Quality of Service in Wireless Ad Hoc Networks through Self Admission Control

Transcription

1 Quality of Service in Wireless Ad Hoc Networks through Self Admission Control Abstract An ad hoc network comprises of mobile nodes forming a temporary network on the fly without the aid of any centralized administration or standard support services. A novel approach is used to provide QoS at the MAC layer by making the nodes possess the self-admission control capability based on the equivalent capacity in the network. Every mobile node incorporates a distributed priority scheduling algorithm that is based on the equivalent bandwidth of its application. IEEE is also retained in this scheme which can be used by nodes not requiring any QoS. An invitation scheme is proposed to enhance the utilization in the network. This scheme is simulated in NS-2 and the behavior is studied under different conditions. A solution for providing QoS at the routing layer that integrates with the proposed scheme at the MAC layer is also discussed. I. INTRODUCTION The mobile nodes in an ad hoc wireless network [1] communicate with each other using multi-hop wireless links. Since the nodes in an ad hoc network act independently and operate in a shared media, there will be collisions and reduction in throughput when two nodes that are close to each other start transmitting at the same time. The MAC protocol at the mobile nodes should be able to minimize collision, allow fair access and transport data efficiently over the wireless links in the presence of rapid topology changes and also tackle hidden and exposed node problems. Some of the MAC protocols used in ad hoc networks are Bluetooth, HiperLAN, IEEE [2] etc. When the number of nodes in an ad hoc network increases, the number of nodes competing for shared bandwidth increases and that leads to the degradation of throughput and mean waiting time of the application. Providing QoS in ad hoc networks is ongoing research. This work uses a novel approach where a portion of the channel capacity is broken into contention-free timeslots. Mobile nodes admit themselves into these timeslots based on the equivalent bandwidth they expect. Some nodes still use the contentionbased approach, but those which need better QoS can use these contention-free slots. In most of the work done so far, QoS is addressed either at the MAC layer or at the routing layer. [3] has an in-depth analysis about the contention protocols and it tries to improve the fairness among the contending applications. [4] discusses the unfairness issues in IEEE and it has a detailed analysis of how the performance of IEEE deteoriates when more than one TCP connection is active. [5], [6], and [7] provide methods for overcoming the fairness problem by making adjustments to the contention window. Other work goes beyond just modifying the contention window. For example, [8] gives priority to Diffserv Expedited Forwarding (EF) and assured Forwarding (AF) classes over Best effort (BE) traffic [2], [21] by allowing EF and AF to grab the channel before BE. Lots of work has been done in the area of providing QoS in wireless networks with base-stations. [9] gives a good overview of the different MAC protocols for providing QoS in wireless networks. Most of the discussed protocols have a similar idea of the channels sending their reservation request in a contention slot and the headend or base station keeping track of the information of all the other end nodes. The Point Coordination Function (PCF) of IEEE does provide some limited QoS in a Wireless LAN. The Point Coordinator (PC) does not have any control over the polled stations about the time duration of transmission or the size of the packet that the polled stations can transmit. Further, the polling procedure itself may get delayed because of the unpredictability of the time used by the DCF procedure. These factors will effect the QoS of the existing applications. Cooperation between DCF and PCF also leads to poor performance [1]. Because of these drawbacks, the IEEE task group came out with IEEE 82.11e for providing QoS [11], [12], [13]. Here, the nodes requiring QoS send a request to the Hybrid Coordinator (HC) during the controlled contention period. If the number of applications requiring QoS is high, new applications or existing applications requiring QoS will find it hard to transmit their request for QoS successfully to the HC. Further, the amount of overhead information that a HC needs to transmit as well as the complexity in the design of the HC requiring it to keep the bandwidth information about all the existing applications is also very high. Further, different traffic classes use different values for the Contention Window (CW), which means less delay and better priority to access the channel but at the expense of more collisions and packet losses when the number of applications increases. Moreover, neither the PCF nor the IEEE 82.11e was designed for ad hoc wireless networks. [14] is another scheme called the blackburst contention scheme, where the station having higher priority traffic to send waits until the channel becomes idle and then completely jams the channel to force all the other nodes into backoff mode so that it can transmit immediately. [15] considers the ability to transmit more than one packet at a time. In [16] transmission of pulses is used to achieve priority in addition to using the blackburst contention mechanism. [17] divides the channel in to contention period (CP) and contention free period (CFP). A node needing to send real time data reserves the channel for CFP during the CP with the Contention Free Period Generator (CFPG). The CFPG keeps track of the number of stations, its QoS requirements and also allocates the residual bandwidth to any stations that require it. The amount of control information

