Provision of Quality of Service Using Active Scheduling S.A Hussain and A.Marshall Advanced Telecommunications Systems Laboratory, School of Electrical & Electronic Engineering, The Queen’s University of Belfast Ashby Building, Stranmillis Road, BT9 5AH Belfast Northern Ireland, U.K.
[email protected],
[email protected] Tel: 44(0) 2890 274248, 0092 42 5168304 Fax: 44(0) 2890 667023
Abstract Active and Programmable networks change the functionality of routers and switches by using agents and active packets. This paper presents a new packet scheduling scheme called Active Scheduling to control and maintain QoS parameters in virtual private networks (VPNs) within the confines of adaptive and programmable networks. In Active Scheduling an agent on the router monitors the accumulated queuing delay for each service. In order to control and to keep the endto-end delay within the bounds, the weights for services are adjusted dynamically by agents on the routers spanning the VPN. If there is an increase or decrease in queuing delay of a service, an agent on a downstream router informs the upstream routers to adjust the weights of their queues. This keeps the end-to-end delay of services within the specified bounds and offers better QoS compared to VPNs using static WFQ. The paper describes the algorithm for Active Scheduling, and presents simulation results and these are compared with WFQ.
1
Introduction
In traditional computer networks, the intermediate nodes (e.g., routers) are vertically closed systems whose functions are rigidly programmed into the embedded software and hardware by the vendors. The Internet also falls in this category. Therefore, the development and deployment of new protocols requires a long standardization process. Active networks [1] allow users or operators to inject customized codes into the network to modify the behaviour of switches and routers. Scheduling of different classes of traffic within the switches and routers has been identified as one of the most important resources to be reconfigured [2]. This paper presents a novel mechanism, “Active Scheduling” for the reconfiguration of scheduling mechanism to modify the queuing strategy in routers [2,3,4]. Active Scheduling allows the introduction of a procedure by which the queue weights of heavily loaded routers are altered dynamically, according to different classes of traffic (voice, video and data), since each class has different QoS requirements. This is achieved by using agents on the routers initiating the reconfiguration of weights. This procedure is called Active
Scheduling. Agents are regarded as software programs designed to carry out specific function or task on behalf of a user or an operator by communicating with other agents [5]. A survey of packet scheduling schemes of active and programmable networks shows that projects such as Tempest, based on switchlets [6], and Spawning networks [7] based on Routelets, use static scheduling mechanisms. One of these well-known scheduling schemes is static WFQ. There are certain problems with static WFQ. Firstly, it allocates a fixed amount of bandwidth over a defined time scale to the sessions over the routers spanning the VPN, that is the assignment of weights to any queue within a scheduler, or across a number of schedulers is essentially static, this leads to bandwidth bottlenecks at the output link of routers for the sessions joining the queues of a router from other sections of the network. Secondly, there is no provision for intelligent load management in static WFQ. This results in higher end-to-end delays for the services and wastage of bandwidth. This is due to the reason that different routers in the network are loaded with different traffic loads. For example if a router in the core network supporting the VPN is heavily loaded and the upstream routers are lightly loaded, then this can lead to wastage of bandwidth and higher end-to-end delays. This paper presents and describes Active Scheduling and compares its performance with static WFQ [8] by presenting a simulation example. The remainder of the paper is organized as follows: section 2 presents the algorithm for Active Scheduling, section 3 presents the performance analysis of Active Scheduling, section 4 describes the fairness analysis for Active Scheduling and static WFQ, and finally it presents concluding remarks.
