Implications of Smoothing on Statistical Multiplexing of H.264/AVC and SVC Video Streams

Abstract

While the hierarchical B frames based scalable video coding (SVC) extension of the H.264/AVC standard achieves significantly improved compression over the initial H.264/AVC codec, the SVC video traffic is significantly more variable than H.264/AVC traffic. The higher traffic variability of the SVC encoder can lead to smaller numbers of streams supported with bufferless statistical multiplexing than with the H.264/AVC encoder (and even less streams than with the MPEG-4 Part 2 encoder) for prescribed link capacities and loss constraints. In this paper we examine the implications of video traffic smoothing on the numbers of statistically multiplexed H.264 SVC, H.264/AVC, and MPEG-4 Part 2 streams, the bandwidth requirements for streaming, and the introduced delay. We identify the levels of smoothing that ensure that more H.264 SVC streams than H.264/AVC streams can be supported. For a basic low-complexity smoothing technique that is readily applicable to both live and prerecorded streams, we identify the levels of smoothing that give (bufferless) statistical multiplexing performance close to an optimal off-line smoothing technique. We thus characterize the trade-offs between increased smoothing delay and increased statistical multiplexing performance for both H.264/AVC, which employs classical B frames, and H.264 SVC, which employs hierarchical B frames. We similarly identify the buffer sizes for the buffered multiplexing of unsmoothed H.264 SVC, H.264/AVC, and MPEG-4 Part 2 streams that give close to optimal performance.

Full text

IEEE TRANSACTIONS ON BROADCASTING, VOL. 55, NO. 3, SEPTEMBER 2009 541
Implications of Smoothing on Statistical Multiplexing
of H.264/AVC and SVC Video Streams
Geert Van der Auwera and Martin Reisslein
Abstract—While the hierarchical B frames based Scalable Video
Coding (SVC) extension of the H.264/AVC standard achieves sig-
nificantly improved compression over the initial H.264/AVC
codec, the SVC video traffic is significantly more variable than
H.264/AVC traffic. The higher traffic variability of the SVC
encoder can lead to smaller numbers of streams supported
with bufferless statistical multiplexing than with the H.264/AVC
encoder (and even less streams than with the MPEG-4 Part 2
encoder) for prescribed link capacities and loss constraints. In this
paper we examine the implications of video traffic smoothing on
the numbers of statistically multiplexed H.264 SVC, H.264/AVC,
and MPEG-4 Part 2 streams, the bandwidth requirements for
streaming, and the introduced delay. We identify the levels
of smoothing that ensure that more H.264 SVC streams than
H.264/AVC streams can be supported. For a basic low-complexity
smoothing technique that is readily applicable to both live and
prerecorded streams, we identify the levels of smoothing that
give (bufferless) statistical multiplexing performance close to
an optimal off-line smoothing technique. We thus characterize
the trade-offs between increased smoothing delay and increased
statistical multiplexing performance for both H.264/AVC, which
employs classical B frames, and H.264 SVC, which employs hier-
archical B frames. We similarly identify the buffer sizes for the
buffered multiplexing of unsmoothed H.264 SVC, H.264/AVC, and
MPEG-4 Part 2 streams that give close to optimal performance.
Index Terms—Delay, H.264/AVC, hierarchical B frames,
smoothing, statistical multiplexing, SVC, video traffic.
I. INTRODUCTION
THE recently standardized Scalable Video Coding ex-
tension (SVC) of the H.264/AVC standard [1]–[3]
with its hierarchical B-frames compresses single-layer
(non-scalable) video significantly more efficiently than the
underlying H.264/MPEG-4 Advanced Video Coding stan-
dard [4] (H.264/AVC for brevity), which is also known as
H.264/MPEG-4 Part 10. H.264/AVC in turn compresses video
significantly more efficiently than MPEG-4 Part 2 (typically
only half the average bit rate with H.264/AVC for same video
quality). H.264/AVC and H.264 SVC video encoding are
expected to be widely adopted for wired and wireless network
Manuscript received May 28, 2008; revised April 20, 2009. First published
August 11, 2009; current version published August 21, 2009. This work was
supported in part by the National Science Foundation through Grants No. Career
ANI-0133252, ANI-0136774, and CRI-0750927.
G. Van der Auwera was with the Department of Electrical Engineering, Ari-
zona State University, Tempe, AZ 85287-5706 USA. He is now with Samsung
Information Systems America, Digital Media Solutions Lab, Irvine, CA 92612
(e-mail: geert.vanderauwera@asu.edu).
M. Reisslein is with the Department of Electrical Engineering, Arizona
State University, Tempe, AZ 85287-5706 USA (e-mail: reisslein@asu.edu;
http://www.fulton.asu.edu/mre).
Digital Object Identifier 10.1109/TBC.2009.2027399
video transport due to their increased compression efficiency
compared to MPEG-4 Part 2 and their widespread inclusion
in application standards and industry consortia specifications,
e.g., DVB, 3GPP2, and MediaFLO.
