Providing QoS Guarantees to Multiple Classes of Traffic in Wireless Sensor. Networks. M. Y Aalsalem, Javid Taheri, Mohsin Iftikhar and Albert Y. Zomaya.
Providing QoS Guarantees to Multiple Classes of Traffic in Wireless Sensor Networks M. Y Aalsalem, Javid Taheri, Mohsin Iftikhar and Albert Y. Zomaya School of IT, University of Sydney 2006, NSW, Australia {aalsalem, javidt, mohsinif, zomaya}@it.usyd.edu.au Abstract different treatment from the network in terms of guaranteed QoS (Quality of Service). Queueing and scheduling have a direct impact on QoS characteristics. There are different types of queueing tools that have been developed to provide different services to heterogeneous traffic classes such as Priority Queueing (PQ), Custom Queuing (CQ), Weighted Fair Queueing (WFQ), Class Based Weighted Fair Queueing (CBWFQ) and Low Latency Queueing (LLQ). On the other hand, the communication between sensor nodes in a sensor network that builds a cluster depends on a number of factors such as communication range, number and type of sensors and geographical location. The efficiency of the network itself depends on the sink location, which directly affects the lifetime of the sensor network. Every cluster has a sink node that is responsible to manage the sensors in the cluster. However, the sensors within a cluster communicate with the sink via short-range wireless communication links as illustrated in Figure 1. The sensor nodes need elegant and uncomplicated queueing technique as they usually work as small routers.
Recent advances in miniaturization and low power design have led to a flurry of activity in wireless sensor networks. However, the introduction of real time communication has created additional challenges in this area. The sensor node spends most of its life in routing packets from one node to another until the packet reaches the sink. In other words, we can say that it is functioning as a small router most of the time. Since sensor networks deal with time-critical applications, it is often necessary for communication to meet real time constraints. However, research dealing with providing QoS guarantees for real time traffic in sensor networks is still in its infancy. In this paper, an analytical model for implementing Priority Queueing (PQ) in a sensor node to calculate the queueing delay is presented. The model is based on M/D/1 queueing system (a special class of M/G/1 queueing systems). Here, two different classes of traffic are considered. The exact packet delay for corresponding classes is calculated. Further, the analytical results are validated through an extensive simulation study. .
1. Introduction
Over the past few years, wireless sensor networks have received a great attention. This technology has changed how we live, work, and interact with the physical environment. Micro electro mechanical systems (MEMS), digital electronics, and wireless communication have enabled the development of a new generation of large-scale sensor networks where small size nodes communicate with each other in short distances with low-power consumption, and therefore are suitable for a wide range of applications [1-3]. Recently, the design of sensor networks has become very important due to several civil and military applications. Emerging sensor applications include habitat monitoring, pollution detection, weather forecasting, and monitoring disasters such as earthquakes, fire and floods. Similar to normal IP network, in this new area of sensor networks, there is real-time and non-real-time traffic; each requiring
978-1-4244-1968-5/08/$25.00 ©2008 IEEE
Figure 1: Multipul sink clustered network sensors
216
In this paper, we present the implementation of Priority Queueing (PQ) in a sensor node. We exploit the M/G/1 queueing system to calculate the queueing delay for two different kinds of traffic in a sensor node. We extract the numerical solution and also conduct an extensive simulation study to verify the analytical results in order to provide guaranteed QoS to different kinds of traffic in sensor networks. The rest of the paper is organized as follow. Section 2 provides detail about the related work. In Section 3, the queueing model and expressions of expected waiting times for two different classes are presented. Details of numerical solution and experimental setup are provided in Section 4 followed by conclusions in Section 5.
