WIRELESS COMMUNICATIONS AND MOBILE COMPUTING

Wirel. Commun. Mob. Comput. 2007; 7:795–808

Published online 7 July 2006 in Wiley InterScience (www.interscience.wiley.com). DOI: 10.1002/wcm.421

A collision detection method for multicast transmissions in

CSMA/CA networks

Thomas Nilsson*,† , Greger Wikstrand and Jerry Eriksson

Department of Computing Science, Umeå University, SE-901 87 Umeå, Sweden

Summary

Compared to unicast traffic, multicast is not protected by any ARQ mechanism in 802.11 networks: collisions with
other multicast and unicast transmissions are not detected and senders will not adapt to the contention situation by
backing off. This results in an unreliable service for multicast transmissions.
We propose early multicast collision detection (EMCD), an algorithm with the purpose of increasing the reliability
of multicast transmissions in the MAC layer of an IEEE 802.11 network. A multicast sender using it will introduce
an early pause in a transmission, perform a clear channel assessment (CCA), and if a collision is detected abort
the transmission after a fixed time and schedule a retransmission. This allows for detecting collisions with both
multicast and unicast transmissions but also adapting to the contention situation.
A probabilistic analysis is provided showing that EMCD is more efficient than ordinary multicast and can be
made even more efficient by tuning parameters.
Simulations show that EMCD leads to increased reliability for multicast transmissions. Copyright © 2006
John Wiley & Sons, Ltd.

KEY WORDS: quality of service; CSMA/CA; medium access control; 802.11; collision detection; reliable multi-
cast; performance evaluation

1. Introduction out ARQ, the back-off mechanism cannot be applied

after a collision. Link adaptation is based on frame er-
It is difficult to provide the same kind of reliable service ror rate, which is not available for the same reason.
for multicast traffic in an IEEE 802.11 network as for Thus, such packets are more likely to be lost.
unicast traffic [1, p. 29]. In an IEEE 802.11 network, One further reason for multicast packet loss is colli-
unicast traffic is protected in three ways: (1) by an auto- sions between access points (AP) in overlapping cells
matic repeat request protocol (ARQ), (2) by backing- on the same channel and in the same extended service
off more in case of congestion, and (3) by adapting set (ESS). Any multicast stream must be transmitted by
the link. These, or other reliability-enhancing mech- each access point in an ESS [1, p. 83] leading to a risk
anisms, are not available for multicast traffic. ARQ for synchronized and therefore colliding traffic. Also,
is problematic because of the ACK-implosion prob- stations are explicitly forbidden to eavesdrop on access
lem, that is multiple stations will send an ACK. With- points with which they are not associated [1, p. 44] thus

*Correspondence to: Thomas Nilsson, Department of Computing Science, Umeå University, SE-901 87 Umeå, Sweden.
†
E-mail: nilsson@cs.umu.se
Copyright © 2006 John Wiley & Sons, Ltd.
796 T. NILSSON, G. WIKSTRAND AND J. ERIKSSON

making it less efficient to use channel-reservation tech- An algorithm, which would enable collision detec-
niques. A configuration like that can be conceived, for tion for multicast frames, was introduced in 1984 by
example, in an arena or an exhibition hall with free Lo and Mouftah [3,4] under the name time split col-
space propagation where a large numbers of access lision detection or CSMA/TCD and again in 1986 by
points have been distributed to maximize coverage and Rom [5]. We call it the ‘the Rom Algorithm’. It is the
association capacity. purpose of this paper to reevaluate their algorithm(s)
This work was done in the context of the Arena in terms of their suitability for solving the problems
project (http://www.cdt.luth.se/projects/arena/). At a outlined above. We investigate the efficiency of the
sport event, spectators were provided with an interac- modified algorithm and the possibility of using it in
tive match program in wireless terminals. Among the an IEEE 802.11 network.
features of the application were the ability to playback The rest of this paper is organized as follows. In
live video and personalized replays in the phones. The Section 2 some other collision detection and avoidance
efficient way to reach the crowds in such a context techniques are described and categorized. In Section 3
would be to use multicast for the multimedia informa- our modified Rom Algorithm, called early multicast
tion. In the case of multicast multimedia this will re- collision detection (EMCD), is described along with
quire minimizing packet losses to provide smooth play- the modifications required for applying it to an IEEE
back and minimizing or bounding delays so that play- 802.11a network. We argue for relevant parameters to
back does not lag too far behind the live event. In par- use and describe various collision scenarios. EMCD
ticular, the system must be robust against unicast inter- was evaluated through probabilistic analysis and sim-
ference as there will be no way to restrict such traffic on ulations. The analysis is presented in Section 4 and the
an unlicensed frequency band. We expect spectators to simulation results are presented in Section 5. Section 6
bring their own equipment from home, so any solution contains our conclusions.
based on changes to their equipment is hardly feasible.
For clarity, we provide the following definition of
2. Related Work
medium access control (MAC)-layer multicast. It is
based on the definition given by Sun et al. [2] but
Much work has been done in the area of channel reser-
adds the requirement of simultaneity. Thus, we do not
vation and collision resolution in MAC protocols for
consider the process in which a station transmits the
unicast traffic, for example, References [6–8]. These
same multicast packet to several stations by sending it
solutions are not directly applicable to multicast traf-
repeatedly, once for each recipient as multicast.
fic where multiple receivers are addressed and normal
Definition 1. Multicast is the process in which a sta- ARQ is not feasible. Below, we review some that have
tion simultaneously transmits the same data packet been developed for multicast. Basically, the problem
to some of its neighbors. A multicast transmission is with lost multicast traffic can be resolved in two ways:
a transmission from one sender, the AP in an IEEE avoiding or detecting packet loss.
802.11 infrastructured network, to a group of zero or
more recipients. Each multicast group has a special 2.1. Channel Reservation and Collision
MAC-address recognized by its members. Detection in IEEE 802.11
Improving the reliability of multicast traffic is im- The MAC scheme used in the IEEE 802.11 standard is
portant since multicast traffic, for example streaming called the distributed coordination function (DCF). It is
video, reaches at least as many users as unicast traf- a distributed contention-based access method based on
fic. Thus, poor reliability for multicast traffic can be the carrier sense multiple access with collision avoid-
expected to have a higher negative impact on user sat- ance (CSMA/CA) algorithm. Carrier sensing, known
isfaction than the converse case for unicast traffic. as clear channel assessment (CCA) in IEEE 802.11,
Not all multicast traffic has the same semantics. is used to determine the status of the medium prior to
Sometimes it is important that all recipients receive attempting to access the channel. If the channel is de-
all frames in a timely and orderly fashion and that termined to be busy or after a collision a station chooses
the sender knows that they have done so, for example a back-off counter (time to wait until retransmission)
configuration and control. In our scenario, it is suffi- randomly from a contention window (CW), which is
cient that most recipients receive most of the packets. doubled after each unsuccessful attempt. The back-off
Streaming video can tolerate and compensate for data counter is decremented by one for each time slot the
loss in several different ways. medium is idle, and as soon as the medium becomes
Copyright © 2006 John Wiley & Sons, Ltd. Wirel. Commun. Mob. Comput. 2007; 7:795–808
 A COLLISION DETECTION METHOD FOR MULTICAST IN CSMA/CA 797

