Traffic Analysis for Voice in Wireless IP Networks

Transcription

1 Traffic Analysis for Voice in Wireless IP Networks TONI JANEVSKI BORIS SPASENOVSKI Deartment of Telecommunicatis, Faculty of Electrical Engineering University Sv. Kiril i Metodij Karos 2 bb, 1000 Skoje Abstract: - In this aer we rovide an extensive analysis of voice service in wireless IP networks. Voice traffic is serviced with riority over the rest of the traffic. Hence, we ccentrate the voice traffic in the analysis. We model the acketized voice traffic with Markov Modulated Poiss Process (MMPP), and we roose an analytical framework for its analysis in wireless networks. Also, we created a simulator in Matlab, with caability for QoS analysis in different network scenarios, csidering user call intensity, voice-encoding rate, link caacity and buffer sizes. The observed QoS arameters are acket loss and delay. Voice traffic is very sensitive to delay, while some low losses may be tolerated. We resent overwhelming QoS analysis of IP telehy traffic at different network setus and give a ccet for dimensiing wireless links for IP telehy under given cstraints the QoS arameters. Key-Words: - Wireless, IP, Quality of Service, Traffic, Voice 1 Introducti Voice service was mainly based circuitswitched technology in the ast, and it is still nowadays. However, develoment of comuter industry and low cost of communicati devices (alm to devices, communicators, mobile hes, la-to comuters etc.) moved the telecommunicatis beyd the voice service. Now, when the enetrati of users in wireless cellular networks almost reached its maximum, telecommunicati industry and telecom oerators are facing with a challenge of extending the market by introducing more additial services besides the traditial voice service. These services that are ered now or those that will be ered in near future, have no other alternative than IP technology and Internet. In such scenario, voice will be just e of the many services ered to the end users. However, it will remain the most used e and the oldest e. Furthermore, users are used to secific quality of the voice service that came from circuitswitched networks, and they will not accet noticeable degradati of that grade of service. So, the questi is how to design cellular access networks based IP that will rovide desired Quality of Service (QoS) for voice service, something to what users are already used to. Let csider the Internet side first. Many researches and committees are addressing QoS issues in Internet. However, main csidered issues are scheduling, differentiati and admissi. However, not much is de csidering the design of IP network carrying real-time traffic. One may justify this by claiming that if enough bandwidth is rovided, then we should not worry about the caacity and the traffic. But, even in the wire Internet voice service end-to-end is still a roblem due to main ccet used at the develoment of Internet i.e. the ctrol is left to the end entities of the communicati (e.g., to the client and the server). In [1] authors rovide analysis of real-time voice service through the wire access network. They have verified the MMPP model with measurements in a laboratory test-bed. On the side of wireless cellular networks, we have rovisi of certain Grade of Service (GoS) by designing cellular networks with given cstraints call-level QoS arameters: new call blocking robability and call droing robability. However, 1G and 2G wireless networks were based circuit switching, where for each call network allocates e channel. There, adated Erlang-B formula was used for the design of cellular network, [2] [3]. But, here we csider acket-based allocati of wireless link. The aer is organized as follows. In Secti 2 we describe the traffic and system model for our analysis. The analytical analyses are given in Secti 3. Next secti shows the results from simulati runs. Finally, cclusis are given in the last secti. 2 Wireless Link Model We erform design of the wireless networks by csidering the caacity er cell (the bandwidth) as well as mobility of the users that results in handovers amg the adjacent cells. The wireless link is based IP. Packets from different sources are multilexed

