Master will transmit a packet from class queues only if the priority queue becomes empty. It serves the class queues in roundrobin manner and the service process is limited- service. We assume that arrival processes of class and class packets to be Poisson with an effective arrival rates d and d, respectively. The overall packet arrival rate is defined to be d = d + d. Here the service time of a class (class) packet is defined as the effective service time during which master sends a class (class) packet in master-to-slave slot s0 slaves s6 s5 s s4 s s3 and receives a packet from the corresponding slave in slave-to-master slot. So the service time corresponds to duration of pair of slots which is random. Let X e and Xe be the first and second moments, respectively of the service time of a class packet. The vacation period is the service time of a class packet. Let V e and V e are the first and second moments, respectively of the vacation period. In the priority queueing discipline, the mean queueing delay of for class packets can be calculated analytically by modeling the class priority queue as M/G/ queue with vacations, if we assume the arrival process of class packets into the priority queue as Poisson and vacation intervals as the intervals during which master spends at a single class queue. Now the results for the M/G/ queue with server vacation model and exhaustive service discipline are used to obtain the mean delay for class packets, W d, in priority queue. Defining ρ d = d X e and if X and X, respectively, are the first and second moments of packet transmission time in a piconet, then W dxe V d = ( ρ d ) + e V e + X () 5. Priority Queueing (PQ) at master with Priority Polling (PP) With a Bluetooth system operating as a adhoc network, a Priority Queueing at master with Priority Polling is developed as the above PQ at master model minimizes the packet delay for traffic only in downlink. We consider a piconet composed of a master and seven slaves, where slaves are sending packets to other slaves in the piconet. Here, master acts as a router receiving packets from the slaves and forwarding them to the destination nodes. Since Bluetooth inherently doesn t support direct slave-to-slave connectivity at the link or radio link level, it will necessarily be done by routing via the designated master of the piconet. There has been a recently proposed algorithm in [4] for routing in Bluetooth system. The proposed algorithm uses Layer 3 header information for intra and inter piconet communication. Here, in a piconet of above type, a packet in slave-to-slave communication may be of class type or class type. To differentiate the packets according to their class, we have added one more bit called

Class of Service (CoS) bit in the Layer 3 header. If CoS =, the payload of the packet is of class type. The value CoS = 0 is meant for class. In this PQ at master with PP model, we divide the slaves into two groups. First group consists of p number of slaves out of seven slaves which are generating class type of data. Second group consists of remaining slaves generating class kind of data. The polling of slaves will be done in two ways: exhaustive polling and limited- polling. Exhaustive polling is done for slaves generating class traffic, i.e., once master starts polling the slave, it continuously polls the same slave until queue of the slave becomes zero. Whereas limited- polling is for slaves generating class traffic i.e., master polls the slave to receive only the first packet from that slave s queue. Here the polling refers to either explict polling or implicit polling. In this model also master maintains single priority queue for class packets destined for all slaves and a per slave class queue as shown in Figure. The scheduling of packets with PQ at master with PP model when p=, is explained in the following. We assume that the binary information regarding the status of the slave queue is available at the master. When a master polling the slaves, it always gives priority to class slave. It polls the class slave only if there is no backlog at class slave. When the master polling the class source exhaustively, all the packets master has received will be queued in priority queue at the master.if the queue of class source becomes zero then the master stops polling the source and starts sending all the packets, that are received during polling, to the destinations. After the priority queue at master becomes zero, master polls the class source only if it is backlogged. If it is not backlogged, master polls the class slaves in round-robin manner. In this model, the arrival process at the slave queues is assumed to be Poisson. Then, the queue of the class source can be modeled as a M/G/ queue with vacations in which first vacation period has a different distribution than the subsequent vacation periods. Here, the first vacation period corresponds to the period during which master sends the packets from it s priority queue to the destinations. The subsequent vacation periods are the periods during which master polls the class slave and receives a packet from corresponding slave. We assume, to be the effective arrival rates of class and class packets, respectively when p=. Here the service time of a class (class) packet is defined as the effective service time during which master polls the class (class) slave in master-to-slave slot and receives a packet from the corresponding slave in the slave-to-master slot. So the service time corresponds to the duration of a pair of master-to-slave and slaveto-master slots which is random. Let X e and Xe are the first and second moments, respectively of the service time of a class packet. The first vacation period is equal to the length of the Busy Period of class slave s queue. Let B and B are the first and second moments of the Busy Period of class slave s queue. The subsequent vacation periods are equal to the service times of class packets. Let V e and Ve are the first and second moments, respectively, of those vacation periods. Defining ρ = X e, from [5], the mean queueing delay of a class packet at class slave is given by- W Xe q = + ρ B +( ρ )Ve () ( ρ )BV e The mean queueing delay of a class packet at the priority queue of the master is equal to (B X e ). where, from [], B = X e ( ρ ) ; B = X e ( ρ ) 3 Therefore, the mean end-to-end delay for class packets is given by W = W q +(B X e )+X (3) In general, for p greater than one: master polls a backlogged class source in an exhaustive manner, i.e., until the queue of the source becomes zero, and receives the packets destined for other nodes in the piconet. Once the class source is unbacklogged, it stops polling and forwards all the packets, that are received, to the destinations. With the same process, master serves all the class sources present in piconet in a round-robin manner. When the master forwarding the packets of a particular class source, it may receive some class packets from destination nodes in the slave-to-master slots. These packets will be queued in the priority queue. Master will start polling the next class source, only if all the packets in the priority queue are cleared. Also,

