Email: {torok, vajda}@ikti.hu ... primary goal of building cheap and simple Wireless Personal ... demands, such as broadband and Quality of Service (QoS).
Techniques to Improve Scheduling Performance in IEEE 802.15.3 based ad hoc networks Attila T¨or¨ok, L´or´ant Vajda
Attila Vid´acs, Rolland Vida
AmI Project Group, Bay Zolt´an Foundation for Applied Research Email: {torok, vajda}@ikti.hu
Dept. of Telecommunications and Media Informatics Budapest University of Technology and Economics Email: {vidacs, vida}@tmit.bme.hu
Abstract— In this paper we propose new techniques to improve scheduling performance in time-slotted superframe based ad hoc networks. Such channel access technology is used in the IEEE 802.15.3 and 802.15.4 standards. Building on the performance analysis of previous proposals, we enhance the scheduling algorithms with flow state signaling and burst eligibility decision, so as to exploit the features of the 802.15.3 architecture. We show by simulation that the scheduling algorithms extended with these special mechanisms achieve higher performance, with better channel utilization and lower power consumption.
I. I NTRODUCTION Nowadays, the wide range of emerging wireless applications triggered the spread of specialized wireless protocols. Bluetooth represents such a specialized solution, with the primary goal of building cheap and simple Wireless Personal Area Networks (WPAN). Targeting applications with different demands, such as broadband and Quality of Service (QoS) requirements, IEEE 802.15.3/4 [1] opts to use a different channel access technology. Similarly to Bluetooth, both protocols are also based on a centralized and connection-oriented ad-hoc networking topology, with a master-slave hierarchy. Nevertheless, unlike Bluetooth, the master node (called PNC) handles only admission control, scheduling and management tasks, without being involved in packet forwarding. The MAC layer of these 802.15.3/4 protocols employs a time-slotted superframe structure, constituted from beacon, optional channel request, and data transmission parts. The beacon is used to carry information about channel time allocations. The channel request is contention-based (called Contention Access Period - CAP), and is used by the nodes to send their requirements to the PNC. Using the gathered information, the PNC schedules the time-slots for the next superframe. For data transmissions time-division multiple access (TDMA) is used. As several papers point out, a dynamic slot reservation MAC protocol with a proper scheduling algorithm is important in such an environment. In [2] the authors developed a general framework for the max-min scheduling problem in static wireless networks. In [3], different slot allocation algorithms are presented for an infrastructure based contention-free MAC protocol (EC-MAC). The authors in [4] proposed a probability model for the channel request part, and a basic scheduling model based on graph-coloring. A simulation-based MAC modeling for 802.15.3 networks is presented in [5]. With the introduction of hierarchical superframes, in [6][7] better channel utilization is achieved for 802.15.3 networks. A novel MAC protocol for WPANs is proposed in [8], with a proper
scheduler and a piggybacking-based state information signaling method. As far as we know, the only work that deals in some degree with state information signaling for scheduling in 802.15.3 networks is [9]. In this paper we investigate performance issues related to scheduling algorithms used in 802.15.3 networks and we propose techniques to improve multimedia scheduling. The scheduling algorithms that we analyze in details are based on the Earliest Deadline First (EDF) [10] and on the Shortest Remaining Processing Time (SRPT) [11] algorithms. SRPT selects for service the pending job in the system with the least remaining service time. The policy is preemptive, so that if a new job arrives with a smaller service time than that remaining for the job currently in service, the scheduler switches immediately to service the newly arriving job. EDF is a dynamic priority scheduler, where the prioritization is based on job arrival times. In EDF the priority of a job increases with the time it spends in the system. By analyzing these algorithms, we have found some drawbacks that lead to network resource wastage. We extend the scheduler algorithms with special state-information signaling and burst eligibility decision. As we show in the paper, better channel utilization is achieved, avoiding the underutilization of the superframes. We demonstrate through simulations that the extended schedulers outperform the previous ones in terms of channel utilization, power-saving, and signaling efficiency. This paper is organized as follows: Section II provides a brief overview of the existing scheduling solutions based on state information signaling, highlighting their lack of efficiency in 802.15.3 networks. In Section III we propose mechanisms to increase scheduling efficiency, while Section IV provides the performance analysis. Finally, Section V concludes the paper. II. E XISTING SOLUTIONS FOR STATE INFORMATION SIGNALING
In the following we overview existing techniques for state information signaling to improve the performance of scheduling algorithms in time-slotted channel-based wireless networks. We consider delay-sensitive (rt-VBR) traffic type sources, modeled with batch arrival processes. The flows consist of packet bursts, each burst being represented by the number of packets that constitute the burst and the packets’ generation time. For bandwidth allocations in a superframe, 802.15.3 uses Guaranteed Time Slots (GTS) [1] with static or dynamic
0-7803-9415-1/05/$20.00 (C) 2005 IEEE This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the IEEE GLOBECOM 2005 proceedings.