busy the countdown is postponed until the medium has to frames being retransmitted after a long time if the
been idle during a DIFS. round-robin queue is long.
Subsequent frame transmissions are separated by In the techniques where all stations respond in turn
short time gaps on the channel called inter frame spaces the responses can either be sequential or simultaneous.
(IFS), used to differentiate control and data frames. The In one method, all recipients send a CTS in an order
Short IFS (SIFS) is used when a station have seized determined by the multicast sender and announced in
the medium and need to keep it for the completion of the RTS packet [2]. This method obviously has a large
a frame exchange sequence, for example between the overhead and will scale poorly. In another scheme, each
data frame and the corresponding ACK frame. Data station waits a random interval in case someone else
frames are separated by a DCF IFS (DIFS) which is responds before sending their own CTS [12]. If the
longer than a SIFS. interval is too long (≥DIFS) the stations will not have
There is an additional inter frame space called ex- time to respond before other transmissions start. If it is
tended inter frame space (EIFS) used by a receiving too short it becomes very probable that several stations
station to defer access to the channel after an unsuc- transmit at once.
cessful reception of a packet. The EIFS is intended to Two methods where all stations send their CTS at
provide sufficient time for another station, that did ex- the same time have been proposed to enable the mul-
perience a correct reception of the frame, to transmit ticast sender to decode the transmissions. In one case,
an ACK to the sender. As a result of this, the length the strongest (closest) sender is assumed to be suffi-
of the EIFS is much longer than other inter frame ciently stronger than all other senders together so that
spaces. the multicast sender can decode the CTS [13]. Others
An optional way of accessing the channel, by an argue that this is highly unlikely and instead propose
exchange of control frames prior to the data transmis- using linearly independent codes for each sender letting
sion, is included in DCF. The sender starts by sending the recipient decode all of them simultaneously [14].
a request to send (RTS) and the receiver responds with This solution might require extensive changes in the
a clear to send (CTS), after first waiting a SIFS. The physical layer.
channel is then reserved and the pending data frame A hybrid method has been proposed where a desig-
may be sent free of collisions. These control frames nated station sends the CTS but where stations who do
contain the duration of the complete frame exchange not ‘agree’ will send a ‘negative’ CTS (NCTS) in order
sequence which allows other stations, in range of the to cancel out the CTS [12]. This solution requires that
sender or the receiver or both, to defer access during all stations agree.
the completion of the frame exchange sequence. The
stations only in range of the receiver are referred to as
2.3. Collision Detection Techniques
hidden terminals (hidden from the sender).
Collision detection techniques can be further subdi-
vided into intransmission carrier sensing and ARQ ap-
2.2. Collision Avoidance Techniques
proaches.
With the exception of Blackburst [9], channel reserva- Kuri and Kasera [12] also proposed an analog hybrid
tion techniques for wireless MAC-layer multicast in scheme for ARQ. In it, a designated station transmits
IEEE 802.11 or other CSMA/CA protocols that we an ACK and any station which has not received the
have found in the literature are based on variations on multicast frame but detects the ACK will transmit a
the RTS/CTS scheme described above. The main dif- ‘negative’ ACK (NAK) to collide with the ACK. This
ference between the proposed methods is the number approach requires all stations to receive a given multi-
of stations which respond with a CTS, one station or cast packet correctly or not at all, that is, this approach
all stations. does not tolerate any packet degradation in marginal
In the techniques with one CTS responder, it can stations. A large number of multicast receivers with
either be a fixed and designated responder as in Robust varying channel qualities will deteriorate the perfor-
Multicast [10], or the stations can take turns in a round- mance of this approach since it is enough that one re-
robin fashion as in BMW [11]. In the latter method, ceiver experiences a packet loss for a retransmission to
the CTS station will also piggy-back information on be triggered.
which frames it has received earlier. The first method Sheu et al. [15] proposed a highly reliable broad-
suffers from the fact that the designated station might cast scheme for ad hoc networks [15]. They mod-
not be representative. The second solution might lead ify the 802.11 MAC scheme to incorporate broadcast
Copyright © 2006 John Wiley & Sons, Ltd. Wirel. Commun. Mob. Comput. 2007; 7:795–808
798 T. NILSSON, G. WIKSTRAND AND J. ERIKSSON

acknowledgments by dividing the DIFS time, follow- Table I. Glossary.

