Quality-Differentiated Video Multicast

in Multirate Wireless Networks

Kate Ching-Ju Lin, Member, IEEE, Wei-Liang Shen,
Chih-Cheng Hsu, and Cheng-Fu Chou, Member, IEEE
AbstractAdaptation of modulation and transmission bit-rates for video multicast in a multirate wireless network is a challenging
problem because of network dynamics, variable video bit-rates, and heterogeneous clients who may expect differentiated video
qualities. Prior work on the leader-based schemes selects the transmission bit-rate that provides reliable transmission for the node that
experiences the worst channel condition. However, this may penalize other nodes that can achieve a higher throughput by receiving at
a higher rate. In this work, we investigate a rate-adaptive video multicast scheme that can provide heterogeneous clients differentiated
visual qualities matching their channel conditions. We first propose a rate scheduling model that selects the optimal transmission bit-
rate for each video frame to maximize the total visual quality for a multicast group subject to the minimum-visual-quality-guaranteed
constraint. We then present a practical and easy-to-implement protocol, called QDM, which constructs a cluster-based structure to
characterize node heterogeneity and adapts the transmission bit-rate to network dynamics based on video quality perceived by the
representative cluster heads. Since QDM selects the rate by a sample-based technique, it is suitable for real-time streaming even
without any preprocess. We show that QDM can adapt to network dynamics and variable video-bit rates efficiently, and produce a gain
of 2-5 dB in terms of the average video quality as compared to the leader-based approach.
Index TermsWireless video multicast, rate adaptation, QoS

1 INTRODUCTION
V
IDEO streaming is arguably one of the most popular
multimedia applications over wireless networks today.
With the ubiquitous of such applications, constant demand
for better wireless access technology has resulted in several
generations of new access point (AP) products, e.g., 802.11 a/
b/g/n, in a relatively short time. Future-generation APs are
expected to have much greater computation capability and
storage capacity. They offer new opportunities to incorpo-
rate even more advanced features to support a variety of
applications (e.g., [1]). In this paper, we consider a new AP
feature, namely quality-differentiated video multicast, to
allow better utilization of limited wireless resources for
video streaming.
Taking advantage of the wireless broadcast nature, a
video source can multicast a video object to a group of
multicast members in order to reduce the bandwidth
requirement, as compared with unicasting the data to each
individual member. However, current commercial network
devices typically transmit multicast packets at the base rate
in the MAC layer, even though 802.11 standard supports
multiple bit-rates up to 11 Mb/s for 802.11b or 54 Mb/s
for 802.11 a/g, each of which has a different modulation
scheme. This is a waste of wireless bandwidth if certain
members in the multicast group are capable of receiving
packets at a higher bit rate, and desire a better visual
quality. To address this problem, we propose a dynamic
rate adaptation scheme with quality-differentiated features
to better support heterogeneity in the clients.
The auto rate selection algorithms for unicast have been
examined in [2], [3], [4], [5], [6], [7], [8]. Other work [9],
[10], [11] studies the rate adaptation schemes for unicast
video streaming. They measure the loss probability accord-
ing to feedback acknowledged from the receiver, and
predict a rate that can achieve the highest throughput. Such
feedback-based schemes cannot be extended to multicast
scenarios because concurrent feedback from several multi-
cast members can lead to severe collision. To avoid this
effect, some multicast rate adaptation schemes [12], [13],
[14], [15], [16] select the member who experiences the worst
channel condition as the leader of the multicast group, and
predict a bit rate that can reach this leader (the worst
node). Such a leader-based approach is particularly
suitable for applications that need to deliver data to all
members reliably, e.g., data dissemination. However, this
approach may not be efficient for video multicast because it
merely selects the rate that maximizes the throughput of
the worst node. In doing so, it penalizes those nodes who
can receive data at a higher bit rate.
The rate selection problem
1
in video multicast scenarios
is more challenging due to heterogeneity of receivers. Since
various members may observe different channel conditions,
they can receive data sent at different bit-rates as illustrated
IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 12, NO. 1, JANUARY 2013 21
. K.C.-J. Lin and W.-L. Shen are with the Research Center for IT Innovation,
Academia Sinica, No. 128, Academia Road, Sec. 2, Nankang, Taipei 115,
Taiwan. E-mail: {katelin, wlshen}@citi.sinica.edu.tw.
. C.-C. Hsu and C.-F. Chou are with the Department of Computer Science
and Information Engineering, National Taiwan University, No. 1
Roosevelt Rd., Sec. 4, Taipei 106, Taiwan.
E-mail: kenneth@cmlab.csie.ntu.edu.tw, ccf@csie.ntu.edu.tw.
Manuscript received 6 Nov. 2010; revised 4 Oct. 2011; accepted 12 Oct. 2011;
published online 11 Nov. 2011.
For information on obtaining reprints of this article, please send e-mail to:
tmc@computer.org, and reference IEEECS Log Number TMC-2010-11-0508.
Digital Object Identifier no. 10.1109/TMC.2011.242.
1. We let video bit-rate denote the streaming rate encoded in the
application layer, and represent the transmission bit-rate, i.e., determined
by the modulation and channel coding rate, used to send MAC-layer
packets as MAC-layer bit-rate or bit-rate.
1536-1233/13/$31.00 2013 IEEE Published by the IEEE CS, CASS, ComSoc, IES, & SPS
in Fig. 1. To leverage this performance factor, an ideal way
is to select a different transmission bit-rate for each video
frame according to its importance (defined as the rate
scheduling problem), so that members can receive differ-
entiated video quality that best takes advantage of their
channel conditions. An intuitive solution of the rate
scheduling problem is to select the rates that can achieve
the maximal peak signal-to-noise ratio (PSNR) value (a metric
standardized by ITU [17] and used to evaluate video
quality [18]). However, under some scenarios (e.g., Fig. 1),
selecting a higher rate to maximize the PSNR may result in
the dilemma that the nodes who have worse channel
conditions (e.g., node D) cannot receive any packet
correctly due to a high bit error ratio (BER) of the MAC-
layer bit-rate and, thus, cannot even obtain the minimal
visual quality.
In this paper, we propose a visual-quality-based rate
adaptation protocol, which makes tradeoff between two
goals: providing differentiated video quality for various
clients to match their heterogeneous channel conditions and
guaranteeing minimum visual quality for each client.
Specifically, our goal is to develop a software-based rate
scheduling protocol in order to produce the maximal total
visual quality for quality-differentiated video multicast
under the constraint of ensuring at least minimum visual
quality for each member. The contributions of this paper
are as follows:
. We model the rate scheduling problem as a variation
of the Knapsack problem, and propose a dynamic
programming solution to solve it optimally.
. We propose a practical protocol, called Quality-
Differentiated Multicast (QDM), which exploits a
sample-based technique to adapt the transmission
bit-rate of each video frame to variable video bit-
rates and client mobility without the need of any
preprocess. Thus, it can be applied to real-time video
streaming.
Our simulation results show that not only can QDM
provide users differentiated video quality matching their
channel conditions, but also produce a better average visual
quality as compared with the leader-based schemes.
The remainder of this paper is organized as follows: We
discuss related work on rate selection for wireless multicast
in Section 2. We formulate the rate scheduling problem for
quality-differentiated video multicast in Section 3, and
introduce the QDM framework in Section 4. In Sections 5
and 6, we present simulation and experimental results,
respectively, to compare the performance of QDM with that
of the leader-based approach. Finally, we give our conclu-
sions in Section 7.
2 RELATED WORK
Due to the heterogeneity among multicast members,
different multicast recipients may observe dissimilar link
qualities. Recently, several works [19], [20] have focused on
how to allocate bandwidth efficiently for broadcasting
video streams to clients either with heterogeneous re-
sources, e.g., screen resolution or decoding capability, or
with multiple access technologies, e.g., 3G and WLAN.
Compared to those works that do no take channel
condition, packet losses and transmission bit-rates into
account, this work allocates bandwidth for video multi-
casting with consideration of multiple available transmis-
sion bit-rates and the corresponding loss probability. Our
goal is to provide clients heterogeneous visual quality
matching their channel conditions.
Most work on rate selection for wireless multicast
focuses on achieving multicast reliability by selecting the
rate that can deliver data reliably to the member with the
worst channel condition. In [12], the Leader-Based Protocol
(LBP) is the first leader-based approach proposed to
overcome the problem of feedback collision. It selects the
worst node as the leader to acknowledge multicast packets.
Other members can issue negative acknowledgements to
collide the acknowledgement sent by the leader and, thus,
trigger the sender to retransmit the lost packets. The goal of
LBP is to support reliability by a single feedback. However,
it does not adapt the transmission bit rate to dynamic
channel conditions, but only sends data at the base rate.
Thus, the rate adaptation algorithms, such as RAM [13] and
ARSM [15], are proposed for the leader-based multicast
protocol. They estimate link quality of the leader and
determine a proper rate that can better reach the leader.
Both of these techniques let each receiver embed the
information about it receiving SNR value in the CTS frame.
The sender can infer the leaders SNR upon receiving
the CTS frames, and predict a suitable rate accordingly.
Some other work [21], [22] extends the leader-based
multicast protocol to an environment with multiple APs
that operate on nonoverlapping channels, and investigates
the optimal client-AP association control problem in such a
multi-AP scenario. A central coordinator is deployed to
connect all APs, and assigns each client to associate with an
AP such that the channel airtime occupied by APs can be
minimized. The association control scheme still requires
each AP to transmit data at the transmission rate that can
reach the worst client associated with it, but attempts to
associate each client with an appropriate AP to minimize
the client heterogeneity. Our work differs from the associa-
tion control scheme along two axis. First, it focuses on the
video streaming application, and provides different clients
heterogeneous visual quality by assigning each video frame
a different transmission bit-rates based on its importance.
Second, our work considers a single-AP scenario where AP
cannot cooperate with each other; however, it can be
integrated with the above association control schemes to
enable each AP to provide its clients differential visual
quality after the optimal client-AP association.
SARM [14] and ARSM [15] further adapt the leader-
based approach to rate selection for video multicast. Based
on experiments, SARM constructs a SRN-PSNR table that
22 IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 12, NO. 1, JANUARY 2013
Fig. 1. Heterogeneity of Multicast Members. Users and 1 can receive
packets sent at 11 mb/s in the MAC layer, while user 1 can only receive
packets at the base-rate.
maps a SNR to a visual quality for each video content.
Hence, it can predict the rate that provides everyone a
minimum visual quality. ARSM proposes a variant, called
H-ARSM, to select another best node who has the best
channel condition. It then selects a bit-rate for the base layer
of a stream according to the channel quality of the worst
node, while transmits the enhancement layer at another bit-
rate based on the channel quality of the best node.
However, since such a two-level scheme does not consider
all multicast clients, it cannot maximize the overall video
quality for the entire group by differentiating the visual
qualities among the individual users based on their channel
conditions.
The work closest to ours is presented in [23], which
formulates the resource allocation problem as an optimiza-
tion model for video multicast in WiMax. Unlike [23] which
utilizes multiple subcarriers in a WiMax to broadcast a
video stream, we consider a single-channel wireless net-
work such as the infrastructure mode in 802.11, where
different frames in a group of pictures (GOP) need to contend
for the channel time before their playback time out and
must occupy various transmission times based on their
importance and the designated transmission bit-rate.
Furthermore, the work [23] formulates the theoretical model
that assumes all information, such as the channel condition
and the best bit-rate of each user, is given and, in addition,
considers a constant scenario, e.g., constant bit-rate video
and stationary topologies. Our goal, however, is to develop
an easy-to-implement protocol that considers the practical
issues, including how to measure the channel conditions,
how to detect channel variation, how to adapt to client
mobility, how to adjust the transmission schedule according
to variable video bit-rate, and how to tradeoff between the
online measurement overhead and system performance.
Another originality of our work is as follows: All the
above schemes predict the bit rate based on SNR of a link.
However, earlier work [24], [25] has shown that not only
may bit error and link quality depend on SNR, but are also
related to environmental factors such as multipath and
interference. To avoid this uncertainty, we exploit a sample-
based approach to measure the bit-rate that produces the
best visual quality for the heterogeneous clients, instead of
predicting a rate based on the theoretical SNR-BER
mapping function.
Finally, some protocols, such as APEX [26] and Softcast
[27], exploit physical-layer information to enable differen-
tial-quality video multicasting. Such a cross-layer design
requires the modification of hardware, and cannot be
deployed in existing WiFi APs. Our work, however, focuses
on developing a software-based solution that can be
realized in current WiFi devices.
3 RATE SCHEDULING MODEL
In this section, we formulate the rate scheduling problem as
a theoretical optimization model, and utilize a dynamic
programming algorithm to compute the solution using
oracle information. This model needs a high computational
complexity and complete information about the packet loss
probability of each wireless link. Our motivation is to use
the model as a reference to assess the performance of our
new technique, called QDM, introduced in the next section.
We will discuss their performance comparison in Section 5.
3.1 Problem Formulation
We consider an environment where the video servers
forward the stream to AP that can help broadcast data to all
multicast members. We assume that AP and the video
server are interconnected by wired networks, which are not
the bottleneck; hence, we only focus on transmission
between AP and multicast members. The rate scheduler
can be installed either in the video server that collocates
with the AP, as shown in Fig. 2, or in multiple proxy servers
that share the workload of scheduling. It can alternatively
be implemented as a driver run in the AP. The rate
scheduler, which is run in user space, can then notify the
network driver of AP, which is usually run in kernel, to
send each packet at the selected transmission bit-rate. Such
information exchange between the user space and the
kernel can be realized by socket programming.
We consider a group of multicast members ` over a
wireless network supporting a set of transmission bit-rates
1 fi
1
. i
2
. . . .g, where i
i
is a lower rate than i
,
if the index
i<,. We let j
i
i denote the packet loss ratio of the link
from the AP to member i 2 `, as the AP sends data at rate
i 2 1. We will describe how to estimate j
i
i in Section 4.1.
Suppose the frames of a video stream are partitioned
temporally into multiple groups of pictures, each of which
contains a set of frames. Let G be this set of GOPs. GO1
i
2
Gdenotes the ith GOP that includes 1 frames, i.e., GO1
i

