VoIP Call Capacity over Wireless Mesh Networks - Semantic Scholar

VoIP Call Capacity over Wireless Mesh Networks Md. Atiur Rahman Siddique and Joarder Kamruzzaman Faculty of IT, Monash University, Australia Email: {Atiur.R.Siddique, Joarder.Kamruzzaman}@infotech.monash.edu.au Abstract—In recent years, research on VoIP over Wireless Mesh Network (WMN) has gained particular attention because of its commercial prospect. This paper presents an analytical method to estimate VoIP call capacity in an WMN employing IEEE 802.11 devices. We used Markov chain analysis of IEEE 802.11 for network delay and loss to estimate the capacity while using Rating factor, R score, defined by ITU-T, to ensure call quality. A detailed analysis of queueing delay and loss in terms of network parameters is also carried out along with their impacts on voice quality. The capacity model estimates call capacity in a single hop WLAN and is extendable to multi hop scenario and for video communications. The theoretical results are verified by simulation and compared to related previous works.

I. I NTRODUCTION VoIP (Voice over IP) is by far the cheapest voice communication mechanism. Although voice quality was initially poor due to low Internet bandwidth, by 2000, broadband was widely deployed and most voice quality issues were resolved. As the vendors (e.g., CISCO & Avaya) and service providers (e.g., Vonage & Time Warner) started competing, service quality increased and VoIP became incredibly popular very quickly. Only Skype to Skype calls reached a 100 billion minutes from 2003 to February, 2008. According to a report by Infonetics1 , in 2009, only north American VoIP market will see a 20 billion dollar revenue. This clearly indicates a thriving market. Yet, not more than 7% of potential customers use VoIP. Most prefer mobile phone since the convenience of mobility is so alluring that it can not be simply set aside. Traditionally, VoIP calls are initiated from fixed lines (wired broadband, PSTN, ATM, DSL or ADSL) since it is the cheapest mechanism to connect to the Internet. Although cellular networks (e.g., GSM and CDMA) can be used to provide mobility to VoIP customers these services incur a high cost which is ultimately pushed down to end customers and the VoIP service becomes costly. Since cheapness of service is the ultimate commercial attraction of VoIP, this solution is neither popular nor commercially viable. While a cheap wireless technology is essential to provide the last mile coverage, IEEE 802.11 might be just up to the task. IEEE 802.11 compatible nodes are self configuring relieving of any fixed maintenance cost. Moreover, these networks do not require deployment of cellular network infrastructure. Only a few AP (Access Points) are required to be deployed which are easy to deploy and maintain. Easy and cheap deployment, operation and maintenance make it a lucrative alternative for the last mile coverage. But as the major part of the WMN consists of client nodes which do not offer much computational power and due to the inherent nature of wireless networks in 1 http://www.infonetics.com

attempting to squeeze in any traffic load, an accurate theoretical estimation of voice capacity is required to guide the WMN design process so that maximum calls can be supported without degrading quality. Moreover, formation of hotspots often creates bottlenecks and requires special attention. In this paper, we develop an analytical capacity estimation model which will be extremely useful in determining call capacity and designing WMNs. The main contributions of this paper is an analytical model to estimate VoIP call capacity over WMN ensuring voice quality. We also analyzed the impact of queue length and queueing delay on call capacity. The analysis can be extended to multihop scenario as well as for video communication. Although there are several works on VoIP capacity over WMN in general but to the best of our knowledge, no work exists in literature that considered voice quality metrics (hence over estimates call capacity) or analyzed the impact of queue in estimating capacity. II. VO IP S YSTEM AND WMN An encoder decoder pair is agreed upon by the talker and listener during the session initiation phase. At the talker end, a voice capture module records the voice and encodes it in PCM data. The encoder reads chunks of PCM data, encodes it into IP packet and hands off to lower layers for transportation. Voice traffic is generally transported over UDP, but UDP does not support sequencing which is very crucial for the decoder. Therefore, RTP is used on top of it. At the receiving end, after stripping off protocol headers, the voice frames are put in a dejitter buffer. The decoder requires data at regular interval but the underlying network can not guarantee their delivery just in time. Therefore, the dejitter buffer plays a crucial role. The decoder picks data frames from it and decodes to PCM and put in a play out buffer from where it is then played out. We assume both MAC and the PHY layers are IEEE 802.11 compliant. IEEE 802.11 standard defines two medium access mechanism, the randomized contention based DCF (Distributed Co-ordination Function) and the time synchronized PCF (Point Co-ordination Function). PCF was designed to support multimedia and other delay sensitive traffic but was made optional in the original standard. It is not a part of the Wi-Fi standard and is generally absent in commonly available 802.11 devices. On the other hand, DCF is popular both in literature and industry. DCF employs CSMA/CA mechanism to access the channel. This paper assumes a DCF based MAC with BEB-m backoff algorithm.