2 sent is more as well as the complexity in the CFPG is also high. If the CFPG fails, the whole of the network will be in an unstable condition. Even if there is a backup CFPG, the amount of information to be transferred is quite high making it un-suitable for Ad Hoc Networks. [18] does admission control and finetuning of application layer parameters by passively monitoring the channel and estimating the delay, packet loss etc through the VMAC algorithm. This requires the VMAC algorithm to run continuously in all mobile hosts. Further, it is not very clear as how application with varying or unknown traffic requirements utilize the VMAC or VS algorithm for its CAC. To summarize the work done so far, [3], [4], [5], [6] and [7] concentrate on fairness among the competing nodesapplications, which is unsuitable for applications with stringent timing requirements or varying QoS requirements. Work similar to [8] achieve QoS by modifying the contention window based on the priority of the application. Reservation based mechanisms [9], [12], [17] are proposed for achieving QoS, which requires complex functionalities at the base-station or at the elected head of the wireless network. [14], [18] differs from the above said schemes. In most of the work discussed so far, no provision is done to limit the number of sources competing for the resources in an Ad Hoc network, which definitely effects the QoS of existing applications. Efficient Connection Admission Control (CAC) and Scheduling are very essential in providing QoS. Papers like [18] do address the CAC, but a simple and efficient in-built CAC and scheduling mechanism is not provided for Ad Hoc Networks so far. Performing CAC and scheduling in a distributed manner with no single node having to burden the task, is the real challenge in Ad Hoc networks. In contrast to many of the above cited works, the work here uses a simple selfadmission scheme to modify the IEEE MAC layer by which all the mobile nodes will start transmitting their traffic only if the channel has sufficient capacity to support their applications. The capacity of the channel depends on the number of competing sources as well as their traffic characteristics. Nodes requiring QoS need not have to contend for the channel; instead the nodes use a distributed algorithm to compute the amount of time they need to transmit as well as the time at which they have to transmit data. By eliminating contention, the throughput reduction occuring due to collissions can be avoided. By limiting the number of sources, service degradation occuring due to the increase in traffic interference can also be avoided. The rest of the paper is organized as follows. Section 2 defines the problem and scope. Section 3 gives an overview of clustering and equivalent capacity. Section 4 describes the IEEE protocol. Section 5 gives the proposed QoS scheme. Section 6 gives the simulation results. Section 7 discusses the solution for providing QoS in a Multi-Hop Ad Hoc network and Section 8 concludes the paper. II. PROBLEM STATEMENT AND SCOPE MAC protocols used in wireless ad hoc networks usually suffer from their lack of service differentiation and degradation of throughput when more nodes contend for the channel, this makes them a poor choice in providing QoS. In a mobile ad hoc network, all the mobile nodes with different traffic characteristics contend for the shared channel which closely resembles traffic from different queues in a router contending for an output link. Unlike a switch, an ad hoc network does not have any central admission controller module for protecting the existing traffic from service degradation or a scheduler for giving preferential treatment to high priority traffic. The following is needed in the MAC layer for supporting applications that have QoS requirements. Each and every mobile node should do a self-admission control to decide whether the ad hoc network has sufficient bandwidth to support its application. For this, a mobile node before it enters the network should know the current bandwidth available in the network. An effective scheme that transmits less number of control messages for getting the available bandwidth in the network should be employed. The mobile nodes should determine the time at which they should emit traffic as well as their draining rate (amount of time they are allowed to transmit). The duration of time available to a mobile node in accessing the shared channel should be proportional to the required throughput of the application. If every mobile node in the network can coordinate their transmission times, the contention window can be eliminated. This increases the bandwidth available in the network. A distributed algorithm is needed for doing all of these. III. CLUSTERING AND EQUIVALENT CAPACITY A. Clustering As discussed in the previous section, the nodes contending for the channel should do self Connection Admission Control (CAC) to decide whether the channel has sufficient capacity to support the application. The nodes contending for the channel are the nodes that are in the sensing range of each other. In order for the nodes to effectively do admission control, a cluster needs to be formed such that the cluster contains nodes that are in the sensing range of each other. Furthermore, clustering has a lot of advantages like frequency reuse, power consumption, robustness, increased system capacity etc. The Cluster Head (CH) is an arbitrarily chosen mobile node which can communicate directly with all the other nodes in the cluster. It will be seen shortly that the functionality of the CH is very limited unlike a base-station or the schemes proposed in [9]. The protocol is implemented such that all the nodes in the cluster receive the bandwidth information about the cluster. So even if a CH fails or a new CH is elected, it can immediately start functioning as a CH without any message transfers. Because the control messages are so important to this scheme, additional power and error control mechanisms may be used for these packets. This will not be discussed further since error control and recovery are outside the scope of this paper. The following assumptions are made about the ad hoc network: Mobility of the nodes are not too high. A mechanism for cluster formation such that the nodes in a cluster are in the sensing range of each other exist. Mechanisms mentioned in [19] can be used.