2. Related Work
Figure 2: Queueing Model in Sensor Network
In many applications, sensor data must be delivered with time constraints to make appropriate real-time actions possible [4]. Most of current QoS provisioning protocols [5-8] in wireless sensor and ad hoc networks are just based on end-to-end path discovery and path recovery. Also, most of existing research is only focused on reliability and lacks the ability to differentiate multiple classes of traffic that have different time constraints [9-11]. Several studies have focused on finding out the maximum reachable throughput and characterizing capacity delay in wireless ad hoc networks [12-14]. Other studies have anticipated queuing models for performance evaluation of the 802.11 MAC. In [15] , the authors evaluate the packet blocking probability and MAC queueing delay in a basic service set with N nodes by using a finite queueing model. The M/MMGI/1/K queuing model has been used in [16] for delay analysis over a single hop in a network. The service times of the node are modeled as a Markov Modulated general arrival process. The difficulty of this approach in finding an accurate parameter description is because of Phase-Type service. In [17] the M/MMGI/1/K queueing model has been used to analyze IEEE 802.11 DFC. This work uses single hop criterion which is an extensible and flexible approach of queueing models and it has found different Applications mainly in routing and admission control. The authors in [18] focus on characterizing the average end-to-end delay and maximum achievable per-node throughput in random access MAC multihop wireless ad hoc networks with stationary nodes for hierarchical network. They present an analytical model that takes into account the random packet arrival process, the extent of locality of traffic, and the back off and collision avoidance mechanisms of random access MAC. They also model random access multi hop wireless networks as an open G/G/1 queueing network and use the diffusion approximation in order to evaluate closed form expressions for the average end-to-end delay. The authors in [19] have proposed different queues for the two different types of traffic with classifier and scheduler. Both classes can have
access to bandwidth from each other. This approach is based on the cost and end-to-end constraints. This work is focused on discovering a least cost, delay constrained path for real time data.
3. Queueing Model Regardless of its numerous limitations, First In First Out (FIFO) is the default queueing algorithm used in several topologies that requires no configuration. Most importantly, FIFO queueing makes no decision about packet priority. FIFO queueing involves storing packets and forwarding them in order of arrival. Explode sources can cause extended delays in delivering time sensitive application traffic, and potentially to network control and signaling messages. Although FIFO queueing was an effective network traffic controller before, but recent intellectual networks need more sophisticated algorithms. Furthermore, in FIFO, a full queue will cause dropping packets; even though it could be a high-priority packet. In fact, the sensor node could not prevent this undesirable packet dropping as it has no room for them in its queue. Furthermore, FIFO cannot differentiate between a high-priority and low-priority packet.
3.1. Implementation of Priority Queueing To overcome the limitations of FIFO queueing discipline, Priority queueing (PQ) is suggested as one of the applicable solutions to meet the desired QoS for real time traffic. In this work, two queues in a sensor node are considered; high-priority and low-priority as illustrated in Figure 2. Here, the scheduler uses the strict priority logic, i.e., it always serves high-priority queue first. If there is no packet waiting in highpriority queue, it will serve the low-priority queue. In this technique, because the scheduler of the sensor node is serving different output queues simultaneously, and hence, behaves similar to a multiple queues/single server system. In this paper, we exploit M/G/1 queueing system to build a queueing model for a
217
sensor node which is behaving like multiple queues/single server system. Before explaining the formulation and notations, it’s worthwhile mentioning the following assumptions. The packets that are related to high-priority queue one (Q1) and low-priority queue two (Q2) are called Class-1 (C1) and Class-2 (C2) packets with the average length of L1 and L2, respectively. Both C1 and C2 packets are coming according to Poisson process with arrival rate O1 and O2 , respectively. The service times are generally distributed and the sensor nodes and the sink are all assumed to be stationary.
The second term in equation (1), E[T1 ] , is the expected total time to serve all C1 packets that are already waiting in Q1 upon arrival of the randomly selected packet. Assume that the expected number of packets already waiting in queue one is E[ N 1 ] . Due to the PASTA property and little’s Law, on average there are E[ N 1 ] O1 E[W1 ] class 1 (C1) packets upon arrival of this randomly selected packet [21]. Since the packets already waiting in Q1 ; each requires on
1
average
service requirement for a C1 packet is E [ S 1 ]
P1
E[T1 ] can be written as:
O1 E[W1 ] P1
P1
.