2 the wireless media, which is shared amg all traffic streams. We assume that voice over IP traffic is differentiated from data traffic, which is based TCP. If IP telehy traffic is mixed with TCP traffic, which is lg-range deendent, then it will add unmanageable acket delays and acket loss. In our revious work [4] we have roosed classificati of IP traffic into two main classes: class-a, for traffic with QoS cstraints, and class-b, for best-effort traffic. We further divide class-a into three subclasses: A1 for real-time traffic with cstant bit-rate (IP telehy), A2 for real-time traffic with variable bit rate (e.g. video streaming), and A3 for best-effort traffic with a cstraint acket delay (e.g. browsing). Today, there do exist mechanisms to differentiate traffic, such as differentiated services models [5]. We assume that IP telehy is differentiated from other traffic the wireless link, and it is not mixed with TCP traffic. Packets from IP telehy are buffered into searate buffers (of course, there are also other mechanisms to bound acket delay or loss). However, we use riority scheme to differentiate IP voice traffic from the rest. Csidering the voice service in a given cell, we may have active users (with going voice cnectis) and idle users. Traffic from all active users within the cell is multilexed to wireless link. We illustrate such scenario in Figure 1. source 1 source 2 source N wireless link to base stati Figure 1. Multilexing voice sources over wireless link The design issue is how many subscribers we can serve with given network resources, or how much bandwidth we need to have service for redicted number of users of the voice service. There are two different level of csiderati: 1. Cnecti level; and 2. Packet level. On a cnecti level voice service is well modeled by using Poiss rocess, i.e. exential distributi of inter-arrival times between csecutive arrivals as well as exential distributi of call durati time. Poiss rocess is well roven for voice service through the usage of Erlang-B formula, which is in fact M/M/c/c queue, where c is the number of channels (servers) the link. Figure 2. Wireless IP network architecture On acket level we rimarily csider ackets from voice telehy streams. However, different traffic tyes are multilexed the wireless link, as shown in a tyical wireless IP network architecture given in Figure 2. When a cnecti is established for a user, we may have a talk eriod (when the user is talking and ackets are generated to the link) and an idle eriod. During talk eriods IP ackets are transmitted back-to-back. We assume that there is no acket generati in idle eriods. We do not discuss the medium access layer and hysical layer. But, we suose that these layers rovide medium access solutis to avoid collisis from different sources when transmitting ackets, based divisi of time into time stes, i.e. semiermanent logical channels. Because we model the sources as - sources, we assume that during eriod each source has dedicated time interval for acket transmissi over the wireless link. 3 Traffic model A. Call level roerties Let csider a single cell in the network. We assume Poiss arrival rocesses in the cell, which is well roven for the voice cnectis. With n and h we denote new call and handover arrival rate in the cell. An going voice cnecti comletes at rate µ c or dearts the cell at rate µ h. But, in ctrary to the circuit-switched cellular networks we do not have a limited number of channels in the networks. We do use statistical multilexing of voice sources and other n-voice traffic over the wireless link. If we denote with the total call arrival rate, then: = n + h (1) Similarly, we denote with µ the total call dearture rate: µ = µ c + µ h (2) Then we obtain a birth-death rocess as shown in Figure 3, where N is the maximum number of active users in the cell.

3 0 1 2 µ 2µ 3 µ ( N 1) µ N-1 Figure 3. Markov chain model So, a user may be an active voice source or idle csidering voice service (in this case are included situatis when users are using n-voice services). B. Single voice-source roerties As for traffic models, voice cnectis arrive according to a Poiss rocess. Once a cnecti (or call) is established, the voice source is modeled as two-state Markov chain with e state reresenting the talk surt (ON) and other state reresenting the silent eriod (OFF) [6]. A simle ON-OFF model accurately models the behavior of a single voice source. During ON (talk) eriods the source is transmitting IP ackets. Most encoding schemes have fixed bit rate and fixed acketizati delay. During (silence) eriods the source sends no ackets. We assume that ON and OFF eriods are exentially distributed, which is well analyzed in [1]. The voice sources can be viewed as two state birth-death rocesses with birth rate (arrival rate for -eriods) and death rate (ending rate for -eriods). Then, 1/ and 1/ are duratis of talk eriod and silent-eriod of a voice source, resectively. Average of talk surt durati is s, and average of silent state durati is s. The talk surt has a cstant bit rate that we denote as b ts. Then, average voice cnecti bi rate is: b T Nµ av = bts = bts (3) T + T + For examle, if talk surt is encoded with 32 kbs, then average bit rate is kbs. Csidering the voice acket sizes there are also different decoders, i.e. in [1] are used encoders with 64 bytes ayload and 40 bytes IP overhead, while in [7] authors use voice stream of 180 byte ackets, with 160 bytes ayload and 20 bytes IP overhead. Csidering the fact that IP telehy is based RTP/UDP/IP rotocol stack, the accurate model should include 20 bytes overhead for IP, and 20 bytes overhead for RTP/UDP (12 bytes for RTP and 8 bytes UDP [1]). Under the assumti of fixed acket size for voice ackets, we may divide the ON-eriod into timestes with durati T, where T is time that voice coder needs to generate e voice acket with length l. Then, e may write: N 1 T = = T N (3) where N is the number of IP ackets during the ONeriod when the ackets are sent to the wireless link back to back. So, during the talk surt ackets are generated at rate b ts : l b ts = (5) T The robability that a user is in ON-eriod is given by: P user _ = Pcall _ Psource _ (6) where: P source 1 P call _ = (7) µ _ = (8) ( + ) where 1 is average call arrival rate for a single source. Csidering Figure 3, we may write: = N 1 (9) The robability that a source is OFF is P =1-P. C. Suerositi of voice sources The suerositi of the voice sources can be also viewed as birth-death rocess, where total incoming rate is sum of incoming rates of individual sources. A cvenient model in teletraffic theory for a suerositi of many - voice sources is Markov Modulated Poiss Process (MMPP). For voice sources with talk surts and silent eriods (without ackets link) it is more cvenient to use the secial case of MMPP, i.e. IPP (Interruted Poiss Process). When the rocess is in state j that means j sources are. Below we show transitistate diagram for suerositi of k active voice sources. k ( k 1) ( k 2) 2 3 k 1) 2 k-1 k ( k Figure 4. State transiti diagram for a suerositi of k voice sources (established voice cnectis) In Figure 4 we assume that all k cnectis are reviously established. Now, we may link this chain with Markov chain for calls given in Figure 3. To