2
Active Scheduling Algorithm
Each VPN in active router environment contains two types of agents on the active routers: 1) queue agents, 2) control agents. Queue agents control the delay of sessions according to the maximum and minimum delay bounds specified by the clients. Control agents control the endto-end delay of a VPN by monitoring the weight of a session on a downstream router. Whenever there is an increase in the weight of a session in the downstream router it informs the upstream router to reduce the weight of the session corresponding to the size of the burst. Q1 in router A and router B forms the VPN for session 1. Similarly other sessions in the network form the VPNs through the queues of active router A and active router B. Active router B is loaded by active
router A and other routers in the network. When the weight of loaded queue Q1 of router B reaches its maximum limit, an agent on router B informs the agent on router A to reduce the weight W1A of its queue Q1 corresponding to the burst size of the bursty session by sending a lightweight weight-change signal (trigger) to it. This reduces the traffic load on Q1 of active router B and its queuing delay decreases. This mechanism is called intelligent load management. When the weight of Q1 in B comes down a weight-change trigger is send again from router B to router A to increase the weight of Q1 in router A. This technique is scalable since the weight change is always between two intermediate routers (upstream and downstream) instead of throughout the routers spanning the VPN. The downstream router sends the weight-change signal to the upstream router by reserving a path between the two routers. This is accomplished by using a high-priority IP packet over that path. Active Scheduling is functionally compatible with the DiffServ architecture. Agents can be used to monitor the queuing delay of aggregates in DiffServ. If more fine grained control is required, agents can be used on a per flow basis, or one agent for N number of sessions within aggregates if the QoS requirements for the N set of sessions is same. The weights will be changed whenever there is congestion on the links or a change in the number of flows. The number of agents on a router depends on the aggregates or classes of service.
Q1 W 1A
VPN 1
Router A QN W NA
Q1 W 1B Router B QN W NB
VPNN
Control agent
weight change
VPN flow Router A
Router B
Queue agent
other flows
Fig.1. Weight adjustments by agents in routers
Max delay bound (client) Increase weight T imax (nominal max delay) No change T imin (nominal min delay) Decrease weight Min delay bound (client)
Fig. 2. Weight change in an Active Router A range or set of limits of queuing delay is defined for each session. This range is based on the lower and upper limits of queuing delays as shown in Fig. 2. The reconfiguration of weights for a session i is based on the delay calculated for the session i. The delay tolerance is given by: Ttol = Ti max − Ti min
(1)
If the measured delay Ti falls out side the limits, then weights for the session are recalibrated to maintain: T i max ≤ T i ≤ T i min
∀i∈B(t)
(2)
Where B(t) is the set of sessions currently backlogged. Timax is the nominal maximum delay, which a session can have. This delay bound is always less than the maximum delay limit set for the client by the network operator. As the queuing delay reaches Timax the weight for the session is increased by the agent on the queue. Timin is the nominal minimum acceptable delay of the session. When Ti is at or less than Timin the weight for the session is decreased by the agent, and the excess bandwidth is shared among other sessions according to their weight ratios. In some cases Timin can be zero (no delay) i.e., no packets in the queues. There is no change in weights when Ti is between Timax and Timin. Whenever a packet joins the queue of a session i, it is enqueued in the queue. The packet waits in the queue till it is served by the scheduler. The queuing delay of the packet is calculated by the following: sum_delay += delay_arrival (i) Ti = sum_delay/delay_arrival_index
Where sum_delay is the accumulated delay, delay_arrival is the delay of each packet, delay_arrival_index is the number of times delay of each packet is accumulated before a weight update occurs, and Ti is the measured delay calculated periodically. If weight (queue) < nominal weight + upper_ weight limit The weight of the queue is increased by ∆w if the queuing delay is more than the upper limit i.e., wi(t) +∆(w+) if