The compression efficiency of a video codec is generally
characterized with a so-called rate-distortion (RD) curve that
shows the bit rate of the compressed video stream as a function
of the video quality (distortion), which is typically measured in
terms of the Peak Signal to Noise Ratio (PSNR). For a given
video quality, the lower the compressed bit rate, the more effi-
cient is the compression. The improvements in rate-distortion
(RD) compression efficiency with H.264 SVC and H.264/AVC
come at the expense of significantly increased variabilities of
the encoded frame sizes (in bits) [5]. Highly variable video
frame sizes, i.e., highly variable video traffic, generally poses
a challenge for efficient network transport [6]–[8]. When the
video frame sizes are highly variable, i.e., when the largest
frames are much larger than the average frame size, then pro-
visioning network bandwidth according to the largest frames
results in inefficient bandwidth usage. The basic idea of sta-
tistical multiplexing is that the largest frames of some video
streams collude with average (or smaller than average sized)
frames of other streams during network transport. With this
statistical multiplexing, the bandwidth requirement is typically
dramatically less than the sum of the peak bit rates of the sup-
ported streams, and may approach the sum of the mean bit rates
of the supported streams. Consequently, statistical multiplexing
is of great interest for network systems transporting video with
variable frame sizes.
However, it was found in [9] that the H.264/AVC encoder can
outperform the H.264 SVC encoder and that even the MPEG-4
Part 2 encoder can outperform both the H.264/AVC and H.264
SVC encoders when multiplexing a small number of video
streams in an elementary bufferless statistical multiplexing
setting. This is due to significantly higher traffic variabilities of
H.264 SVC encoded video streams compared to H.264/AVC
encoded streams, as well as the significantly higher traffic
variabilities of both H.264 SVC and H.264/AVC encoded video
streams compared to MPEG-4 Part 2 encoded streams. The
higher traffic variabilities can compensate the lower average
bit rates achieved with H.264 SVC encoding compared to
H.264/AVC encoding, as well as the lower average bit rates
achieved by both H.264 SVC and H.264/AVC compared to
MPEG-4 Part 2.
In this paper we examine the effectiveness of two elementary
techniques for mitigating high traffic variability, namely (i)
video traffic smoothing, i.e., the averaging of several successive
frame sizes before sending them into the bufferless multiplexer,
and (ii) buffered multiplexing of unsmoothed video streams.
0018-9316/$26.00 © 2009 IEEE
Authorized licensed use limited to: Arizona State University. Downloaded on August 31, 2009 at 13:04 from IEEE Xplore. Restrictions apply.

542 IEEE TRANSACTIONS ON BROADCASTING, VOL. 55, NO. 3, SEPTEMBER 2009
From the wide spectrum of video traffic smoothing techniques
we consider two extreme approaches: optimal smoothing
[10], [11], which minimizes the traffic variabilities, and basic
smoothing, which simply averages (aggregates) the sizes of a
prescribed number of successive video frames, whereby the
number of averaged video frames is denoted by the aggrega-
tion level . Optimal smoothing achieves the minimal traffic
variability subject to given smoothing (receiver) buffer and
start-up delays by computing offline the transmission schedule
that delivers each video frame by its playout deadline while
avoiding overflows of the smoothing buffer and minimizing
transmission rate changes. Optimal smoothing has a computa-
tional complexity of , whereby denotes the number
of frames in the sequence and can not be directly applied to
live streams. In contrast, basic smoothing is computationally
very simple (has complexity ) and can directly be ap-
plied to live streams. For a range of numbers of statistically
multiplexed streams and video (texture/motion) complexities,
we provide guidelines for (i) setting the aggregation levels
of basic smoothing that ensure that more H.264 SVC streams
than H.264/AVC streams are supported, and (ii) setting the
aggregation levels that provide similar statistical multiplexing
performance with basic smoothing as with optimal smoothing.
We find that generally SVC requires larger aggregation levels
to overcome its higher traffic variabilities. We also examine
the delay introduced by the hierarchical B frame predictions in
H.264 SVC in conjunction with the aggregation levels for the
traffic smoothing and compare with the corresponding delays
for H.264/AVC.
We also examine elementary taildrop buffered statistical mul-
tiplexing of unsmoothed video streams. We identify the multi-
plexer buffer sizes required to support close to the maximum
number of streams (given by the link capacity divided by the av-
erage stream bit rate). We find that H.264 SVC streams require
roughly twice the buffer size of H.264/AVC streams, while in
turn H.264/AVC streams require approximately twice the buffer
size of MPEG-4 Part 2 streams.
This paper is structured as follows. In Section II, we review
related work. In Section III, we present our evaluation set-up,
including the examined H.264 SVC, H.264/AVC, and MPEG-4
Part 2 encoders and their settings, as well as the video sequences
used for the evaluations. In Section IV, we first describe the em-
ployed basic and optimal smoothing techniques and the consid-
ered bufferless statistical multiplexing setting. We then present
simulation results for optimal smoothing, followed by simula-
tion results for basic smoothing. In Section V, we first describe
the examined elementary buffered statistical multiplexing sce-
nario, and then present simulation results. We summarize our
conclusions in Section VI and analyze the delays for smoothed
transmission of video encoded with classical and hierarchical B
frames in the Appendix.
II. RELATED WORK
For MPEG-4 Part 2, H.263, and preceding codecs, the bit
rate-distortion characteristics and rate variability characteristics
have been extensively studied, see for instance [12]–[14] and
references therein. Similarly, the video traffic of these codecs
has been extensively studied, see for instance [15]–[19], and
they have been used as a basis for the existing studies on video
traffic smoothing, as reviewed in Section IV-A, and buffer man-
agement, as reviewed in Section V.