978-1-4244-2324-8/08/$25.00 © 2008 IEEE. This full text paper was peer reviewed at the direction of IEEE Communications Society subject matter experts for publication in the IEEE "GLOBECOM" 2008 proceedings.

III. R ELATED W ORKS Several works analyzed VoIP performance over WMN. Most authors have used simulation and testbeds in their studies. Awoniyi & Tobagi analyzed effect of fading on VoIP performance [1]. VoIP capacity in presence of real data traffic is studied by Anjum et al. in [2]. A measurement based capacity estimation model is explained in [3]. The authors have used an interference map based on which VoIP traffic is routed through the least interfered zones. All of these works have analyzed results from simulation or testbeds and none provided any analytical model to give a precise estimate of VoIP capacity which can be used in network designing and planning. Some authors have provided analytical models to estimate VoIP call capacity. One of the earliest and most cited works in determining VoIP call capacity over single hop WLAN is presented by Garg & Kappes [4]. The authors have used an 802.11b testbed to assess the call capacity of G.711 a-Law codec. A similar and simplistic study by Hole & Tobagi is presented in [5]. But both [4] and [5] over estimated capacity by ignoring queueing delay, queueing loss, dejitter buffer delay and most importantly the effect of packetization delay. In fact, we show that when the packetization delay is very high, it brings capacity down. Both of these models are based on over simplified assumptions, for example, BEB (Binary Exponential Backoff) is always at its first retry stage [5] or average backoff delay is constant irrespective of number of contending nodes [4]. Moreover, both models have estimated capacity merely by bandwidth consumption of codecs which is inadequate to ensure voice quality. We put utmost importance on voice quality, because if quality is dissatisfactory, customers will never use VoIP. The focus of this paper is emphasis on QoS assurance and incorporation of codec parameters in theoretical capacity estimation and extendibility to multihop scenario. We considered the characteristics of the queue to analyze their effect of voice quality. The model is developed so that it can be used for one/m way voice/video calls and extendable to multihop and for any medium access control mechanism. In the following section we present the capacity model for contention based medium access. IV. C APACITY A NALYSIS M ODEL To develop the capacity model we follow a top down approach. We assume that the call quality will be maintained and deduce the requirements to ensure it following ITU-T guidelines We then formulate an optimization problem in terms of network parameters with the quality ensuring requirements as constraints.. We assume silence suppression is not used since its use incurs inconvenience to the listener and most popular VoIP systems like Skype does not use it [6]. Therefore, the traffic can be modeled as two CBR (Constant Bit Rate) connections in opposite directions i.e., each encoder generates a fixed length packet at regular interval and completely independent of the talk/silence pattern of either talker and listener. With