ing a broadcast transmission, into several mini-slots
and require each station to contend for transmitting ρ channel utilization
short acknowledgment frames. This ARQ approach l contention window size
k number of stations
does not scale to the number of multicast receivers and n number of slots
is limited by DIFS. τ probability for a station to have bo = 0
bo current back-off of a station
p(i, k, l) probability that i stations finish back-off simultaneously
w(r, i) probability that i stations back-off simultaneously in the
3. Early Multicast Collision Detection r:th attempt
P(r, i) probability of i contending stations in the r:th attempt
E[U] expected useful time
We have adopted the algorithm proposed by Rom [5] E[B] expected busy time
for the purpose of detecting collisions experienced by E[I] expected idle time
CWmin minimum back-off window size
multicast senders in an IEEE 802.11 network and thus TCDI duration of the collission detection interval
enabling retransmissions of lost packets. This will in- Tm duration of main transmission
crease the reliability for the otherwise unprotected mul- Tmin duration of shortest possible transmission
Ts duration of a successful transmission
ticast traffic but also minimize the overhead associated TCND duration of an unsuccesful transmission
with collisions. Tv duration of vanguard transmission
The Rom algorithm differs from non-persistent T smallest detectable time difference
CCA clear channel assessment
CSMA in its ability to detect collisions during the trans- CIFS collision inter frame space
mission. The main difference from CSMA is that a SIFS short inter frame space
sender transmitting a packet pauses during the trans- DIFS DCF inter frame space
EIFS extended inter frame space
mission and performs an additional carrier sense op-
eration to detect other potential transmitters. A sender
detecting a collision does not abort the transmission
immediately but continues for a predetermined period, 3.1. Protocol Description
referred to as the collision detection interval (CDI), to
assure that a later scheduled pause does not miss the EMCD operates in three phases, see Algorithm 1. (A
collision. If the channel is sensed idle during the CDI glossary of our notation is provided in Table I.) The
the transmission continues as normal. first phase is the vanguard transmission, the second
Lo and Mouftah [4] and Rom [5] have proposed sim- phase is the carrier sensing phase, and the third phase is
ilar algorithms for collision detection for the CSMA the main or jamming transmission. Since stations can
family of protocols. Using the categories in Section 2, transmit data at varying rates, frames of the same size
these algorithms and EMCD are collision detection may require different transmission durations. For this
techniques. The work presented by Rom [5] is a more reason, EMCD is based on timing intervals rather than
elaborate algorithm in the sense that the pausing sub packet size. This is different from the Rom algorithm
interval is chosen from a uniform distribution, whereas which implicitly assumes a fixed transmission rate for
Lo and Mouftah assume a non-slotted CSMA with the all transmitters [5].
same pausing sub interval for all senders. We have chosen not to double the contention window
for EMCD stations in order to prioritize the multicast
traffic carried in the downlink. This is motivated by
the unbalanced contention between up and downlink
Algorithm 1 Early Multicast Collision Detection reported by Pilsof et al. [16]. The access point, alone
success = false, bo = 0, retries = 0 contends for all the downlink traffic, whereas each sta-
while ¬ success AND retries≤ Rmax do
Perform backoff/collision avoidance tion individually contends for its uplink traffic. This
Perform Vanguard Transmission Tv , duration from (3.1.2) results in an unbalanced contention since the access
if CCA() = busy then point have no higher priority than a normal station in
retries=retries+1
jam until end of CDI accessing the channel. Hence, unicast stations contend-
bo ∈ U {0, CWmin } ing for the uplink can easily starve the downlink traffic
else under high load. Under saturated conditions with a total
Perform main transmission Tm
success=true of N contending senders including the access point, the
end if access point will only recieve 1/N of the total capacity,
end while that is, the same as all the other stations.
Copyright © 2006 John Wiley & Sons, Ltd. Wirel. Commun. Mob. Comput. 2007; 7:795–808
 A COLLISION DETECTION METHOD FOR MULTICAST IN CSMA/CA 799

