An integrated IP QoS architecture - performance

zyxw zyxwv AN INTEGRATED IP QOS ARCHITECTURE - PERFORMANCE Byungsuk Kim Ipl Sebiiktekin Telcordia Technologies, Inc. 445 South Street, Momstown, NJ 07960 zyxwvutsrqpon zy zyxwvutsr zyxwvu zyxw zyxwvu Admission control of applications into each service class based on the bandwidth capacity allocated to each service class and its particular QoS objectives that sufficiently adapt to the dynamic state of the underlying network which could be a highly mobile ad-hoc network in a lossy wireless environment. ABSTRACT In this paper, we discuss empirical, analytical and simulation-based performance studies f o r an integrated IP QoS architecture implementing QoS resource management over a heterogeneous wireless and wireline network. The intearated IP 00s - architecture is based on Assured Forwarding (class-based) Differentiated Services and admission control of individual flows into each service class by a centralized server, called the Bandwidth Broker (BB) managing the network. The results we present in this paper are used as guidelines in designing the capacity estimation algorithm for admission control and optimizing QoS resource management within our integrated IP QoS architecture. These results serve as an instrument to understanding how to perform effective QoS resource management, using class-based differentiated services and admission control to guarantee class-appropriate end-toend QoS over IP networks. I I. INTRODUCTION We have developed and implemented an integrated IP QoS architecture that assures QoS to real-time and non-realtime IP traffic with varying performance requirements (including mission critical voice and data) within a BBmanaged network. Our approach integrates two fundamental QoS functionalities that utilize the available network resources and adapt to the particular performance characteristics of the underlying networks in providing edge-to-edge QoS assurances. These two complementary QoS functions are: The very early design and implementation specifics of our integrated IP QoS architecture are discussed in [5], [6]. There have been significant modifications to the architecture to support multi-class services and ad hoc wireless environments, which we will detail in an upcoming publication. In this paper, we focus on presenting the results of empirical, analytical and simulation-based performance studies that we have performed to use as guidelines in designing the capacity estimation algorithm for admission control and optimizing QoS resource management within our integrated IP QoS architecture. First, as a baseline study, we present an extensive simulation analysis of VoIP traffic using an ITUT P.59 artificial speech model [7] over various link speeds. Our end-to-end performance objectives are: PSTN equivalent VoIP quality (less than 2% packet loss rate (PLR) and 150msec end-to-end delay) and 70-90% link utilization. We describe how these objectives can be met, using a combination of several networking technologies. Comparison analysis with theoretical results is also provided. Then, we present empirically, the efficacy of service differentiation by provisioning bandwidth for the real-time traffic flows, while non-real-time non-critical traffic gains access to the unused portion of the link bandwidth. zyxwvut zyxwvutsr Differentiated Services based QoS resource management to support traffic marking, classifcation, and differentiated and class-appropriate treatment of aggregated classes of traffic that are adequately segregated from one another Ill, [21, [31, [41. 0 2002 Telcordia Technologies, Inc. All Rights Reserved. This research has been sponsored by the U S . Defense Advanced Research Projects Agency under contract MDA972-00-9-009. 0-7803-7625-0/02/$17.00 @ZOO2 IEEE. Finally, we present empirical test results for multiple traffic classes mapped to one of our four Assured Forwarding [3] per-hop behaviors (PHBs), including mission-critical TCP data, VoIP, Video and UDP data, in the presence of aggressive best-effort congestion. We provide empirical results for two scheduling algorithms (WRR and Priority-based) used for link sharing and describe the impact of the scheduler selection and the parameters on the performance of each class. 1189 ~ zyxwvutsrqp zyxwvutsrqp zyxwvuts zyxwvutsrq zyxwvut These results serve as an instrument to understanding how to perform effective QoS resource management, using class-based differentiated services and admission control to guarantee class-appropriate end-to-end QoS over IP networks. 