this independence of traffic in opposite direction we model one way traffic and divide end capacity by two since two way traffic simply produces twice the number of packets on the same path compared to a one way call and at each collision domain on the end-to-end path twice the number of transmissions are required. Thus the model is also useful for 1/2/m way multimedia traffic e.g., VoIP, video chat, IP TV or IP radio. In case of m way traffic the end capacity is to be divided by m. To ensure voice quality we start with voice quality assessment method. A. Call Quality Assessment The most accepted voice quality assessment method, MOS (Mean Opinion Score) [7], requires human experimenters to listen to test voice signal and rate its quality in an scale of 1 to 5. The other widely used methods are E-model [8] and PESQ (Perceptual Evaluation of Speech Quality) [9]. PESQ requires a reference signal to be injected in one end of the network, while at the other end, the received signal is compared to the reference signal to assess voice quality. Since both MOS and PESQ requires an actual network to be used, they are unsuitable to be used in network planning. On the other hand, E-model does not require any of these, accounts for network parameters to assess voice quality for a proposed network and is an appropriate method to be used in network designing. E-model represents voice quality with a scalar value called R-score. Since MOS is the most accepted voice quality assessment method we start with MOS and then map it to R. MOS requires the M OS level to be 3.60 or above for a medium quality call and 4.03 or above for a high quality one. Using R to M OS level mapping [8] the R-score requirement for different quality calls canbe represented as shown in (1). 70.07, medium R≥ (1) 80.16, high Cole [10] and Meddahi [11] simplified the E-model to make it usable for IP based packet network. Cole’s model, being more simplistic, is used by most researchers. We follow a method similar to Cole. ITU-T Recommendation G.107 [8] defines R as a sum of various impairments. R = Ro − Is − Id − Ie ef f + A Here, Ro is the basic SNR including room noise, Is is the simultaneous impairment which is a combination of impairments occurring more or less simultaneously with the voice signal (e.g., SLR, TELR, etc.), Id is the delay impairment factor, Ie ef f is the effective equipment impairment factor and A is the advantage factor which allows for compensation of impairments when there are other advantages to the user. Many parameters of these impairments are irrelevant in an IP packetized network and ITU-T recommends default values when these parameters are not of interest. Using ITU-T recommended default values, we find Ro = 94.7688 and Is = 1.4136. Advantage factor A has a default value of 0 for wired network, 5 for mobility within building and 10 for mobility in geographic area. In fact, an advantage factor of around 10 is cited as appropriate for cellular network through private communication in [10]. In this work, we assume limited, non-frequent mobility


Table I: Co-efficients of Id (d) approximation

Table III: Encoder characteristics

Coefficient

d < 125

125 ≤ d ≤ 225

225 < d

Encoder

lF (ms)

dL (ms)

A B C D

3.84e-7 4.12e-4 -0.1048 8.959

5.62e-7 -1.63e-4 0.036 0.1476

-4.45e-7 -2.55e-4 0.08968 -17.22

G.729A G.711 G.723.1 G.726

10 10 30 5

5 0 7.5 0

50 45

Delay Impairment, I

d

40

ITU−T[8] Cole et. al. [10] Our approx.

35 30 25 20 15 10 5 0 0

100

200

300

400

500

End to end delay, d (ms)

Figure 1: Approximation of Id (d) Table II: Codec Parameters for Determination of Ie Codec

Ie

Bpl

G.729A+VAD G.711 G.723.1+VAD

11 0 15

19.0 4.3 16.1

ef f

and use A = 5. This reduces the requirement on R to a budget of only Ie and Id as shown in (2) for different quality calls. 28.3553, medium Id (delay) + Ie ef f (loss) ≤ (2) 13.1952, high ITU-T [8] defined Id by a set of 29 functions of various delay parameters which are not of interest in a packetized network. Using default values for these parameters we approximate Id with a stepwise third order function of only end-to-end delay, d. Cole [10] has shown a similar approximation with a stepwise linear function [10]. But at certain ranges it is not accurate. Our approximation is shown in (3) with co-efficients defined in Table I. A comparison of the approximations in Fig. 1 shows our method closely follows ITU-T recommended values. (3) Id (d) = Ad3 + Bd2 + Cd + D On the other hand, equipment impairment factor Ie ef f depends on the codec and is a function of end-to-end loss. Previously it was produced as tabulated values depending on the packet loss as presented in Appendix I/G.113 [12] of which Cole [10] used a log function to approximate. In 2005, ITUT defined Ie ef f as a function of both end-to-end loss and its randomness. Packet loss randomness depends on specific implementation. In absence of such information, we assume random and independent loss and present Ie ef f as in (4) [8]. e Ie ef f = Ie + (95 − Ie ) (4) e + Bpl Here, Ie is the equipment impairment at zero loss, e is the end-to-end loss and Bpl is the packet loss robustness factor. Ie and Bpl depends on the codec and their standard values are shown in Table II. B. End-to-end delay, d and loss, e End-to-end delay d is calculated as the delay from mouth to ear. With aggregation level nA the encoder will read nA number of voice frames, each of length lF , encode them and put