The retransmission mechanism used in IEEE 802.11

uses a short and long retry limit for short control pack-
ets (CTS and RTS) and longer data packets, respec-
tively. Recommended values are four for data packets
and seven for control packets [1]. The retransmission
strategy is important to multimedia applications, that
are more sensitive to delay and jitter. Retransmitting a
data packet carrying video information may not make
sense if the play out deadline has passed. There is a Fig. 1. Different Tv generated by three senders.
clear trade-off between tolerance to packet losses and
delay. We are assuming a lower retransmission limit Tmin is the minimum allowed or possible transmission
(Rmax ) for multicast packet to minimize delays, see time. T is the smallest time difference that can be
Section 5. Furthermore, as mentioned earlier EMCD detected between two vanguard transmission from dif-
is not doubling the contention window to prioritize the ferent senders by the CCA and TCDI is the maximum
multicast traffic, this will give the multicast senders duration of the vanguard transmission plus CIFS (see
higher capacity and lower delays. below). Figure 1 illustrates how three senders have se-
lected different Tv . All EMCD stations in an ESS, that
3.1.1. Collision detection interval is, stations including access points that implement and
use EMCD, should use the same value for TCDI .
The length of the CDI, that is TCDI , affects the perfor-
mance of EMCD in three ways. First, TCDI determines 3.1.3. Duration of the carrier sensing phase
the maximum number of unique transmission times
that are selectable by the multicast senders. If the CDI After the vanguard transmission, the station performs
is very short there will only be a few unique trans- a CCA to detect other transmissions, see Algorithm 1.
mission times and hence the probability for deriving We refer to the interval in which the CCA is performed
the same transmission time will be greater. Second, as the collision inter frame space (CIFS). Selecting a
TCDI has an effect on the detection probability for short CIFS minimizes overhead.
collisions with unicast senders. If the derived vanguard A lower bound on the CIFS is defined by the time
transmission is longer than the unicast transmission, required to perform a correct CCA (ordinarily < 4 ␮s
the multicast sender will not detect the collision. Third, [17, pp. 27 & 40]) the signal propagation time, and the
a long detection period will add additional delay and time required to switch from transmitting to receiving
more overhead to the retransmission, mainly because and back again, that is the receive/transmit turnaround
the collision is detected later. time. The first term is partially hardware dependent,
So the CDI should be as short as possible to detect the second term depends of the physical configuration
a majority of the unicast transmissions, but still large of the arena, and the third term is hardware dependent.
enough to yield several unique transmission times. An upper bound on the CIFS is defined by the SIFS,
A static solution might be envisaged where every PIFS, and DIFS timing intervals (16, 25, and 34, re-
multicast sender would be assigned a Tv,i = Tv,j for spectively in IEEE 802.11 [1, p. 85] [17, p. 40]). It
any stations i = j to be used for every vanguard trans- must be shorter than PIFS and DIFS of them to prevent
mission; no collisions would go undetected. However, a other stations from detecting the channel idle, and con-
static solution would require configuration of the mul- sequently begin their transmissions in between Tv and
ticast senders, which is contrary to the basic idea of Tm . So, in an 802.11a WLAN the CIFS can be in the
decentralization in IEEE 802.11. It will not be consid- range 11 ≤ CIFS ≤ 25.
ered in this paper.
3.2. Collision Scenarios
3.1.2. Duration of the vanguard transmission
Figure 2 illustrates a collision between two multicast
The duration of the vanguard transmission is derived senders (1 and 2) starting their transmissions at the
from the expressions below: same time. Sender 2 will detect the collision before
Sender 1, and continue by jamming the medium until
Tv = Tmin + XT the end of the CDI to assure that Sender 1 also detects

 the collision. At the end of the CDI they will end their
X ∈ U 0, TCDIT−T

min
(1) transmissions and reschedule the collided packets
Copyright © 2006 John Wiley & Sons, Ltd. Wirel. Commun. Mob. Comput. 2007; 7:795–808
800 T. NILSSON, G. WIKSTRAND AND J. ERIKSSON

Fig. 2. Collision between two multicast senders is detected by EMCD. Retransmissions are scheduled.

according to Algorithm 1. The access point generating Table II. Protocol overhead in ␮s for multicast using EMCD,
multicast, and unicast. The payload length is 1496 bytes and the
the shortest back-off time will win the next contention rate is 6 Mbps.
phase (Sender 2 in Figure 2).
In Figure 3 a multicast sender (1) and a unicast sender
EMCD Multicast Unicast
(2) start their transmissions simultaneously resulting
in a collision. The carrier sense operation performed Preamble 40 20 20
by station 1 will detect the collision. Station 2 will ACK — — 44
detect the collision when it does not receive an ACK. SIFS — — 16
CIFS 11 — —
Both stations will back-off before retransmitting their Data 1996 1996 1996
collided packets. If the packet had been shorter than Mac header 104 52 52
Tmin the collision would not have been detected. Total ␮s 2151 2068 2128
% 7.2 3.55 6.22

3.3. Compatibility
It is sufficient to implement EMCD in the multicast packets are split into two transmissions by the EMCD
senders. Multicast recipients need no further infor- algorithm and the overhead per packet is doubled. The
mation than what is required to reassemble the two overhead for EMCD is slightly larger than for unicast.
parts of the packet. The multicast senders could make
that process entirely transparent by refragmenting the
packet at a higher protocol level. That would also cir- 4. Analytical Evaluation
cumvent the injunctions in the 802.11 standard against
fragmenting multicast packets [1, p. 71] and mak- The Rom algorithm was evaluated analytically in the
ing earlier fragments of a packet shorter than later original paper [5]. The analysis uses an infinite popu-
fragments [1, p. 93]. lation of users which as an aggregate form a Poisson
arrival process. In general the time to transmit a packet
is shorter than the arrival rate so the channel is far from
3.4. Protocol Overhead
saturation [18].
The extra carrier sense operation in EMCD imposes Here it is assumed that there is a finite population
an additional overhead on transmissions. In Table II of k stations. Each station operates under saturation
the protocol overhead is shown for unicast, multicast, conditions, that is, it always has a packet ready for
and multicast using EMCD, under the assumption that transmission. We further assume that there are only
the data payload is 1496 bytes and the rate is 6 Mbps. EMCD stations. We also assume that no packets are
Note that only the packet overhead is considered in the lost because of noise and that the EIFS mechanism
table and not the back-off time, DIFS, etc. Multicast works perfectly.

Fig. 3. Collision between a multicast and a unicast transmission. Both stations detect it and schedule retransmissions.

Copyright © 2006 John Wiley & Sons, Ltd. Wirel. Commun. Mob. Comput. 2007; 7:795–808
 A COLLISION DETECTION METHOD FOR MULTICAST IN CSMA/CA 801

The first two assumptions are fairly standard in this and we write
type of analysis, cf. [19]. The following three assump-  
k
tions are made for simplicity of analysis, the impact of i τ(l)i (1 − τ(l))k−i i≤k
p(i, k, l) = (4)
breaking them are considered in the simulation section 0 i>k
later.