f)
i
1
. )
i
2
. . . . . )
i
1
g. Based on the loss probability, the scheduler
can determine the rate schedule

)i
i
<)i
i
1
. )i
i
2
. . . . . )i
i
1
,
where )i
i
/
is the bit rate assigned to transmit frame )
i
/
, for
each GO1
i
such that the frames can be transmitted to the
receivers before the playback deadline T
GO1
i .
The goal of this paper is to find a rate schedule

)i for a
given video stream such that the maximal video quality can
be achieved for the multicast group, subject to the
constraint that each member must get at least the minimum
video quality 1o`1
iii
. Even though we formulate the rate
scheduling problem for a single video stream, the rate
scheduler can still solve the rate schedule for each of
multiple video streams separately by considering the other
streams as background traffic. We model the rate schedul-
ing problem as follows:
max 11o`1
toto|

X
i2`
11o`1i. 1a
subject to
LIN ET AL.: QUALITY-DIFFERENTIATED VIDEO MULTICAST IN MULTIRATE WIRELESS NETWORKS 23
Fig. 2. Framework architecture.
11o`1i ! 1o`1
iii
. 8i 2 `. 1b
X
)
i
/
2GO1
i
|ci

)
i
/
.
i
/
)i
i
/
T
GO1
i 1 c
/
. 8GO1
i
2 G. 1c
The constraint in (1b) ensures that each member i can
obtain a higher expected visual quality 11o`1i than
the minimum video quality 1o`1
iii
. The PSNR of a ii
video frame [17], [18] is computed as follows:
`o1
1
ii
X
i
i1
X
i
,1
1i. , 1i. ,
2
.
1o`1 10 log
10
1
2
ior
`o1

.
where 1 is the original frame, 1 is the distorted frame,
and 1
2
ior
is the maximum pixel value of the image. For each
GO1
i
2 G, frame )
i
/
2 GO1
i
is only useful if it is delivered
to the receivers successfully before the playback deadline
T
GO1
i . Besides, we let c
/
denote the proportion of time that
the channel is occupied by the background traffic; therefore,
the residual time available for transmitting GO1
i
is
T
GO1
i 1 c
/
. Since sending a frame that arrives after the
deadline does not help improve the video quality, the
constraint in (1c) determines whether each frame )
i
/
2
GO1
i
with a packet size |ci)
i
/
is tardy and should be
dropped; otherwise, this constraint selects a suitable
transmission rate for this frame. In (1c), .
i
/
is a binary
variable that indicates whether )
i
/
is dropped or not. That
is, .
i
/
1 if the AP can transmit )
i
/
at rate )i
i
/
; otherwise,
.
i
/
0 and )
i
/
must be dropped. This constraint ensures
that the AP does not transmit frames that cannot be
delivered to the group members on time. Given a feasible
solution of the rate schedule

)i and the loss probability
j
i
i for all i 2 `, the scheduler can compute the
expected PSNR 11o`1i and estimate the overall
video quality 11o`1
toto|
by
P
i2`
11o`1i. Finally,
the goal is to find the optimal rate schedule