The bit rate-distortion characteristics of H.264/AVC and
H.264 SVC have been examined in a few studies [3], [4], [20]
and the rate variability characteristics of H.264/AVC and H.264
SVC have been investigated in [5], [9], [21]. The study of
network transport mechanisms for H.264/AVC and H.264 SVC
has just begun to attract interest, see for instance the studies
[22]–[27], all of which are complementary to our study exam-
ining the fundamental statistical multiplexing characteristics.
We note that the traffic characteristics of individual smoothed
H.264/AVC and H.264 SVC streams have been studied in [9];
furthermore, the bufferless statistical multiplexing of un-
smoothed H.264/AVC and H.264 SVC streams has been
examined in [9]. To the best of our knowledge, the fundamental
bufferless statistical multiplexing characteristics of smoothed
H.264/AVC and H.264 SVC video and buffered multiplexing
characteristics of unsmoothed H.264/AVC and H.264 SVC
video are for the first time examined in this paper.
III. EVALUATION SET-UP
A. Video Encoding Set-Up
We employ the H.264/AVC encoder [4], [20], [28]–[30] in
the Main profile with all compression tools enabled, including
spatial intra frame prediction, variable block sizes, three refer-
ence frames for the past and the future, referenced B frames,
P and B frame weighted prediction, Context Adaptive Binary
Arithmetic Coding (CABAC), and Lagrangian based rate-dis-
tortion optimization (RDO). In particular, we employ the JM
reference software (version 10.2), which is the official MPEG
and ITU reference implementation for the H.264/AVC Main
profile. For the H.264 SVC encodings, we used the SVC refer-
ence software named JSVM (version 5.9), and similar settings
as for H.264/AVC.
Throughout, we employ H.264/AVC with classical B frame
prediction, where a B frame is predicted only from the preceding
I or P frame and from the subsequent I or P frame; other B
frames are not referenced. In contrast, H.264 SVC [1]–[3] em-
ploys the hierarchical B frame structure which uses B frames
for the prediction of B frames, as illustrated in the Appendix.
More specifically, with the employed dyadic B frame hierarchy,
the number of B frames between successive key pictures (I or
P frames) is
(1)
of so-called temporal layers of B frames.
We use the MPEG-4 Part 2 encoder [31], specifically the
MPEG-4 Part 2 Microsoft v2.3.0 software, in the Advanced
Simple profile (ASP), which includes B frames. We employ half
pixel motion compensated prediction; RDO is not supported by
the reference encoder implementation. The MPEG-4 Part 2 en-
coder uses one reference frame for the past and one for the fu-
ture, and 16 16 blocks for motion estimation that can be split
into 8 8 blocks.
Authorized licensed use limited to: Arizona State University. Downloaded on August 31, 2009 at 13:04 from IEEE Xplore. Restrictions apply.

VAN DER AUWERA AND REISSLEIN: STATISTICAL MULTIPLEXING OF H.264/AVC AND SVC VIDEO STREAMS 543
For the H.264/AVC encodings and the MPEG-4 Part 2 en-
codings, which are both based on classical B frames, we em-
ploy GoP structure IBBBPBBBPBBBPBBB (16 frames, with 3
B frames per I/P frame) denoted by G16-B3. For the H.264
SVC encodings (hierarchical B frames), we employ GoP struc-
ture IBBBBBBBBBBBBBBB (16 frames, with 15 B frames per
I frame) denoted by G16-B15. The statistical video traffic anal-
ysis in [5], [9] demonstrated that these encoding parameter set-
tings and GoP structures result overall in very good rate-distor-
tion (RD) efficiencies for the respective encoders. The analysis
in [5] also indicated that encoding parameter settings that result
in lower RD efficiency generally reduce the traffic variability;
conversely, settings that further increase the RD efficiency gen-
erally increase the traffic variability (further increasing the need
for traffic smoothing). In addition, the considered GoP struc-
tures provide identical random access functionalities (I frame
period). We consider quantization parameters that correspond
to the range of average PSNR qualities from either 30/32 dB
(acceptable quality) or 35 dB (good quality) to at least 40 dB
(high quality).
Throughout this study, we consider single-layer (non-scal-
able) encoding and encode the video with fixed quantization
scales, which results in nearly constant video quality and vari-
able video traffic bit rates. By considering variable bit rate en-
coding without the use of rate control mechanisms we are able
to examine the fundamental traffic characteristics of the H.264
SVC and H.264/AVC video coding standards, which do not
specify a normative rate control mechanism. An additional mo-
tivation for the focus on variable bit rate video encoded with
fixed quantization scales is that the variable bit rate streams
allow for statistical multiplexing gains that have the potential
to improve the efficiency of video transport over communica-
tion networks [6].