3 6 The nodes move inside the cluster freely and also the nodes move together as a cluster. Inter cluster movement is limited. Power aware mechanisms, error recovery and acknowledgement schemes are not considered in this work. B. Equivalent Capacity The equivalent capacity is the bandwidth requirement of a single application 1 in case of a single node and is the bandwidth requirement of multiplexed connections in the case of a network. The equivalent capacity of one application is independent of traffic generated by the other sources. Lots of work has been done to effectively characterize the sources and compute the equivalent capacity. [22] contains an excellent overview of the different call admission control schemes and also the methods employed to compute the equivalent capacities. The scheme proposed in [23] is used here for computing the equivalent capacity ( ) of a source. This scheme has the following features: The equivalent capacity of a Markov modulated fluid source is approximately the maximal real eigenvalue of a matrix derived from source parameters, buffer requirements and cell loss probability. The equivalent capacity is the maximum real eigen value of the matrix where and!#"$&%(' and M is the infinitesimal generator of modulating Markov chain that governs the transitions between the states of the arrival process. p is the loss probability of the application and B is the buffer size. If there are N sources, the equivalent capacity is asymptotically equal to )+*!,.-., where is the equivalent Capacity of the i1 source. This sum must always be less than the capacity of the channel. This scheme is chosen because of the following reasons: The computation of the equivalent capacity of a single source is not computationally intensive and it has a closed form solution. For computing the equivalent capacity of the network, a node has to know only the existing total equivalent capacity of the network with which it can add its equivalent capacity to obtain the new total equivalent capacity of the network. The equivalent capacity of the network seen by all these nodes is the same. IV. DESCRIPTION OF IEEE The IEEE standard is described in this section as it is the basis upon which the new scheme is built. IEEE 82.11, when used in its Distributed Coordination Function (DCF) mode, is based on a Carrier Sense Multiple Access (CSMA) 2 Please note that the terms application and connection are used inter-changeably in this paper Collision Avoidance (CA) scheme. A node that intends to transmit a packet waits until the channel is sensed idle for a time period equal to Distributed Inter-Frame Space (DIFS). After this, in order to avoid collisions a node does backoff by choosing a random number between to CW as the backoff interval. If the node detects a transmission during its backoff waiting time, it will freeze its backoff operation until the channel becomes free for DIFS and then continue its backoff procedure again from where it left. The mobile nodes implement the backoff procedure by decrementing a counter from the chosen random number to. The counter is decremented as long as the channel remains idle, and is stopped when the node senses a transmission. The timer is reactivated after the channel is again sensed idle for a DIFS. When the backoff procedure terminates, the node can start transmitting its traffic after it completes its Request to SendClear to Send (RTSCTS) procedure. RTSCTS procedure is required because without it the nodes suffer from what is called as the Hidden Node Problem. A node after receiving a packet waits for a Short Inter-Frame Space (SIFS) before transmitting the next packet. A. Structure of a Cycle V. PROPOSED SCHEME For this work, an entirely different approach is used than the MAC. The channel is divided into cycles or frames of equal duration. Every node sends their packets in their respective slots of a cycle. The slots repeat after a fixed duration. The structure of a cycle is shown in Figure 1 and it has the following properties: Fig. 1. Free Channel Duration PREAMBLE CBR SOURCES VBR SOURCES PREAMBLE CBR1 CBR2 DIFS Structure of a Cycle CYCLE DURATION VBR2 DIFS VBR1 The beginning of a cycle is marked with a '<: to synchronize. The packet from the CH. This will be useful for the new nodes 6 packet at the start of every cycle has the following information: Start of the free channel from the beginning of the cycle. Duration of the free channel. The CBR connections transmit at the beginning of the cycle. The VBR connections transmit at the end of the cycle. There exists a Free channel at the middle of the cycle which is occupied by the connections that do not ';:

4 6 Y ; D D ; # need reservation of the channel like the best effort services (UBR), packets sent for admission control, and other control packets. This is basically the unused bandwidth. When a new CBR or VBR connection comes up, the Free Channel at the middle shrinks. Some bandwidth is always reserved at the Free Channel so that the best effort applications do not starve. Transmission in the free channel is done according to IEEE DCF procedures (contention window procedures) with slight modifications. The nodes sending UBR traffic wait for the start of the free channel and start their contention window procedure. The modified contention window procedure is such that if the whole of the backoffrtsctsdataack packet sequence doesn t get completed within the cycle ( The VBR sources would have started transmitting before the node could count down to zero), the node does the following: Continue its contention window procedure during the free channel of the next cycle Continue its contention window procedure when the slot of any of the VBR source is idle. This is the invitation scheme and is explained more in Section V.D The control packets needed for clustering are sent during the free channel at the middle of the cycle. Since, the inter cluster movement is assumed to be limited, the overhead due to clustering will also be less. However, the reserved bandwidth at the free channel should be configured to be sufficient for clustering messages. Priority can be given to clustering traffic over the best effort traffic by making their inter-frame space smaller than DIFS. Clustering will not be discussed further since the scope of this paper