4 rovide desired Quality of Service (QoS) for the voice calls we need to determine maximum number of active voice calls at a time. This number is denoted with N in Figure 3. So, at a given moment of time we may have k voice cnectis, where k n. If we have k active cnectis then we have the situati given in Figure 4 csidering ON and OFF eriods of sources. IP ackets arrive from k inut lines/sources. Each of the k calls occuying the inut lines i.e. wireless link alternates between talk surts and silent eriods. Assuming state-equilibrium in Figure 4, we may derive the distributi of active voice sources, i.e. the robability that j sources are ON out of k active voice cnectis is: b j k j = 1 + k j (10) Suerositi of many Poiss rocesses is also a Poiss rocess. So, for the wireless link we may csider e Poiss rocess where arrival rate for ON eriods is a sum of ON arrival rates of individual sources. One may notice that the same statement is valid for call arrival rocess from users. Csidering the assumti for exential distributi of call arrivals and call duratis, as well as exential distributi of ON and OFF eriods during e call, and by using (5) and (6), e may write: P{at least e source is ON}= (11) µ + ) ( 4 Simulati analysis We imlemented our model in Matlab rogramming envirment. As inut arameters we use number of IP voice sources, bandwidth of the wireless link, buffer length in the network nodes, voice encoding rate, call arrival and dearture rates. We observe acket loss and mean acket delay. Various voice coders for IP networks exist the market today. All of them are using RTP/UDP/IP rotocol stack. For the inut arameter values, csidering the simulator setu, we used the same values as found in [1], i.e., 32 kb/s ADPCM voice encoding with 16 ms acket inter-arrival time, which results in 64 bytes of voice ayload er acket, a rotocol overhead of 40 bytes (12 bytes for RTP, 8 for UDP and 20 bytes for IP). Link headers are not included. So, the total acket size is 104 bytes. We use fixed size for all voice ackets. The durati of talk-surts and silence eriods is exentially distributed the ositive integers with a mean of 0.351s for ON-eriods and 0.650s for the OFFeriods. The results from the simulatis are divided into two main grous. The first grou csiders voice sources that are ermanently active (during the whole simulati eriod). This assumti does not reflect the real situatis in the network. However, it is useful for calculating the limits of the network caabilities. The number of sources, the number of buffers and the bandwidth of the outut link are the inut arameters while the average acket loss and the average acket delay are the outut arameters of the simulati rocess. Figures 5 and 6 show the average acket loss and the average acket delay as a functi of the number of buffers (measured in ackets). The number of sources is used as a arameter. The bandwidth of the outut link is 2 Mb/s and the voice sources have bit rates of 52 kb/s in ON eriods. The simulati eriod is 1 hour. We can see from these grahs that the average acket loss, regardless of the number of sources, raidly decreases until certain buffer length. For 40 and 80 voice sources, the number of buffers that is of articular imortance is 20 and 40 buffers, resectively. Further enlargement of buffer length does not lead to better results, and the average acket delay increases with increasing the number of buffers. This means that the revious buffer length gives otimal results, regarding the average acket loss and average acket delay. When the network link is loaded with 110 voice sources, after reaching buffer length of 40 ackets, the average acket loss stays retty much the same while the average acket delay increases linearly. According to the recommended QoS arameters [8] and csidering the results of the simulati runs, the otimal buffer length for this articular situati is around 45. With roer analysis of the simulati results, we can give certain guarantees for the quality of the voice services that are being ered by the roviders to their csumers. Also, we can determine the margins of the csumers that can receive adequate quality regarding the network infrastructure. The secd set of results reflects the behavior of the real systems. This means that the activity of the voice sources is modeled csidering statistical analysis of the real voice traffic in the acket-based networks. We csider the erformance model of a single cell in a cellular wireless communicati network. We use Poiss arrival stream of new calls at the rate n and the Poiss stream of handover arrivals at the rate h. An going call (new or handover) comletes service at the rate µ c and the mobile engaged in the call dearts the cell at the rate µ d. These rates are used as inut arameters in the simulati. The bandwidth of the outut link, the number of the buffers and the sources bit rates are