11. CAPACITY ESTIMATION OF VOIP TRAFFIC We have performed extensive simulations using the ns-2 [SI simulator to identify appropriate packet size, buffer size and number of VoIP calls that can be multiplexed over a given link bandwidth for various types of links and targeted per-node loss, delay and jitter objectives. Due to the space limitation, we only present a subset of the results in this paper. goal including the over-the-air loss, our objective is to limit the queuing loss to less than 0.2% per node. C. Single Link Performance Results Figure 2 and 3 show the transmission + queuing delay and packet loss rate over a T I (1544Kbps) link vs. the number of 16Kbps VoIP sources for various buffer sizes. The packet size used is here is 80bytes and the transmission portion of the delay over the T1 link is approximately 0.4 msec. A. Simulation Environment Each VoIP flow was modeled as an exponentially modulated on-off process, with the mean on and off times, as per the F U P.59 [7] recommendation, being 1.008 sec and 1.587 sec respectively with hangover time of 200 msec. All VoIP flows are modeled as 16Kbps voice over RTP/UDP/IP. Without header compression, RTF'+UDP+IP headers impose 40-byte overhead to each VoIP packet. However, with CRTP header compression, the 40-byte overhead is reduced to 4 Bytes. In this paper, we consider two types of packet sizes as shown in Figure 1. Figure 2: Transmission and Queuing Delay Figure 1: Packet Sizes B. End-to-End Delay and Loss Budget The end-to-end performance objectives are: PSTN equivalent VoIP quality (less than 2% PLR and 150msec end-to-end delay) and 70-90% link utilization. End-to-end delay includes access delay, packetization delay, propagation delay and per-node queuing and transmission delay. Here, except for the number of hops and per-node queuing delay, other factors are mostly fixed and rely on physical distance, underlying networking technologies, and voice CODEC in use. Considering the fixed portion of the delay, our objective is to limit pernode maximum queuing delay to less than l0msec. End-to-end PLR includes over-the-air loss and queuing loss due to buffer overflow. In order to meet the 2% PLR Figure 3: Packet Loss Rate The results show that with increased number of VoIP flows, the delay and loss grow exponentially. Therefore, in order to meet our performance objectives, the Bandwidth Broker (BB) in this example should not admit more than 87 flows on to the T1 link. Increasing the buffer size helps lower the PLR up to a certain point. However, it should be noted that it also contributes to increased delay and jitter in packet delivery. Therefore, unless link utilization is significantly lower than the goal (70-90%), the buffer size should be kept less than 10 msec. 1190 D. Packet Size Comparison zy zyx zyxwv zyx zyxwvut E. Comparison with Capacity Scaling Equation Figure 4 and 5 summarize the results for various link speeds for 80-byte and 44-byte packet sizes respectively. The results illustrate that with 44-byte packets (less overhead than with SO-byte packets), not only more flows can he admitted on to the link, but also higher utilization could be achieved. This is due to higher statistical multiplexing gain obtained with smaller packet sizes. The difference becomes more significant at lower link speeds. In either case, utilization of low-speed links (lower than 512 Kbps) is lower than higher-speed links. Utilization on lower speed links can be improved by: 1) decreasing the packet size by compression or by selecting different types of CODEC, and 2) increasing the buffer size. However, as illustrated in Figure 2 and 3 above, there is a delay tradeoff with increasing the buffer size. Another alternative is to select less stringent PLR objectives if the number of hops on the end-to-end path is expected to he small or high quality links (low bit error rate, low loss) are used along the path. In this section, we compare our simulation results with the Capacity Scaling Equation [9]. The capacity scaling equation (1) helow estimates the capacity C, required to serve n, users when the capacity C, to serve n, users is known. Here, m is average transmission rate per user. H is the Hurst parameter, which represents the "degree" of selfsimilarity in the traffic - For Poisson-like traffic H i s 0.5, and for Self-similar traffic it is usually 0.9. (0.5 for VoIP, and 0.9 for data traffic). Figure 