is limited to QoS. B. Procedure for a New Application When a node wants to start an application that needs some quality of service, it waits for the beginning of the cycle, which is recognized by the node when it receives a packet from the CH. The source then waits for the free channel which lies in between the CBR and VBR sources. This information is sent by the CH as explained previously. The source waits for the channel to become free for a period of DIFS which will be during the free channel, after which it communicates with the CH using the modified contention window procedures as explained for UBR. After the backoff procedure is completed, if there is sufficient free channel duration, the ';: 6 packet asking the equivalent capac- node sends a78 4 ity in the network from the CH. The CH after getting a packet from any of the mobile nodes, will broadcast a packet to the network, which has the following information: Total Equivalent bandwidth of the CBRVBR connections (6 6 ) Total number of CBRVBR connections ( ) After getting the packet, the node computes the equivalent bandwidth of the new application (6 ) as 6 = Peak rate of transmission, if the application is of type CBR 6 = Equivalent bandwidth as calculated from section III.B. [23], if the application is of type VBR. The node also keeps track of its position in the cycle with the help of a variable called the node number or sequence number. This variable is computed from the value of the total number of CBRVBR connections sent by the CH. The node then computes the total Equivalent bandwidth of a CBRVBR connection in the network which is given by: 6"! 6 6 $# 6 '% &# The condition for self Connection Admission Control is as follows: ' ( *) 6"! 4,+.- " 3,1 is the " Channel Capacity is the reserved channel capacity which is the minimum allowed for the free channel and lies at the middle for best effort and control packets and is assumed to be a well known value. 3,23 " is the capacity used by the preamble packets. It also includes the capacity wasted due to guard time, which will be introduced shortly. The node after finding that the channel has sufficient capacity to support its application, sends the following information 4 '% to the CH in the form of a packet: ' ( '% 6"! Type of connection (CBRVBR) After the CH gets the message from the node, it updates the total count of CBRVBR connections as well as the equivalent capacity in the network for CBRVBR connections. A MAC acknowledgement message is sent to the node after this. The total duration of transfer specified in the RTS packet takes into account the complete sequence of messages that includes CTS, CAC REQ, CAC REP, TOT BW and ACK messages. If the procedure doesn t get completed in the present cycle, it is continued (not restarted) in the next cycle. This can happen when there are too many sources transmitting in the Free Channel such that this node cannot complete its contention window mechanisms in that cycle. The node can start transmitting its data packets after getting the ACK message from the CH. The duration of time a node is allowed to transmit is given by :<;8 = 6 " '% >? The node, '<: after the cycle starts (as indicated by the 6 packet of the CH at the beginning of the before transmitting its first is given by: cycle), waits 6@AB A@AB for a duration of packet. 6@AB '<: #EDGF HIJ$K2.LNMGOQPSRUTVXW $# 6 1+Z=, if connection type is CBR DQF H[J\KL]MQOGP^RU_ VXW Y 9 +:<;C # 1+"=, if connection type is VBR (1)

5 + + ; 6 # - CycleDur is the duration of the cycle. GuardTime is the time between adjacent slots and is equal to -. Refer Figure 1. PREAMBLE Dur is the duration of the PREAMBLE packet sent by CH. After the first transmission, the node waits every time for duration equal to 6@AB 9 +2: ;C 587 This equation holds good for both CBR and VBR connections. This also includes the guard time and the PREAMBLE packet transmission time by the CH. Thus, all the nodes have their own slots and the waiting time for their slots is also computed independently. Even though it is computed independently, there won t be any overlapping of slots. Since the slot is completely dedicated to a node, there won t be any collision and hence no need of having the RTSCTS, backoff procedures. This will greatly enhance the throughput in the network and the nodes will also get a guaranteed throughput as well. Here, the destination does not send ACK packets for every data packet received. Instead, the source signals the end of its transmission in the cycle by sending a packet. After receiving this packet, the destination sends an ACK packet that also indicates the number of data packets received in this cycle. By doing this, the transmission of the data packets can go at a stretch which will increase the throughput. Because of the collision free transmission, the chances that a destination will not receive a packet sent by the source is less unless the destination moves out of range or the channel is highly error prone. Any acknowledgement scheme can be employed such as the Selective Repeat or Go Back N. The '% computation of 6 also takes into account the overhead due to MAC header and Tx CompleteACK packets. C. Termination of an Application When an application terminates, its slot goes unused. This will create holes in the cycle as well as wastage of bandwidth. This scenario is similar to the concept of holes created during paging and segmentation of the user memory in an Operating System (OS). In the case of an OS, when a hole is created in the memory, coalescing is done to combine the holes and the data together. coalescing the free channel is done by advancing the slot occurence for the connections whose slots occur after the slot of the terminated connection in the cycle to close the gap in the cycle. Thus, the waiting time will be modified only for the next cycle. This operation does not effect the data transmission or the throughput in any way. Thus coalescing of the free channel is not as inefficient as coalescing of holes and memory in an operating system. The detailed protocol operation is explained below: When a node wants to terminate its application it broadcasts a control packet at the beginning of its slot indicating that its connection is closing and it also includes the new equivalent "C capacity in the network. The node computes the new equivalent capacity as follows: 6"! - + 6! 