)i

that can
maximize 11o`1
toto|
as shown in (1a).
3.2 Dynamic-Programming-Based Solution
Note that the model in (1) is a variation of the 0-1 knapsack
problem. Given a set of items (frames) in each GOP, the
scheduler must determine the number of frames to transmit
so that the total weight (transmission time) does not exceed
T
di))
T
GO1
i 1 c
/
T
iii
, where T
iii
is the time used to
send a subset of frames in each GOP at the rate of the worst
node in order to satisfy (1b); and, hence, T
di))
is the time
allowed to send the remaining frames in the GOP at a
higher rate to provide heterogeneous clients differentiated
video quality. Nevertheless, the difference between the rate
scheduling problem and the knapsack problem is that each
item (frame )
i
/
) has j1j different weights (transmission
time
|ci)
i
/

i
i
. 8i
i
2 1) and values (total incremental quality
Q)
i
/
. i
i
) because the number of clients who can receive
)
i
/
is determined by the rate i
i
used for the packet
transmission. Specifically, since each client i 2 ` can
receive a frame sent at rate i
i
with the probability j
i
i
i,
the expected value of Q)
i
/
. i
i
can be approximated by
P
i2`
)
i
/
j
i
i
i, where )
i
/
denotes the incremental
quality if a client successfully receives )
i
/
. )
i
/
can be
estimated by the rate-distortion function studied in [28],
[29]. In other words, sending a frame at a higher rate
spends less transmission time. However, this results in a
lower incremental quality Q for a multicast group
because only a few clients who have good enough channel
conditions can receive that frame.
Thus, the rate scheduling model must not only deter-
mine whether a frame should be transmitted, but also select
a suitable rate for that frame. Consequently, the rate
scheduling problem is at least as hard as the knapsack
problem, which is NP-complete, because the knapsack
problem is a special case of the rate scheduling problem as
j1j 1. In this work, we propose a novel dynamic
programming solution that extends the conventional
knapsack algorithms to solve the rate scheduling problem
in pseudopolynomial time. Our solution considers the
following two types of coding techniques:
1. Predictive coding: Predictive coding (e.g., MPEG4) is
a coding scheme that predicts a frame by referencing
a previously coded frame. Thus, if a reference frame
is not received correctly, those frames that reference
it cannot be decoded either. In general, the playback
order and the decoding order of the video frames
are not the same. To facilitate decoding, the
transmission order of the frames is the same as
their decoding order. We can sort the frames in
GO1
i
in the decoding order such that frame )
i
/
can
help improve the video quality by Q)
i
/
. i only if
both of )
i
/
and )
i
|
are reconstructed correctly for all
| < /. Hence, we let the video server send the first
/

frames of GO1
i
when T
di))
is only available for
transmitting /

frames. Because the number of

frames /

that can be sent before T

GO1
i depends
on the transmission time required for each frame,
which is determined by the transmission bit-rate
used to send that frame, we must select the optimal
/

and rates for each frame so as to maximize the

total video quality.
2. Multiple description coding (MDC): MDC encodes a
media stream into multiple independent substreams
(i.e., descriptions), any part of which can be decoded
independently. Receivers can get minimum quality
by decoding one arbitrary description, and achieve
incremental improvement by receiving additional
descriptions. For simplicity, we deem each descrip-
tion as a frame. That is, description )
i
/
2 GO1
i
can
help improve the video quality by Q)
i
/
. i if it is
delivered to the client correctly. Hence, we can
determine whether each frame should be trans-
mitted independently.
Algorithm 1 shows the proposed algorithm for two
coding schemes. To apply the dynamic programming
solution, we represent the transmission time of each frame
)
i
/
2 GO1
i
as a discrete variable, which is approximated
by d
|ci)
i
/

i
e if it is sent at rate i 2 1. Define Q/. t as the
maximum visual quality that can be obtained with weight
(time slots) less than or equal to tbT
di))
c by sending up to
24 IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 12, NO. 1, JANUARY 2013
/ frames. When using MDC, the solution is similar to the
dynamic programming method used to solve the knapsack
problem except that, as computing the value of Q/. t, it
must consider each frame )
i
/
by comparing j1j 1 alter-
natives, i.e., dropping it or sending it at j1j various rates,
instead of only 0-1 choices in the knapsack problem. It can
finally get that the maximal 1o`1
toto|
of the rate scheduling
problem equals Q1. bT
di))
c.
Algorithm 1. Dynamic programming solution to the rate
scheduling problem
1: Initialize Q/. t 0 for 0 / 1. 0 t bT
di))
c
2: for t 1 to bT
di))
c (Test each size of knapsack) do
3: for / 1 to 1 (Go through each frame )
i
/
) do
4: if Use MDC then
5: Q/. t maxfQ/ 1. t. Q

)
i
/
. i
i

Q/ 1. t d
|ci)
i
/

i
i
e : i
i
2 1. t ! d
|ci)
i
/

i
i
eg
6: else if Use predictive coding then
7: i arg
i
i
maxfQ

)
i
/
. i
i

Q/ 1.
t d
|ci

)
i
/

i
i
e:i
i
2 1. t !
d
|ci)
i
/

i
i
e. Q/ 1. t d
|ci)
i
/

i
i
e6 1g
8: if i 6 c then Q/. t Q

)
i
/
. i

Q/ 1. t d
|ci)
i
/

i
e; else Q/. t 1; endif
9: end if
10: end for
11: end for
12: if Use MDC then1o`1
toto|
Q1. bT
di))
c;
else 1o`1
toto|
max
0/1
Q/. bT
di))
c; endif
By contrast, in predictive coding, since frames in a GOP
are dependent on their reference frames, we must include
)
i
/
only if its reference frames )
i
|
(| < /) are also transmitted.
To enforce Algorithm 1 to select the first /

frames, as
computing Q/. t, we do not consider the alternative of
dropping the current frame )
i
/
(i.e., Q/ 1. t, which
indicates that it only includes the first / 1 frames, while
discards )
i
/
). This is because we can always benefit from
sending it if there is enough time to transmit it. In other
words, if there is still enough time to transmit some frames,
dropping the current frame ()
i
/
) and sending the following
ones ()
i
/
0 . /
0
/) does not help improve the video quality
due to the lack of the reference frame )
i
/
. Therefore, we set
Q/. t to 1 if the set of finding the maximum shown in
line 7 of Algorithm 1 is empty, which means that there is not
enough time to send )
i
/
no matter which rate i 2 1 is used.
Otherwise, as computing Q/. t, we only consider the
rates i
i
2 1 whose Q/ 1. t b
|ci)
i
/

i
i
c is not equal to 1,
and select the one that can produce the maximal incre-
mental quality. Because Q/ 1. t b
|ci)
i
/

i
i
c 1 means
that )
i
/1
is already excluded by Q/ 1. t b
|ci)
i
/

i
i
c, in
this case, we do not need to include )
i
/
as well. Finally,
unlike in the MDC scheme, since there is no guarantee that
Q/. t is always larger than or equal to Q/ 1. t in
predictive coding, 1o`1
toto|
is set to the maximum of
fQ/. bT
di))
c : 0 / 1g. That is, after solving each element
Q/. t, Algorithm 1 must finally find the number of frames
required to be transmitted so that 1o`1
toto|
can be
maximized.
The running time of Algorithm 1 equals OT
di))
1j1j.
The algorithm is polynomial only if T
di))
is polynomial.
Therefore, the proposed dynamic programming based
scheme can be solved in pseudopolynomial time. Since the
available transmission time, T
di))
, is limited, the algorithm
can be run with a low complexity.
4 QDM FRAMEWORK
The model in the last section is proposed to find the most
efficient solution, which needs complete information about
the loss probability j
i
i, and incremental quality of
each frame. However, to realize a practical protocol, we
must consider the following problems.
. How can the scheduler collect the visual quality
perceived by each receiver efficiently?
. How can the scheduler determine a suitable rate
schedule for real-time streaming if the information
about incremental quality is not computed by
preprocessing?
. How can the scheduler adapt the rate schedule to
network dynamics, such as channel variation, node
mobility, and variable video bit rates?
To cope with the above practical issues, we propose
QDM, a practical video multicast framework including
three components: 1) cluster construction: it clusters clients
according to their channel conditions in order to character-
ize the heterogeneity of clients; 2) sample-based rate
scheduling: it predicts the rate schedule by real-time
sampling, and, thus, can estimate visual quality even if
information about is not given; 3) two-stage rate
adaptation: using the finite-state machine, it adapts the rate
schedule to variable video bit rates and channel conditions,
and, at the same time, avoids the unnecessary sampling
overhead.
4.1 Cluster Construction
A key challenge to characterizing heterogeneous channel
conditions of multicast members is that simultaneous
feedback from all members causes collision. Even though
traditional delayed-based methods, which schedule time-
division-based feedback, can avoid collision, it is still time
consuming for collecting responses from all members.
Thus, the design of QDM is to separate members to
multiple clusters C such that the receivers with similar
channel conditions can be classified into the same cluster.
Specifically, we let ii denote the best rate of member
i 2 ` if it can receive the maximal throughput from AP at
ii. Those who have the same best rate i
i
2 1 can be
grouped into cluster C
i
fi : i 2 `. ii i
i
g. Hence, a
multicast group can be divided to j1j clusters, each of
which can select the one in C
i
who experiences the worst
channel condition as the cluster head (CH) of C
i
(denoted
by CH
i
) to represent all members in C
i
. The scheduler can
collect heterogeneous channel conditions from the CHs,
and reduce the overhead of information exchange sig-
nificantly. Based on the performance reported by the
LIN ET AL.: QUALITY-DIFFERENTIATED VIDEO MULTICAST IN MULTIRATE WIRELESS NETWORKS 25
representative nodes (i.e., CHs), the scheduler can approx-
imate the overall video quality of a multicast group by
~
1o`1
toto|