U1 E[W1 ]
(3)
Substituting E[TR ] and E[T1 ] in equation (1) E[W1 ] can be calculated as follows A very similar expression has been given in [22] for M/G/1 with priority as well:
The second moment of service requirement for a C1 2
packet is E[ S1 ] . Here, the aim is to calculate the queueing delay for each C1 packet. Because a packet is randomly selected, its arrival time can be analyzed using the PASTA property of Poisson arrival streams [20]. Here, the queueing delay is defined as the expected waiting time E[W1 ] in Q1 for a C1 packet before its being serviced. Because Q1 is the highpriority queue, according to strict priority scheduler logic, the expected waiting time of C1 packet consists of two components; (1) remaining service time of a packet in service, and, (2) the time needs to serve all the packets with the same priority (C1) that are already present in the system at the arrival of this new randomly selected packet. In equation form, the expected waiting time can be written as: E[W1 ] E[TR ] E[T1 ] (1)
2
E[ S k2 ] 2 E[ S k ] 1 U1
¦U k 1
E[W1 ]
k
(4)
The property of deterministic service time of scheduler in a sensor node is used. since (1) the scheduler in the
Lk time units to serve a Ck packet R with the transmission rate of R , and (2) the average service requirement for a Ck packet 1 Lk , the second moment of service is E[ S k ] Pk R requirement of a Ck packet can be expressed as sensor node needs
Where E[TR ] is the expected remaining time for a packet in service when the scheduler is busy. The probability that the scheduler (server) is busy is U . A packet of C1 is in service with probability U1 O1 E[ S1 ] , which is the utilization of Class one packets. Since the arrival time is randomly selected, the remaining service time can be viewed as that obtained for a renewal sequence consisting of generic random variables S [20]. Thus, the remaining
follows:
2
E[ S k ] Var[ S k ] ( E[ S k ]) 2
(5)
In our approach, because the scheduler is having a fixed service time to serve its packets and hence functioning similar to a M/D/1 queueing system, the service time would be deterministic with zero variance, i.e., Var[ S k ] 0 [23]. Thus, (5) can be simplified as: 2
E[ S k ]
2
processing time of a C1 packet is equal to
E[ N 1 ]
E[T1 ]
Starting from queue one; we assume that the average
1
service time on,
P1
3.1.1. High Priority Queue (Packets C1)
E[ S1 ] . 2 E[ S1 ]
( E[ S k ]) 2
(
Lk 2 ) Based on the above R
assumption, the remaining service time of a C1 packet will be
However, because at the arrival time of the randomly selected packet, the class (either C1 or C2) which is already being served is unknown, the final equation should be modified to: E ª S k2 º 2 «¬ »¼ (2) E TR ¦ Uk ª º k 1 2 E «S » ¬ k¼
L1 on average with the probability of U1 2R
when a C1 packet is in service. However, since upon arrival of the randomly selected packet, it is not clear which packet (either C1 or C2) is in service, equation (4) is modified as:
> @
2
E[W1 ] 218
¦U k 1
k
Lk 2R
1 U1
U1
L1 L U2 2 2R 2R 1 U1
(6)
For example: Let Mj denote the number of type j arrivals over Zi, j=1,2 then
Where U1 and U 2 are the utilizations caused by C1 or C2 packets with average lengths of L1 and L2, respectively.