6 The node broadcasts the following information when closing its connection: Type of connection, CBRVBR New Equivalent Bandwidth in the network (6"! '% -2+ ) Equivalent bandwidth of the closing application (6 + ) Sequence number of the closing applicationnode. The CH as well as the other mobile nodes update their database containing the information about the network after getting the above message. The nodes which have CBR type of connections decrement the wait time for their transmission in the next cycle to 9 +:<;8 = 6 : The nodes which have VBR type of connections increment their wait time for their next cycle to # Y 1+ '% 9 +:<;C = 6 where 6 is the equivalent bandwidth of the closed connection. This is done only for the subsequent cycle and only for those nodes whose node number (sequence number) is greater than the node number of the connection that closed. Afterwards the wait time is the same as above, 9 +:<;8 587 The above equations imply that the starting time of the old CBR connection becomes the starting time of the CBR connection after it and the ending time of the old VBR connection becomes the ending time of the connection before it. D. Invitation Mechanism Here we assume that VBR applications are ONOFF sources, so it is possible that a VBR source may not have anything to transmit in its slot, leading to wastage of bandwidth. This bandwidth wastage can be avoided if other VBR or Best-Effort sources can be made to utilize the slot. So, during its OFF periods, a VBR node sends a 6 packet inviting the other nodes to transmit their traffic. The other VBRbest-effort sources grab the channel according to the IEEE standards by following the RTSCTS, backoff procedures. But, the backoff window is computed differently. The backoff window is computed taking the following factors into consideration Nodes transmitting VBR type of traffic will have priority over the nodes transmitting best effort traffic. A node having more packets in its queue should be given more priority. After a node receives the? 6 packet, it performs the following:

6 6 If it has a CBR type of application, it ignores the packet. If it has a VBR type of application, it starts the backoff window procedure for a duration which is obtained picking a random number between [.. ], the value of depends on the occurrence of its slot with respect to the node that has sent the packet as well as on the length of its queue. Essentially, nodes with more packets in their queues andor nodes that have already completed their transmissions in this compute a lesser value of and hence have a better priority to access the channel. This is done to improve jitter. Nodes having UBR type of applications pause for a # tion of and then continue their backoff window procedures from where it left in the free channel. is any small arbitrary number and can be equal to SIFS. This is done to give prefferential treatment to VBR traffic over UBR traffic. As explained before, a node sending best effort traffic can transmit during the free channel at the middle of the cycle and also when it receives a? 6 packet from any VBR source. Thus it is ensured that the channel capacity doesn t go unused in any way. Now, it is also apparent why CBR and VBR sources are grouped together as in Figure 1. This is done for the VBR sources, because the slot duration for the VBR sources is proportional to the equivalent capacity which is not an exact estimate of the traffic generated by them. It is possible that at some point in time, some VBR sources generate data whose rate is higher than the equivalent capacities computed for them. In such cases, the amount of traffic sent by the source in its slot duration is less than the amount of traffic generated. So, when the VBR sources are grouped together, there is a possibility that some sources are in the OFF state whose slot can be made use of by the sources whose rate at that time exceeds the bandwidth (slot) allocated to them. This will prevent the excessive delays as well as the buffer overflow which might have been experienced by the VBR sources otherwise. For CBR sources no special provisions are needed as their slots will always be occupied. VI. SIMULATION Extensive simulation study was done using NS-2. A flat topology was assumed to exist and the nodes were placed in such a way that all the nodes were in the sensing range of each other. The nodes were made to move randomly in a 25X25m area. The MAC protocol used was IEEE and the routing protocol was DSR. Simulation was done for different cases using the default IEEE procedures as well as the new scheme using the Self-Connection Admission Control. Simulation was run for a period of 15 seconds and throughput, mean waiting time, number of packets lost and number of retransmissions are calculated. The channel capacity is 1Mbps. Throughput is measured after every.5 seconds. A. Simulation Study A number of simulations were performed with different traffic sources like CBR, VBR and UBR. Different scenarios were studied by varying the cycle duration, peak rate of the sources and the packet sizes. This sub-section does a comparison of the behaviour of IEEE and the self-cac scheme with emphasis on throughput and mean waiting time for both CBR and VBR traffic in a highly loaded channel. The buffer size for all the nodes is set to 1, bytes. The cycle duration is kept at.1s and the packet size is set to 4 bytes. UDP is employed as the transport layer for carrying the CBR traffic. All the sources start transmitting at the same time. Simulation is performed with 2 CBR sources. One source transmits at a constant rate of 512 kbps and the other source transmits at a constant rate of 256kbps. Figure 2 shows the throughput of all flows. Self-CAC flows obtain 512kbps and 256kbps while flows obtain around 39kbps(instead of 512kbps) and 256kbps. This behaviour of IEEE is due to its backoff, RTSCTS procedures and collisions. For the sources whose arrival rate is high, the number of packets getting queued also increases, which increases the waiting time of a packet as the simulation progresses. But the throughput attained by the self-cac scheme remains nearly at the peak rate of the source throughout the simulation as shown in Figure 2. This is not