X
C
i
2C
1o`1CH
i
jC
i
j.
where jC
i
j is the number of members in cluster C
i
.
In order to collect information about jC
i
j and ii for
each i 2 `, AP can send a probe packet at each rate
i 2 1 before broadcasting the video stream. Based on
those probe packets, each member i can estimate the bit
error rate of each rate i 2 1, denoted by 111
i
i. An
100-byte probe packet for each rate is long enough to
capture the average bit error rate accurately. Each client
can then estimate the throughput of each rate based on [5],
which is computed by i 1 111
i
i
`
, where 1
111
i
i
`
is the probability that the `-bits probe packet
is received correctly without any bit errors. Each client can
then select its best rate ii that can produce the highest
throughput, i.e., ii arg max
i21
i 1 111
i
i
`
,
and responds a control message <i. ii. o`1
i
to
the rate scheduler, where o`1
i
is the average SNR
observed by i. The scheduler can then use the feedback
of control messages to construct the clusters, select the
CHs, and broadcast the information <i
i
. CH
i
. o`1
CH
i
to
each cluster C
i
2 C. Probe packets are only sent before
cluster construction, so the overhead is relatively smaller
than periodical channel estimation.
However, we note that the channel condition of each
client might change dynamically due to channel variation
or client mobility. To adapt the rate schedule to such
channel dynamics, the scheduler needs to update the
clusters as the channel condition of certain nodes changes
substantially. For a multicast application, it is however
difficult for the AP to monitor the channel condition of
every member all the time. Therefore, we let each member
i 2 C
i
measure the SNR value of its received data packets
(i.e., video frames), and issue a reclustering request to the
scheduler if it detects that its average SNR is smaller than
c o`1
CH
i
, where 0 c 1 is a sensitivity coefficient. As
a result, detecting channel variation can be done by each
client without any extra overhead. Once the scheduler
decides to perform reclustering, it executes the same
probing procedure as what it did in the initialization stage
such that the clients who experience channel variation can
update its channel condition to the scheduler. During the
probing period, the scheduler pauses the transmission of
video stream, and resumes the streaming after the new
transmission schedule is recomputed. We note that, since
we only need to perform clustering in the start-up stage or
when the network changes significantly, the overhead of
clustering should be negligible for the static environments
and reasonable for the dynamic environments.
2
4.2 PSNR-Guaranteed Quality-Differentiated Rate
Selection
We note that even though the solution proposed in Section 3
can provide an efficient rate schedule, it may only be suitable
when the information about is computed by preproces-
sing. Specifically, if the information about is given, we
can vary Algorithm 1 to solve the rate schedule and
maximize
~
1o`1
toto|
for the cluster-based framework. How-
ever, if the information about is not given, it is
challenging to online analyze the incremental quality
and compute 11o`1
toto|
for various rate schedules

)i as
solving the dynamic programming algorithm. Therefore,
instead of online computing incremental quality costly,
we propose QDM, a heuristic two-level rate scheduling
scheme that can adapt the rates dynamically based on a
sampling technique. The goal of QDM is to find a suitable
rate for each frame efficiently for real-time video streaming.
To reduce the complexity of rate scheduling, we sort the
frames of each GOP according to their decoding order, and
partition each GOP into three sets: essential frames (1
c
),
supplementary frames (1
:
), and discarded frames (1
d
), as
shown in Fig. 3, where variables i
c
, i
:
, and i
d
are the
number of frames in the sets 1
c
, 1
:
, and 1
d
, respectively. The
main idea is that, first, we tend to transmit the frames in 1
c
at rate 1
c
to all i 2 `, so that each member can achieve at
least the minimum visual quality 1o`G
iii
. We note that 1
c
must be set to a fixed rate min
i2`
ii because the essential
frames should be received by every member i 2 `.
Second, the frames in 1
:
can be sent at a higher rate (say
i
:
) in order to provide differentiated quality for hetero-
geneous clients. Finally, the rest of frames are put in 1
d
and
must be discarded if the available bandwidth is not enough
to send every frame in a GOP. Therefore, we can relax the
original problem shown in (1) as the following one:
max
~
1o`1
toto|
max
X
C
i
2C
1o`1CH
i
jC
i
j. 2a
subject to
1o`1i ! 1o`1
iii
. 8i 2 `. 2b
X
)21
c
|ci)
1
c

X
)21
:
|ci)
i
:
T
GO1
i 1 c
/
. 8GO1
i
2 G. 2c
In the relaxed problem, the goal is to find the required
number of essential frames (i
c
) and the transmission rate of
supplementary frames (i
:
) such that the approximate overall
visual quality
~
1o`1
toto|
can be maximized. Equation (2c)
bounds the time used to transmit the frames in each GOP up
to the available transmission time T
GO1
i 1 c
/
. Once the
scheduler determines i
c
to satisfy 1o`1
iii
, as shown in
(2b), it computes the time used to send the essential frames,
which equals
26 IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 12, NO. 1, JANUARY 2013
2. For 802.11b, the overhead of sending 100-byte probe packets equals
1,1 1,2 1,5.5 1,11 800 10
6
1.42 ms. Even if probing is per-
formed periodically per GOP, the overhead is about 2.8 percent if each GOP
is 0.5 second long. The response from multicast clients might need a higher
overhead. However, we note that only the first clustering in the start-up
stage requires every client to report its channel condition; for reclustering,
only the clients that experience channel variation need to report their
updated channel conditions to the scheduler. Thus, the response overhead
is relatively small for reclustering.
Fig. 3. Three blocks in each GOP.
T
iii

P
)21
c
|ci)
1
c
.
where |ci) is the number of bits in frame ). The AP can
then send as many supplementary frames as possible at rate
i
:
within the residual available time T
di))
T
GO1
i 1
c
/
T
iii
. Specifically, since each supplementary frame ) 2
1
:
takes |ci),i
:
(seconds) for transmission, the AP can
send at most i
:
(i.e., j1
:
j) frames such that
P
)21
:
|ci)
i
:
T
di))
.
Finally, we discard the rest of frames (i.e., 1
d
) because
there is no time to transmit any data before the playback
deadline T
GO1
i .
Instead of solving the optimal solution of i
c
and i
:
directly, we propose a sample-based scheme that enables
the scheduler to explore the suitable value of i
c
and i
:
based on feedback from the cluster heads. As Fig. 4 shows,
we divide a video stream into multiple sampling intervals,
each of which contains more than three GOPs. For each
sampling interval, the scheduler sends 1
:
of some GOPs at a
bit-rate other than the current i
:
to gather information about
other rates. Specifically, we let the first three GOPs of a
sampling interval be the samples that help the scheduler
figure out a suitable rate i
:
. The scheduler sets the
transmission bit-rate of 1
:
in GO1
1
, GO1
2
, and GO1
3
at
the current rate i
:
, the next higher rate i
0
:
i
iidcri
:
1
, and
the next lower rate i
00
:
i
iidcri
:
1
, respectively, where
iidcri
:
is the index of the current rate i
:
, e.g., iidcri
:

1 if i
:
i
1
. We only send three sample GOPs, i.e., sampling
the next higher rate and next lower rate, such that the
system can conservatively adapt the transmission bit-rate
and gradually converge to the optimal solution.
Instead of requesting each cluster head to compute the
average visual quality of all clients in the same cluster, we let
it respond its own reception quality to the scheduler. The
rational behind this design is that each client in the same
cluster has a similar channel condition and thereby could
receive a comparable visual quality. Therefore, we can use
the reception quality of the cluster head to estimate that of
other clients in the same cluster. To be more conservative
about the estimation reported by the cluster heads, we select
the client that has the worst channel condition in a cluster to
act as the representative cluster head who would likely to
have the worst visual quality in the corresponding cluster.
To this end, upon receiving the sample GOPs, each CH
responds the scheduler a mask, which is an array of binary
variables where the ith bit equals 1 if the CH receives the ith
packet of a GOP correctly, and 0 otherwise. We use each bit
in the mask to acknowledge the reception of each packet,
instead of each frame, because a frame may be segmented
into several packets. The scheduler can compute
~
1o`1
toto|
of each sample GOP by decoding the stream with the packet
losses pattern logged in the masks reported by CHs and
comparing the decoded GOP with the original stream. We
believe that the scheduler has strong enough computation
capability for real-time decoding few sample GOPs re-
ported by j1j representative CHs. By comparing
~
1o`1
toto|
of three sample GOPs, the scheduler can set i
:
to the rate
used by the sample GOP that achieves the highest
~
1o`1
toto|
(called GO1
/c:t-:oij|c
for short). For example, if the channel
rate of GO1
/c:t-:oij|c
is i
0
:
, the scheduler updates i
:
i
0
:
. In
addition, if any of CHs cannot achieve 1o`1
iii
in
GO1
/c:t-:oij|c
, it means that the current number of essential
frames (i
c
) is not enough to provide everyone 1o`1
iii
.
Therefore, the scheduler will increase i
c
by 1 (i.e.,
i
c
i
c
1) in order to improve the basic video quality
for each multicast member. After the sampling stage (i.e.,
the first three sample GOPs), the scheduler can use the
updated i
:
and i
c
for the rest of GOPs in the current
sampling interval. The whole procedure repeats for each
sampling interval.
4.3 Two-Stage Rate Adaptation
Since the sampled rate (i.e., a higher and a lower rate of the
current i
:
) may not be suitable for the multicast group,
QDM should avoid unnecessary samples if both of the
network topology and the video rate are static. Specifically,
the scheduler must adjust the sampling interval so as to
adapt the rate to network dynamics and variable video bit
rates; however, it must keep the selected rate unchanged if
the environment is stable. Here, we propose a two-state rate
adaptation scheme, which is a finite-state machine designed
to determine the size of the sampling interval that can
reflect variation of channel conditions or video bit rates.
As Fig. 5 shows, there are two states and two transition
functions in the proposed two-state machine. We expect
that the system stays in the active state if it still needs to
search a suitable rate due to unstable environments (i.e.,
network dynamics or variable video rates); otherwise, the
system can stay in the static state, and use the currently
selected i
:
and i
c
. The functions of two states and two
transitions are specified as follows:
. Active state: The system stay in the active state if the
clients are mobile or the video bit-rate varies with
time. In this state, we repeat the sampling procedure
for each sampling interval, which is set to a fixed
size (e.g., set to six GOPs in our simulations). Hence,
the system can sample a better rate periodically.
. Active state ! static state: If the system selects the
same rate i
:
for / (set to 2 in our simulations)
continuous sampling intervals, it then switches from
the active state to the static state because the selected
LIN ET AL.: QUALITY-DIFFERENTIATED VIDEO MULTICAST IN MULTIRATE WIRELESS NETWORKS 27
Fig. 4. Rate selection by sampling.
Fig. 5. Two-state rate adaptation.
rate works well for the current environment. As the
transition occurs, the system records the PSNR of
GO1
/c:t-:oij|c
of the last sampling interval in the
active state as 1o`1
octi.c
CH
i
for each CH
i
.
. static state: The system stops sampling, and uses the
selected i
:
and i
c
for the following GOPs.
. static state ! active state: When the system is in the
static state, it notifies the cluster heads to report the
mask of each GOP such that it can track the visual
quality of each CH. Once the system detects that any
CH gets a 1o`1CH
i
less than 1o`1
octi.c
CH
i

, where is a constant set to 10 dB in our
simulations, the system switches from the static
state back to the active state.
Even though the cluster heads must respond the mask
more frequently, however, the size of a mask (i.e., r bits for
r packets) is much smaller than the size of supplementary
frames 1
:
in a sample GOP. If the scheduler samples an
inappropriate i
:
during the sampling procedure, receivers
may not achieve an adequate video quality. In this case,
transmission of 1
:
in sample GOPs at an inappropriate rate
can be deemed as a sampling overhead. That is, the
scheduler may waste a significant sampling overhead if it
executes the sampling procedure when the environment is
stable and suitable to use the current rate i
:
. Thus, we
prefer to let the system enter the static state, which may
need a few overhead of responding the masks, but
significantly save the sampling overhead. By controlling
the size of the sampling interval, the two-state machine can
adapt i
:
to network dynamics with a reasonable sampling
overhead.
5 SIMULATION-BASED EVALUATION
We first conduct simulations to evaluate the performance of
QDM in various scenarios using simulations, and then
check the feasibility of QDM using trace-based implementa-
tion in the following section. The simulations are imple-
mented in NS2 [30], which is modified to support multiple
rates. In each simulation, AP is deployed in the center of the
topology as the video source; 30 clients are uniformly
distributed at random over the square field topology. Each
client is equipped with an 802.11b interface, which supports
rates 1, 2, 5.5, and 11 Mb/s. Since the default NS2 does not
consider bit error of each packet transmission, we use the
802.11b PHY Simulink Model [31] as the channel error
model, and set the transmission power of each node to
10 dB. We use JM15.1 to encode and decode the H.264 video
streams, each of which has 15 frames in each GOP and a
duration of 30 seconds. Each video is encoded as a variable
bit-rate stream with the average video bit rate 300 Kbps. We
test several video sequences, including forman qcif.yuv,
akiyo qcif.yuv, and grandma qcif.yuv, each of which is
encoded as the QCIF format at a video frame rate of
30 (frames/s). The background traffic is set to 750 Kb/s, so
the available bandwidth is not enough to multicast every
frame at the base rate, i.e., 1 Mb/s. 1o`1
iii
is set to 30 dB.
The sensitivity coefficient c used to recluster clients is set to
70 percent. In the two-state adaptation scheme, the thresh-
old of duplicated samples / is set to 2, and is set to 10 dB.
The default sampling interval (in the active state) is six
GOPs. All figures represent the average results over
10 random topologies.
We evaluate the performance of QDM in terms of
1o`1
toto|
and the CDF of PSNR perceived by each client.
We compare the following schemes:
1. Oracle: the dynamic-programming-based solution
with oracle information, as shown in Algorithm 1;
2. QDM: the proposed quality-differentiated multicast;
3. ARSM: the leader-based solution, which selects the
transmission bit-rate based on the worst node;
4. H-ARSM: the two-level solution, which selects the
rate for the base layer according to the channel
condition of the worst node, while selects the rate for
the enhancement layer based on the best node;
5. BASE-RATE: the baseline scheme that always sends
packets at the base rate.
5.1 Video Quality Comparison
We uniformly distribute 30 clients at random in a 100 m by
100 m field. Fig. 6 shows the CDF of PSNR perceived by
each client. Since the best rate of the worst node in this
scenario is the base rate (1 Mb/s), ARSM performs the same
with BASE-RATE. Thus, we exclude the results of BASE-
RATE from Fig. 6. In ARSM, because the stream is sent at
the fixed base rate, all clients can receive the same subset of
frames and get the same quality (about 32 dB). On the other
hand, since H-ARSM sends the base-layer and the enhance-
ment-layer at two different transmission bit-rates, it can
provide a better video quality for the clients locating nearby
AP. However, the side effect is that it instead penalizes
most of clients who cannot receive the enhancement layer,
which is sent at a higher rate (i.e., the transmission bit-rate
of the best node). Therefore, in H-ARSM, 47 percent of
clients cannot achieve 1o`1
iii
(i.e., 30 dB), and, even
worse, get lower PSNR than those in ARSM. Unlike H-
ARSM that produces two extreme qualities (extremely high
and low PSNR), QDM can exploit the cluster-based
structure to characterize the heterogeneity of clients, and,
thereby, provide each user a visual quality matching its
channel condition. Besides, almost all clients in QDM can
achieve 1o`1
iii
.
In Figs. 7a, 7b, and 7c, we compare the performance of all
schemes under three different node distributions, i.e.,
uniform, normal, and zipf, respectively. The zipf distribu-
tion is defined as 120 i 1 di:t 20 i
c
i
, for
i 1. 2, . . . , 1, where di:t is the distance between the client
and AP, c
P
1
i1
1
i