Ti≥Timax
(3)
This weight increase dequeues the packet and brings down the sum_delay and correspondingly Ti -> Timin. Where ∆w can be piecewise linear or step weight change. When the weight of the queue in a heavily loaded router reaches its maximum limit, the agent on the queue informs the upstream router to reduce the weight of the same queue in the upstream router, by sending a weight-change signal to it. When the weight of the loaded queue comes down to the nominal weight, a weight-change signal is send again to the queue of the upstream router to increase the weight of the queue of the upstream router from the queue that was loaded. If weight (queue) > nominal weight + lower_ weight limit The weight of the queue is decreased by ∆w if the queuing delay is less than the upper limit i.e., wi(t) - ∆(w-) if Ti≤Timin , Ti ≠0
(4)
If delay of the queue remains between the max and min limits then the weight of the queue remains at the nominal value i.e., wi(t+τ) = wi(t)
if Ti max ≤ Ti ≤ Ti min
(5)
Where wi(t) is the nominal weight, ∆(w+) and ∆(w-) are the weight increment and decrement values whenever there is an increase and decrease in the queuing delay of a session. If Dimax is the end-to-end delay for a session,
D i max =
N
∑T i =1
i max
(6)
Where N is the number of routers spanning the VPN. The worst-case queuing delay bound of a session when there is high congestion on the link is calculated by the equation
Ti max =
burstsize Packetsize MTU + + ri ri r
(7)
Where burstsize and Packetsize are an application’s burst size and packet size respectively, ri is the bandwidth allocated to the application. The weights of the application vary between maximum, minimum and nominal bounds. r is the link rate, MTU is the maximum transfer unit (MTU) of traffic on the link. The values of burst size, Packet size and MTU are SLA (Service Level Agreements) parameters. The weights of connections across multiple active routers can be worked out by using a scalable QoS Policy Information Model (QPIM) [9]. QPIM establishes a standard framework for specifying and representing policies that administrator, manage and control access to network QoS resources. QPIM is independent of any particular data storage mechanism and access protocol. The weights of routers will be changed according to the set of rules (e.g., delay limits etc) stored in a policy repository.
3
Performance Analysis of Active Scheduling
3.1
Simulation Environment
A prototype network consisting of a single active and legacy routers has been implemented using a simulation environment as shown in Fig. 3. The traffic in the network is constituted by FTP over TCP, VOIP, and MPEG2 video subnets running over UDP. The aim of simulations is to study and compare the performance of VPNs using legacy and active routers in terms of queuing delay, end-to-end delay and fairness. The simulations are based on OPNET (Optimized Network Engineering Tool) [10] software. weight-change
Loaded Router Core Network weight-change
Fig 3 Simulation Environment for Multiple Active Routers
The traffic between the clients goes through the active and legacy routers. The link between the routers is 2.0Mbits/sec. It is a typical bottleneck since all the traffic is passing through it. This link is a typical VPN’s dedicated bandwidth over a high-speed link between the routers in the core network. The simulation measurements are taken when there is congestion in the network i.e., the link between the routers is highly congested and the traffic is flowing away from the active router (towards the legacy router). The measurements for static WFQ are taken by replacing the active router with a legacy router. These clients compete for the bandwidth over the link between the routers. The link is loaded by different percentages of traffic over a period of time. The clients send the traffic to the ingress and egress routers through 10Mbits/sec links. Active router 2 is the loaded router. MPEG2 video subnet B1 sends the traffic in the direction of egress switch 5, whereas FTP remote server A traffic flows in the direction of egress switch 2 loading the active router 2. The delay for a voice packet is calculated as follows: Number of bytes/packet on the queue = voice packet_ size + UDP header+ IP header = 32+8+20+ = 60 bytes =480 bits Packet Rate = 250 packets per second. Talk Spurt Length=0.352 seconds Average traffic =480*250=120,000 bits/sec =120000*0.352= 42240 bits/sec Since there are 11 sessions. Average traffic =11*42240=464640 bits/sec In the worst-case delay bound calculation, a session is always assigned bandwidth equal to or greater than its average rate, i.e., ri ≥ raverage [6]. If the nominal and maximum weight bounds for the VOIP clients are 12 and 24 (i.e., 12% and 24% of 2.0 Mbit/sec) respectively, the link rate is 2.0 Mbit/sec, and MTU is 12000 bits then,