5 also used as inut arameters. The outut arameters are the same as in the revious set of results. Our main task was to create a set of grahs that can be used for dimensiing wireless cellular networks for IP telehy. Figures 7 and 8 show the average acket loss and the average acket delay as a functi of the number of total calls (new and handover). The number of buffers is used as a arameter. The bandwidth of the outut link is 2 Mb/s and the voice sources have bit rates of 52 kb/s in active eriods. From these grahs we can see that buffer length of 20 ackets is comletely unusable. Also, we can determine the range of the inut voice traffic that is eligible to certain, desirable quality. After we determined this range, in order to find the otimal buffer length, we made two more grahs, i.e. Figures 9 and 10, which show the average acket loss and the average acket delay as a functi of the number of buffers. The number of total calls is used as a arameter. With roer analysis of these grahs, according to the required quality of service, we can easy inoint the most effective buffer length that can be used in the articular network. Usually we have given cstraints both arameters, i.e. acket loss and delay. While real-time alicatis (such as IP telehy) have stringent demands acket delay, they can tolerate reasable acket loss (e.g., losses < 1%). We should choose the buffer size as a balance to either of the two cstraints. Of course, e should aly admissi ctrol mechanism that will satisfy the cstraints loss and delay, i.e. reject the incoming calls to rovide the desired quality to the going calls. Figures 11 and 12 show the average acket loss and the average acket delay as a functi of the voice source s bit rate (this means different voiceencoding techniques). The number of buffers is 40 and the link is loaded with 1.0 calls/s (the bandwidth of the outut link is 1.5 Mb/s). We may notice that the average acket loss decreases u to 90 % going from 52 kb/s to 28 kb/s rate of the voice-encoding scheme, as given in Figure 11. Also, the average acket delay also decreases significantly, Figure 12. According to these results we may cclude that by reducing the encoding bit rate we can either increase the quality of the ered services (the network QoS) for the same number of csumers, or increase the number of the csumers that will be able to receive the revious quality of service. 5 Cclusis In this aer we analyzed the QoS for voice service over the wireless IP networks. However, the analysis may be extended to the wired es. We assumed that voice is serviced with riority over other services the link. Figure 5 Average acket loss vs. number of buffers for different number of sources Figure 6 Average acket delay vs. number of buffers for different number of sources Figure 7 Average acket loss vs. call intensity for different buffer sizes Figure 8 Figure 7 Average acket delay vs. call intensity for different buffer sizes

6 One model that rovides such differentiati of voice traffic is differentiated services model. We resented an analytical framework for the analysis of the voice service in wireless IP networks. We used ON-OFF MMPP model, which is well roven for modeling acketized voice traffic [1]. We imlemented our model in Matlab. Using the simulator we erformed analyses of acket loss behavior and delay for different traffic and network scenarios, i.e. by using different inut arameters, such as call intensities of the users, number of voice sources multilexed the link, voice-encoding rates, link caacity, buffer sizes, to obtain their influence the QoS arameters. Csidering the results shown in the revious Secti, this model can be alied in several different situatis. One tyical scenario is when we have a given bandwidth and we need to rovide certain QoS to the users, according to the delay requirements secified in [8] as well as a given cstraint the acket losses. In such case, we may use the simulati results for access network dimensiing wireless links for IP telehy, or, as an inut to an admissi ctrol algorithm. References [1] B. Ahlgren at el., Dimensiing Links for IP Telehy, SICS, CNA Laboratory, Sweden. [2] L. Gavrilovska, T. Janevski, Modeling Techniques for Mobile Communicatis Systems, GLOBECOM 98, Sydney, Australia, November [3] G. Harring, Loss Formulas and Their Alicati to Otimizati for Cellular Networks, IEEE Transactis Vehicular Technology, Vol. 50, No. 3, May [4] T. Janevski, B. Sasenovski, QoS Provisiing for Wireless IP Networks with Multile Classes Through Flexible Fair Queuing, GLOBECOM 2000, San Francisco, USA, November [5] C. Dovroolis et al, Proortial Differentiated Services: Delay Differentiati and Packet Scheduling, SIGCOMM, [6] M. Schwartz, Broadband Integrated Networks, Prentice Hall, [7] D. Cavendish, A. Dubrovsky, M. Gerla, G. Reali, S.S. Lee, Statistical Internet QoS Guarantees for IP Telehy, UCLA CSD TR# [8] ITU-T, One-way transmissi time, Recommendati G.114, Geneva, Switzerland, March [9] T. Janevski, B. Sasenovski, Admissi Ctrol for QoS Provisiing in Wireless IP Networks, Euroean Wireless 2002, February Figure 9 Average acket loss vs. buffer sizes for different call intensities Figure 10 Average acket delay vs. buffer sizes for different call intensities Figure 11 Average acket loss at different voiceencoding rates 2 Mb/s link Figure 12 Average acket delay at different voiceencoding rates 2 Mb/s link