Can I add a VoIP call?

Sachin Garg Martin Kappes

Avaya Labs Avaya Labs
Basking Ridge, NJ 07920 Basking Ridge, NJ 07920
Email: sgarg@avaya.com Email: mkappes@avaya.com

Abstract— In this paper, we study the inherent limitations A naive calculation might indicate that given the 11Mbps
of the 802.11 (a/b) distributed coordination function (DCF) in peak rate and 128Kbps needed for a full duplex VoIP call,
supporting VoIP calls over a wireless LAN. Specifically, we approximately 85 simultaneous calls can supported. The real
evaluate the upper bound on the number of simultaneous VoIP
calls that can be placed in a single cell of an 802.11(a/b) network. motivation for the analysis comes from a simple experiment,
Making one additional VoIP call in that cell would degrade the goal of which was to determine that number. We briefly
the quality of all VoIP calls. The upper bound is calculated describe the experiment, setup and its surprising outcome
as a function of the choice of VoIP codec and the length of before moving on to the analysis.
the audio payload. As an example, when a G711 codec with 20 In the experiment, multiple Wireless PCs running Windows
millisecond audio payload is used, an 802.11b cell can support
only 3 to 12 simultaneous VoIP calls. The actual number depends 2000, were associated with the same 802.11b AP, which was
on the effective transmission rate of the wireless station, which connected to a 100Mbps Ethernet. The setup was used to make
for 802.11b can be 1Mbps, 2Mbps, 5.5Mbps and 11Mbps. We also full-duplex VoIP calls between a wireless PC and a wired PC
study the effect of spatial distribution of the wireless stations on using IP phones. For each call, we used the ITU G711 a-
the upper bound which is the dominant factor in determining Law codec where frames are sent out every 10 milliseconds.
the effective transmission rate of a station.
Each call results in two RTP streams, from wired to wireless
I. I NTRODUCTION and vice-versa. We tested the number of VoIP connection with
acceptable voice quality by successively establishing new calls
The IEEE 802.11 wireless LANs [12] have been commodi- in addition to the ongoing calls. The quality of the connections
tized for data applications in enterprise networks. Meanwhile, was monitored through measurements of loss, jitter and round-
VoIP deployment in an enterprise mostly involves VoIP termi- trip time by a commercially available tool.
nals connected to its wireline IP network. However, as VoIP For the first five calls, the quality of all the calls was fine.
gains more traction, wireless VoIP terminals will also begin Loss (0%), round-trip time (around 5 ms) and jitter (around
to be deployed to leverage the pervasive 802.11 networks for 7 ms) were all in acceptable ranges for a good quality VoIP
caller mobility. call. When the sixth call was placed, except for a sporadic
Ensuring the Quality-of-Service for VoIP in wireless LANs increase in the round-trip time for some of the connections
is a big concern as the performance characteristics of their the quality of all six simultaneous connections was still
Physical and MAC layers is much worse than their wireline acceptable. As soon as the seventh call was placed, all seven
counterparts. In particular, lower peak transmission rate, lossy “wired to wireless” streams started suffering approximately
medium, interference problems etc. are some drawbacks of the 16% loss and the call quality became unacceptable for all
PHY layer. At the MAC layer, 802.11 offers two choices. The calls in this direction. All “wireless to wired” streams still
first, called Distributed Coordination Function (DCF) belongs exhibited acceptable call quality. In short, the outcome of the
to the CSMA/CA family of protocols, where each station experiement indicated that given the codec setting, only six
determines the access to the channel on its own without calls can be placed on a single 802.11b AP.
the involvement of any central coordinator. The second is The rest of the paper provides an explanation of the ex-
called the Point Coordination Function (PCF) in which the perimental observation and is organized as follows. Section II
access point (henceforth referred to as AP) determines which consists of a brief background on DCF for the sake of self
connected station gets to transmit at any time. While the PCF containment and also gives a short summary of the related
was designed to support real-time services, its implementation work. In Section III, we develop the simple analytical model
in an access-point and/or client cards is optional. As a result, to determine the upper bound on the number of simultaneous
the majority of 802.11(b) networks currently deployed do not VoIP connections over DCF and in Section III-A, the model
support PCF. is used to tabulate and explain the results for an 802.11b AP.
As wireless VoIP takes hold, especially in public hot-spots In Section III-B, we evaluate the effect of spatial distribution
such as airports, hotels etc., the question of how suitable DCF of the stations within the cell on the upper bound. The base
is for supporting VoIP traffic becomes important. That is the model also qualifies to evaluate the upper bound for 802.11a
focus of this paper. In particular, we analyze the number of [13] as only some parameter values change in 802.11a while
simultaneous VoIP calls a single AP running DCF can support. the underlying DCF remains unchanged. These results are