surprising as the connections are assigned a dedicated slot in which there won t be any collisions. Thus the throughput of a connection, once accepted into the network, doesn t depend on the network load. The throughput of the 512kbps source using the scheme is affected as long as the 256kbps source is active, after which its throughput shoots up as can be seen in Figure 2 after 13s. Throughput (Mbps) Throughput Comparison of CBR Sources Self-CAC, 512kbps flow IEEE 82.11, 512kbps flow self-cac, 256kbps flow IEEE 82.11, 256kbps flow Time (sec) Fig. 2. Throughput Comparison in the presence of 512 Kbps and 256 Kbps CBR Sources Figure 3 shows the waiting time for the packets observed for the first 4 packets. The graph for the self-cac scheme exhibits a uniform saw-tooth figure. This is because the packets that arrived closest to their slots will have to wait for a lesser duration than the packets that arrived farthest from their slots. Packets that arrived after the slot will nearly have to wait for the duration of the cycle to be transmitted in their slot of the next cycle. Thus the mean waiting time for a connection is roughly

7 half the duration of the cycle irrespective of the network load. The curve in Figure 3 for the scheme clearly shows that cannot handle the 768 kbps total input rate. Hence the queues continue to fill. Time (sec) Waiting Time Comparison of CBR Sources Packet Number (sec) Self-CAC Scheme IEEE Fig. 3. Waiting Time Comparison in the presence of 512 Kbps and 256 Kbps CBR Sources Another set of simulations were performed with VBR traffic alone. The size of the packets was set to 386 bytes. An exponential ON-OFF traffic generator was used as the source with mean ON period of 8ms and mean OFF period of 2ms. UDP is employed as the transport layer for carrying the VBR traffic. A total of 11 nodes were made to move inside the flat topology. Five of the nodes acted as VBR sources generating traffic at a constant rate of 128kbps during the ON period and no traffic during the OFF period. Five of the other nodes acted as the destination for the VBR traffic. One node acted as the cluster head. All the sources started transmitting at the same time. Figure 4 shows the throughput comparison for the two schemes for individual connections as a function of time. The Fig. 4. Throughput (Mbps) Throughput Comparison of 128 Kbps VBR Sources Time (sec) Self-CAC Scheme IEEE Throughput comparison of 128Kbps VBR Sources (5 Sources) following can be observed about the measured throughput for the self-cac scheme: Throughput measured over some intervals is zero or less than the equivalent capacity of around 118kbps. This occurs when the source is in an OFF state and not having enough packets in its queue to fill the slot. The measured throughput over an interval is around 118kbps. This equals the equivalent capacity computed by the source as well as the capacity of the slot. This means the slot at the source is completely occupied. The measured throughput at some intervals exceeds 128kbps in spite of them being allocated less than 128kbps. This is because the VBR sources grab the unused bandwidth in other timeslots which occurs when the other VBR sources are in the OFF state. The throughput attained by the VBR source in case of IEEE is much less. The reason is the same as explained for CBR sources. Another set of simulations were performed with UBR traffic. The cycle duration is kept at.25s and the packet size is set to 4bytes. A total of 7 nodes were made to move inside a flat topology. Three of the nodes acted as UBR sources generating traffic at a constant rate of 128kbps. Three of the other nodes acted as the destination for the UBR traffic. One node acted as the cluster head. All of the sources started transmitting at the same time. Since the nodes do not need any QoS guarantee, they do not perform any self-cac procedures. The nodes transmit in the Free Channel as explained before. The throughput attained was observed and was found to be the same for both IEEE and the self-cac scheme. Figure 5 shows the aggregated mean waiting time computed after every 5 packets throughout the simulation for both the cases. It can be seen that the mean waiting time for the self-cac scheme is almost the same as of IEEE This proves that for lower network loads the performance of the self-cac scheme using the contention mechanism is as good as the IEEE scheme. The significance of this result will be explained in Section V1.B.2. Mean Waiting Time (Sec) Aggregate Mean Waiting Time Comparison of UBR Sources Number of Packets self-cac Scheme IEEE Fig. 5. Aggregated Mean Waiting Time Comparison of 128Kbps UBR Sources (3 Sources) Another simulation was performed with a highly loaded channel. A total of 17 nodes were made to move inside a flat topology. Three of the nodes acted as CBR sources sending traffic at 128kbps, three of the nodes acted as VBR sources sending traffic at 128kbps during ON period. The ON and OFF periods were exponentially distributed with a mean of.8s and.2s respectively. Two of the nodes acted as UBR sources generating traffic at a constant rate of 16kbps. The throughput attained by a UBR source is shown in Figure 6. It is seen that the achieved throughput is around 17kbps. This graph shows that the throughput of the UBR sources doesn t get effected in the

8 presence of high priority sources as long as there is sufficient free channel capacity at the middle of the cycle. The modification done to the IEEE scheme whereby the sources perform their contention procedure at the middle of the cycle, doesn t affect the throughput of the nodes. The mean waiting time acheived for the UBR nodes is not very good since transmission is done only during the middle of the cycle. The three CBR and VBR sources have a dedicated slot due to which their throughput and mean waiting time is always good. The throughput achieved by the CBRVBR nodes almost equals the rate of transmission of the source and the mean waiting time depends on the cycle duration. The throughput and mean waiting time pattern exhibited by these sources is the same irrespective of the load in the network and was already discussed before. Fig. 6. Fig. 7. Throughput (Mbps) Throughput (Mbps) Throughput of a UBR source at High Load Time (Sec) Throughput Throughput of a UBR Source