B. Video Sequences
The five CIF (352 288 pixels) resolution video sequences
employed in the statistical multiplexing simulations presented
in this study are the ten minute Sony Digital Video Camera
Recorder demo sequence (17,682 frames at 30 frames/sec),
which we refer to as Sony Demo sequence, the first half
hour of the Silence of the Lambs movie (54,000 frames at
30 frames/sec), the first half hour of the Star Wars IV movie
(54,000 frames at 30 frames/sec), and the first hour of the To ky o
Olympics video (133,128 frames at 30 frames/sec). We also
use about 30 minutes of the NBC 12 News (49,523 frames at
30 frames/sec), including the commercials. These sequences
were obtained with the MEncoder tool through decoding the
original DVD sequences into the uncompressed YUV format
and subsampling to CIF resolution. The video sequences Si-
lence of the Lambs,Star Wars IV,Tokyo Olympics, and NBC 12
News can respectively be described as drama/thriller, science
fiction/action, sports, and news. The Sony Demo sequence
is documentary style, and is a mixture of detailed scenes
(textures) and various motion activities. The NBC 12 News
and Sony Demo videos have relatively higher motion and
texture complexity than the other three videos and pose more
challenges for statistical multiplexing as we demonstrate in
Section IV-C-1.
In order to facilitate further research on network transport of
H.264 SVC, H.264/AVC, and MPEG-4 Part 2 encoded video, all
encodings presented in this study are publicly available as video
traces from the video trace library at: http://trace.eas.asu.edu.
Frame size video traces [32] are files mainly containing video
frame time stamps, frame types (e.g., I, P, or B), encoded frame
sizes (in bits), and frame qualities (PSNR). Video traces are
employed in simulation studies of the transport of video over
communication networks, see e.g., [33]–[37], and as a basis for
video traffic models, as for instance in [12], [15], [16], [19],
[38]–[41]. Traffic modeling of H.264/AVC and H.264 SVC
video traffic is a nascent research area, see e.g., [21], [42]–[44],
and we directly employ the video traces for a realistic repre-
sentation of H.264 video traffic in our simulations. Generally,
advantages of using video traces over using regular encoded
bit streams in simulations are the availability of a large number
of traces of long and real video sequences, the fact that video
traces are not copyrighted, and that only knowledge of basic
concepts of video encoding are required.
IV. BUFFERLESS STATISTICAL MULTIPLEXING OF SMOOTHED
VIDEO TRAFFIC
A. Frame Size Smoothing
A wide variety of frame size smoothing mechanisms have
been developed and studied in the context of the MPEG-4
Part 2, H.263, and preceding video standards. Broadly, these
smoothing mechanisms can be classified into non-collabora-
tive mechanisms that smooth a single video stream, see for
instance [10], [45]–[56], and collaborative mechanisms that
jointly smooth several streams sharing networking resources,
see for instance [33], [57]–[64]. We focus on non-collaborative
smoothing in this study and leave evaluations of collaborative
smoothing for H.264 SVC and H.264/AVC for future work.
Among the non-collaborative smoothing mechanisms, we
first consider basic smoothing of the sizes (in bit) of the video
frames over non-overlapping blocks of frames each. More
specifically, for the aggregation level , the sizes of consec-
utive frames are averaged and transmitted at the corresponding
average bit rate. Given the original (unsmoothed) frame size
sequence , , we obtain the smoothed frame
sizes
(2)
for . The aggregation level can be varied, with
larger values resulting in lower video traffic variabilities at the
expense of increased delay, which is analyzed in the Appendix.
We also consider optimal smoothing [10], [11], which is op-
timal in the sense that it minimizes the bit rate variability and the
peak bit rate of the video traffic subject to prescribed smoothing
(receiver) buffer and start-up delays. Optimal smoothing en-
sures that the given receiver buffer does not underflow nor
overflow, while sending video frame bits ahead of the decoding
Authorized licensed use limited to: Arizona State University. Downloaded on August 31, 2009 at 13:04 from IEEE Xplore. Restrictions apply.

544 IEEE TRANSACTIONS ON BROADCASTING, VOL. 55, NO. 3, SEPTEMBER 2009
times of the corresponding video frames. The optimization al-
gorithm computes the transmission schedule of the video frame
bits in piecewise constant bit rate segments that are as long as
possible and have the smallest rate changes possible, without
overflowing the client buffer, and while delivering the video
frames by their playout deadlines. Optimal smoothing takes
as input the frame sizes of the pre-encoded video stream and
computes the transmission schedule off-line. With denoting
the number of video frames in a pre-encoded video sequence,
the computational complexity of a basic implementation of
optimal smoothing is (whereby a complexity reduction
to is possible with a more involved implementation)
[10]. For our simulations we set the client buffer size to 48 KB
and set the (additional) start-up delay (see Appendix) to .
The 48 KB buffer ensures that for the highest quality streams
in our experiments (approximately 40dB), the largest frames
can fit into the client buffer.
Although many more video traffic smoothing techniques are
available, we focus on basic smoothing and optimal smoothing,
because these two techniques represent extreme situations, i.e.,
lowest computational complexity ( with basic smoothing)
and lowest achievable rate variability (with optimal smoothing).
B. Bufferless Statistical Multiplexing
In the real-time frame-based video streaming scenario based
on a bufferless statistical multiplexer [56], [65]–[67], a channel
with bandwidth capacity [bit/s] connects a streaming video
server with a bufferless statistical multiplexer to receivers.
Each video frame is transmitted during one frame period (e.g.,
33 ms for a frame rate of 30 frames/s). Let [bit] de-
note the frame size of frame , , of stream ,
. Then, the bit rate required to transmit frame
of stream during one frame period of length is given by
. Let be a random variable denoting the index
of the frame of stream transmitted during frame period .