Work
3.1.2. Low Priority Queue (Packets C2) We obtain the expected waiting time for a randomly selected C2 packet arriving to the low priority queue by analyzing the events that constitute this delay. The amount of work in the system at any time is defined as the (random) sum of all service times that will be required by the packets in the system at that instant. The waiting time of a C2 packet (which is the low priority queue) can be written as E[W2 ] E[ Z 1 ] E[ Z 2 ] E[ Z 3 ] .... (7)
Z3 Z1
W
E[T2 ]
k 1
Pk
¦
Ok E[Wk ] ¦ Pk k 1 2
2
¦U k 1
k
Since the service times and the arrival process are independent. For a stationary packet arrival process, this can be:
E[M j ] E[E[M j | Z j ]] E[c1Z j ] c1 E[Z j ] Due to mentioned independence, where c1 ! 0 is a constant particular to the arrival process. That is, expectation of the number of arrivals in any period of time is proportional to the length of that period because of stationarity in time and linearity of expectation. In our stationary poisson traffic input process, C1 is the expected number of arrivals per unit time of C1 packet
1
U1 E[W1 ] U 2 E[W2 ]
Pk
(which can be called the arrival rate) each requiring
1
P1
E[Wk ]
service time: Hence the expected waiting time
reduces to:
E[W2 ] E[Z1 ] E[S1 ]c1E[Z1 ] E[S1 ]c1E[Z2 ]
By putting the values of E[TR ] and E[T2 ] , we can write equation (8) as follows:
E[Z 1]
L E[ Z 1 ] ¦ U k k U1 E[W1 ] U 2 E[W2 ] 2R k 1 Now, E[ Z 2 ] is the expected amount of work 2
E[Z 1]
associated with higher priority C1 packets arriving during E[ Z 1 ] , E[ Z 3 ] is the expected amount of work
c1
P1 c1
P1
(E[Z 1 ] E[Z 2 ] ) E [W 2 ]
In other words, during E[W2 ] time units, the C2 packet has to wait;
associated with C1 packets arriving during E[ Z 2 ] and so on. As illustrated in Figure 3, the waiting time of an arriving packet of C2 is indeed given by the total workload building in front of it [24]. The arrows in the figure denote the arrival times of C1 packets, and all the oblique lines have 45 degrees angle with the time axis. In this figure the expected waiting time is
E[W2 ]
M 2
S1
E[W2 ] E[Z1 ] E[S1 ]E[M1 ] E[S1 ]E[M2 ]
service time on average, E[T2 ] can be calculated as
E[ N k ]
M 1
S1
1
M
service (if any), which can be calculated in the same way, as we did for Q1 and E [T 2 ] is the time needed to serve all the packets of the higher priority class C1, and equal priority class C2 upon the arrival of the randomly selected C2 packet. E[T2 ] is related to the number of packets per class in the both queues (Q1 and Q2) upon arrival of the C2 packet. Referring to the PASTA property and the Little’s law, there are E[ N k ] O k E[Wk ] Ck packets on average upon
2
Z
2
Where S1 j denotes the random sum of Mj independent service times of C1 packets. Then,
E[ Z 1 ] E[T2 ] E[TR ] (8) E[TR ] is the remaining service time of the packet in
follows.
Z4
time Figure 3: waiting time of a C2 (type 2) packet in terms of Zj’s.
Where E[ Z 1 ] is the expected amount of work seen by the arriving C2 packet in Q1 and Q2 (i.e, higher priority and equal priority), plus the work needed to finish the service of a packet, which is already in service (if any), E[ Z 1 ] can be further written as:
arrival of a new C2 packet. Since each requires
Z2
O1 E[W2 ] packets of C1 arrive on average,
each requiring
c1
P1
1
P
service time. Hence 1
E[W2 ] can be written like this: O 1 E [W 2 ] P1
E[ Z 1 ] E[ Z 2 ] E[ Z 3 ] E[ Z 4 ]
219
U 1 E [W 2 ]
simultaneously, the expected delays for a Q1 and Q2 would be 0.069189 ms and 0.076876 ms, respectively. Now the arrival rate to a sensor node has been increased from 78pps to 156pps, which causes the utilization of the system to become 0.2 and the expected delay for Q1 and Q2 as 0.150588 ms and 0.188235 ms, respectively. Similarly the gradual increase in the arrival rate to a sensor node to 234pps then to 312pps, causes the utilization of the system to become 0.3 and 0.4, respectively, hence increasing the expected delays for the corresponding classes as shown in Table 1.