position. Due to its characteristics, optimal scheduling of rtVBR traffic requires allocation of variable length GTS. Timeslots are wasted if a flow does not have packets to transmit but the scheduler (implemented in the PNC) still allocates GTS for it. Thus, to achieve better resource utilization, the scheduler must be informed about the internal state (e.g., instantaneous queue length, packet arrival rate) of mobile nodes. Its performance highly depends on how fast and accurately the state information can be obtained from the nodes. Nodes in wireless ATM networks use piggybacking techniques to transfer the internal state information [12]. The base station forwards all the data packets of the communicating nodes; thus, it can analyze the piggybacked state information. In [9], a similar technique is used for 802.15.3 networks. Here the piggybacked internal state is the number of packets in each queue. These solutions involve an excessive overhead to transmit the information in a timely manner, especially in case of traffic with highly dynamic and delay sensitive nature (rt-VBR flows). On the other hand, receiving every packet overwhelms the PNC in terms of processing power and energy consumption. In C-FD3R wireless ATM MAC [13], in order to minimize the signaling overhead the authors use the residual lifetime of the rt-VBR flows as a parameter to determine the transmission time of reservation requests. Only the number of requested slots and the number of superframes after which the deadline expires is sent to the base station. Each burst contains this piggybacked information about the reservation request of the subsequent burst in the queue. The drawback of this solution is the over-simplified modeling of traffic characteristics. The authors suppose that there are always cells to transmit in the buffers; thus, the case when the subsequent reservation cannot be propagated to the base station, due to buffer depletion (we call silent period), does not occur. As can be seen, none of the presented solutions is efficient in 802.15.3 networks. Piggybacking-based solutions are not good in systems where the PNC participates only in scheduling and piconet management, but not in packet forwarding. In C-FD3R MAC the amount of piggybacked information is reduced, but the traffic model is over-simplified. During a silent period, when the queue of the flow is empty, it should be allowed for the node to signal an upcoming burst. We consider that a contention based solution (in CAP) would introduce different problems (e.g., significant delay at high network load, hidden terminal problem); therefore, it is not suitable for signaling. III. T ECHNIQUES TO IMPROVE SCHEDULING In this section we propose techniques to improve scheduling performance in 802.15.3 networks. The system is enhanced with special flow state signaling and burst based eligibility decision. With these, higher scheduling performance with better channel utilization can be achieved. A. Flow-state signaling using control packets In order to avoid the excessive overhead caused by piggybacking, we propose to use special control (CTRL) packets
(see Fig. 1); they are sent by the nodes only when it is necessary to update the internal state information of the PNC. First, we present the kind of information that should be conveyed in the CTRL packets. Then, we investigate how CTRL packet time-slots must be allocated. Flow state signaling can be reduced by burst aggregation, which we discuss at the end of this subsection. 1) Information sent in CTRL packets: In [9] bandwidth wastage could appear due to a wrong scheduling decision. Sending just an arbitrary number of packets from each flow can lead to situations when half of a burst is scheduled, while the other half is left in the queue. This could result in massive packet dropping. As opposed to this, sending the first burst’s lifetime and length could be used for an optimized scheduling; the PNC should first allocate only the necessary amount of time-slots (one burst) for each flow.