Dvoice =
σi ri
Dvoice =
+
Pi MTU + ri rlink
464640 480 12000 + + = 0.975 sec 480000 480000 2000000
This gives an average Dvoice per user=
0.975 = 88.6ms 11
(8)
If the maximum weight bound and nominal weight bounds of MPEG2 video is 47 and 30 respectively, the worst-case delay bound (client) is calculated as: Dmpeg 2 =
σi ri
+
Pi max MTU + ri rlink
(9)
Since any packets more than 1500 bytes are fragmented by the IP layer into 1500 bytes packets, the size of the burst received at the router is 1.1 Mbits/sec depending on the size of I, B and P frames. DMPEG 2 =
1.1M 12000 12000 + + = 1.016 sec 1100000 1100000 2000000
This produces an average DMPEG2 per user =
1.016 = 203ms 5
The maximum and nominal delay bounds for FTP clients are 21 and 10 respectively. The worstcase delay bound of FTP clients is calculated as: DFTP =
DFTP =
σi ri
+
Pi MTU + ri rlink
1000000 12000 12000 + + = 2.4 sec 420000 420000 2000000
This produces an average DFTP per user =
3.2
(10)
2.4 = 800ms 3
Effect of Weight Update Intervals and Weight Change Profile on the Queuing Delay
The weights of the sessions in active routers are changed according to the delay bounds calculated in section 3.1. The queuing delay of sessions in active routers is measured by changing the weight update interval to 0.01 sec, and 0.5 sec. These two values of weight update interval have been chosen to show the effect of weight update interval on the performance of active scheduling. The 0.5 sec period represents the burst duration of FTP and MPEG2 clients. The reason of selecting 0.5 sec as the update interval is to make sure that the burst from these clients is served within this period. The weights during congestion are changed according to 1) piecewise linear, 2) step change i.e., the weights are increased and decreased as a step function. The sessions with step weight increase perform better than piecewise linear change. With step weight increase, weights increase and decrease quickly and keep the delay within bounds, whereas the change is sluggish in the case of piecewise linear weight changes.
The queuing delay of these sessions for 0.5 sec update interval is lower compared to 0.01 sec. For 0.01 sec update interval the weights are updated frequently, this creates instability in the system and the queuing delay of sessions increase. The reason for this instability is the high degree of delay variability of packets and competition for the bandwidth among the sessions. Hence it becomes difficult for the system to keep up with the frequency of fast weight updates. ete delay (piecew ise linear)
2.5
ete delay(step function)
1.4 1.2 ete delay(sec)
ete delay(sec)
2 1.5 1 0.5
1 0.8 0.6 0.4 0.2 0
0 0
1000
2000
3000
0
4000
1000
2000
3000
4000
time(sec)
time(sec)
Active Scheduling
Static WFQ
Fig 4 (a) End-to-End delay of MPEG2 for static WFQ and active routers ete delay(piecew ise linear)
3
0.7
2.5
0.6
2
ete delay(sec)
ete delay(sec)
ete delay(step function)
1.5 1 0.5
0.5 0.4 0.3 0.2 0.1 0
0 0
1000
2000 time(sec)
3000
4000
0
1000
2000
3000
time(sec)
Active Scheduling Static WFQ Fig 4 (b) End-to-End delay of VOIP for static WFQ and active routers
4000
ete delay(piecew ise linear)
2.5
ete delay(piecew ise linear)( RT services step function) 7 6
1.5
ete delay(sec)
ete delay(sec)
2
1 0.5
5 4 3 2 1
0
0
0
1000
2000
3000
4000
0
1000
time(sec)
2000
3000
4000
time(sec)
Static WFQ
Active Scheduling
Fig 4 (c) End-to-End delay of FTP Subnet A for static WFQ and active routers ete delay(piecew ise linear)
6
ete delay(piecew ise linear)(RT step function) 8
4
7
3
ete delay (sec)
ete delay(sec)
5
2 1 0
6 5 4 3 2 1
0
1000
2000 time(sec)
3000
4000
0 0
1000
2000
3000
4000
time(sec)
Static WFQ
Active Scheduling