Substituting all the values in equation (7), E[W2 ] can be calculated as follows:
E[W2 ]
2
¦U k 1
k
Lk U1E[W1 ] U2 E[W2 ] U1E[W2 ] 2R
Bringing E[W2 ] to one side and by simplifying, then it can be: 2
E[W2 ] 4.
Lk U1E[W1 ] 2R 1 U1 U 2
¦U k 1
k
(9)
Table 1: Expected Delay for Q1 and Q2 packets Utilization Delay (Q1) Delay (Q2) 0.1 0.069189 ms 0.076876 ms 0.2 0.150588 ms 0.188235 ms 0.3 0.247700 ms 0.353800 ms 0.4 0.365700 ms 0.914200 ms
Experimentation and Evaluation
A sample network has been generated to show the performance of the presented model. Then, its analytical results have been validated through simulation.
4.2.2. Simulation Results
4.1. Environmental Setup
J-Sim has been used to simulate 100 nodes; the software which provides a high fidelity simulation for wireless communication with detailed propagation, radio and MAC layers. The delay has been calculated for real time traffic (Q1) and non real time traffic (Q2) on a node 5; which has been selected at randomly position out of 100 nodes during the 15 randomized runs of the simulation.
A square shape network (200m, 200m) is selected as the benchmark to validate the accuracy of the presented formulations. 100 nodes with 40m radio range are generated to operate in this network. For each node, a free space propagation channel model is assumed with the transmission speed of 200kbps, total packet length of 32 Kbit for both C1 and C2 with 15 packet capacity as their buffer sizes. Also, for each node in the sensing state, packets are generated at a constant rate of 1 packet/sec. The real-time packet generation rate is 3 packets/sec to prevent the nodes of being congested and/or over loaded [25] [26] [27].
4.3. Analysis The Table 1 shows the differences in delays for Q1 and Q2 for several utilizations of the sensor nodes. As can be seen, Q2 delays exceed those of Q1 delays in all positions. As shown in Figure 4, we can notice the characteristic of PQ, as the utilization increases, there is a sharp increase in the queueing delay of class 2 ( C 2 ) packets. The horizontal axis in Figure 4 shows the variable of arrival rates in units of utilization, while the vertical axis shows the range of the variable of delay in units of milliseconds. Thus, the graph shows the change in delays over the arrival rates of the packets for both of the numerical and simulation results. The straight lines are used to show the numerical result of Q1 and Q2 while the dotted lines are used to illustrate the simulation results. From utilization 0.1 to 0.4 the impact of priority queueing implementation in the sensor node increased sharply in the delay for Q2 at higher arrival rate for the numerical and simulation results. The slight difference between the numerical and simulation results in both of the Q1 and Q2 is because of several factors such as, the retransmission of some packets and the time it takes to find new path to forward the packet. Figure 5 shows the on time delay. The horizontal axis shows the variable of arrival rates in times of seconds, and the vertical axis, shows the range of the
4.2. Results The results have been validated in this paper by comparing result obtained from the equations (numerical result) and the result obtained from the simulation (simulation results). The mathematical delay calculation has been explained in the numerical result section. The simulation technique has been clarified in the simulation result section. The result has been breakdown in analysis section. 4.2.1. Numerical Results
The total arrival rate for each sensor node is assumed to be 78 packets per second with 32Kbit packet length. The scheduler of the sensor can serve 780 packets per second. Therefore, the total utilization of the sensor node will be 0.1. Now the utilization has been calculated of both the queues (Q1 and Q2). For Q1, real time packets are arriving at a rate three times more than the non real time packets, which is arriving to Q2. If the arrival rate to real time queue is three times more than non real time queue, the utilization of Q1 will be 0.75 and utilization of Q2 will be 0.25. Using these values in equation (6) and (9) and solving them
220
the most appropriate queueing scheme implementation for wireless sensor networks. In our future work, we will concentrate to calculate the end-to-end delay for different classes of traffic passing through multiple sensors in a sensor network.