Fig. 1.
Possible burst behavior scenarios
However, the overall queue size of a flow is also a useful information that can optimize the scheduling. When there is more than one burst in the node’s queue, and the superframe is not overloaded, it is advisable to schedule also the remaining packets from the queue. Thus, the information sent in the CTRL packets is: • lifetime: the residual lifetime of the first burst. It is calculated at the sender node and it is relative to the end of the superframe when the burst was generated. It represents the deadline in number of superframes; • nr. of packets: the number of packets in the first burst; • queue size: the overall queue size of the flow. 2) Time-slot allocation for CTRL packets: Allocating timeslots for CTRL packets will lead to a certain signaling overhead. To reduce it we classify the flows into different states. According to the flow states, the control information needed for the PNC will be different. At any given time, the flows in the scheduler can be in one of the following states (see Fig. 2): Blind state: In this state the PNC does not have valid information about the flow. This state follows the initial flow admission (CAC) or the transfer of a burst. As it was previously pointed out, a time-slot for CTRL packets must be reserved when there is a silent period of the respective flow. This state lasts until a new burst is signaled; it is then followed by the starting state. Starting state: Is triggered when the PNC received information in a CTRL packet. While the flow is in the starting state, it does not get any time-slot for CTRL packets, as the PNC has already all the information necessary for scheduling. Until the deadline expiration of the current burst there is no need for information about the next burst.
This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the IEEE GLOBECOM 2005 proceedings.
Middle state: The flow enters the middle state after the PNC has scheduled it, and the flow’s packets are started to be transmitted. At the end of this state (last packet of the burst), a time-slot is allocated for a CTRL packet. A possible scenario is illustrated in Fig. 1. If there is a new burst in the queue at the end of this state, the flow switches back to the starting state. If not, i.e., there is a silent period of the sender, the flow enters the blind state.
Fig. 2.
as it always selects the burst with the smallest length, not taking into consideration the residual lifetime. If, during the serving of a burst, a new burst with smaller length appears, the scheduler switches and starts to serve the newly arrived burst. Thus, the shorter bursts have higher probability for successful transmission. In both cases a flow with partially transferred bursts will also affect the other flows (inter-flow effect), which causes bandwidth wastage.
Flow states at the scheduler
3) Burst aggregation in CTRL packets: Until this point, only the case of one burst per superframe was taken into consideration. Nevertheless, in certain cases it can happen that during one superframe more bursts of the same flow are generated. In such cases, sending information only about the first burst can lead to the other bursts timing out, as without the additional information the scheduler will presume that there is enough time to serve the next bursts. To prevent this, we use burst aggregation. In case of burst aggregation, each burst generated during the current superframe gets the same residual lifetime. (The residual lifetime of the bursts is calculated relatively to the end of the ongoing superframe, when the burst is generated.) The number of packets signaled in CTRL packets is the sum of the packets of the aggregated bursts. By using aggregation, signaling overhead is reduced; in the meantime, the deadline limit of the bursts is kept under control.
Fig. 3.
The shadow effect
A partially transferred burst can also affect the following burst of the same flow (intra-flow effect), when these two have a high-paced arrival pattern. As it was presented above, until the flow is in starting state, it will not get any CTRL timeslots. This leads to the appearance of the shadow effect: when a new burst arrives after the first burst was already signaled but before it’s lifetime expired, the second burst will not be signaled to the scheduler until the first burst is transmitted, or it turns out that it must be dropped.