0-7803-7802-4/03/$17.00 (C) 2003 IEEE

tabulated and explained in Section III-C. The paper concludes random backoff determines the number of time slots the client
in Section IV. defers its transmission in addition to the DIFS time.
If the medium is sensed idle in such a “slot”, the backoff
II. R ELATED W ORK AND BACKGROUND timer is decreased by one. If the random backoff has decreased
to 0, the node starts transmitting. If another node starts trans-
With respect to 802.11 networks, prior performance studies, mitting before this happens, the node continues to count down
such as [3], have mainly focussed on analyzing the behavior the backoff timer after the medium has been sensed idle for the
of the MAC protocol itself. Stochastic Petri Nets are used DIFS period. Thus, if multiple clients want to transmit a frame,
in [7] to model the behavior of DCF and then performance the one with the lowest random backoff time will win the
measures such as effective channel throughput are derived. contention for the medium. However, if more than one node
Bianchi, in [2], used a Markov process to model DCF and happens to choose the same backoff time, they will start to
evaluated the channel throughput, and frame loss as a function transmit at the same time and a collision will occur. The clients
of the number of wireless stations. In all of the above, the assume that the frame was lost if an acknowledgment is not
objective has been to study the performance of DCF itself received within SIFS. In case of an unsuccessful transmission,
while being agnostic to any protocols running over the DCF the CW value doubles until a CWmax value is reached. The
MAC. Hence the assumptions made on the traffic load do not CW parameter is reset to the CWmin after each successful
correspond to the load generated specifically by VoIP traffic. (i.e. acknowledged) transmission.
Prior work that pertains to studying VoIP over 802.11 has The IEEE 802.11/802.11b standard defines SIFS to be 10
focussed mainly on PCF as the protocol of choice, for instance µs. A slot time is 20 µs and the value of DIFS is 50 µs.
in [4]. Veeraraghavan et. al., in [10], present an analysis of the The size of an acknowledgment frame is 14 bytes which take
delay and loss characteristics for voice traffic over PCF as a about 10 µs to transmit at 11 Mbps. However, each transmitted
function of the inter-poll period. A simulation approach was frame also needs some physical layer overhead (PLCP header
used in [11] to study and analyze variations in the polling of 48 µs and a preamble of 144 µs) which is about 192 µs.
schemes for PCF. Thus, the total time to transmit an acknowledgment is 203 µs.
Enhancements to the MAC protocols that are neither DCF CWmin is defined to be 31.
nor PCF have also been proposed [9] so that real-time QoS
guarantees can be provided to the VoIP calls. No one, to III. A NALYSIS FOR N UMBER OF VO IP CONNECTIONS
the best of our knowledge, has quantified the the maximum
number of simultaneous VoIP calls that can be supported on With the basic understanding of DCF, we move on to