4.1.3. `Winners' in each step

4.1. Channel Utilization
Recall that the EIFS affects what happens after a colli-
Consider a transmission cycle consisting of zero or sion. Non-colliding stations will defer until a successful
more unsuccessful transmission attempts followed by a transmission occurs or to the end of the EIFS. Let e be
successful transmission. The channel utilization ρ will the length of the EIFS in slots and P(r, i) be the prob-
be the time required for the transmission of useful data ability that i stations are left before contention phase
in a cycle divided by the length of the cycle. (A glossary r begins. First, consider the case where the EIFS is
of our notation is provided in Table I.) Let E[U], E[I], longer than the contention window, that is, e ≥ l, then
and E[B] be the expected useful, idle and busy times the probability w(r, i) for i ‘winning’ stations is given
of a cycle then [4,18,19] by

E[U] k

ρ= (2) p(i, j, l)P(r, j)
E[I] + E[B] w(r, i) = (5)
1 − p(0, j, l)
j=1
E[U] is easily obtained from the average packet length
m in units of time. The expected idle and busy times Under our assumptions we start with k stations in the
are given in Equations (13) and (11), respectively. Let n first step. The number of stations left is basically the
denote the number of slots in the CDI and k the number same as the number of winning stations in the preced-
of stations. ing steps except for the following: if there is a single
winning station in the earlier step then there will be no
stations left in the following step and we can write
4.1.1. Single station back-off
P(r, 1, . . . , k) =
Under the assumption of independent back-off between 
stations the back-off of a single station can be described (0, . . . , 0, 1) r=1
(6)
using a Markov-chain. By solving for the stationary (0, w(r − 1, 2), . . . , w(r − 1, k)) r > 1
probability distribution it is possible to obtain the prob-
ability that a single station has a back-off counter of If e < l two cases can occur, either at least one station
zero and will transmit. It was given as τ = 2/(W + 1) finishes backing-off before the end of the EIFS or the
by Bianchi [20, Eq. 8], where W is CW+1. We shall EIFS ends with no stations having finished backing-
use a slightly different notation l = W − 1, and write off. The probability that a station i will have a back-off
counter boi higher than the EIFS timeout is
2
τ(l) = (3) l−e e
l+2 P(boi > EIFS) = =1−
l l
The probability that none of the stations have fin-
4.1.2. Multiple stations back-off ished backing-off before the end of the EIFS is
 

For Bianchi, it was sufficient to determine the proba- k l−e k
bilities that there was a transmission from one station P min boi ≥ EIFS = (7)
i=1 l
and if so, if that station was the only one transmitting.
In our case we must know how many stations collide
where k is the number of initially participating stations,
to be able to derive the probability that a collision will
1 ≤ k ≤ k. So the probability that at least one station
be detected.
finishes backing-off is
The probability that i out of k stations have finished


backing-off at the same time will be given by the bi- k l−e k
nomial distribution with τ(l) as the success probability P(min boi < EIFS) = 1 − (8)
i=1 l
Copyright © 2006 John Wiley & Sons, Ltd. Wirel. Commun. Mob. Comput. 2007; 7:795–808
802 T. NILSSON, G. WIKSTRAND AND J. ERIKSSON

Considering the two cases, the probability that there 4.1.5. Idle time
will be i ‘winning’ stations in step r is given by com-
Remember that the expected value of a discrete stochas-
bining Equations (5), (7), and (8). After some trivial
tic variable is the sum of the probabilities of each out-
simplifications we have
come times its numerical value. The channel will be
k
 
 idle until at least one station has finished backing off.
 p(i, j, l) l−e j
w(r, i) = P(r, j) 1 − The probability that at least one station has finished
1 − p(0, j, l) l backing off is 1 − p(0, k, l) from Equation (5). The ex-
j=1

k

 pected number of back-off steps before that happens is
p(i, k, l)  l−e j given by the geometric distribution
+ P(r, j)
1 − p(0, k, l) l
j=1
p(0, i, l)
(9) E[back-off] = (12)
1 − p(0, i, l)

4.1.4. Busy times Each transmission attempt will be proceded by a

DIFS and with a certain probability an EIFS. Hence
In each transmission attempt, there are three possible
the expected idle time in a cycle is
outcomes: Only one station transmits and the transmis-
sion is considered to be successful (s). Two or more sta- 
∞ 
 k
i
tions transmits and either the collision is detected (CD) l−e
or it is not detected (CND). The busy times are derived E[I] = q(r, i) DIFS + etslot
l
from Algorithm 1 and the IEEE 802.11a standard [17]. r=1 i=1

The busy time will be the same if the transmission p(0, i, l)tslot
is successful or not, and will be given by Ts = TCND = + (13)
1 − p(0, i, l)
m + 155, see Table II. If the collision is detected the
busy time will be TCDI .
The channel utilization is easily found by substitut-
The probability of not detecting a collision is the
ing Equations (11) and (13) in Equation (2).
probability that all stations will choose the same slot
to pause in and is

4.2. Results
P(detect) = n1−k
The analysis was validated by comparison to simula-
tions in GloMoSim see Section 5 for details. Figure 4
if there are k transmitting stations. The risk of not
shows that the analytical results are a good match for
detecting a collision will decrease with increasing n
the simulated results as long as the number of stations
and k. The expected busy time for a single transmission
is small.
attempt is
The lowest curves, for n = 1, show what perfor-
mance could be expected from normal multicast but
E[busy] = n1−k TCND + (1 − n1−k )TCDI (10) with some additional overhead. Channel utilization
decreases assymptotically towards a value around 0.35.
The expected busy time for all attempts in a cycle is The upper two pairs of curves show an interesting be-
havior in that they first decrease then increase slightly.
k
∞  The effect is more pronounced in the simulated curves.