Fig 4 (d) End-to-End delay of FTP Subnet B for static WFQ and active routers Figures 4 (a), (b), (c) and (d) show the worst-case end-to-end delays for MPEG2 video, VOIP and FTP clients for multiple legacy and active routers for piecewise linear and step function weight change when the weight update interval is 0.5 seconds. These results show that due to static allocation of weights, the end-to-end delays for MPEG2 and VOIP sessions for static WFQ is high. In Fig 4 (a) and 4 (b) the worst-case end-to-end delay for any MPEG2 and VOIP client for piecewise linear weight change is 0.252 sec (1.26/5=0.252 sec) and 0.058 sec (0.64/11 =0.058 sec) respectively. The worst-case end-to-end delay for any MPEG2 and VOIP client for step weight increase is 0.158 sec and 0.035 sec respectively. The worst-case end-to-end delay of MPEG2 and VOIP sessions for static WFQ is 0.42 sec (2.1/5=0.42 sec) and 0.24 sec (2.7/11=0.24 sec) respectively. The worst-case end-to-end delay for MPEG2 and VOIP sessions for active routers is also within the accepted bounds as described by [11,12,13,14] and [15,16,17]
respectively. These results show that active routers perform better than legacy routers using static WFQ. Figures 4 (c) and 4 (d) show the performance of FTP sessions when the weights of FTP sessions are changed as piecewise linear function and the weights of real time (RT) are changed as a step function. The worst-case end-to-end delay of FTP sessions is higher for active routers as compared to static WFQ. The reason for this increase of delay of FTP is the bandwidth usage by MPEG2 video and VOIP clients up to their limits. Since FTP is a best effort service, complete delivery of data without packet loss is more important since a higher delay does not violate the QoS requirements specified in the SLA.
Active Router 2
Active Router 1 weight
Active Router 1
39m 25.5s
Active Router 2
39 26.5s
weight
39m 25.5s
39m27.5
39 26.5s
39m27.5
time(sec)
time(sec)
(a) Weight Change of MPEG2 Subnet A1 (active Router 1&2)
(b) Weight Change of VOIP (Active Router 1 &2)
Active Router 1 Active Router 2
weight
weight
Active Router 1
Active Router 2
39m 25.5s
39 26.5s
time(sec)
(c) Weight change of FTP Subnet A (Active Router 2 &3)
39m27.5
39m 25.5s
39 26.5s
39m27.5
time(sec)
(d) Weight Change of FTP Subnet B (Active Router 1& 2)
Fig 5 Step Weight Change of MPEG2 and VOIP Sessions & Piecewise Linear Weight Change of FTP Sessions (0.5 sec weight update interval)
The weights of MPEG2 video and VOIP sessions in Fig 5 are changed as a step function as these are delay sensitive services as compared to FTP which is a best-effort service. The weights of FTP sessions are increased as a piecewise linear function. When the weight of FTP or MPEG2 video session reaches its maximum value in active router 2, the agent monitoring the queue of that session informs the control agent of upstream active routers to reduce its weight. This reduces the traffic load of active router 2, and brings down the queuing delay of sessions. Tables 1, 2 and 3 show the weight bounds for the sessions assigned according to the equations in section 2.1. Sessions MPEG2 Subnet A1 VOIP FTP Subnet B Sessions FTP Subnet A FTP Subnet B MPEG2 Subnet A1 VOIP Sessions MPEG2 Subnet B1 FTP Subnet B
Maximum Weight Nominal Weight Minimum Weight 55 28 12 24 12 6 21 10 4 Table 1 Weight Bounds for sessions in Active Router 1 Maximum Weight Nominal Weight Minimum Weight 11 5 3 10 5 3 55 28 12 24 12 6 Table 2 Weight Bounds for sessions in Active Router 2 Maximum Weight Nominal Weight Minimum Weight 70 35 25 30 20 10 Table 3 Weight Bounds for sessions in Active Router 3
The overhead of signaling has been measured for a weight update period of 0.5 seconds and piecewise linear weight change profile. The size of IP header (160 bits) was taken as the minimum amount of bits used for weight-change in the upstream router by the downstream router. The minimum amount of bandwidth used for weight change during a 36 sec measurement period when there is burst from MPEG2 video is 4 bits/sec. The total bandwidth used by weightchange signaling during one hour is 28 bits/sec (since there are 7 weight updates for MPEG2 video). Similarly for FTP1 and FTP2 the minimum amount of bandwidth used by weight-change signaling during one hour is 92 bits/sec (23 weight updates) and 88 (22 weight updates) bits/sec respectively. This bandwidth is quite low as compared to the bandwidth between the links of active routers (2 Mbits/sec).