B. Burst transfer eligibility decision (BTED) Until now we considered cases when a burst follows the previous one only after the first burst’s deadline expired. However, in certain scenarios it can happen that bursts are arriving with a higher pace. In a highly loaded network wrong scheduling decisions can occur if the scheduler does not get information about these new bursts soon enough. This situation leads to the appearance of the shadow effect, described in the next subsection. Then we describe our proposal to reduce the shadow effect using the BTED mechanism. 1) The shadow effect: Bandwidth can be wasted when a burst is scheduled, but some of its packets cannot be transmitted due to some reason. One such reason could be the residual lifetime expiry of the burst. In this case the partially transferred burst unnecessarily occupies the channel. This can happen, for example, in the case of deadline based scheduling (EDF), when the scheduler selects a burst based on its residual lifetime. If there is congestion, the scheduler makes bursts with high residual lifetime to wait until their lifetime becomes the smallest one. Therefore, in many cases the residual lifetime of longer bursts will expire before the last packet can be sent. SRPT suffers less from bandwidth wastage than EDF,
Fig. 4.
BTED - Decision 1
Fig. 3 presents the shadow effect for a flow with two consecutive bursts. The scheduler learns about the first burst and enters in the starting state; thus, it does not get information about the second burst. It can be seen that the length of the shadow interval depends on the network load, the lifetime of the first burst, etc. The problem appears when at the end (6th superframe) it turns out that the first burst cannot be transferred entirely. Unfortunately, until this moment there is no information about the second burst. In such a case both bursts will be dropped with high probability. 2) The BTED algorithm: In order to prevent the above presented problems, we propose a mechanism called burst
This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the IEEE GLOBECOM 2005 proceedings.
transfer eligibility decision (BTED). Using this mechanism the scheduler first determines if the burst is eligible, i.e., if all its packets can be transmitted; the burst is scheduled for transmission only afterwards. For better decision making, we use a two level decision process (see Fig. 5a). The first level of eligibility decision (Decision 1, see Fig. 4) is applied at the beginning of each superframe construction, for every flow which is in starting state. The purpose of this process is to eliminate, as soon as possible, the flows which already have no chance of sending their entire first burst. During the decision the possible bandwidth demand until the deadline expiration of the respective burst will be estimated. In the calculations the different flow priorities and the possible amount of CTRL time-slot allocations are considered. This is a best-case type of decision, so it underestimates the possible system load in order to avoid burst dropping due to wrong decisions. Its accuracy increases each time when it is applied. The second level of eligibility decision (Decision 2) is similar to the first one and is used at the begining of the middle state. Before actually selecting and scheduling a burst, it is checked again whether its lifetime limit would be violated or not. This second decision level can be more accurate than the first one, since the burst in question will have the highest scheduling priority. Therefore, only the CTRL time-slot allocations must be estimated for the upcoming superframes.
Fig. 5.