the encoded data in a single UDP packet. Therefore, an initial a delay of nA lF is introduced. Additionally, some encoders look into the succeeding frame to improve compression efficiency for the current one. This introduces a look ahead delay denoted as dL . The delay in dejitter buffer in the receiving end is denoted as bg [10] where 2b is the length of dejitter buffer in number of packets and g is the packet inter arrival delay. Any extra time taken to encode the voice data is taken to be close to zero assuming a high speed computing system. Average network delay is a sum of delays in protocol stack, the queue, dQ and medium access, dM . Assuming delay in protocol stack to be close to zero, the end-to-end delay can be represented as (5) d = dL + nA lF + dI + dQ + dM + bg where dI is the delay in the Internet and depends on the distance and route. The network designers can predict its value depending on the expected call pattern. In absence of such information we assumed dI = 0 as done in previous related works [4], [5], however, our model accommodates such delay when an estimated value is available. The decoder is reasonably fast compared to the network. So, no queueing delay is expected at the receiver node. Our definition of d defers from that of [10], which did not consider queueing delay at all. As discussed later in details, we found queueing delay to be a limiting factor for voice quality, specially for short frequent packet and in presence of regular data traffic. Moreover, the break down of d is an essential part to keep provision for extension to multihop. Sample values of some parameters for different codecs are given in Table III. Dejitter buffers can be static or dynamic. We assume a static dejitter buffer which uses a fixed buffer length and tries to minimize the loss. Following the work in [10], we assume a dejitter buffer of bg is sufficient to keep dejitter buffer loss close to zero. We use dejitter buffer loss, eJ as shown in [10] and derive end-to-end loss, e from MAC layer loss, eM and Queue loss, eQ and eJ as shown below. e = eM + (1 − eM )eQ + (1 − eM − eQ )eJ (6) C. MAC layer delay, dM and MAC layer loss, eM DCF protocol employing BEB-∞ (BEB with infinite retry limit) was analyzed by Bianchi [13] using a 2-D Markov chain which was later extended by Wu et al. [14] to incorporate retry limit representing BEB-m (BEB with m retry limit) and further extended by Chatzimisios et al. [15] to account for channel error. We used the model in [15] and made appropriate extensions/modifications as required. We assume p(n, P ER) is the probability of a transmission being failed and dC (n, P ER) is the channel access delay as defined in [15] where n is the number of contending stations and P ER is the packet error rate. All three models in [13]–[15]


eM (n, P ER) is given by [15], eM (n, P ER) = pm+1 (n, P ER). (10) Assuming the queue in routing/IP layer to be sufficiently long ensuring zero packet loss, the network loss is a sum of MAC layer loss and the loss in the interface queue which is shown in the next subsection.

0.05

Chatzimisios [15] Our approx., eqn (9) Simulation

MAC layer delay, dM (second)