4
Fairness Analysis of Static and Active Scheduling
The WFI for a service discipline is given by [18]
C i ,WFQ = d k i ,WFQ − a k i −
( )
Qi , s a k i
(11)
ri
Where ri is the throughput guarantee to session i, Qi,s(aki) is the queue size of session i, at time aki, dki, WFQ is the time at which kth packet of the ith session departs, under WFQ and Ci,WFQ is the
Worst-case Fairness Index for the session i at server s employing static WFQ. Intuitively Ci,,WFQ is the maximum time a packet coming to an empty queue needs to wait before it starts receiving its guaranteed service. The lower the WFI the more fair the scheduling discipline is. In Active Scheduling the weights are increased to keep the queuing delay within bounds, this produces a new packet order. This dependence upon the upper delay bound (maximum allowable delay) is included into the fairness measure of the Active Scheduling. Thus the equation for the WFI of Active Scheduling can be written as, C i ,act = d ik,act + Ti kmax − a ik −
( )
Qi , s a k i ri
(12)
Where Ci,act is the WFI of the session, Tki,max is the upper delay bound (maximum allowable delay) session, Qi,s(aki) is the queue size of session i, at time aki and ri is the service rate allocated to the session. When the weight of the session is at its maximum limit and its delay is the worst-case delay, the WFI in this situation always represents the worst situation regarding the fairness. The Worst Case Fairness Index (WFI) has been measured for active router 2 and legacy router 2. Active 2 and legacy 2 are selected for WFI since these are the loaded routers and there is competition for the bandwidth among the services at this router. Hence the WFI of this router plays a very important part in the fairness for the sessions. Fig 6 shows that the active routers perform better than static WFQ when there is a bursty traffic from MPEG2 video and when there are large size files being downloaded by the FTP server.
WFI of VOIP(static WFQ)
WFI of MPEG2 video subnetA1(Static WFQ) WFI of MPEG2 video subnet A1(Active Router 2)
WFI of VOIP(Active Router 2)
0.8 0.7
2
0.6 0.5
WFI
WFI
1.5 1
0.4 0.3 0.2
0.5
0.1
0 0
1000
2000
3000
0
4000
0
1000
time(sec)
2000
3000
4000
time(sec)
Fig 6 (a) Worst Case Fair Index of MPEG2 video and VOIP WFI of FTP subnet A (static WFQ)
WFI of FTP subnet B(static WFQ)
WFI of FTP subnet A (Active Router2)
WFI of FTP subnet B(Active Router 2) 3.5
6
3
5
2.5
4
WFI
WFI
7
3
2 1.5 1
2
0.5
1
0
0
0 0
1000
2000
3000
4000
1000
2000
3000
time(sec)
time(sec)
Fig 6 (b) Worst Case Fair Index of FTP Subnet A and FTP Subnet B
5
Conclusions
This paper has presented a new scheduling scheme, called Active Scheduling. Dynamic virtual private networks consisting of active routers have been implemented in OPNET. Active routers reconfigure the weights of sessions on-the-fly, based on the acceptable queuing delay limits of clients. The results have shown that active routers offer better performance in terms of end-to-end delays and fairness. Acknowledgements The authors gratefully acknowledge, support and financial assistance provided by Nortel Networks under the JIGSAW project.