Then, the aggregated bit rate in frame period when statisti-
cally multiplexing all streams is given
(3)
If the aggregate bit rate exceeds the link capacity , then
loss occurs, which we measure as the information loss proba-
bility [66], [67], i.e., long-run fraction of lost video bits:
(4)
where . For a given experiment, we stream
identical video sequences, whereby the starting phase for each
stream is randomly selected according to a uniform distribution
over all frames of the sequence [32], [66]. The streams are
wrapped around to obtain streams of equal lengths.
Aside from providing an appropriate model for low-delay,
low-buffer transmission systems [65], [67], bufferless statistical
multiplexing provides a “ground truth” for studying the funda-
mental implications of the bit rate variabilities associated with
the H.264 SVC, H.264/AVC, and MPEG-4 Part 2 video en-
coders and with the video content. By considering the outlined
elementary bufferless statistical multiplexing scenario, we avoid
introducing confounding parameters, such as network buffers,
cross traffic, and network topology. Only the video encoder (and
its encoding settings), the video content, and the link capacity
(along with the number of streams ) influence the outcome
of the experiment and we are thus able to uncover the funda-
mental statistical multiplexing characteristics of the smoothed
H.264/AVC and H.264 SVC streams.
We note that predicting the loss probability of statistical
multiplexing from statistical descriptors of the video traffic
has been extensively studied for MPEG encoded videos and
verified through simulations with traces of MPEG encoded
videos, see e.g., [56], [65]–[70]. Generally, such prediction
works relatively well when the number of multiplexed streams
is high and the streams are relatively smooth. Predicting the
loss probability when multiplexing few streams as well as for
the new H.264/AVC and H.264 SVC encodings with their high
variability is a largely open research area. In this study we
conduct extensive simulations with traces of H.264/AVC and
H.264 SVC videos for a wide range of numbers of multiplexed
streams, which can be used as a baseline for assessing the
accuracy of novel prediction mechanisms.
C. Simulation Results
In the first set of simulations we estimate the maximum
number of video streams that can be accommodated by
the given link capacity , while constraining the information
loss probability to a value smaller than a prescribed small
constant . Many independent replications of each simulation
were run until the 90\% confidence interval of the information
loss probability estimate was less than 10\% of the corre-
sponding sample mean. In the second set of simulations, we
estimate the minimum link capacity that accommodates a
prescribed number of streams subject to . For each
estimate we perform 500 runs, each consisting of 1000
independent video streaming simulations. We do not include
the 90\% confidence intervals in the plots, because the
confidence intervals are very small ( 1\% of sample mean) and
would clutter the figures.
1) Simulations With Optimal Smoothing: Fig. 1
gives the curves and simulation curves, ob-
tained with and , for the five
video sequences. The simulation curves are,
respectively, named as SIM-G16B3-H.264-unsm for un-
smoothed H.264/AVC streams with GoP structure G16-B3,
SIM-G16B15-SVC-unsm for unsmoothed H.264 SVC streams
with GoP structure G16-B15,SIM-G16B3-MP4-unsm for un-
smoothed MPEG-4 Part 2 streams with GoP structure G16-B3,
SIM-G16B3-H.264-48KB for optimally smoothed H.264/AVC
streams, SIM-G16B15-SVC-48KB for optimally smoothed
H.264 SVC streams, and finally SIM-G16B3-MP4-48KB for
optimally smoothed MPEG-4 Part 2 streams. For reference, we
plot the curves corresponding to the multiplexing of
Authorized licensed use limited to: Arizona State University. Downloaded on August 31, 2009 at 13:04 from IEEE Xplore. Restrictions apply.

VAN DER AUWERA AND REISSLEIN: STATISTICAL MULTIPLEXING OF H.264/AVC AND SVC VIDEO STREAMS 545
Fig. 1. simulation and curves for five long CIF sequences encoded with H.264/AVC (G16-B3), H.264 SVC (G16-B15), and MPEG-4 Part 2
(G16-B3). The channel capacity is and the bit loss probability is . curves are provided for unsmoothed and optimally
smoothed traffic with client buffer size 48 KB . Perfect CBR curves are included for comparison: (a) Silence of the Lambs; (b) Star Wars IV;
(c) Sony Demo; (d) NBC 12 News; (e) Tokyo Olympics.
perfect constant bit rate traffic, denoted by . We define
PCBR video traffic as the sequence of frame size values that are
equal to the average frame size of the video stream. Hence, the
rate variability of a PCBR video stream is zero and is com-
puted by dividing by the stream’s average bit rate, resulting
in the theoretical maximum value for .
The values for the unsmoothed streams are strongly af-
fected by the rate variability of the video traffic. Toillustrate this
effect, we compare the curves of the unsmoothed traffic
with those of the PCBR video traffic. The unsmoothed traffic
clearly results in fewer supported streams than the PCBR video
traffic, which is only attributable to the rate variability. In ad-
dition, the gap between the PCBR curves of the H.264
SVC and the H.264/AVC encodings is much wider than the gap
between the corresponding unsmoothed traffic curves, e.g., see
Fig. 1(a) and (b). This is also evidence of the profound impact of
the rate variability increase of H.264 SVC traffic on com-
pared to H.264/AVC traffic.
Very interesting is that for the Sony Demo (Fig. 1(c)) and NBC
12 News (Fig. 1(d)) sequences, which have relatively high tex-
ture and motion complexity, the curve of the unsmoothed
H.264 SVC traffic is below the curve of the H.264/AVC traffic.