at High Network Load Offered Load vs Throughput Offered load vs Throughput B. Performance Analysis Offered Load IEEE Self-CAC Scheme 1) Offered Throughput: For providing Quality of Service, throughput guarantees should be provided to connections throughout their duration irrespective of the load in the network. Figure 7 shows the throughput vs normalized offered load for an individual connection on a 1Mbps channel. Nodes transmit CBR traffic at a rate of 64kbps. The number of nodes is increased and throughput is observed for both the IEEE and self-cac scheme. The size of the packets is set to 4 bytes and the cycle duration for the self-cac scheme is set to.25s. It is seen that when the load on the network crosses 6 percent, the throughput per connection for the IEEE scheme drops. But, the throughput of the connection for the self-cac scheme remains at nearly 64kbps irrespective of the load in the network. For the self-cac scheme, new connections are accepted as long as the load is less than.9. New connections are rejected when the load exceeds.9. For all connections that pass the admission control, throughput guarantee is given. The remaining 1 percent of the capacity is the overhead incurred because of the self-cac scheme and is discussed in the later sections. But in IEEE like any other contention protocol, the throughput of all the connections drops when the load on the network increases, making it very difficult to achieve QoS guarantees. Most of the existing MAC protocols in the literature as well as IEEE 82.11e suffer from this scalability issue, wherein no provision is done to protect the QoS of the nodes when the load in the network exceeds beyond a certain value. 2) Cycle Duration: The cycle duration is one of the key parameters for this system. If the cycle time is set too low, a large portion of the cycle time will be spent on overhead bits. If the cycle duration is set to high, the delay for each packet might be too large. Therefore, the tradeoff is between lost bandwidth due to overhead versus delay. It was seen in the last sub-section that once a flow has entered the network, throughput guarantee can be given to the flow as long as the flow is active. It was also seen in the previous sections that the mean waiting time observed for a connection is approximately half the duration of the cycle. Of course, the duration of the cycle can be reduced in order to get a better mean waiting time for the connections. But reducing the cycle duration increases the overhead proportion incurred due to preamble packets. This alone is not a major limitation as the size of the preamble packets is very small (6 bytes). But, the lesser the cycle duration, the lesser the slot duration for a node in the cycle. This may increase the amount of fragmentation done, thereby increasing the overhead due to the MAC header. The equivalent capacity calculation also takes the overhead due to the MAC header into consideration. Thus, reducing the cycle duration will reduce the size of the packets or increase the amount of fragmentation, both of which increases the MAC overhead. This will reduce the capacity available for transmitting useful data which in turn will increase the rejection ratio or reduce the number of connections that can be accepted into the network. The overhead incurred in this scheme is due to the preamble packets that are transmitted at the beginning of every cycle and the guard time between the slots of every connection. The more the number of users, the more the amount of guard time between slots and more the capacity wasted. The capacity wasted due to guard slots also increases with decreasing cycle durations. Figure 8 shows the graph which plots the overhead and mean waiting time vs the cycle duration. The overhead calculated is due to the preamble packets and the guard time as mentioned above for twelve users generating traffic at a constant rate of 64kbps. Both the curves seem to intersect when the cycle duration is between.4s and.5s, but this is to some

9 degree just an artifact of scales chosen for the plot. Possibly a more important consideration is that the curve for the overhead has flattened by this point. The mean waiting time shown is for a single 64kbps CBR source. As mentioned, this shows that the mean waiting time for a connection largely depends on the duration of the cycle time choosen. When the cycle duration is.25s, the mean waiting time acheived by the connections is around 12ms which is very good for voice traffic [1]. It was found that the mean waiting time in case of IEEE is less when the load is light. When the load increases to around 65 percent, the performance drops and waiting times grow without bound. Fig. 8. Overhead (kbps) and Mean Waiting Time (sec 1) Comparison of Cycle Duration vs Overhead and Mean Waiting Time Cycle Duration vs Overhead Cycle Duration vs Mean Waiting Time Cycle Duration (sec) Comparison of Cycle Duration vs Overhead and Mean Waiting Time A major benefit of this approach is that this scheme could be implemented to use the best of both self-cac and IEEE The node in the self-cac scheme has the option of reserving bandwidth or can occupy the free channel at the middle using the IEEE contention procedures. When the load in the network is light, the duration of the free channel is large and thus the nodes can attain their required throughput and mean waiting time without doing any reservation. This was evident in Figure 5. When the load increases, sources can switch to slotted reservations. Thus it is possible to use the best of both schemes depending on the state of the network. During low loads, the contention based approach would be used. 3) Utilization: Another important performance measure is the utilization of the network which is the total fraction of the capacity utilized for sending useful data. The factors that reduce the throughput of the system are the number of control packets, backoff schemes etc, since during those periods data packets are not transmitted. This section analyzes the utilization of the network for the self-cac scheme. The duration of the slot reserved is proportional to the equivalent capacity of the application. For a CBR source, the equivalent capacity is the peak rate of the source. The slot for a CBR source is always occupied and hence