E[B] = w(r, i)(n1−i TCND + (1 − n1−i )TCDI ) This is an effect of the back-off no longer being inde-
r=1 i=1
pendent for each station when there are a large number
of stations as explained in Subsection 4.3.
(11) The optimal number of steps nopt as a function of the
number of stations and the effect on channel utilization
If n = 0, that is ordinary multicast, then the right- is presented in Figure 5. The figure shows two things:
hand factor will reduce to TCND which is obviously first, the optimal number of slots decreases with the
larger than n1−i TCND + (1 − n1−i )TCDI for any posi- number of stations and second, the choice of n has
tive n and i. smaller impact for lower numbers of stations.
Copyright © 2006 John Wiley & Sons, Ltd. Wirel. Commun. Mob. Comput. 2007; 7:795–808
 A COLLISION DETECTION METHOD FOR MULTICAST IN CSMA/CA 803

0.85

0.8

0.75

0.7
Channel utilization ρ

0.65

Simulated, n=1
0.6
Simulated, n=3
Simulated, n=5
0.55
Analytical, n=1
Analytical, n=3
0.5
Analytical, n=5

0.45

0.4

0.35
0 10 20 30 40 50
Number of stations

Fig. 4. Channel utilization ρ as a function of the number of stations k with CW l = 15, simulated and analytical. The relative
mean square error is ≈ 2.1%.

4.3. A Note on Asymptotic Behavior the assumption that almost all channel allocations will
start with a competition for the medium during at least
As the number of stations k grow in relation to the num- one back-off time slot.
ber of available back-off slots l + 1 the preceding anal- What will happen instead is that we will end up with
ysis will become increasingly inaccurate. It is based on 1 ≤ k ≤ k stations with bo = 0. All of them will try

20 0.9
n=nopt
n=2
n=4

15 0.8
Channel utilization ρ
nopt

10 0.7

nopt
5 0.6
0 10 20 30 40 50
Number of stations

Fig. 5. Selecting the optimal number of slots n.

Copyright © 2006 John Wiley & Sons, Ltd. Wirel. Commun. Mob. Comput. 2007; 7:795–808
804 T. NILSSON, G. WIKSTRAND AND J. ERIKSSON

to transmit (and fail). After that they will choose a new tion is kept during the whole simulation. All stations in
backoff value so that on average k /(l + 1) stations will a BSS are members of a single multicast group.
remain with bo = 0. This will reoccur until zero or one Both the multicast and unicast traffic assumes a con-
stations are left. In the first case, all stations will back- stant bit rate model with a packet size of 1496 bytes.
off once. In the second case a successful transmission Two different scenarios are simulated. In the first
will occur. scenario the multicast senders have a fixed bit rate of
500 kbps and the aggregated load from unicast senders
is gradually increased by introducing more senders un-
5. Simulation Results til the saturated condition is achieved. In this scenario,
each unicast sender adds a packet rate of four packets
We extended the analytical evaluation from the previ- per second to the aggregated load. In the second sce-
ous section with simulations in order to examine factors nario with saturated conditions each sender constantly
which are beyond the scope of the analysis. The simu- has a new packet to send.
lations were performed in the discrete event simulator For the modeling the path loss we have used a noise
GloMoSim [21]. accumulating two-ray path loss model, and for fading
We are mainly interested in reliability and efficiency. we assume a Rician distribution with K = 5. This rep-
Efficiency is often measured as throughput which is resents a strong line of sight path between the trans-
hard to define. The aggregated throughput for multicast mitter and receiver, resembling an open arena environ-
can not be easily calculated because there are multiple ment. The output power is set to 15 dBm. All access
recipients and some of them may receive the packet points are assumed to use the same frequency channel
while other do not. Consequently, it is difficult to know resulting in overlapping basic service sets. The data
if a multicast packet should be included in the aggre- rate for all transmissions is set to 6 Mbps.
gated throughput or not. For instance, Tourhilles [10]
considered only a single broadcast receiver. Another
5.2. Variable Load Scenario
way is to consider the average reception. We consider
the average packet delivery ratio which is more a mea- In the first simulation scenario EMCD is evaluated un-
sure of reliability than efficiency. der variable load. In the simulation there are four and
The following metrics are used: delay and ratio of nine overlapping cells (access points) with both multi-
assumed success to measure efficiency and packet de- cast and unicast senders.
livery ratio to measure reliability. The delay is the time In Figure 6, the average packet delivery ratio is plot-
from when a packet arrives at the MAC layer in the ted against the number of unicast senders. For standard
sender until it is successfully received. The average multicast (802.11) the packet delivery ratio decreases
packet delivery ratio is defined as the average ratio rapidly as the offered load increases, when more
of multicast packets received by the multicast group unicast senders join the network. When the medium
to the unique packets offered for transmission. Ra- becomes saturated, for around 70 stations, the rate of
tio of assumed success is defined as (#main transmis- decrease becomes smaller mainly as a result of much
sions)/(#vanguard transmissions). When a main trans- higher CW values for the unicast senders. EMCD is ca-
mission begins the transmission is assumed successful. pable of maintaining a relatively high packet delivery
More than one vanguard transmission per main trans- ratio for an increasing load. The upper curves are for
mission will result in a ratio of assumed success smaller four and nine access points, and two values of n. The
than one. effect of increasing n is more evident in the case
of nine access points where more collisions are
between the access points and more slots will increase
5.1. Traffic and Radio Model
the detection probability. Moreover, the saturation
In each simulated scenario a specific number of access points is reached more quickly for EMCD since
points are evenly distributed onto a quadratic area with retransmissions are performed that increase the
sides 200 m long. The specific number of stations are load.
varied for the different simulation scenarios and the In Figure 7 the ratio of assumed success is shown for
position of each station is randomly set within the sim- an increasing number of unicast senders. Increasing the
ulation area. number of unicast senders and/or access points require
We assume that there is no mobility. Each station is more vanguard transmissions per main transmission
associated with the closest access point. The associa- due to collisions. The fraction of collisions experienced
Copyright © 2006 John Wiley & Sons, Ltd. Wirel. Commun. Mob. Comput. 2007; 7:795–808
 A COLLISION DETECTION METHOD FOR MULTICAST IN CSMA/CA 805