This is a very important observation, since this means that the
RD efficiency gain of H.264 SVC is completely canceled out by
the associated increased rate variability. For very high quality
(38 dB), the H.264 SVC curve for the Sony Demo
Authorized licensed use limited to: Arizona State University. Downloaded on August 31, 2009 at 13:04 from IEEE Xplore. Restrictions apply.

546 IEEE TRANSACTIONS ON BROADCASTING, VOL. 55, NO. 3, SEPTEMBER 2009
Fig. 2. simulation and curves for the Sony Demo and NBC 12 News sequences encoded with H.264 SVC (G16-B15), H.264/AVC (G16-B3),
and MPEG-4 Part 2 (G16-B3) for unsmoothed traffic and for optimally smoothed traffic . The channel capacity is , the bit loss
probability is . Perfect CBR curves are included for comparison: (a) Sony Demo; (b) NBC 12 News.
Fig. 3. Minimum channel capacity simulation results for the Silence of the Lambs and NBC 12 News sequences encoded with H.264/AVC (G16-B3), H.264
SVC (G16-B15), and MPEG-4 Part 2 (G16-B3) for unsmoothed video traffic. The bit loss probability is and the numbers of streams are , 16, and
64: (a) Silence of the Lambs; (b) NBC 12 News.
sequence even approaches the MPEG-4 Part 2 curve, and
surprisingly, for the NBC 12 News sequence the H.264 SVC
curve is below the MPEG-4 Part 2 curve. The reason
is that for these two relatively complex sequences, the number of
streams that can be supported by the link is small ( 20 streams)
and as a result the statistical multiplexing effect that copes with
the rate variability of the streams is reduced.
Next, we study whether traffic smoothing would bring out
the gains in the number of supported streams that one
would expect from the RD efficiency gains of H.264 SVC over
H.264/AVC. We initially employ optimal smoothing with a
client buffer size of 48 KB. We observe that all curves
for the optimally smoothed traffic in Fig. 1 have significantly
increased values compared to the values for the unsmoothed
traffic, and that they are much closer to the theoretical maximum
values given by the PCBR curves. (In additional experiments
with the Sony sequence, we found that optimal smoothing
with a larger, 128 KB buffer increases by one to five
streams; generally, for very large smoothing buffers the PCBR
is approached [71].) When examining the gaps between the
curves of H.264 SVC and H.264/AVC, we notice that
the gaps have increased and approach the theoretical max-
imum gaps of the PCBR curves or equivalently the maximum
gain in number of supported streams. We conclude from this
initial analysis that optimal smoothing effectively mitigates
the effects of the increased variability of H.264 SVC traffic
on the maximum number of streams supported in a bufferless
statistical multiplexer. Interesting is that for the relatively lower
complexity (texture, motion) Silence of the Lambs,Star Wars
4, and Tokyo Olympics sequences, the curves of the
smoothed MPEG-4 Part 2 traffic approach the curves of
the unsmoothed H.264 SVC and H.264/AVC traffic in the very
high quality region. For the relatively higher complexity Sony
Demo and NBC 12 News sequences, the curves of the un-
smoothed H.264 SVC and H.264/AVC traffic are considerably
below the curves of the smoothed MPEG-4 Part 2 traffic.
The above observations are clearly dependent on the video
content, but also on the chosen link capacity . Clearly, if the
link can only support a small number of streams, then the statis-
tical multiplexing effect is small, resulting in a strong impact of
the rate variability on the number of multiplexed streams. The
impact is particularly significant when multiplexing high quality
H.264 SVC encodings of the relatively complex Sony Demo and
NBC 12 News sequences in the scenario with
considered in Fig. 1. In order to examine the statistical multi-
plexing of these two sequences with a higher link capacity, we
plot in Fig. 2 curves for and .
First, we observe that the values are much larger than for
Authorized licensed use limited to: Arizona State University. Downloaded on August 31, 2009 at 13:04 from IEEE Xplore. Restrictions apply.

VAN DER AUWERA AND REISSLEIN: STATISTICAL MULTIPLEXING OF H.264/AVC AND SVC VIDEO STREAMS 547
Fig. 4. Minimum channel capacity simulation results for the Sony Demo sequence encoded with H.264/AVC (G16-B3), H.264 SVC (G16-B15), and MPEG-4
Part 2 (G16-B3) for unsmoothed video traffic. The bit loss probabilities are and the number of streams are , 16, and 64: (a)
Sony Demo,; (b) Sony Demo, ; (c) Sony Demo, .
the experiments, as we expected. Second, the
curves for the unsmoothed traffic are closer to the theoret-
ical upper boundary given by the PCBR curves. The optimally
smoothed traffic is particularly close to this theoretical upper
limit, again illustrating that even for large there is still a sig-
nificant impact of smoothing on the values. Nevertheless,
in both cases, unsmoothed and smoothed, H.264 SVC clearly al-
lows for more statistically multiplexed streams than H.264/AVC
and MPEG-4 Part 2.