Utilization of System (Sensor node) 1.2
Delay (msec)
1 0.8
6.
0.6
[1]
0.2 0 0.1
0.2
0.3
0.4
Utilization Queue 1
Queue 2
Queue 1 (simulation)
Queue 2 (simulation)
[2]
Figure 4: Utilization of system (Sensor Node) On time Delay
0.9
[3]
Delay (msec)
0.8 0.7
[4]
0.6 0.5 0.4 0.3
[5]
0.2 0.1
Queue 1
Queue 2
63.8
62.8
61.1
59.9
58.8
57.6
53
56.3
51.1
49.9
43
48.5
41.5
39.9
38.5
0
Time
[6]
Figure 5: The on Time Daley [7]
variable of delay in units of milliseconds. Thus, the graph shows the change in delays over the arrival rates of the packets for node 5 on the simulation. The on time delay from 38.5s to 63.8s in Q1 and Q2 shows that the delay has been increased and dropped periodically. The reason of the difference between the delays in Q1 and Q2 is because some packets have been dropped for different reasons and some packet chose other path and did not pass though this node. However, because each sensor node has a small buffer and the delay inside the node is gradually increasing, as shown Figures 4 and 5, the probability of dropping the packets would be increasing accordingly.
5.
References
0.4
[8]
[9]
[10]
Conclusion
This paper have presented the closed-form expressions of the queueing delay for multiple (real time and non real time) classes of traffic in a senor node through the implementation of priority queueing based on M/G/1 queueing system. The analytical results have been verified through numerical and simulation studies. The results presented provide a way to analyze the performance of priority queueing implementation in a sensor node. Our measurement data can be useful as input to simulation study of sensor networks. Also, our analytical modeling technique and its verification through numerical and simulation results is the first step towards finding out
[11]
[12] [13]
[14]
221
Kahn, J. M., R. H. Katz, and K. S. J. Pister, "Next Century Challenges: Mobile Networking for Smart Dust," in Proceedings of the Fifth Annual ACM/IEEE International Conference on Mobile Computing and Networking (MOBICOM '99), Seattle , Washington , USA, August 1999, pp. 271-278. I. F. Akyildiz, W. Su, Y. Sankarasubramaiam, and E. Cayirci, "A Survey on Sensor Networks," in IEEE Communication Magazine. vol. 40, no 8, August 2002, pp. 102-114. G. J. Pottie and W. J. Kaiser, "Wireless Integrated Network Sensors," Communication of the ACM, vol. 43, pp. 51-58, May 2000. J. A. Stankovic, "Research Challenges for Wireless Sensor Networks," SIGBED Review: Special Issue on Embedded Sensor Networks and Wireless Computing, 1(2), July 2004. T. Chen, J. Tsai, and M.Gerla, "QoS Routing Performance in Multihop Multimedia Wireless Networks," in IEEE Sixth International Conference of Universal Personal Communications, 1997, pp. 557– 561. R. Sivakumar, P. Sinha, and V. Bharghavan, "CEDAR: Core Extraction Distributed Ad Hoc Routing Algorithm," IEEE Journal on Selected Areas of Communications, vol. 17, pp. 1454–1465, 1999. S. Chen and K. Nahrstedt, "Distributed Quality-ofService Routing in Ad Hoc Networks," IEEE Journal on Selected Areas of Communications, vol. 17, pp. 1488–1505, 1999. B. Hughesand and V. Cahill, "Achievingrealtimeguaranteesinmobileadhocwirelessnetworks," in