Effect of BTED on shadow
The flows whose bursts cannot be scheduled due to lifetime expiry are announced by the PNC in the upcoming beacon; a CTRL time-slot is also allocated for each of them. Based on the announcement, the sender node is able to drop the ineligible packets and send its new state information. By using this technique, the PNC gets fresh information about the state of the flow; thus, more efficient scheduling is possible. The BTED mechanism can also be used to reduce the shadow effect. Fig. 5b presents our case when the first burst is rejected by BTED. In this superframe a CTRL time-slot is allocated for the respective flow. The sender is able to send soon enough CTRL information about the second burst; thus, the PNC can schedule it before its lifetime expires. IV. P ERFORMANCE ANALYSIS This section presents the investigations on how the presented techniques increase the performance of IEEE 802.15.3 systems. We apply these techniques for the SRPT and EDF
algorithms in the same simulation scenarios, and evaluate the resulting differences in system performance. In the following we call the extended versions of the scheduling algorithms SRPT+T and EDF+T, respectively. A. Simulation scenario Simulation results were obtained using ns-2 [14], with the 802.15.3 module presented in [9]. We implemented the proposed techniques and enhanced the simulator in order to support our analysis demands. The simulation topology consisted of 10 nodes (9 slaves and 1 PNC) organized in one piconet, which remained stationary during all the simulation runs. Each simulation was run 20 times, for a duration of 200 seconds. To provide a realistic nature to our investigation, on the one hand we used 8Mbps real-time video flows as traffic patterns generated by an MPEG4 traffic generator that used the Transform Expand Sample (TES) method [15]. On the other hand, a worst case analysis is also presented, where the traffic pattern is different from the realistic one (see Section IV-D). We assumed the wireless channel to be ideal; therefore, no noise, distortion, and other interference is present. The physical layer’s data rate is 100Mbps. In simulation runs the superframe length is assumed to be 4ms; the rt-VBR burst generation interval is 30fps, with a burst deadline of 33ms. The EDF and SRPT schedulers were used with the original internal state signaling presented in [9], which was based only on queue size information. On the contrary, SRPT+T and EDF+T used both deadline, packet number and queue size as state signaling information. B. Performance metrics The first and most obvious aspect in network characterization is the efficiency of channel utilization. Thus, during the performance analysis of the proposed techniques we considered this parameter as being the most important one. The main goal of all simulation runs was to analyze the service quality provided in the case of an increase in network load. In order to analyze the overall system performance, we used the following performance metrics: Job Failure Ratio (JFR) is a drop ratio metric; all the bursts dropped because of an expired deadline are characterized by this metric. For an effective channel utilization the JFR should be as low as possible. Response Time (RT) is the time between passing a packet from the upper layer to the MAC layer, sending it, and receiving back a MAC layer acknowledgement. This metric is measured only for the successfully transferred bursts. Therefore, there can be cases when the system presents a good RT value, while JFR is very high. Consumed Energy is the amount of consumed energy for every packet the PNC receives. This metric basically depends on the packet size to be received, which affects the usage time of the radio receiver of the PNC. C. Analysis results As the network load increases, the JFR shows an increasing tendency in all cases (see Fig. 6). The particularly bad behavior
This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the IEEE GLOBECOM 2005 proceedings.
of EDF can be explained by the fact that, during scheduling the bursts chosen to be sent have their residual lifetime almost expired. This leads to the previously presented problem, called partially transferred bursts. By constantly switching the serving among bursts if a smaller one arrives, SRPT avoids at a certain level the burst dropping. On the other hand, in EDF if a burst is started to be served, other bursts are not allowed to be scheduled. Therefore, the probability to have a partially transferred burst increases. It is important to observe that using the EDF+T algorithm the system performance is highly improved in terms of JFR. Comparing the results of the SRPT+T and EDF+T algorithms, relevant performance difference can be seen under medium network load. The difference between these two shows a significant performance increase in favor of EDF+T. (e.g., under 72% network load SRPT+T: 6.3%; EDF+T: 2.3%)
Fig. 6.
Fig. 7.
Response Time - RT
the other algorithms, it does not violate the maximum allowed burst deadline; therefore, it is still suitable for real-time traffic. In the following, we investigate the energy consumption of the PNC nodes if piggybacking or CTRL packet allocation strategy is used for state information signaling. If a simple piggybacking technique is used (EDF, SRPT) the signaled information is located in the header of each packet. Meanwhile, if SRPT+T and EDF+T are considered, the status information is coded and sent only in the CTRL packets, the header of the Data packets remaining unmodified. It turned out (Fig. 8) that both algorithms present an increasing energy consumption as the network load is increased. This tendency is mainly due to the amount of the generated status information.