0.045 0.04 0.035 0.03 0.025

Packet generation interval, n l

0.02

AF

0.015 0.01

Queue starts growing ⇒dQ and eQ increase

0.005 0

⇒Voice quality degrades 0

5

10

15

20

25

30

35

40

45

50

Number of nodes, n

Figure 2: Medium access delay assumed saturated nodes which is feasible only if the packet 1 generation rate, rD = nA1lF ≥ d (n,P ER) where every node C always has some data to transmit. When this is not the case then not all nodes will contend all the time and the number of nodes contending at any point in time depends on CBR starting time distribution. In the worst case, all of the nodes start together. After dC (n, P ER), one node succeeds with probability n1 (since BEB chooses the value of the counter as a uniform random variable). Now, (n − 1) nodes contend and a given node succeeds after dC (n − 1, P ER) with probability 1 n−1 . In this way it goes on until all n nodes succeed. We represent the expected channel access delay as shown in (7). n n n { 1i (1 − 1j )} dC (j, P ER) dC (n, P ER)= i=1 j=i+1 j=i (7) n n 1 = d (j, P ER) C n i=1

j=i

The generalized channel access time, dC (n, P ER) is now given by dC (n, P ER)=dC (n, P ER)S + dC (n, P ER) (1 − S) (8) where dC and dC are used appropriately depending on whether the queue shall grow by use of the step function S such that S(x) = 1 for X ≥ 0 and S(X) = 0 otherwise. Here, S and S (dC (n, P ER) − nA lF ) are used interchangeably. While dC is important to model the queue, it does not represent the delay faced by the packet. When the packet is received by the receiving node, it is sent up to the upper layers and put into dejitter buffer (and may already be played out) but the receiving node will wait for a SIFS (10µs for 802.11b) and send an acknowledgement packet (requiring additional 304µs). The SIFS and acknowledgment delay is not the delay faced by the packet in accessing medium and should therefore be excluded from (8), which gives medium access delay, dM (n, P ER) as shown in (9), dM (n, P ER) = dC (n, P ER) − (tSIF S + tA ) (9) where tSIF S is the length of SIFS and tA is the time to transmit acknowledgement frame. From simulation we show that our modification of dM for unsaturated network holds true as illustrated in Fig. 2 for a 11 Mbps WMN with G.729A codec and nA lF = 20 ms. A packet not acknowledged i.e., failed in transmission is retransmitted until number of failures exceed m + 1 times, where m is the retry limit as defined in 802.11 standard. The packet loss in MAC layer due to retry limit exceed,

D. Queueing delay, dQ and Queueing loss, eQ When the arrival rate to the queue is less than the service rate i.e., rD dC ≤ 1, the queue will not grow and there will be no delay except for enqueueing and dequeueing delay which should be minimal for high speed computing system. Since every newly generated packet will find the queue empty, there is no loss in the queue either. Therefore, dQ = 0 and eQ = 0. When rD dC > 1, the queue starts growing and the system state becomes undefined in the long run and can be defined only if the time is known. In the following we use the notation dC interchangeably with dC (n, P ER) for convenience. After a period of t seconds, there are t(rD − d1C ) packets in the queue. The Head of the queue (HoQ) packet, which is in service now, was generated rtD (rD − d1C ) seconds ago. Taking time starts at 0, its generation time = t − rtD (rD − d1C ) = t(1 − rD dC −1 t rD dC ) = rD dC sec. Then the sequence number (assuming starts from 1) of HoQ packet is given by

t rD dC 1 rD

+1 =

t dC

+ 1.

Replacing i = + 1, we have t = (i − 1)dC . The queueing delay, as experienced by the ith packet, is then given by t rD dC − 1 (rD dC − 1) = (i − 1). (11) diQ = rD dC rD If the expected voice call length is T seconds, a total of rD T voice frames will be generated yielding the average queueing delay, dQ as rD T rD dC −1 dQ = E[diQ ] = rD1T i=1 (i − 1) rD (12) (rD dC −1)(rD T −1) = . 2rD However, if the call is sufficiently large ie, T >> 0 then we should consider dQ for each packet. Since the queueing arrivals are deterministic, dQ will be an increasing function. Therefore, the last packet of the call (or talk spurt) should suffice to be considered for delay calculation. This can be calculated as shown below. The generation time of the last packet is T which gives T = rDtdC and t = T rD dC . Here t represents the time when the last packet becomes HoQ. Its index can be denoted by i = T rD + 1. Using (11), the queueing delay of the last packet of a talk spurt of length T seconds reduces to t dC

diQ = (rD dC − 1) T.