a single AP running the basic, standards compliant DCF. The calculate the maximum number of VoIP clients a single AP
upper bound is in the sense that adding one more VoIP call will can support. The analysis is based on the assumption that
disrupt the quality of all ongoing calls. Further, as the purpose one end-point of each VoIP call is a wireless client, while
in this paper is to study the inherent limitations of DCF vis-a- the other end-point is on the wired network. This number
vis VoIP, it is assumed that no data traffic is present. Presence depends on the maximum throughput the channel can achieve,
of any data traffic will only reduce the number of simultaneous which is a function of the packet size. Other factors effecting
calls. the channel throughput include the byte overheads of RTP,
Before the analysis itself, we give a brief introduction to UDP, IP, MAC and physical layers. Further, the channel access
DCF as follows. As collisions in the wireless medium cannot mechanism (CSMA/CA) imposes an overhead due to the back-
be detected, the MAC protocol is designed to prevent collisions off procedures between successive packet transmissions from
from occurring and required sensing of the wireless medium. the same station.
All unicast frames are acknowledged by the receiving station Assume the following terminology. Let P be the size of
within a certain duration of receiving the frame. This duration voice payload. For G711 a-Law codec, this payload is 80 bytes
is called the short inter frame space (SIFS). A node may for 10 ms of audio. Let Ravg be the average data transmission
transmit a frame if it senses the medium idle for a certain rate of the access point. Note that 802.11b allows for multiple
duration of time called the DCF inter frame space (DIFS). data rates and most implementations support data rates of
Since DIFS is longer than SIFS, a correctly received frame is 11Mbps, 5.5Mbps, 2Mbps and 1Mbps. The actual data rate
always acknowledged before the channel is used again. of a client (or an access point sending to this client) depends
If a node wants to start transmitting while the medium is on the signal-strength of the client from the access point as
busy or if it wants to transmit another frame after just finishing well as other factors such as interference. The signal-strength
a transmission, it also waits for the medium to be idle for in turn depends on the distance between the client and the
the DIFS period. Then, the node does not begin to transmit access point. In general, depending on the special distribution
immediately but enters a contention phase for the medium. of clients Ravg varies from 1 Mbps to 11 Mbps.
Contention is done by choosing an integer random backoff Table I shows a refined view of the overhead per packet in
between 0 and a parameter CW (CW stands for contention bytes or microseconds also taking the average data rate into
window size) which is initialy set to a value CWmin. The account. Let TP be the time taken to transmit the VoIP payload
probability distribution among these values is uniform. The of P bytes. Further let Toverhead be the average overhead per