Although the simulations provide insight into the sig-
nificant effects of the increased rate variability of H.264 SVC,
they are dependent on the prescribed link capacity and re-
sult in varying numbers of multiplexed streams (i.e., varying
levels of statistical multiplexing) across the range of average
PSNR video qualities. Therefore, in the next section we per-
form a second set of simulations that estimate the minimum link
capacity required for supporting a prescribed number of
streams . These simulations allow us to study the effects
of the rate variability for a fixed number of multiplexed streams
across the range of PSNR video qualities.
2) Simulations With Optimal Smoothing: Fig. 3 depicts
the curves for unsmoothed traffic of the sequences Silence
of the Lambs and NBC 12 News for multi-
plexed streams for . In general, for , we ob-
serve that the values are somewhat lower for the H.264
SVC streams than for H.264/AVC streams. This link capacity
difference is particularly significant for Silence of the Lambs
in the high quality range ( 35 dB), otherwise the differ-
ences become relatively small. However, both encoders have a
clear advantage over MPEG-4 Part 2. For , the statis-
tical multiplexing effect is less able to compensate for the bit
rate variabilities. Overall, the H.264/AVC streams are accom-
modated by values that are smaller than or nearly equal
to the values for the H.264 SVC streams, despite the higher av-
erage bit rates of the H.264/AVC streams. H.264 SVC still out-
performs MPEG-4 Part 2 over the entire quality range.
For , the increased rate variability of H.264 SVC results
in values that are overall comparable to those of multi-
plexed MPEG-4 Part 2 streams. For the Silence of the Lambs
sequence, we observe the surprising result that H.264 SVC re-
quires the highest values over the entire quality range and
MPEG-4 Part 2 even outperforms H.264/AVC below 38 dB. For
the NBC 12 News sequence, H.264 SVC has worst performance
in the quality range above 35 dB. The conclusion is that for a rel-
atively small number of multiplexed streams ( 16), H.264/AVC
generally results in lower requirements, while depending
on the video sequence, H.264 SVC can even be outperformed
by MPEG-4 Part 2 streams.
Next, we examine the impact of the information loss
probability on in Fig. 4. The unsmoothed Sony
Demo streams are multiplexed with maximum losses
, respectively, for , 16,
Authorized licensed use limited to: Arizona State University. Downloaded on August 31, 2009 at 13:04 from IEEE Xplore. Restrictions apply.

548 IEEE TRANSACTIONS ON BROADCASTING, VOL. 55, NO. 3, SEPTEMBER 2009
Fig. 5. simulation results for five long CIF sequences encoded with H.264/AVC (G16-B3), H.264 SVC (G16-B15), and MPEG-4 Part 2 (G16-B3) for
optimally smoothed video traffic with a 48KB client buffer. The bit loss probability is and the number of streams are , 16, and 64: (a)
Silence of the Lambs; (b) Star Wars IV; (c) Sony Demo; (d) NBC 12 News; (e) Tokyo Olympics.
and 64 streams. The values are significantly lower when
the allowable losses are larger, as we expected, and this is the
case for all encoders and numbers of streams . Interesting is
that overall the relative order of the curves, corresponding
to the different encoders for each value of , is preserved.
In Fig. 5, we examine the values for optimally smoothed
streams (client buffer size 48 KB). Overall, optimally smoothed
H.264 SVC traffic has lower values for , 16, and 64
over the entire quality range. The quality range above 35 dB is
particularly favorable for optimally smoothed H.264 SVC over
H.264/AVC. Optimally smoothed MPEG-4 Part 2 traffic clearly
requires substantially more network bandwidth resources.
In summary, we conclude from the and simula-
tions with optimally smoothed traffic that optimally smoothed
H.264 SVC streams clearly have an advantage over optimally
smoothed H.264/AVC and MPEG-4 Part 2 streams. In partic-
ular, the simulations indicate that close to optimal results
(PCBR) are achievable with optimally smoothed traffic. Optimal
smoothing [10], [11] is an off-line technique designed for prere-
corded video streams. Optimal smoothing can been adapted for
live video through appropriate traffic descriptors and predictors,
which have so far only been examined for MPEG-4 Part 2 and
preceding MPEG codecs [72]. Researching appropriate traffic
descriptors and predictors for the new H.264/AVC and H.264
SVC encoders with their more bursty traffic is an open problem.
On the other hand, basic smoothing, which is computationally
significantly less complex than optimal smoothing, can easily be
implemented for live video. We are therefore motivated to com-
Authorized licensed use limited to: Arizona State University. Downloaded on August 31, 2009 at 13:04 from IEEE Xplore. Restrictions apply.

VAN DER AUWERA AND REISSLEIN: STATISTICAL MULTIPLEXING OF H.264/AVC AND SVC VIDEO STREAMS 549
Fig. 6. simulation results for the Silence of the Lambs and Sony Demo sequences encoded with H.264 SVC (G16-B15) and H.264/AVC (G16-B3) for basic
smoothed traffic with aggregation level and for optimally smoothed traffic . The bit loss probability is and the number of streams
are , 16, and 64: (a) Silence of the Lambs, H.264 SVC (G16-B15); (b) Silence of the Lambs, H.264/AVC (G16-B3); (c) Sony Demo, H.264 SVC (G16-B15);
(d) Sony Demo, H.264/AVC (G16-B3).
pare the bufferless statistical multiplexing performance
of basic smoothing with optimal smoothing.