6 compares the simulation results with the capacity scaling cuwe. The scaling curve is derived from above equation using the simulation result for 1024 Kb/s link as a pivot point. The figure shows that the estimated results from the capacity scaling equation closely match the simulated results. Therefore, since simulations do not practically scale to very high capacity links, the capacity scaling equation can be used instead to estimate the optimal number of flows to be admitted over high-speed links. However, for lower capacity links, the estimated results deviate from the simulated results. This is due to the diminished statistical multiplexing gain, and it is desirable to rely on simulated or actual measured results than on analytically estimated results for lower capacity links. zyxwvuts Figure 4: Results Summary - SO-Byte Packets Figure 6 Capacity Scaling for VolP zyxwvuts In. EFFICACY OF DIFFSERV AND ADMISSION CONTROL BASED QoS ARCHITECTURE Figure 5: Results Summary - 44-Byte Packets So far, we have presented ways to estimate maximum number of single-type flows that can be admitted (i.e., maximum utilization) on a given link in order to assure the 1191 zyxwvutsr zyxwvutsrq zyxwvutsr zyxwvut In this section, we present empirical results obtained over our integrated QoS testbed as shown in Figure 7, to show the efficacy of service differentiation coupled with admission control to assure the QoS of each traffic type using DiffServ-based classification, conditioning and class-by-class segregated treatment of different types of traffic aggregates, and BB-managed admission control of traffic into each service class based on available capacity. Within each traffic class, methods we discussed in previous section can be used to estimate and limit the maximum number of flows to be admitted for the particular traffic type. A. Integrated QnS Testbed *nq* Figure 8: Service Class Description B. Empirical Performance Results To verify the efficacy of service differentiation, we ran performance tests with the following scenarios, empirical results of which are summarized in Figure 9 below: Single Class Test: Test with each class separately, using the class-specific bandwidth share as described in Figure 8. Multi-Class Test Without DiffServ: Test with multiple classes simultaneously without any DiffServ mechanisms using the class-specific bandwidth share. zyxwvutsr zyxwvu zyxwvut Multi-Class Test With DiffServ: Test with multiple BEUOsETR~C classes simultaneously with DiffServ mechanisms in place using the class-specific bandwidth share. The DiffServ mechanisms include per-flow conditioning at the ingress edge router and packet scheduling over all network routers. We have tested with two types of scheduling - Priority-based and WRR scheduling. Figure 7: Testhed Configuration All machines in the testbed illustrated in Figure 7 are running the Linux Operating System. There are 4 routers in the network and a combination of 100/10 Mbps Ethernet and 5 Mhps 802.11 links are used. We have implemented over our testbed IETFs Assured Forwarding (AF) [3] Expedited Forwarding (EF) [4] and standards as well as the hest-effort service as required [Z]. We have used the EF service class for control and BB signaling traffic to enable expeditious handling of traffic admission control into each one of the four AF service classes, and configuration of per-flow DiffServ policies to classify, condition and mark admitted flows at the network ingress. We have used the four AF service classes to map into each and segregate TCPIIP, VoIP, Videom, and other UDPIIP based applications traffic, for which admission control, i.e., QoS, is requested. Unadmitted flows and unmarked packets are treated as Best Effort (BE) traffic. Figure 8 describes the specifics of each service class. Perflow conditioning is performed at the ingress edge router and packet scheduling (Priory-based or WRR) is done at every router in the DiffServ (DS) domain. Figure 9 Summary of Performance Tests and Results Results show that without service differentiation and admission control (No DiffServ), each service class experiences significant packet loss and reduction in throughput in the presence of aggressive hest effort traffic. However, with DiffServ mechanisms in place, dramatic 1192 zyxwvuts zyxwvutsrq zyxwvutsrqponmlkjihgfe zyxwvutsrqpon zyxwvutsrqponmlkjih zyxwvutsrqponmlkjih zy zy concluded that the capacity scaling equation can be used for high-speed links instead of