0.9
Average packet delivery ratio

0.8

0.7

0.6
802.11, 4 AP
0.5 802.11, 9 AP
EMCD, 4 AP, n = 2
0.4 EMCD, 4 AP, n = 13
EMCD, 9 AP, n = 2
0.3 EMCD, 9 AP, n = 12

0.2
10 20 30 40 50 60 70 80 90 100
Number of stations

Fig. 6. The average packet delivery ratio for EMCD and standard 802.11 for four and nine access points.

between multicast senders are higher for low and high 5.3. Saturated Load Scenario
loads. For lower loads, few unicast senders are present
and a majority of the collisions are between multicast In the second scenario, a fully saturated network with
senders. For high loads, the unicast senders have much four access points, the aim is to illustrate the trade-off
higher values of CW then the multicast senders (hav- between packet losses and delay when changing some
ing CWmin ) resulting in more collisions between the of the parameters.
multicast senders. This is why the effect of increasing In Figure 8 the average packet delivery ratio is shown
n is more distinct for these two load levels. for three different retransmission limits Rmax and two

1
EMCD, 4 AP, n = 13
EMCD, 4 AP, n = 2
0.9
EMCD, 9 AP, n = 12
EMCD, 9 AP, n = 2
Ratio of assumed success

0.8

0.7

0.6

0.5

0.4
10 20 30 40 50 60 70 80 90 100
Number of stations

Fig. 7. The ratio of assumed success versus an increasing number of unicast sender for four and nine access points.

Copyright © 2006 John Wiley & Sons, Ltd. Wirel. Commun. Mob. Comput. 2007; 7:795–808
806 T. NILSSON, G. WIKSTRAND AND J. ERIKSSON

0.95
Average packet delivery ratio

0.9

Rmax =2, n = 2
Rmax =2, n = 13
0.85
Rmax =3, n = 2
Rmax =3, n = 13
Rmax =4, n = 2
0.8
Rmax =4, n = 13

0.75
10 20 30 40 50 60 70 80 90 100
Number of stations

Fig. 8. The average delivery ratio for three different values of Rmax and two values of n.

sizes of n. The delivery ratio is increasing for higher percentage of the collisions will be between multicast
values of Rmax . The effect of changing the number of senders and increasing n has smaller effect.
slots n depends on the fraction of unicast senders in the When Rmax = 4 and only a few slots exists the
system. When the fraction of unicast senders are low, packet delivery ratio may actually increase when more
the majority of collisions will be between multicast unicast senders enters the network. This is because all
senders and if n is increased some gain in performance collisions with unicast senders will be detected, lead-
is achieved. However, this effect is less significant when ing to more retries and so the success probability will
the fraction of unicast senders are higher. Only a small increase.

Rmax = 2, n = 2
0.028 Rmax = 2, n = 13
Rmax = 2, n = 2
0.026
Rmax = 2, n = 13
Rmax = 4, n = 2
0.024
Average delay (s)

Rmax = 4, n = 13

0.022

0.02

0.018

0.016

0.014
10 20 30 40 50 60 70 80 90 100
Number of stations

Fig. 9. The average delay for three different values of Rmax and two values of n.

Copyright © 2006 John Wiley & Sons, Ltd. Wirel. Commun. Mob. Comput. 2007; 7:795–808
 A COLLISION DETECTION METHOD FOR MULTICAST IN CSMA/CA 807