0-7803-7802-4/03/$17.00 (C) 2003 IEEE

 fixed overhead variable

SIFS DIFS back off

OTHER FRAME ACK MAC/IP/UDP HDR UDP payload

time
Fig. 1. IEEE 802.11 CSMA/CA medium access scheme.

Overhead Bytes Time (µs)

RTP Layer 12 12 ∗ 8/Ravg these clients. Assuming G711 a-Law codecs, the access point
UDP Layer 8 8 ∗ 8/Ravg handles upstream traffic of n × 64 kbps and sends n × 64Kbps
IP 20 20 ∗ 8/Ravg downstream traffic. The total channel throughput is therefore
MAC 34 34 ∗ 8/Ravg
Physical 24 192µs n × 128Kbps. The maximum channel efficiency is given by
SIFS 10µs TP
DIFS 50µs .
ACK 14 14 ∗ 8/Ravg TP + Tlayers + TSIF S+DIF S+ACK + Tdcf
The maximum number of VoIP connections, nmax , all using
TABLE I
G711 codecs is therefore given by
OVERHEAD PER SENT FRAME TAKING THE AVERAGE DATA RATE INTO
ACCOUNT.

TP ∗ RAvg
RTP packet. The channel throughput is simply given by .
128000 × (TP + Tlayers + TSIF S+DIF S+ACK + Tdcf )
TP
× Ravg . In our VoIP setup, each of the n clients has data to transmit
TP + Toverhead
very infrequently (every 10 ms for 80 byte payload), while the
While TP is independent of the number of clients transmitting, access point sends n times the data throughput downstream to
Toverhead varies with the number of clients. Specifically, the clients. Under this scenario, we argue that Tdcf can be
Toverhead can be divided into two components. First, the approximated, even for large n, by the value of Tdcf , which
overhead incurred in transmitting the extra bytes of various results when an access point and a single other client always
networking layers and second, the overhead imposed by have a packet to transmit. In other words, from the perspective
the Distributed Coordinated Function (DCF) of 802.11. Let of the channel, the value of Tdcf is approximated by the
Tlayers and Tchannel denote these overheads respectively. value when there are exactly two active senders. The argument
Moreover, Tchannel comprises one SIFS interval, one DIFS proceeds as follows: Assume that each of the n clients is using
interval, time to send an ACK and the average idle slots a 10 ms audio payload per packet. This implies an 80 byte RTP
(Tdcf ) per frame as seen on the channel. It is apparent that payload. The time to transmit this payload is 58 µs. In other
only Tdcf depends on the number of clients in the network. words, with Ravg = 11M bps, TP = 58. The total overhead,
Moreover, as previous analysis studies have evaluated [1], [2], not counting Tdcf is Tlayers + TSIF S+DIF S+ACK = 500µs.
[7], [3], the dependence is non-linear. As the number of clients Assuming that the clients are not synchronized, the probability
increases, the average idle slots per frame as seen on the that a client, which is sending packets only every 10 ms,
channel decrease. For example, for a single client, assuming attempts to transmit a frame within 558µs of an ongoing
that the source always has a frame to send, there are 15.5 transmission by another client is very low. In previous work
idle slots (i.e. 310µs) (Mean of uniform distribution U [0, 31]) on throughput analysis of 802.11 networks, Tdcf has been
per frame. If there are n clients, then the average number of evaluated both via analytical approximations as well as via
idle slots are determined by two factors with opposing effects. simulations. From these as well as from our own simulation,
First, after a successful frame transmission, the second frame is for the two sender scenario, the average back-off window size
put on the channel by the client which calculated the minimum is determined to be 8.5 SLOTs and the collision probability
backoff among the n clients. However, it is also possible that is 0.03. Therefore, Tdcf for this case is given by Tdcf =
more than one client calculated the same backoff resulting in a 8.5 × 20 + TW × 0.03 µs., where, TW = TP + Tlayers +
collision of the next frame or some subsequent frame, reducing TSIF S+DIF S+ACK .
the channel throughput. The collision probability increases
with increase in n, counteracting the first effect. Further, upon A. Maximum number of VoIP connections and effect of codec
collision, the contention window size is doubled, which leads selection
to an increase in the average idle slots per frame. In effect, We use the above equations to compute the maximum
Tdcf is a concave function of n. If n = 1, however, there are number of VoIP connections a single 802.11b access point can
no collisions and Tdcf = 310µs. support. These calculations are done for three standard codecs.
For VoIP connections, the access point receives frames from ITU’s G711 a-Law, G729 and G723.1. G711 allows 10 ms of
each of the n clients and sends VoIP packets to each of audio data in 80 bytes of payload, without any compression. A