(13)

When rD dC > 1, the loss eQ depends on the queue size and the call length. Let us assume the outgoing queue has length lQ (in number of packets). After a time of t seconds, there will be a total of t(rDddCC −1) packets in the queue. When dtC (rD dC − 1) = lQ then the l d queue is full and the packet, generated at time t = rDQdCC−1 + 1 rD , will be dropped. However, at time t + dC a packet is


serviced and room is made for another in the queue. From C −1 packets in every dC number of newly then, onward, rD drD generated packets will find the queue to be full and will be dropped. If the call runs for a length of T seconds, we calculate the queueing loss, eQ as shown in (14). T (rD dC (n,P ER)−1)−dC (n,P ER)lQ eQ (n, P ER, T ) = rD dC (n,P ER)T (14) lQ dC (n,P ER) ·S T − rD dC (n,P ER)−1 · S(rD dC (n, P ER) − 1) E. VoIP Call Capacity Using (2)∼(4) and subsequent derivations that take network and codec parameters, we estimate VoIP capacity over WMN by formulating an optimization problem as shown in (15). M aximize: n Subject to: Id (d) + Ie ef f (e) − 28.3553 ≤ 0 d =dL + nA lF + dI + dM (n, P ER) (15) +dQ (n, P ER, T )S (dC (n, P ER) − nA lF ) + bg e =pm+1 (n, P ER) +eQ (n, P ER, T )S (dC (n, P ER) − nA lF ) Here we try to maximize the number calls, n i.e., we add calls as long as the sum of the impairments due to end-to-end delay and loss are less than the limit of total impairments as shown in (2). The impairments are calculated from network and queue delay and loss factors. Delay and loss in the queue are conditioned on whether the queue shall grow and is a function of queue length and call length. We can support any call of undefined length if the queue does not grow ie, rD dC (n, P ER) < 1 and the expected delay and loss are maintained within their respective upper limit. But once the queue starts growing, we can keep the quality only as long as the queueing delay, dQ (n, P ER, T ) and queueing loss, eQ (n, P ER, T ) does not overwhelm the system. In this case, a number of calls can be supported but for a length of T seconds which ensures an upper limit for both dQ and eQ . V. A NALYTICAL AND S IMULATION R ESULTS We carried out extensive simulation in NS-2 v2.32. The different modules in NS-2 have been extended to better simulate the VoIP system and a separate tracing mechanism is developed to retrieve information on delay and loss at each protocol layer. A typical 802.11 WMN is simulated for different data rates with the UDP, RTP, IP and MAC headers of length 8, 12, 20 and 28 bytes respectively. Long preamble is assumed giving a total of 192 bits of PLCP and preamble. SIFS and idle slots are of length 10µs and 20µs, respectively. The optimization problem (15) has computational complexity of O(n) and is solved numerically. Theoretical and simulation results for call capacity of two way calls over a single hop WMN are plotted in Fig. 3(a)∼3(c) for G.729A, G.711 and G.723.1 codec respectively at 11Mbps data rate and in Fig. 4(a)∼4(c) at 54Mbps. For comparison purpose, the figures also present call capacity estimated in [4], [5]. In all cases our theoretical estimation shows a close match with simulation results validating our approach. The results show the overestimation in [4], [5] which is due to ignoring the