1
, and 1 is the network size (defined
28 IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 12, NO. 1, JANUARY 2013
Fig. 6. CDF of PSNR in uniform distribution.
as the radius of a topology) divided by 20 m. The
characteristic of the zipf distribution is that the probability
of placing a client nearby AP is higher than that of placing it
far from AP; hence, it can represent an environment where
most of clients are close to AP, while only a few clients are
distant from AP and experience a worse channel condition.
On the other hand, the normal distribution means that the
distance between each client and AP follows the normal
distribution, which means that only few clients are
extremely close to or far from AP.
In the uniform and zipf distributions, we vary the
network size from 20 m to 160 m; in the normal distribution,
we fix the mean of the network size to 60 m and vary the
standard deviation from five to 40 since the maximal radius
cannot be known in the normal distribution. We note that,
in all distributions, BASE-RATE provides clients similar
video qualities because AP always sends the stream at the
base rate without considering various channel conditions of
clients. On the other hand, ARSM can outperform BASE-
RATE when the network size is small because it could select
the rate according to the channel condition of the worst
node. Hence, when the network size is small, the rate used
by the leader could be higher than the base rate. Thus,
ARSM allows clients to receive a higher throughput and
video quality as compared to BASE-RATE. However, when
the network size grows, the sender will transmit at the
lowest base-rate even if there is only a single client, the
worst client, that can only receive at the base-rate. In this
case, ARSM performs exactly the same with BASE-RATE
when the network size is larger than 80 m.
Even though H-ARSM notices the heterogeneity of
clients, it enables clients using the same rate with the best
node to receive a better visual quality by penalizing other
clients. Such a two-level strategy could be suitable for the
zipf distribution, where most nodes gather nearby the
sender. However, it may perform worse than BASE-RATE
when most of nodes locate between the best node and the
worst node, such as in the normal or uniform distribution.
By contrast, the proposed QDM can provide a higher
average PSNR than other schemes no matter in which node
distribution. This is because the cluster-based structure can
capture the heterogeneity of clients, i.e., the number of
clients in each cluster; hence, it can select an appropriate rate
to transmit each video frame, and maximize video quality
for each cluster. Besides, unlike H-ARSM, which does not
consider the minimum video quality requirement, QDM
first provides each node a minimum video quality, and then
attempts to improve the total PSNR. Hence, the worst
nodes in QDM do not suffer from the starvation problem.
Finally, the proposed protocol, QDM, can provide
almost the same visual quality as the oracle solution when
the network size is smaller than 100 m. It indicates that the
scheduler can use the efficient sampling mechanism to
figure out the appropriate transmission bit-rates for each
individual frame. The reason of the small gap between
QDM and the oracle solution is because of the estimation
error of the incremental quality used in the oracle
solution. Specifically, the dynamic programming solution
considers the influence of missing each individual frame
based on its incremental quality . However, such an
incremental quality can only be estimated by existing
distortion functions. In addition, the incremental quality
of each frame can only be used as the first-moment
estimate; that is, the actual video quality is not exactly
equal to the summation of the incremental quality of all
received frames. As a result, the performance of the
dynamic programming-based solution is determined by
accuracy of the estimate, . By contrast, our cluster-
based algorithm applies the sampling mechanism to real-
time compute the decoded visual quality of the represen-
tative cluster heads, and adjusts the transmission schedule
based on online feedback. Therefore, it can avoid the
estimation error of and, in a few cases, can slightly
outperform the dynamic-programming-based solution.
The performance gap increases slightly when the net-
work size grows and, in turn, the heterogeneity of clients
becomes more obvious. This is because the cluster-based
method used in QDM can only select representative cluster
heads to estimate channel quality of all users. By contrast,
the dynamic-programming-based algorithm proposed in
Algorithm 1 can use the oracle information to explicitly
calculate visual quality of each multicast member and solve
the rate schedule explicitly. However, we note that the
performance gap between the oracle solution and QDM is
at most 1.5 dB. Another observation is that the oracle
solution estimates incremental visual quality based on
the existing rate-distortion functions, which still have some
limitation to reflect real visual quality accurately. There-
fore, it sometimes performs worse than QDM because
QDM inspects the actual visual quality by computing
PSNR based on the acknowledgement responded by CHs.
5.2 Impact of Rate Adaptation
Next, we evaluate the impact of rate adaptation on the
performance of QDM. In this simulation, 30 nodes are
LIN ET AL.: QUALITY-DIFFERENTIATED VIDEO MULTICAST IN MULTIRATE WIRELESS NETWORKS 29
Fig. 7. Impact of node distribution.
uniformly distributed in a 140 m by 140 m field. We let
location of clients and background traffic keep static.
Figs. 8a and 8b compare the performance of QDM with
adaptation and that of QDM without adaptation in terms of
the average PSNR of a multicast group and the PSNR of the
worst node, respectively. Adaptation means that QDM
adapts the rate dynamically for each sampling interval; no
adaptation means that AP uses the fixed rate selected in the
start-up stage without updating the rate afterward. We plot
the results of first 20 seconds of a video stream, and
illustrate how video quality varies with time.
We note that even though the environment is stable
(i.e., static nodes and background traffic), the average PSNR
in Fig. 8a and the PSNR of the worst node in Fig. 8b still
degrade because the video bit rate may vary with time. The
rate selected in the initial sampling intervals is only suitable
for the video bit rate in that period, but may not provide an
adequate video quality when the video bit rate changes.
Specifically, when the video bit rate increases, the scheduler
may need to increase the number of essential frames (i
c
) in
order to provide a minimum video quality 1o`1
iii
for
every member. Hence, QDM can produce a better video
quality than no adaptation and provide the worst node at
least an adequate quality because it adapts the rate to
variable video bit rates even when the wireless environment
is static. We note that sometimes the PSNR of the worst
node is still lower than 1o`1
iii
. This is because the video
bit rate in some periods is burst so that the available
bandwidth is not enough to deliver all essential frames 1
c
used to satisfy 1o`1
iii
. Sometimes, adaptation might
perform worse than no adaptation, e.g., at 12th second,
because it is unnecessary to waste the sampling overhead
when the initial rate happens to be suitable for the current
video rate. However, we note that, for most of the time,
adaptation can gain higher visual quality than no adaptation
because the bit-rate of a video stream usually fluctuates.
5.3 Impact of Reclustering
We next examine whether the system can gain from
reclustering. We consider a scenario that clients are mobile
based on three different mobility patterns. First, we
distribute nodes uniformly in the topology, and let each
30 IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 12, NO. 1, JANUARY 2013
Fig. 9. Impact of reclustering.
Fig. 8. Impact of rate adaptation.
node move according to the random waypoint model [32].
Fig. 9a presents the performance of various mean mobility
velocities, with the standard deviation 0.5. The figure does
not include the results of the dynamic-programming-based
solution because it can only be solved in preprocess, and,
thus, cannot adapt to dynamic scenarios. In this simulation,
we tend to evaluate the effect of reclustering clients as the
topology changes due to the client movement. The figure
shows an interesting phenomenon that video quality does
not decline much if QDM does not recluster the clients
when they move to the new locations. This is because, in
the random waypoint model, the node distribution during
the movement is similar to the original one. That is, each
node moves to a new location selected randomly from the
topology, so the locations of clients during/after movement
still follow the uniform distribution. Thus, the number of