0-7803-7802-4/03/$17.00 (C) 2003 IEEE

single G711 VoIP stream therefore constitutes 64Kbps. G729 traffic are very much in the acceptable ranges even as this
uses the same sampling rate of audio stream, but compresses maximum number is crossed. In other words, the dominant
the digitized data to pack 10 ms of audio in 10 bytes reducing factor that makes DCF unsuitable for VoIP is not that delay and
the bandwidth of a single stream to 8Kbps. G723.1 further jitter values are high, but that the effective available bandwidth
reduces the bandwidth requirement by compressing 30 ms of is too low.
audio data into 24 bytes, which amounts to 6.4 Kbps per VoIP Another interesting fact to note is that for VoIP with small
stream. packet sizes, and because of the DCF behavior, delay and jitter
Table II tabulates the maximum number of VoIP connections values for these packets do not become unacceptable.
for the three codecs. In these calculations, Ravg = 11 Mbps,
which is the maximum possible transmission rate. B. Effect of spatial distribution of clients in 802.11b
802.11b allows support for multiple data-rates and most
Audio (ms) G711 G729 G723
10 6 7
commercial implementations support data transmission rates
20 12 14
of 11Mbps, 5.5Mbps, 2Mbps and 1Mbps. The actual transmis-
30 17 21 21 sion rate of a client depends on the strength of the radio signal
40 21 28 to/from the access point and other factors such as interference.
50 25 34 For lower observed signal strengths, a client will transmit at a
60 28 41 42
70 31 47 lower transmission rate to reduce the bit error rate. Since signal
80 34 54 strength depends on the distance, Ravg depends on the spatial
90 36 60 61 distribution of the clients. As the actual function is vendor
100 39 66
dependent and Ravg is influenced by other factors such as
TABLE II interference, we study the effect on nmax for the whole range
M AXIMUM NUMBER OF VO IP CONNECTIONS FOR DIFFERENT CODECS of Ravg .
Figure 2 shows nmax plotted against Ravg , where Ravg
is varied between 1Mbps and 11Mbps for G711 and G729
The first column shows the audio payload sent in every RTP codecs. The scale of the vertical axis in both is same for
packet. For G.711 and G.729 codecs it is varied in increments comparison purposes. The thick lines in each represent the
of 10 milliseconds whereas for G.723 codec, it is varied in typical payload implementation in commercial IP phones. The
increments of 30 ms. The most important observation is the first obvious observation is the significant reduction in nmax
very low number of VoIP calls for lower payloads per RTP for all payload sizes for lower values of Ravg . For instance,
packet. For a payload of 10 ms audio, only 7 connections with a payload of 30 ms, nmax reduces from 17 at 11Mbps
can be supported by a single access point. It is also worth to only 4 at 1Mbps. This implies that the physical location of
noting that most commercial implementations of IP phones an access point in enterprises and public hot-spots is crucial,
use a payload size of either 20 or 30 ms audio in each RTP at least as far as supporting VoIP connections are concerned.
packet. With G711 and 20 ms audio payload, the maximal
number of VoIP connections is 12, and for 20 ms audio it is 17. C. Maximum VoIP connections in 802.11a
In other words, the choice of payload size in IP phones results
The maximum data-rate in 802.11a is 54Mbps, approxi-
in very inefficient use of the available bandwidth in 802.11b.
mately 5 times that of 802.11b. The value of the overhead
An obvious solution as shown by the numbers is to use larger
Tlayers of transmitting a total of 74 bytes for MAC, IP,
payload per RTP packet. However, VoIP calls traverse both
UDP and RTP headers is therefore reduced proportionally.
wireless and wired networks and the larger the payload, the
The physical layer overhead (PLCP header and preamble)
worse are the delay, loss and jitter characteristics adversely
is 24µs compared to 192µs for 802.11b. Further SIF S =
affecting the VoIP call quality. Another fact worth observing
16µs and each SLOT = 9µs. Tdcf can be approximated by
is that for 20 and 30 ms payloads, choosing G729 or G723
using exactly the same arguments as for the 802.11b case.
codecs increases the maximum number of calls supported only
In 802.11a, each client chooses a random backoff from a
by a small number. For instance, use of G729 over G711 with
uniform distribution U [0, 15] as opposed to the upper limit
20 ms payload only increases nmax by 2. This brings out a key
of 31 for 802.11b for successful transmissions. The average
limitation. While in wired networks (such as Ethernet), choice
backoff when two 802.11 clients always contend for the
of codecs is highly effective in dealing with network load, in
medium is shown via simulations to be 41µs (4.5 SLOTs).
802.11 networks, this is not the case. Comparing G729 and
The collision probability however is double that in 802.11b to
G723, even for higher payload sizes, nmax is almost equal.
0.06. Therefore,
Therefore G729 should always be the preferred codec as it
uses less compression. Tdcf = 4.5 × 9 + TW × 0.06 µs.
As seen above, the total available payload bandwidth at
small packet sizes is exhausted even for a small number VoIP Using the maximum value of Ravg = 54 Mbps, Table III
of calls. An interesting observation, which is explored in detail tabulates the maximum number of VoIP calls that a single
in [6], is that the delay and jitter characteristics of the VoIP 802.11a access point can support.