queueing delay, queueing loss and dejitter buffer delay. Moreover, capacity in those models increases monotonically due to ignoring the effect of packetization delay, nA lF . In our study we find that when packet generation rate is high, the capacity is low. This is because then, with even a smaller number of nodes, rD dC becomes close to 1 and as soon as rD dC becomes greater than 1 the queue starts growing (explained in Section IV-D). As a result queueing delay and queueing loss becomes very high resulting in drastic degradation of voice quality. With increase in nA , rD decreases (rD ∝ n1A ) and call capacity increases but only up to a certain point and then decreases again. This is because the gain achieved by keeping rD low is overwhelmed by the increase in nA lF which contributes to d and Id , in turn. In the 11 Mbps case, G.729A supports more calls than G.711 (45 vs. 34) since G.711 generates larger packets and consumes greater bandwidth. Although G.723.1 uses more compression than G.729A, it supports less number of calls (40) due to its high value of initial equipment impairment at zero loss (Ie = 15). In the 54 Mbps case also, a similar trend is followed by each codec, except the peak is now reached for a lower value of nA lF indicating increase in its effect. In this case, the maximum number of medium quality two way calls are 49, 58 and 44 for G.729A, G.711 and G.723.1, respectively. Comparing results for 11Mbps and 54 Mbps we find that although data rate is increased almost 5 times, the increase is capacity is minimal. In case of G.729A only 4 additional calls can be supported. This is due the fact that VoIP packets are small and hence the gain by increasing data rate is minimal. Only for G.711, which uses less compression and generates larger packets, the gain in capacity is relatively higher i.e., 24 additional calls are now supported. Although G.729A performs better than G.711 in 11 Mbps case, it performs worse in 54 Mbps due to high value of Ie which is 11 and 0 for G.729A and G.711, respectively (Table II). Performance of VoIP employing RTS/CTS mechanism is presented in Fig. 5(a) which shows theoretical and analytical results over an 11 Mbps RTS/CTS WMN. Although RTS/CTS mechanism minimizes the effect of hidden node and exposed node problem, the overhead of RTS/CTS packets does not pay off for smaller VoIP packets. It is efficient only for larger packets. Therefore, RTS/CTS is found to provide less number of calls for all codecs. Only G.711 shows less degradation in capacity as it generates larger packets keeping the impact of overhead relatively low. We also investigated the impact of call quality on VoIP capacity. Comparison of high quality calls versus medium quality calls for different codecs are shown in Fig. 5(b). A moderate number of high quality calls is supported by G.711 where G.729A and G.723.1 supports close to zero high quality calls. This is because the impairment budget for high quality calls is 13.1952 as shown in (2) but the initial equipment impairment for G.729A and G.723.1 are very high even in the case of 0 loss. The effect of queueing loss for different queue lengths and packetization intervals is shown in 5(c) for a call length of 50s. It presents the threshold queue length, lQ (the abscissa) to ensure eQ = 0 for different nA lF .


70

Analytical model Simulation Garg et. al. [4] Hole et. al. [5]

100

60

Number of Two Way Calls, n


120

80

60

40

40

30

20

10

0

0

40

60

80

100

120

140

160

180

200


100

50

20

20

120 Analytical model Simulation Garg et. al. [4] Hole et. al. [5]


140

80

60

40

20

50

100

150

nAlF (ms)

200

250

0

300

40

60

80

100

nAlF (ms)

(a) G.729A

120

140

160

180

nAlF (ms)

(b) G.711

(c) G.723.1

Figure 3: Simulation and analytical results over 11 Mbps WMN with basic DCF using a) G.729A, b) G.711 c) G.723.1 140

150

100

120

140



120


200

160




250

100

80

60

40

100

80

60

40

50 20

20

0

50

100

150

200

250

0

300

50

100

150

nAlF (ms)

200

250

0

300

40

60

80

100

nAlF (ms)

(a) G.729A

120

140

160

180

nAlF (ms)

(b) G.711

(c) G.723.1

Figure 4: Simulation and analytical results over 54 Mbps WMN with basic DCF using a) G.729A, b) G711, c) G.723.1 45

1

G.729A, Medium G.729A, High G.711, Medium G.711, High G.723.1, Medium G.723.1, High

40



30

25

20

15

10

5

0

G.729A, Analytical model G.729A, Simulation G.711, Analytical model G.711, Simulation G.723.1, Analytical model G.723.1, Simulation 50

100

150

35 30 25 20 15

250

300

nAlF=30 ms

0.7

n l =40 ms

0.6

nAlF=50 ms

AF

nAlF=60 ms 0.5 0.4

0.2

0

0.1

50

100

nAlF (ms)

150

200

250

300

nAlF (ms)