members in each cluster C
i
2 C does not change much. This
is the reason why AP can still estimate
~
1o`1
toto|
correctly
based on the original cluster sizes jC
i
j, even if it does not
recluster the clients.
We consider another mobility pattern that the network
size expands from60 m 60 mto 150 m 150 m. We let each
node move away from AP so that the number of clients in
eachcluster couldchange withtime. Whennodes move away
fromAP, there may be more members that canonly receive at
a lower rate. Thus, without reclustering, AP may get an
incorrect estimate of
~
1o`1
toto|
by
P
C
i
2C
1o`1CH
i
jC
i
j
because jC
i
j has changed when the network size grows. The
CDF of PSNR perceived by each member in Fig. 9b indicates
that, without reclustering, the AP keeps using the out-of-date
rate such that only few nodes staying nearby AP can get
satisfactory PSNR. Those members moving away from AP
cannot receive data sent at the original rate anymore, and
thereby perceive a degraded video quality. However, by
reclustering the clients, AP can characterize the up-to-date
heterogeneity of clients after movement, andadapt the rate to
producing an adequate quality under mobile environments.
Besides, QDM with reclustering achieves the average PSNR
33.18 dB, which is higher than that without reclustering.
Finally, we consider an extreme scenario where the
clients are originally uniformly distributed with a distance
to the AP smaller than 50 m. During the streaming period,
the cluster heads do not move, while the other clients move
away from the AP. Such a scenario occurs in a conference or
a hotspot setting when attendees move from one place to
another. We let the distribution of the mobile clients speed
follow the uniform distribution. We then check the gain of
reclustering when the mean moving speed of the mobile
clients varies from 1 to 20 m/s. The speeds are selected to
simulate different mobile scenarios, such as 1 m/s for the
walking speed, 5-10 m/s for the running speed, and 20 m/s
for the driving speed. We notice that 20 m/s might not be a
reasonable setting for the conference scenario, but is only
used to test the limitation of our protocol. Moreover, to
prevent a user from moving quickly to the area out of the
transmission range, we let all the users keep static for the
first 10 seconds, and start moving at the 10th second. If
the users move to the boundary of the transmission range,
they stop moving and keep staying in the same location.
The PSNR value is computed over the entire video
streaming duration.
Fig. 9c shows that enabling reclustering produces a
better visual quality because, even if the cluster heads are
stable, the mobile clients can trigger reclustering when their
SNRs change significantly. The figure also shows that the
gain of reclustering is not obvious when the moving speed
is slow because the structure of clusters does not change
much during the streaming period. Reclustering however
plays a key factor that improves the visual quality when the
moving speed increases to 10 m/s, resulting in a smooth
change in the cluster structure. When the speed keeps
increasing up to 20 m/s, the gain of reclustering is limited.
The reason of this limitation is two-fold: 1) when the
structure of clusters changes too fast, the cost of recluster-
ing could offset the gain of reclustering; 2) more impor-
tantly, when the moving speed becomes 20 m/s, all clients
could move to the region where only the lowest rate is
available. As a result, there is no need to trigger reclustering
because most of clients, except cluster heads, can only
receive at 1 Mb/s. In this case, the system cannot gain from
reclustering.
We next investigate how the value of the triggering
threshold (c) affects the overhead of reclustering in the
conference scenario, i.e., the third mobile scenario, with the
mean speed 10 m/s. We then change the value of c from 0.1
to 0.7. Fig. 9d indicates that when the threshold is set to a
too small value, the system cannot quickly adapt to network
changes. On the contrary, when the threshold increases, the
cost of performing the probing procedure for reclustering
could occupy a few more airtime and, thus, slightly sacrifice
the transmission opportunities for the useful video stream.
To adapt to network dynamics, setting c to a value between
0.4 and 0.7 is a reasonable setting.
5.4 Impact of Two-State Adaptation
Table 1 shows the overhead of QDM based on the two-state
machine and periodical sampling (called one-state for short
due to always in the active state). We mentioned in Section 4
that the sampling overhead is much higher than the
feedback overhead. Hence, by applying the two-state
machine, QDM can adapt the size of the sampling interval
dynamically, and reduce the sampling overhead from 16.21
to 6.9 percent. Fig. 10a also shows that PSNR in the one-state
scheme sometimes drops, e.g., at 26th and 44th seconds,
because the scheduler may waste some unnecessary sample
GOPs to test unsuitable rates. On the contrary, the proposed
two-state mechanism can avoid this situation by letting the
system switch to the static state when the environment is
stable. Thus, in the two-state scheme, AP can send the
stream at the previous selected rate, which is suitable for the
current environment, instead of testing the wrong rates for
sample GOPs. Besides, for most of the time, the two-state
scheme outperforms the periodical sampling scheme. This
LIN ET AL.: QUALITY-DIFFERENTIATED VIDEO MULTICAST IN MULTIRATE WIRELESS NETWORKS 31
TABLE 1
Overhead Comparison (Video Size: 8.8 Mb)
means that the proposed two-state machine can assist the
scheduler in distinguishing whether the environment is
dynamic, and allow the system to switch to the correct state.
Otherwise, the two-state mechanism may produce a worse
quality than periodical sampling if it keeps the system
staying in a wrong state.
We next examine the effect of two-state adaptation in a
mobile scenario, where clients have the same uniform
mobility pattern with the scenario shown in Fig. 9a. Fig. 10b
plots the comparison between two-state adaptation and
periodical sampling in such a mobile scenario. The figure
shows that, when the environment is dynamic, the sender
should stay in the active state to keep tracking the best
transmission bit-rate. Our two-state adaptation scheme can
detect the network dynamics automatically and tune itself
to operate similar to periodical sampling. As a result, it
can perform as well as periodical sampling in a dynamic
environment.
6 EXPERIMENTAL EVALUATION
We use real traces to examine the performance of QDM.
Experimental setup. We implement a trace-based test
environment to evaluate the performance of QDM in real
wireless channels. To do so, we deploy a sender at the
location marked as X, as shown in Fig. 11, and put a
receiver at one of the 30 marked locations in Fig. 11. For
each receiving location, we make the sender transmit a
stream of 1,000-byte packets. The packet stream cycles
between transmitting at each of the 802.11b rates, i.e., 1, 2,
5.5, and 11 Mb/s. Each packet is marked with a sequence
number. The receiver logs the packets as a trace file using
Wireshark [33]. We offline process the traces, and produce a
combo-mask that represents the actual packet loss pattern
of using each different rate on the air. For each location, we
repeat the experiment to collect 10 traces. We then use each
trace to compute the best bit-rate that can achieve the
highest throughput for each marked location. Given the
best rate of each different location, we feed the correspond-
ing combo-mask to our program that performs our
clustering-based algorithm and compute the total PSNR
based on the packet loss pattern of the real traces.
Results. For each trace-based evaluation, we randomly
assign / multicast clients to / randomly selected locations
marked in Fig. 11, where / varies from 5 to 25. For each
selected location, we run our trace-based evaluation 10 times
using 10 different traces. The results are the average over
20 random location assignments. Fig. 12a shows the PSNR
comparison between our QDM and the previous schemes,
such as ARSM and H-ARSM. The figure shows that our
QDM enables heterogeneous visual quality and therefore
improve the average PSNR. Fig. 12b plots the CDFs of PSNR
when there are 25 clients. The figure shows a trend similar
to our simulation results. This figure also indicates that, in
H-ARSM, most of clients that cannot receive at the highest
rate cannot receive part of video frames and thereby cannot
get a satisfactory visual quality. This is also why H-ARSM
produces a lower average PSNR than ARSM.
32 IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 12, NO. 1, JANUARY 2013
Fig. 11. Testbed. Dots refer to client locations. Mark X refers to sender
location. Fig. 12. Trace-based evaluation.
Fig. 10. Impact of two-state rate adaptation.
7 CONCLUSIONS
This paper investigated the rate scheduling problem of