The average delay is shown in Figure 9. The delay is 3. Lo WF, Mouftah HT. Carrier sense multiple access with colli-
evidently increasing for higher values of Rmax . Chang- sion detection for radio channels. In International Communica-
tions and Energy Conference. IEEE: New York, NY, USA, 1984;
ing n has much smaller impact on the delay, especially pp. 244–247.
for smaller values of Rmax . The two curves at the bot- 4. Lo WF, Mouftah HT. Collision detection and multitone tree
tom are almost indistinguishable while there is more search for multiple-access protocols on radio channels. IEEE
Journal on Selected Areas in Communications 1987; 5(6):
difference in the two upper curves. 1035–1040.
Clearly, there is a trade-off between delay and packet 5. Rom R. Collision Detection in Radio Channels. In Local Area
delivery ratio which is determined by Rmax . In the case and Multiple Access Networks, Pickholtz RL (ed.). Computer
Science Press, Inc.: Rockville, MD, USA, 1986; 235–249.
of a streaming application the choice of Rmax may be 6. Kwon Y, Fang Y, Latchman H. Design of MAC protcols with
important to maximize the utility of the service. fast collision resolution for wireless local area networks. IEEE
Transactions on Wireless Communications 2004; 3(3): 793–807.
doi:10.1109/TWC.2004.827731.
6. Conclusion 7. Sagduyu YE, Ephremides A. Energy-efficient collision res-
olution in wireless ad hoc networks. In INFOCOM 2003,
Vol. 1. IEEE: Piscataway, NJ, USA, 2003; pp. 492–502.
We have adapted the Rom algorithm to an IEEE 802.11 doi:10.1109/INFCOM.2003.1208700.
network with the purpose of detecting and recovering 8. Fullmer CL, Garcia-Luna-Aceves JJ. FAMA–PJ: a chan-
nel access protocol for wireless LANs. In MOBICOM’95.
from multicast packet losses caused by collisions. We ACM Press: New York, NY, USA, 1995; pp. 76–85.
call the algorithm EMCD. We have used probabilistic doi:10.1145/215530.215559.
analysis and simulations to study the performance of 9. Sobrinho JL, Krishnakumar AS. Real-time traffic over the IEEE
802.11 medium access control layer. Bell Labs Technical Journal
EMCD. 1996; 1(2): 172–187.
The main contributions of this paper are (a) that 10. Tourrilhes J. Robust broadcast: improving the reliabil-
we have adapted the Rom algorithm to contemporary ity of broadcast transmissions on CSMA/CA. In PIMRC,
Vol. 3. IEEE: New York, NY, USA, 1998; pp. 1111–1115.
CSMA/CA networks and MAC layer multicast, (b) that doi:10.1109/PIMRC.1998.731349.
we have provided a probabilistic analysis of its perfor- 11. Tang K, Gerla M. MAC reliable broadcast in ad hoc networks.
mance under saturated load conditions, and (c) that we In MILCOM 2001, Vol. 2. IEEE: Piscataway, NJ, USA, 2001;
pp. 1008–1013. doi:10.1109/MILCOM.2001.985991.
have provided simulation results for the algorithm un- 12. Kuri J, Kasera SK. Reliable multicast in multi–access
der both saturated and unsaturated conditions in mixed wireless LANs. Wireless Networks 2001; 7(4): 359–369.
networks. doi:10.1023/A:1016631911947.
13. Tang K, Gerla M. MAC layer broadcast support in 802.11 wire-
Some of the key findings are that EMCD signif- less networks. In MILCOM 2000, Vol. 1. IEEE: Piscataway, NJ,
icantly reduces undetected packet losses caused by USA, 2000; pp. 544–548. doi:10.1109/MILCOM.2000.905017.
collisions and reduces the time wasted on collisions 14. Lott M, Sitalov A, Linsky E, Li H. Performance analysis
of multicast transmission in WLAN. In VTC 2003-Spring,
(see Eq. (11)). The net effect is higher efficiency and Vol. 2. IEEE: Piscataway, NJ, USA, 2003; pp. 1223–1227.
reliability. doi:10.1109/VETECS.2003.1207822.
Only a relatively low number of slots are needed to 15. Sheu ST, Tsai Y, Chen J. A highly reliable broadcast scheme for
IEEE 802.11 multi–hop ad hoc network. In 2002 IEEE Inter-
achieve optimal channel utilization. With a large num- national Conference on Communications. Conference Proceed-
ber of stations, n = 7 is optimal. With fewer stations n ings, Vol. 1. IEEE: Piscataway, NJ, USA, 2002; pp. 610–615.
is higher but the system is also less sensitive to changes doi:10.1109/ICC.2002.996925.
16. Pilsof S, Ramjee R, Raz D, Shavitt Y, Sinha P. Understanding
in n. Adding unicast to the mix changes the situation TCP fairness over wireless LAN. In IEEE INFOCOM 2003,
slightly, on one hand fewer slots are needed to detect a Vol. 2. IEEE: Piscataway, NJ, USA, 2003; pp. 863–872.
collision as long as there is at least one unicast station 17. O’Hara B (ed.). Wireless LAN Medium Access Control (MAC)
and Physical Layer (PHY) Specifications—Amendment 1: High-
involved, on the other hand ending the transmission at speed Physical Layer in the 5 GHz band. IEEE: New York, NY,
TCDI has less effect as the unicast transmission will con- USA, 1999.
tinue anyway. An acceptable choice in many situations 18. Kleinrock L, Tobagi FA. Packet switching in radio channels:
Part I—carrier sense multiple-access modes and their
could be n = 7. throughput-delay characteristics. IEEE Transactions on Com-
munications 1975; 23(12): 1400–1416.
References 19. Calı̀ F, Conti M, Gregori E. Dynamic tuning of the IEEE 802.11
protocol to achieve a theoretical throughput limit. IEEE/ACM
Transactions on Networking 2000; 8(6): 785–799.
1. O’Hara B (ed.). Wireless LAN Medium Access Control (MAC) 20. Bianchi G. Performance analysis of the IEEE 802.11 distributed
and Physical Layer (PHY) Specifications. IEEE: New York, NY, coordination function. IEEE Journal on Selected Areas in Com-
USA, 1999. munications 2000; 18(3): 535–547.
2. Sun M, Huang L, Wang S, Arora A, Lai TH. Reliable MAC 21. Zeng X, Bagrodia R, Gerla M. Glomosim: a library for the par-
layer multicast in IEEE 802.11 wireless networks. Wireless allel simulation of large-scale wireless networks. In PADS ’98.
Communications and Mobile Computing 2003; 3(4): 439–453. IEEE Computer Society: Washington, DC, USA, 1998; pp. 154–
doi:10.1002/wcm.129. 161. doi:10.1145/ 278008.278027.

Copyright © 2006 John Wiley & Sons, Ltd. Wirel. Commun. Mob. Comput. 2007; 7:795–808
808 T. NILSSON, G. WIKSTRAND AND J. ERIKSSON

Authors’ Biographies a doctoral student at Linköping University during 1998 and

1999. In 2000, he was with Procter & Gamble in Stockholm,
Sweden. During 2001–2002, he worked as a research project
Thomas Nilsson is a Ph.D. student at
manager at Ericsson AB. Greger Wikstrand is a student mem-
Umeå University, Sweden. He focuses
ber of the ACM and the IEEE.
on resource optimization and Quality of
Service in wireless networks. Thomas
received his Master of Science degree Jerry Eriksson is a senior lecturer at
in computing science in 2002 and his Umeå university. He obtained the Master
licenicate degree in 2005, both from of Science in Computing Science from
Umeå University. He was born in Umeå, Umeå University in 1990 and the Ph.D.
Sweden in 1978. Thomas Nilsson is a from the same university in 1995. His
student member of the IEEE. doctoral dissertation (1995) treats the op-
timization and regularization of nonlin-
ear least squares problems. He worked at
Greger Wikstrand is a doctoral student
Ericsson from 1997 to 2000. His current
at Umeå University. He was born in
research interests involves telecommunication and resource
Uppsala, Sweden in 1972. He obtained
optimization.
the Master of Science in Engineering
Physics degree from Uppsala University
in 1998, and the Licenciate of Engi-
neering degree in Computing Science
from Umeå University in 2003. He was

Copyright © 2006 John Wiley & Sons, Ltd. Wirel. Commun. Mob. Comput. 2007; 7:795–808