simulations to estimate the maximum number of flows that can be multiplexed into a service class since the simulations do not scale well to very high capacity links. improvement in both packet loss rate and throughput (i.e., utilization) is observed. Results also highlight that WRR scheduling performs slightly better than Priority-based scheduling. zyx C. Effects of Increased Buffer Size on PLR and Utilization for VOW I . I I t We also presented empirically, the efficacy of service differentiation and admission control by provisioning and utilizing bandwidth optimally across multiple traffic classes, even in the presence of aggressive best-effort congestion. We provided empirical results for two scheduling algorithms (WRR and Priority-based) used for link sharing across service classes and describe the impact of the scheduler selection and the parameters on the performance of each class. .: :m.Ecoa9p. iUaiZalh”280% ~ .............. .........a$ ....... ............. rP1:A;IC ... .... . . ! Iu-z79pa , I ’.:’: ~ . . . . ” ’ I ............................................ . ......... I ................... ................... ..... .....: ..................... . 1.00 . 0.00 0 5 ..... .... y.; ...... , , 10 ( 5 ’. ,..........., ........................... , , , P 30 35 , 40 , , ;j , , , -,- , ,: , The results we presented in this paper serve as an instrument to understanding how to perform effective QoS resource management, using class-based differentiated services and admission control to guarantee classappropriate end-to-end QoS assurances over IF’networks. ~ ... ,- , d. YI 15 €3 65 70 75 80 85 80 ,. 5 os rm Maximum ousuing Delay (mxc) -PLR (107VolPca15) -‘.~lilabon(107VolPcallr) Figure 10 Delay vs. PLR and Utilization for VoIP ACKNOWLEDGMENT Figure IO illustrates the effect of increasing the buffer size (thus the maximum queuing delay) on PLR and utilization for VoIP traffic. It can be observed that with increased buffer size, both the PLR and utilization performance improves. However, considering the end-to-end delay, the buffer size should not be increased excessively. For the example in Figure 10, the advisable size of the buffer would be approximately between 20 m e c and 60 msec where reasonable PLR and utilization performance could be achieved. We gratefully acknowledge Ken Young for his oversight and support on this project. REFERENCES [l] SBlake, D.Black, M.Carlson, E.Davies, Z.Wang. W. Weiss, ?in Architecture for Differentiated Services,” RFC 2475, December 1998. [2] K.Nichols, %Blake, F.Baker, D.Black, “Definition of the Diflerentiated Services Field (DS Field) in the IPv4 and IPv6 Headers,” RFC 2474, December 1998. N. CONCLUSIONS [3] J.Heinanen, F.Baker, W.Weiss, J.Wroclawski, ”‘Assured Forwarding PHB Group,” RFC 2597, June 1999. [4] V.Jacobson, K.Nichols, K.Poduri, “An Expedited Forwarding PHB,” RFC 2598, June 1999. zyxwvutsrq In this paper, we presented the results of empirical, analytical and simulation-based performance studies that we have performed to use as guidelines in designing the capacity estimation algorithm for BB admission control and optimizing QoS resource management within our integrated IP QoS architecture. [ 5 ] K X m , P.Mouchtaris, SSamtani, R.Talpade, L.Wong, “A Bandwidth Broker Architecture for VolP QoS,” Proceedings of SPIES Intl Symposium on Convergence of IT and Communications (ITCOM). . . Aumst 2001. I First, we presented extensive simulation analysis results for V o P traffic over various link speeds. Our end-to-end performance objectives were: PSTN equivalent VoIP quality (less than 2% packet losses and 150msec delay) and 70-90% link utilization. We described how these objectives could be met, using a combination of several we then the - technologies, simulation results with analytical estimations using the Capacity Scaling Equation. The capacity scaling equatiouestimated results, especially for higher capacity links closely matched the simulated results. Therefore, we [6] K.Kim, P.Mouchtaris, SSamtani, R.Talpade, L.Wong, ‘ X Simple Admission Control Algorithm for IP Networks,” Proceedings of Intl Confon Networking, July 2001. [7] International TelecommunicationsUnion Recommendation P.59, “Artificial Conversational Speech,’’ 1993. [8] “the network simulator - ns2,” htto:l/www.isi.edulnsnan~nsl. [9] K.Krishnan, A.Neidhardt, A.Erramilli, “Scaling Analysis in 1193 Traflc Management of Self-similar Processes.” Proceedings of the 15th International TeletrafficConference (ITC), June 1997.