master polling a class source can send a class packet destined for that source in the master-to-slave slot. This model can be explained logically as follows: Consider, two logical states 0 and for class sources and master. Class source remains in state 0 as long as it is unbacklogged. Master is also in state 0, if it s priority queue is empty. A class source will enter into state, if it is backlogged. The master will also enter into state if the priority queue is non-empty. Master maintains a separate counter corresponding to each class source whose values are also can be either zero or one. Initially assume that master is at state 0, since it only acts as router, and all the counters are also in zero state. Step: Now the master polls the class source whose state is until source enters into state 0. Once master started polling, the counter value corresponding to that source becomes. Step: Then the master will enter into state. Step3: Master keeps on sending the packets that are arrived at it s priority queue until it enters into state zero. Step4: Once the master enters into state 0, it polls the class source whose state is and whose counter value is zero. If two or more slaves with state are having counter values equal to zero, master will give equal priority to those slaves. Go to Step and the procedure follows. Here, note that if all the counters are at state, they will be reset to zero. Master polls the class slave if all the class slave-master pairs are in 0-0 state, i.e., if all the class sources are unbacklogged. It polls the class slaves following the round-robin manner. Here also, the polling can be explicit polling or implicit polling. Thus, a Priority Queueing at master with Priority Polling model can give end-to-end QoS gaurantees by effectively using the bandwidth. This model can be used to provide end-to-end QoS gaurantees in a Bluetooth scatternet. 6. Simulation Results Discrete Event Simulations have been performed for the scheduling policies proposed in the paper. 6.. Piconet with PQ at master model We simulate a piconet consisting of master and seven slaves. The data arrival process at the master and the slave queues is assumed to be Poisson. The overall packet arrival rate in the uplink is defined as u. The arrival rates (packets/sec) at the slaves are given by si = u =7, 0» i» 7. The offered load in the uplink is ρ u = u X. The overall packet arrival rate in the downlink is given by d. where, d = d + d. The offered load in the downlink is given by ρ d = d X. The data arrival rates at the class output queues at the master are msi = d =7, 0» i» 6. The service time of a data packet depends on the packet length. A packet size is chosen uniformly from, 3 and 5 slot lengths with equal probablity. We keep the buffer sizes sufficiently large to hold the packets. Discrete Event Simulations were run for 00 TDD slots. Figure shows the mean delays for both class and class packets in the downlink with PQ at master model when ρ d = ρ u. Here, half of the offered load in the downlink (ρ d ) is from class. From the figure, class s delay goes to infinity at around offered load ρ d = 0:68. This means that class traffic is rarely transmitted for ρ d greater than 0.68. The class s delay is bounded even at ρ d equal to one. Figure Mean Delay (ms) 0 0 0. 0.0 Offered Load ( ρ ) d class Figure : Mean Delay Vs Offered Load for both class and class in downlink with PQ at master model. 3 shows the throughput values for both class and class in downlink with the model. Here, throughput is defined as defined as the rate (packets/sec) at which packets undergone service. From the figure, it can be observed that beginning at load ρ d =0:7,the throughput for class starts to decrease while that of class continues to increase. This is reasonable since half of the offered load is from class and the priority is given to class packet transmission.

Throughput 300 50 00 50 50 class Mean End-to-End Delay (ms) 0 0 0. class 0 Offered Load ( ρ ) d Figure 3: Throughput Vs Offered Load for both class and class in downlink with PQ at master model. 6.. PQ at master with PP We simulate a piconet consisting of master and p number of slaves out of seven slaves generating class traffic and other slaves generating class traffic. Slaves are communicating with other slaves through the master, where the master acting as a router. The data arrival process at the slaves is assumed to be Poisson. The arrival rates (packets/sec) at the slaves are equal and given by si = =7, 0» i» 6. The overall packet arrival rate in piconet is defined as. The offered load in the piconet is given by ρ = X. A packet size is chosen uniformly from,3 and 5 slot lengths with equal probability. Discrete Event Simulations were again run for 00 TDD slots. Figure 4 shows the delay performance for both class and class when p=3. In can be observed from the figure that the delay for class packets is increased as the offered load from class sources is more in this case. Figure 5 shows the throughput for both class and class when p=3. From the figure throughput for class continues to increase, whereas the throughput for class is decreasing starting at around ρ =0:55. 7. Conclusion In this paper we have proposed two new scheduling policies with the aim of minimizing delay for interactive sessions like class. It is observed that a PQ at master with PP gives the end-to-end QoS guarantees. 0.0 Offered Load Figure 4: Mean End-to-End Delay Vs Offered Load for both class and class packets with PQ at master with PP model when p=3 Throughput 350 300 50 00 50 50 class 0 Offered Load Figure 5: Throughput Vs Offered Load for both class and class with PQ at master with PP model when p=3 8. References [] J.C. Haartsen, The Bluetooth Radio System, IEEEPersonal Communications, pp. 8-36, Feb000. [] L. Kleinrock, Queueing Systems, Vol, Wiley 975. [3] D.Bertsekas and R. Gallager, Data Networks, Prentice Hall, 994. [4] P. Bhagwat and A. segall, A Routing Vector Method(RVM) for Routing in Bluetooth Scatternets, MoMuC 99, pp. 375-379, 999. [5] D. Raveendra Babu Priority-based scheduling policies for Bluetooth, M.Tech Thesis, IIT Kanpur, June, 00.