the utilization of the network during a CBR slot is high. But for a VBR source, the equivalent capacity computation is not always exactly equal to the traffic transmitted by the source and it has some approximations in it [23]. The same was also mentioned in the previous sections. There is a possibility that the slot may go idle due to which the channel capacity may go wasted. As already mentioned, when a VBR node doesn t have anything to transmit in its slot, it invites the other nodes to send traffic in its slot. The other VBR nodes compete among themselves for the slot. Figure 9 shows the mean waiting time comparison of a connection with and without this scheme. The figure shown is for five VBR sources transmitting at 128kbps. The cycle duration is set to.25s. It can be seen that the achieved mean waiting time is better using this invitation scheme. The achieved throughput for the connection was found to be almost the same with and without the scheme proving the effectiveness of the equivalent capacity computation. The mean waiting time achieved is significantly better using the invitation scheme. This is because when the source is in the ON state, it transmits at its peak rate. It is known that the channel allocated to a VBR source is less than the peak rate of the source and is equal to the equivalent capacity computed in [23]. So, when the source is in the ON period, the queue starts to build slowly. When the source goes to the OFF period, the packets in the queue gets drained. Aggregate Mean Waiting Time (sec) Mean Waiting Time Analysis of Self-CAC Schemes With Invitation Scheme Without Invitation Scheme Number of Packets Fig. 9. Mean Waiting Time comparison with and without the Invite Packet Scheme When the VBR sources are made to occupy the unused slot, the throughput and mean waiting time improves dramatically. This is because the packets that were not transmitted in the current cycle do not have to wait till the next cycle. When another VBR source is in the OFF state, this source has the oppurtunity to grab the channel and transmit the remaining packets in the queue. Thus the queue doesn t build up as much as in the case without the invitation scheme. Figure 1 shows the bandwidth wasted with and without this scheme for five VBR sources transmitting at 128kbps. The cycle duration is set to.1s for this case. The figure shown is the bandwidth wasted due to a single connection. The amount of bandwidth wasted is measured after every second, which in this case is ten cycles. This result showing a better bandwidth utilization using the invitation scheme is not surprising. This is because, without the invitation scheme, the whole slot will go unused when a VBR source does not have anything to transmit in its slot. The interesting thing to note is that there is some bandwidth wastage even in the presence of the invitation scheme. Let s see why this happens. Bandwidth wastage occurs due to the following reasons: Insufficient number of packets in the slot, but sufficient enough to fill the slot in such a way that the remaining

10 : D Bandwidth Wastage (bps) Without Invitation Scheme With Invitation Scheme 1 Goodput vs Load Comparison Self-CAC (CBR) IEEE82.11 (CBR) Bandwidth Wastage (bps) Goodput Time (sec) Fig. 1. Comparison of bandwidth utilization with and without the Invite Packet Scheme Fig Load Goodput Comparison of CBR Traffic duration in the slot is not sufficient for the other nodes to utilize effectively. Occurence of collission when more than one node tries to access a slot left by a VBR source which is in the OFF state. A source will invite the other nodes to transmit only if the remaining slot has sufficient time to transmit a pre-determined number of bits in its slot. 4) Goodput: Goodput is defined as the percentage of total channel capacity utilized for sending data traffic. The control packets, guard slots, header overhead and anything that utilizes the channel for not sending the data packets are omitted from the goodput computation. The cycle duration for the goodput simulation is kept at.25s. Figure 11 shows the goodput achieved for CBR type of sources. It can be seen that the goodput achieved in case of self-cac scheme goes to as high as 8 percent when the load in the network is slightly more than.94. The new connections will be rejected after the load in the network crosses.94. The goodput for the IEEE scheme doesn t exceed 65 percent. Another simulation is performed with a equal number of CBR, VBR and UBR sources. The load in the network is increased and the goodput is observed for the self-cac and IEEE scheme. Figure 12 shows the goodput achieved for both the schemes with 95% confidence interval. It can be seen from the graph that the achieved goodput using the self-cac scheme is better. For the self-cac scheme, the achieved goodput for mixed traffic is lesser than the achieved goodput for CBR sources alone. This is due to the presence of VBRUBR traffic and the approximation done for the equivalent capacity computation of VBR traffic. The factors that affect the goodput are Overhead due to header Overhead due to preamble and guard slots Overhead due to self-cac scheme Overhead due to invitation scheme The expressions for the above parameters are as follows: Let the overhead due to MAC header be x bits. Overhead due to MAC headers in a cycle is given by: 4 B *, - = Goodput Fig Goodput vs Load Comparison Load Self-CAC (Mixed) IEEE82.11 (Mixed) Goodput Comparison of CBR, VBR and UBR Traffic is the total number of connections requiring QoS is the average number of packets for the 1 connection in a cycle and is given by: 6"! 3 6"! " =1 9 +2: ;C = 3 + " is the equivalent capacity of 1 connection + is the packet size of 1 connection Overhead due to preamble, guard slots and MAC acknowledgement packets is given by: 4 H L 5 T 5 W! V#"$&%L * DQF HI J\K B # B ) & is the time taken for the acknowledgement procedures in self-cac, which includes the time taken for sending the transmission complete and the acknowledgement messages. This procedure is done once per slot per cycle. +* B, HI is the time taken for sending a PREAMBLE packet -' is the duration of DIFS is the channel capacity Overhead due to admission control assuming that no collission of RTS packets occur during admission control, is given =1: (' =1