(a) RTS/CTS based DCF

nAlF=20 ms

0.8

0.3 10 5

200

nAlF=10 ms

0.9

Queuing loss, eQ

35

(b) Effect of call quality

0

100

200

300

400

500

600

700

800

900

1000

Length of queue, lQ

(c) Queueing loss

Figure 5: Simulation and analytical results over 11 Mbps WMN a) RTS/CTS DCF b) Call quality c) Queueing loss VI. C ONCLUSION In this paper, we developed an analytical model to estimate VoIP call capacity over IEEE 802.11 based single hop WMN. The model ensures the call quality by employing R-score defined in ITU-T E-model. It analyzes various delay components and their impact on voice quality and, in turn, call capacity. In particular, queueing delay is found to have significant impact on capacity. The model is extendable to explore requirements for multihop WMN which must be met to maintain call quality throughout the lifetime of a call. The model is validated by extensive simulation using the widely used NS-2 simulator. This model will be highly useful in designing the WMN infrastructure network. R EFERENCES [1] O. Awoniyi and F. Tobagi, “Effect of fading on the performance of VOIP in IEEE 802.11 WLANs,” IEEE ICC, vol. 6, pp. 3712–3717, jun 2004. [2] F. Anjum, M. Elaoud, D. Famolari, A. Ghosh, R. Vaidyanathan, A. Dutta, P. Agrawal, T. Kodama, and Y. Katsube, “Voice performance in WLAN networks - an experimental study,” in IEEE GLOBECOM, vol. 6, 2003, pp. 3504–3508. [3] A. Kashyap, S. Ganguly, S. R. Das, and S. Banerjee, “Voip on wireless meshes: Models, algorithms and evaluation,” in IEEE INFOCOM, 2007, pp. 2036–2044. [4] S. Garg and M. Kappes, “Can i add a VoIP call?” in IEEE ICC, vol. 2, 2003, pp. 779–783.

[5] D. Hole and F. Tobagi, “Capacity of an IEEE 802.11b wireless LAN supporting VoIP,” in IEEE ICC, vol. 1, 2004, pp. 196–201. [6] S. Guha, N. Daswani, and R. Jain, “An Experimental Study of the Skype Peer-to-Peer VoIP System,” in IPTPS, feb 2006, pp. 1–6. [7] ITU-T, “Methods for subjective determination of tranmission quality,” ITU-T Recommendation P.800, aug 1996. [8] ——, “The e-model, a computational model for use in transmission planning,” ITU-T Recommendation G.107, mar 2005. [9] ——, “Perceptual evaluation of speech quality (PESQ), an objectie method for end to end speech quality assessment of narrow-band telephone networks and codecs,” ITU-T Recommendation P.862, 2001. [10] R. Cole and J. Rosenbluth, “Voice over IP performance monitoring,” ACM SIGCOMM Computer Communication Review, vol. 31, no. 2, pp. 9–24, apr 2001. [11] A. Meddahi, H. Afifl, and D. Zeghlache, ““Packet-e-model”: e-model for wireless VoIP quality evaluation,” in IEEE PIMRC, vol. 3, 2003, pp. 2421–2425. [12] ITU-T, “Transmission impairments - appendix i: Provisional planning values for the equipment impairment factor ie,” ITU-T Recommendation G.113 - Appendix I (2001), oct 2001. [13] G. Bianchi, “Performance analysis of the IEEE 802.11 distributed coordination function,” IEEE Journal on Selected Areas in Communications, vol. 18, no. 3, pp. 535–547, 2000. [14] H. Wu, Y. Peng, K. Long, S. Cheng, and J. Ma, “Performance of reliable transport protocol over IEEE 802.11 wireless LAN: analysis and enhancement,” IEEE INFOCOM, vol. 2, pp. 599–607, 2002. [15] P. Chatzimisios, A. Boucouvalas, and V. Vitsas, “Performance analysis of IEEE 802.11 DCF in presence of transmission errors,” IEEE ICC, vol. 7, pp. 3854–3858, jun 2004.