0-7803-7802-4/03/$17.00 (C) 2003 IEEE

 ITU G711aLaw ITU G729.1
50 50
Audio = 10 ms
Audio = 20 ms (typical)
Audio = 30 ms (typical)
40 40

Number of VoIP Connections

Number of VoIP Connections

Audio = 40 ms
Audio = 50 ms
Audio = 60 ms
Audio = 70 ms
30 30

20 20

10 10

0 0
1.0 3.0 5.0 7.0 9.0 11.0 1.0 3.0 5.0 7.0 9.0 11.0
Average Data Transmission Rate (Mbps) Average Data Transmission Rate (Mbps)

Fig. 2. Effect of spatial distribution/Load on the number of VoIP connections

Audio (ms) G711 G729 G723

10 30 32
Another conclusion drawn from the study was that the effect
20 56 64 of codec selection does not help much in 802.11 networks,
30 79 95 96 which is in contrast with wired Ethernet.
40 98 126
50 116 156 R EFERENCES
60 131 185 187 [1] G. Anastasi and L. Lenzini. QoS provided by the IEEE 802.11
70 145 214 wireless LAN to advanced data applications: a simulation study, In
80 158 243 Wireless Networks, Vol. 6, pp 99–108, J.C. Baltzer AG, Science
90 169 270 275 Publishers, 2000.
100 180 298 [2] G. Bianchi. Performance Evaluation of the IEEE 802.11 Distributed
TABLE III Coordination Function. In IEEE Journal on Selected Areas in
Communication, Vol. 18, No. 3, March 2000, pp. 535–547.
M AXIMUM NUMBER OF VO IP CONNECTIONS FOR DIFFERENT CODECS
[3] H. S. Chayya and S. Gupta. Performance modeling of asynchronous
WITH 802.11 A (54 M BPS ) data transfer methods of IEEE 802.11 MAC protocol. In Wireless
Networks Vol 3, 1997, pp. 217-234.
As is apparent by comparing these numbers from those [4] b. P. Crow, I. Widjaja, J. G. Kim and P. Sakai, “Investigation of
the IEEE 802.11 medium access control (MAC) sublayer functions,”
of Figure II, nmax is approximately 5 times higher. The In Proceedings of INFOCOM’97, volume 34, pages 126-133, April
reason is simply the proportional increase in bandwidth and 1997.
approximately the same channel efficiency. [5] P. Fergusan and G. Huston Quality of Service, Wiley Computer
Publishing, John Wiley and Sons, ISBN 0-471-24358-2, 1998.
IV. C ONCLUSIONS [6] S. Garg, M. Kappes, “An experimental study of throughput for
UDP and VoIP traffic in IEEE 802.11b Networks”, Submitted for
In this paper, we studied the behavior of VoIP over 802.11 publication
[7] A. Heindl and R. German. Performance modeling of IEEE 802.11
networks, from the perspective of number of connections that wireless LANs with stochastic Petri nets. Performance Evaluation,
a single base-station can support. An important conclusion is 44 (2001), 139-164.
that given the choice of payload sizes in IP phones, 802.11b [8] J. Schiller. Mobile Communications, Addison Wesley, ISBN 0-201-
39836-2, 2000.
base-stations prove to be inadequate in handling large number [9] J. L. Sobrinho and A.S. Krishnakumar “Real-Time traffic over the
of VoIP calls. In fact, the inherent channel inefficiency of IEEE 802.11 MAC protocol”, In Bell Labs Technical Journal, Autumn
802.11b, at smaller frames sizes, limits the maximum number 1996.
[10] M. Veeraraghavan, N. Cocker and T. Moors “Support of voice services
of VoIP calls to a very low number. The actual number of VoIP in IEEE 802.11 wireless LANs”, In Proceedings of INFOCOM’01,
calls are further reduced by factors such as spatial distribution 2001.
of the clients. [11] J.-Y. Yeh and C. Chen “Support of multimedia services with the IEEE
From the study, the use of larger payload per frame becomes 802.11 MAC protocol.” In Proceedings of ICC’02, New York, April
2002.
apparent as the solution to increasing the number of VoIP [12] ISO/IEC and IEEE Draft International Standards, “Part 11: Wireless
calls. However, this needs to be balanced against the adverse LAN Medium Access Control (MAC) and Physical Layer (PHY)
effect of larger payloads on the multi-hop wired network in Specifications,” ISO/IEC 8802-11, IEEE P801.11/D10, January 1999.
[13] http://grouper.ieee.org/groups/802/11/
terms of delay, jitter and loss. This naturally lends to an
optimization problem which is not addressed in this paper.

0-7803-7802-4/03/$17.00 (C) 2003 IEEE