Proceedings of the Work-in-Progress session of the 24th IEEE Real-Time Systems Symposium, December 2003. B. Deb, S. Bhatnagar, and B. Nath, "ReInForm: Reliable Information Forwarding Using Multiple Paths in Sensor Networks," in Proceedings of IEEE International Conferenceon Local Computer Networks., pp. 406 – 415. T. He, J. Stankovic, C. Lu, and T. Abdelzaher, "SPEED: AStateless Protocol for Real-Time Communication in Sensor Networks," in Proceedings of IEEE International Conference on Distributed Computing Systems, 2003, pp. 46 – 55. F. Stann and J. Heidemann, "RMST: Reliable Data Transport in Sensor Networks," in Proceedings of IEEE International Workshop on Sensor Network Protocols and Applications, 2003, pp. 102–112. A. E. Gamal, J. Mammen, B. Prabhaker, and D. Shah, "Throughput-delay trade-off in wireless networks," in IEEE INFOCOM, March 2004. J. Li, C. Blake, D. S. D. Couto, H. I. Lee, and R. Morris, "Capacity of ad hoc wireless networks," in MobiCom 01, Pages 61-69 New York, USA: ACM Press, 2001. P. Gupta and P. R. Kumar, "Capacity of wireless networks," IEEE Trans. on Information Theory, pp. 388-404, March 2000.
[15] G. zeng, H. Zhu, and I. Chlamtec, "A Novel Queueing model for 802.11 Wireless LANs," in WNCG Wireless Network Symposium, 2003. [16] O. Tickoo and B. Sikdar, "A Queueing model for finite load IEEE 802.11 random access MAC," in IEEE ICC, Paris, France, 2004. [17] M. Ozdemir and A. B. McDonald, "An M/MMG1/1/K queuing model for IEEE 802.11 ad hoc networks," in Proceeding of PE-WASUN'05, 2004, pp. 107-111. [18] N. Bisnik and A. Abouzeid, "Queuing Network Models for Delay Analysis of Multihop Wireless Ad Hoc Networks," in The 2006 International Wireless Communications and Mobile Computing Conference (IWCMC 2006) Vancouver, Canada: IEEE Communication Society and ACM, July, 2006. [19] K. Akkaya and M. Younis, "An Energy-Aware QoS Routing Protocol for Wireless Sensor Network," in Proceedings of the 23 rd International Conference on Distributed Computing Systems Workshops (ICDCSW'03), 2003. [20] R. Nelson, "Probabililty, Stochastic Processes, And Queueing Theory, The Mathematics of Computer Science Modeling," Springer-Verlag New York,Inc. , pp. 283-323, 1995. [21] J.D. Little, "A Proof of Queueing Formula L=ȜW," Operations Research, 9, pp 383-387, 1961. [22] B. R. Haverkort, "Performance of Computer Communication System," John Wiley and Sons, 1998, pp. 115-132. [23] D. Gross and C. M. Harris, "Fundamentals of Queueing theory," vol. Third Edition, pp. 211-213, 1998. [24] M. Iftikhar, M. Caglar, and B. Landfeldt, "An Analytical model based on G/M/1 with Self-Similar input to provide end-to-end QoS in 3G networks " in IEEE/ACM Mobiwac 06 (MSWIM), Torremolinos, Malage, Spain, 2nd Oct 2006. [25] J. B. Andresen, T. S. Rappaport, and S. Yoshida, "Propagation Measurements and Models for Wireless Communications Channels," in IEEE Communications Magazine. vol. 33, January 1995. [26] M. Gerla, G. Pei, and S.-J. Lee, "Wireless, Mobile AdHoc Network Routing," in Proceedings of IEEE/ACM FOCUS'99, New Brunswick, NJ, , May 1999. [27] "Data sheet for the Acoustic Ballistic Module," SenTech Inc., http://www.sentechacoustic.com.
222