Job Failure Ratio - JFR
In case of the SRPT and EDF algorithms the BTED mechanism is missing, which concerns the scheduling of those bursts that cannot be fully transferred before their residual lifetime expires. It is interesting to note that there is no significant performance increase when the BTED mechanism is used in SRPT+T. This can be attributed to the switching behavior of SRPT, which will decrease the accuracy of the BTED decisions. By analyzing these results, it can be seen that BTED is an important performance increasing mechanism in terms of channel utilization. Fig. 7 presents the Response Time (RT) measurements for the basic and the enhanced scheduling algorithms. As we can see, the basic SRPT algorithm presents an almost constant RT value, in spite of the network load increase. If we apply the SRPT+T scheme, the RT can be even more reduced. Meanwhile, if the EDF or especially the EDF+T algorithms are applied, the RT shows an increasing tendency. However, we have to keep in mind that RT is calculated only for successfully transferred bursts. With EDF and SRPT, under high network load the JFR is also high; however, those bursts that are not dropped have to wait only for a short time in the queue. In EDF+T the JFR is lower, more bursts are transmitted, but some of them have to wait for a longer period of time. Nevertheless, even if the EDF+T algorithm presents higher RT values than
Fig. 8.
Energy consumption at PNC
We note that the energy consumption in case of piggybacking is much higher than in case of CTRL packet signaling. Piggybacking does not use states for flows; instead, it sends the status information in every superframe, even if there is no change in the flow’s internal status. Therefore, the radio receiver of the PNC is used with higher frequency, which implies a higher energy consumption. D. In-depth analysis of the proposed techniques The results of this subsection represent a worst-case analysis. In this way we have analyzed the effectiveness of the proposed mechanisms (enhanced control information signaling and burst aggregation, BTED) and the interaction between them (see Section III). During all the simulation runs only one flow was applied; it had a constant burst size and a periodic arrival pattern.
This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the IEEE GLOBECOM 2005 proceedings.
The inter-arrival time between successive bursts is uniformly distributed in [0, tmax ]. In order to achieve different network loads tmax is increased with steps of half of the length of the superframe. In this way the smaller tmax is applied, the higher the network load is. From the JFR graph presented in Fig. 9 we can observe where the different techniques influence the performance of the scheduling algorithm. As far as the SRPT and EDF algorithms are concerned, at high burst intensity, the system has a 100% JFR. Meanwhile, if SRPT+T and EDF+T are applied at the same traffic load, a performance increase can be observed. It can be seen that the enhanced algorithms deliver the best results if both the BTED and the burst aggregation mechanisms (BTED+AGGR) are applied. We can also see that at high burst intensity it is important to use the burst aggregation mechanism. Meanwhile, as the burst intensity is decreasing, the algorithms (BTED+AGGR and BTED+noAGGR) present similar results. If no BTED mechanism is applied (noBTED+noAGGR, noBTED+AGGR) it becomes indifferent if the bursts are aggregated or not. From the graph it can be seen the importance of knowing the queue size state information at the scheduling. Even if both BTED and the burst aggregation mechanisms are used, the performance of the algorithms is heavily decreased if no queue size information is sent (BTED+AGGR+noQSinfo). Since there will be numerous bursts is the sender’s queue, without the overall queue size information the PNC must be informed about each burst in a separate CTRL packet. This will induce delayed signaling and wasted time-slots in the superframes; thus, many times the burst deadlines will expire. If the queue size information is conveyed, the PNC can use it to schedule more than one burst and avoid the waste of time-slots. As a conclusion, we can say that the BTED mechanism is useful in an overall manner, while the burst aggregation method should be applied at high burst arrival intensity. It seems that the queue size state information of the flows is mandatory in such cases.
Fig. 9.