video multicast for heterogeneous clients over wireless
environments. We have proposed a rate scheduling model
that solves the theoretical optimal solution by dynamic
programming. A practical protocol, called QDM, was
further presented for real-time video streaming even with-
out preprocess on computing the rate-distortion function
and estimating the loss probability of each wireless link. In
QDM, we exploit a cluster-based structure to provide
differentiated qualities for heterogeneous clients. Based on
the information reported by cluster heads, the sender can
estimate the total video quality and explore a suitable rate
for each video frame based on a sample-based scheme. The
performance evaluation shows that QDM can produce a
gain of 2-5 dB in terms of the average video quality as
compared to the leader-based scheme, while also guarantee
that each client perceives at least a minimum video quality.
In addition, the two-state rate-adaptation scheme in QDM
can adapt the rate to network dynamics and variable video
bit rates with a reduced sampling overhead.
ACKNOWLEDGMENTS
This work was partially supported by the National Science
Council of ROC under contract No. NSC99-2218-E-001-005
and NSC100-2221-E-001-005-MY2.
REFERENCES
[1] K.A. Hua and F. Xie, A Dynamic Stream Merging Technique for
Video-on-Demand Services over Wireless Mesh Access Net-
works, Proc. IEEE CS Seventh Ann. Conf. Sensor Mesh and
Ad Hoc Comm. and Networks (SECON), 2010.
[2] A. Kamerman and L. Monteban, WaveLAN-II: A High-Perfor-
mance Wireless LAN for the Unlicensed Band, Bell Labs Technical
J., vol. 2, no. 3, pp. 118-133, 1997.
[3] G. Holland, N. Vaidya, and P. Bahl, A Rate-Adaptive MAC
Protocol for Multi-Hop Wireless Networks, Proc. ACM MobiCom,
2001.
[4] B. Sadeghi, V. Kanodia, A. Sabharwal, and E. Knightly, OAR: An
Opportunistic Auto-Rate Media Access Protocol for Ad Hoc
Networks, Proc. ACM MobiCom, 2002.
[5] J. Bicket, Bit-Rate Selection in Wireless Networks, PhD
dissertation, Massachusetts Inst. of Technology, 2005.
[6] J. Camp and E. Knightly, Modulation Rate Adaptation in Urban
and Vehicular Environments: Cross-Layer Implementation and
Experimental Evaluation, Proc. ACM MobiCom, 2008.
[7] G. Judd, X. Wang, and P. Steenkiste, Efficient Channel-Aware
Rate Adaptation in Dynamic Environments, Proc. ACM MobiSys,
pp. 118-131, 2008.
[8] M. Vutukuru, H. Balakrishnan, and K. Jamieson, Cross-Layer
Wireless Bit Rate Adaptation, Proc. ACM SIGCOMM, 2009.
[9] Z. Lei and N.D. Georganas, Rate Adaptation Transcoding for
Video Streaming over Wireless Channels, Proc. IEEE Intl Conf.
Multimedia and Expo (ICME), 2003.
[10] S. Pal, S.R. Kundu, A.R. Mazloom, and S.K. Das, Video Rate
Adaptation and Scheduling in Multi-Rate Wireless Networks,
Networking: Proc. Sixth Intl IFIP-TC6 Conf. Ad Hoc and Sensor
Networks, Wireless Networks, Next Generation Internet, pp. 475-487,
2007.
[11] S. Lee and K. Chung, Combining the Rate Adaptation and
Quality Adaptation Schemes for Wireless Video Streaming,
J. Visual Comm. and Image Representation, vol. 19, no. 8, pp. 508-
519, 2008.
[12] J. Kuri and S. Kasera, Reliable Multicast in Multi-Access Wireless
LANs, Proc. IEEE INFOCOM, Mar. 1999.
[13] A. Basalamah, H. Sugimoto, and T. Sato, Rate Adaptive Reliable
Multicast MAC Protocol for WLANs, Proc. IEEE Vehicular
Technology Conf. (VTC Spring), May 2006.
[14] Y. Park, Y. Seok, N. Choi, Y. Choi, and J.-M. Bonnin, Rate-
Adaptive Multimedia Multicasting over IEEE 802.11 Wireless
LANs, Proc. IEEE Consumer Comm. and Networking Conf., 2006.
[15] J. Villalon, P. Cuenca, L. Orozco-Barbosa, Y. Seok, and T. Turletti,
Cross-Layer Architecture for Adaptive Video Multicast Stream-
ing over Multirate Wireless LANs, IEEE J. Selected Areas in
Comm., vol. 25, no. 4, pp. 699-711, May 2007.
[16] O. Alay, T. Korakis, Y. Wang, and S. Panwar, Dynamic Rate and
FEC Adaptation for Video Multicast in Multi-Rate Wireless
Networks, Mobile Networks and Applications (Online Version),
vol. 15, pp. 425-434, 2009.
[17] Objective Perceptual Multimedia Video Quality Measurement
in the Presence of a Full Reference, ITU recommendation J.247
(08/08), http://www.itu.int/rec/T-REC-J.247/en, 2012.
[18] S. Gringeri, R. Egorov, K. Shuaib, A. Lewis, and B. Basch, Robust
Compression and Transmission of MPEG-4 Video, Proc. ACM
Seventh ACM Intl Conf. Multimedia, pp. 113-120. 1999,
[19] D.-N. Yang and M.-S. Chen, Bandwidth Efficient Video Multi-
casting in Multiradio Multicellular Wireless Networks, IEEE
Trans. Mobile Computing, vol. 7, no. 2, pp. 275 -288, Feb. 2008.
[20] C. Hsu and M. Hefeeda, Flexible Broadcasting of Scalable Video
Streams to Heterogeneous Mobile Devices, IEEE Trans. Mobile
Computing, vol. 10, no. 3, pp. 406-418, Mar. 2011.
[21] A. Chen, D. Lee, and P. Sinha, Optimizing Multicast Performance
in Large-Scale WLANs, Proc. IEEE 27th Intl Conf. Distributed
Computing Systems (ICDCS), 2007.
[22] R. Chandra, S. Karanth, T. Moscibroda, V. Navda, J. Padhye, R.
Ramjee, and L. Ravindranath, DirCast: A Practical and Efficient
Wi-Fi Multicast System, Proc. IEEE 17th Intl Conf. Network
Protocols (ICNP), 2009.
[23] S. Deb, S. Jaiswal, and K. Nagaraj, Real-Time Video Multicast in
WiMAX Networks, Proc. IEEE INFOCOM, 2008.
[24] K. Leung, M. Clark, B. McNair, Z. Kostic, L. Cimini, and J.
Winters, Outdoor IEEE 802.11 Cellular Networks: Radio and
MAC Design and Their Performance, IEEE Trans. Vehicular
Technology, vol. 56, no. 5, pp. 2673-2684, Sept. 2007.
[25] D. Cox, Delay Doppler Characteristics of Multipath Propagation
at 910 MHz in a Suburban Mobile Radio Environment, IEEE
Trans. Antennas and Propagation, vol. AP-20, no. 5, pp. 625-635,
Sept. 1972.
[26] S. Sen, S. Gilani, S. Srinath, S. Schmitt, and S. Banerjee, Design
and Implementation of an Approximate Communication System
for Wireless Media Applications, Proc. ACM SIGCOMM, 2010.
[27] S. Jakubczak, H. Rahul, and D. Katabi, One-Size-Fits-All Wireless
Video, Proc. ACM SIGCOMM HotNets Workshop, 2009.
[28] J. Chakareski, J. Apostolopoulos, W.-T. Tan, S. Wee, and B. Girod,
Distortion Chains for Predicting the Video Distortion for General
Packet Loss Patterns, Proc. IEEE Intl Conf. Acoustics, Speech, and
Signal Processing (ICASSP 04), 2004.
[29] A. Wyner and J. Ziv, The Rate-Distortion Function for Source
Coding with Side Information at the Decoder, IEEE Trans.
Information Theory, vol. IT-22, no. 1, pp. 1-10, Jan. 1976.
[30] The Network Simulator NS-2, http://www.isi.edu/nsnam/ns,
2012.
[31] 802.11b PHY Simulink Model, http://www.mathworks.com/
matlabcentral/fileexchange/2262, 2012.
[32] D.B. Johnson and D.A. Maltz, Dynamic Source Routing in
Ad Hoc Wireless Networks, Mobile Computing, vol. 353, pp. 153-
181, 1996.
[33] Wireshark, http://www.wireshark.org, 2012.
LIN ET AL.: QUALITY-DIFFERENTIATED VIDEO MULTICAST IN MULTIRATE WIRELESS NETWORKS 33
Kate Ching-Ju Lin received the PhD degree
from National Taiwan University in 2009. She
was a visiting scholar in the Computer Science
and Artificial Intelligence Laboratory (CSAIL) at
MIT in 2007. After her graduation, she joined the
Research Center for Information Technology
Innovation at Academia Sinica, Taiwan. She is
now an assistant research fellow. Her current
research interests include wireless systems,
wireless mesh networks, and sensor networks.
She is a member of the IEEE.
Wei-Liang Shen received the BS degree from
National Central University, Taiwan, in 2008 and
the MS degree from National Taiwan University
in 2010. He is also working toward the PhD
degree in the Department of Electrical Engineer-
ing, National Taiwan University. He is currently a
research assistant in the Research Center for
Information Technology Innovation at Academia
Sinica, Taiwan. His current research interests
include multiuser MIMO systems and peer-to-
peer streaming.
Chih-Cheng Hsu received the MS degree from
National Cheng Kung University, Taiwan, in
2006 and the PhD degree from National Taiwan
University in 2012. After his graduation, he
joined the Computer Science and Information
Engineering Department at the National Taiwan
University and he is now a postdoctoral fellow.
His research interests include wireless sensor
networks, underwater sensor networks, P2P
computing, and cloud networks.
Cheng-Fu Chou received the PhD degree from
the University of Maryland, College Park. After
his graduation, he joined the Computer Science
and Information Engineering Department at the
National Taiwan University, where he is now an
associate professor. He was a visiting scholar at
the University of Southern California. His current
research interests are in peer-to-peer networks,
distributed multimedia systems, multi-hop wire-
less networks, sensor networks, and their
performance evaluation. He is a member of the IEEE.
> For more information on this or any other computing topic,
please visit our Digital Library at www.computer.org/publications/dlib.
34 IEEE TRANSACTIONS ON MOBILE COMPUTING, VOL. 12, NO. 1, JANUARY 2013