4000
6
References 1.
TENNENHOUSE, D.L, SMITH, J.M, SINCOSKIE, W.D, WETHERALL, D.J, and MINDEN, J.G.: ‘A survey of active network research’, IEEE Comm. Mag., 1997, vol. 35, no. 1 pp. 80-86.
2.
HUSSAIN, S.A, and MARSHALL, A.: ‘Transport of Real time and Non Real time Traffic Using Active Routers in Adaptive Open Network Environment’, Proceedings of IST, Sept 2001, Tehran, Iran, pp. 65-70.
3.
HUSSAIN, S.A, and MARSHALL, A.: ‘A programmable Scheduling Paradigm for Routers in an Adaptive Open Network Environment’, Proceedings of the IASTED International Symposia Feb 2001, Innsbruck, Austria, pp. 153-158.
4.
HUSSAIN, S.A, and MARSHALL, A.: ‘An Active Scheduling Policy for Programmable Routers’, Sixteenth UK Teletraffic Symposium-Management of QoS – the New Challenge May 2000, Harlow, pp. 1-8.
5.
HAYZELDEN, A, and BIGHAM, J (EDS).: ‘Software Agents for Future Communication Systems’ (Springer-Verlag, ISBN 3-540-65578-6, Germany, 1999).
6.
ISAACS, R, and LESLIE, I.: ‘Support for Resource-Assured and Dynamic Virtual Private Networks’, IEEE journal on Selected Areas in Communications., 2001. 19(3) pp. 460-472.
7.
CAMPBELL, A.T, VICENTE, J.B, DE MEER, H.G, MIKI, K, and KALAICHELVAN, K.S.: ‘Spawning Networks’, IEEE Network., 1999, 1(4) pp. 16-29.
8.
PAREKH, A.K, and GALLAGER, R.G.: ‘A Generalized Processor Sharing Approach to Flow Control in Integrated Services Networks-The Single Node Case’, IEEE/ACM Trans on Networking., 1993, 1(3) pp. 344-357.
9.
SNIR,Y.: ‘Policy QoS Information Model’, IETF Internet Draft (draft-ietf-policy-qosinfo-model-04.txt), November 2001.
10. MIL3.: www.opnet.com. 11. ITU.: ‘Transmission systems and media, general recommendations on the transmission quality for an entire international telephone connection, Recommendation G.114, Telecommunication Standardization Sector of ITU’, Geneva, Switzerland.
12. JIANG, W and SCHULZRINNE, H.: ‘QoS Measurement of Internet Real-time Multimedia Services’, Technical Report (CUCS-015-99), Department of Computer Science Columbia University, 1999. 13. VERSHEURE, O, FROSSARD, P, and HAMDI, M.: ‘User-Oriented QoS Analysis in MPEG-2 Video Delivery’, Journal of Real-time Imaging., 1999, 5 (5) pp. 305-314. 14. WU, D, HOU, Y.T, and ZHANG, Y.: ‘Transporting Real-time Video over the Internet: Challenges and Approaches’, Proceedings of the IEEE, 2000, 88(12) pp. 1855-1874. 15. KOSTAS, T. J, BORELLA, M.S, SIDHU, I, SCHUSTER, G.M, GRABIEC, J and MAHLER, J.: ‘Real-time Voice Over Packet-Switched Networks’, IEEE Network., 1998 Vol. 12, No 1, pp. 18-27.
16. HASSAN, M, NAYANDORO, A, and ATIQUZZAMAN, M.: ‘Internet telephony: services, technical challenges, and products’, IEEE Commun Mag., 2000, 38(4) pp. 96103. 17. MARK, A AND MILLER, P.: ‘Voice Over IP’ (M & T Books, CA USA, 2000). 18. BENNETT. J AND ZHANG, H.: ‘WF2Q: Worst-case weighted fair queuing’, Proc IEEE INFOCOM’96, March 1996, San Francisco, CA, pp 120-128.