JFR for one flow
V. C ONCLUSION In this paper we investigated the scheduling issues in 802.15.3 networks, IEEE’s emerging wireless standard. Besides the used scheduling algorithm, the performance of such a system is highly dependent on additional mechanisms, such as state information signaling. However, the existing mechanisms are not efficient in 802.15.3 networks. We proposed a special flow state signaling and a burst eligibility decision mechanism in order to increase the efficiency of the scheduling. Using the ns-2 network simulator, we analyzed the performance increase of the proposed techniques and compared them to the previous solutions. During the analysis, different performance metrics were investigated, showing that the schedulers enhanced with the proposed techniques outperform the previous solutions in terms of channel utilization and power efficiency. Future work includes the improvement of the BTED mechanism to better cope with SRPT, and further optimization of the CTRL time-slot allocation strategies. R EFERENCES [1] IEEE 802.15 working group for Wireless Personal Area Networks, http://grouper.ieee.org/groups/802/15/. [2] A. Rajeswaran, G. Kim, R. Negi, A scheduling framework for UWB and cellular networks, in: Proc., 1st First International Conference on Broadband Networks (BROADNETS’04), San Jose, California, USA, 2004. [3] A. Rangnekar, K. Sivalingam, Multiple access protocols and scheduling algorithms for multiple channel wireless networks, in: Handbook of Algorithms for Mobile and Wireless Networking and Computing, Azzedine Boukerche, Ed., 2004. [4] Y.-H. Tseng, H.-K. Wu, C. Gen-Huey, Maximum traffic scheduling and capacity analysis for IEEE 802.15.3 high data rate MAC protocol, in: Proc., IEEE Vehicular Technology Conference VTC’2003, Orlando, Florida, USA, 2003. [5] K.-W. Chin, D. Lowe, Simulation study of the IEEE 802.15.3 MAC, in: Proc., Australian Telecommunications and Network Applications Conference (ATNAC), Sydney, Australia, 2004. [6] L. Vajda, A. T¨or¨ok, L. Kyu-Jung, J. Sun-Do, Hierarchical superframe formation in 802.15.3 networks, in: Proc., IEEE International Conference on Communications (ICC), Paris, France, 2004. [7] A. T¨or¨ok, L. Vajda, Y. Kyu-Jung, J. Sun-Do, Superframe formation algorithms in 802.15.3 networks, in: Proc., IEEE Wireless Communications and Networking Conference (WCNC), Atlanta, Georgia, USA, 2004. [8] R. Cusani, M. Torregiani, F. D. Priscoli, G. Ferrari, A novel MAC and scheduling strategy to guarantee QoS for the new-generation WINDFLEX wireless LAN, IEEE Wireless Communications Magazine 9 (3) (2002) 46–56. [9] R. Mangharam, M. Demirhan, Performance and simulation analysis of 802.15.3 QoS, IEEE 802.15-02/297r1 (Jul. 2002). [10] C. Lu, J. a. Stankovic, G. Tao, S. H. Son, Design and evaluation of a feedback control EDF scheduling algorithm, in: Proc., 20th IEEE RealTime Systems Symposium, Phoenix, AZ, USA, 1999, pp. 56–67. [11] Z. Shao, U. Madhow, QoS framework for heavy-tailed traffic over the wireless Internet, in: Proc., IEEE Military Communications Conference (Milcom), Vol. 2, Anaheim, CA, 2002, pp. 1201–1205. [12] N. Passas, L. Merakos, D. Skyrianoglou, F. Bauchot, G. Marmigere, S. Decrauzat, MAC protocol and traffic scheduling for wireless ATM networks, Mobile Networks and Applications 3 (3) (1998) 275–292, iSSN:1383-469X. [13] C. G. Kang, C. W. Ahn, K. H. Jand, W. S. Kang, Contentionfree distributed dynamic reservation MAC protocol with deterministic scheduling (C-FD3R MAC) for wireless ATM networks, IEEE Journal on Selected Areas in Communications 18 (9) (2000) 1623–1635. [14] http://www.isi.edu/nsnam/ns/. [15] A. Matrawy, I. Lambadaris, C. Huang, MPEG4 traffic modeling using the transform expand sample methodology, in: Proc., 4th IEEE International Workshop on Networked Appliances (IWNA4), Gaithersburg, MD, 2002.
This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the IEEE GLOBECOM 2005 proceedings.