3) Simulations With Basic Smoothing: Fig. 6 depicts
the curves for the Silence of the Lambs and Sony Demo
H.264 SVC video traffic (G16-B15) that is smoothed with ag-
gregation level (GoP size), and for H.264/AVC video
traffic (G16-B3) smoothed with . We also include the
results obtained for optimal smoothing. The basic smoothing
curves are only very slightly above the curves for
optimally smoothed traffic. This indicates that basic smoothing
with is almost as effective as optimal smoothing in re-
ducing the rate variability for efficient bufferless statistical mul-
tiplexing.
4) Basic Smoothing Delay Implications: The simulation re-
sults in the preceding sections together with the delay analysis in
the Appendix establish a reference framework for evaluating the
traffic smoothing versus delay trade-off. In this section, we in-
vestigate the choice of the basic smoothing parameters that en-
sure that (i) the link capacity requirements for H.264 SVC traffic
(hierarchical B frames) are reduced compared to H.264/AVC
traffic (classical B frames), and (ii) the link capacity required
with basic smoothing closely approaches the link capacity re-
quired with optimally smoothed traffic.
Fig. 7 depicts simulation curves for unsmoothed and
smoothed (basic) traffic with aggregation levels ,4,8,
and 16. The experiments cover the five sequences that are en-
coded with H.264 SVC (G16-B15) and H.264/AVC (G16-B3).
We present illustrative results for , 16, and 64 streams,
while the bit loss probability is restricted to ;wehave
also analyzed identical experiments with , which we
can not include due to space constraints. Fig. 7(a) and (b) present
the case with multiplexed streams, Fig. 7(c) and (d)for
streams, and Fig. 7(e) and (f) for streams. We
present illustrative results for videos with relatively low texture
and motion complexity in Fig. 7(a), (c), and (e), while illustra-
tive results for videos with relatively high texture and motion
complexity are presented in Fig. 7(b), (d), and (f).
For streams, in general, the unsmoothed H.264
SVC streams require smaller values than the H.264/AVC
streams. This is explained by the relatively large number of
streams that are statistically multiplexed. Ideally, the
values should be close to the values for optimally
smoothed streams or, equivalently, close to the values
for basic smoothed streams with aggregation , which
is the GoP size, as we illustrated in Fig. 6. In Fig. 7(a), for
example, the simulation curves for increasing aggregation
levels approach the curves of the traffic smoothed with
(which gives very close to optimal smoothing results).
This observation holds for all test sequences and numbers of
multiplexed streams , although for the Sony Demo sequence
the convergence is slower than for the other four sequences.
Overall, when the aggregation level should be set
Authorized licensed use limited to: Arizona State University. Downloaded on August 31, 2009 at 13:04 from IEEE Xplore. Restrictions apply.

550 IEEE TRANSACTIONS ON BROADCASTING, VOL. 55, NO. 3, SEPTEMBER 2009
Fig. 7. simulation results for unsmoothed and basic smoothed traffic with aggregation levels , 4, 8, and 16. The five sequences are encoded with
H.264 SVC (G16-B15) and H.264/AVC (G16-B3). The bit loss probability is and the number of streams are , 16, and 64: (a) Silence of the Lambs
; (b) Sony Demo ; (c) Star Wars 4 ; (d) NBC 12 News ; (e) Tokyo Olympics ; (f) NBC 12 News .
to or for H.264/AVC stream multiplexing to
approach the optimal performance, and to or for
H.264 SVC streams. The choice between the two values for
each encoder depends on the content type, with the larger value
meant for the most complex sequences.
Analogously, we analyzed the cases with , ,
and multiplexed streams. Table I enumerates aggregation
levels that when applied to both H.264 SVC and H.264/AVC
video streams result in lower requirements for H.264 SVC
streams (G16-B15) than for H.264/AVC streams (G16-B3) for
both examined loss probabilities and . Table II gives
basic smoothing aggregation levels that achieve close to op-
timal smoothing values for H.264/AVC and H.264 SVC,
respectively. For the cases with two values, we recommend
the higher value for sequences with relatively high texture and
motion complexity. The corresponding end-to-end delays, cal-
culated based on the delay analysis in the Appendix, are pro-
vided in Table II for live video streaming (middle two columns)
and for prerecorded video streaming (right two columns).
From this analysis we conclude that the H.264 SVC streams
generally require aggregation levels twice as large as the
H.264/AVC streams to obtain close to optimal statistical multi-
plexing performance. The corresponding end-to-end-delays are
approximately two to three times larger for H.264 SVC than
for H.264/AVC.
The preceding analysis considers one video sequence (out of
the five sequences) in a given multiplexing experiment. Next,
we examine whether the recommendations for the choice of
Authorized licensed use limited to: Arizona State University. Downloaded on August 31, 2009 at 13:04 from IEEE Xplore. Restrictions apply.

VAN DER AUWERA AND REISSLEIN: STATISTICAL MULTIPLEXING OF H.264/AVC AND SVC VIDEO STREAMS 551
TABLE II
AGGREGATION LEVELS [( FOR LOW COMPL.SEQ.) (FOR HIGH COMPL.SEQ.)] FOR BASIC SMOOTHING SUCH THAT FOR BASIC SMOOTHING VERY
CLOSELY APPROACHES FOR OPTIMAL SMOOTHING FOR H.264/AVC AND H.264 SVC, RESPECTIVELY;C
ORRESPONDING DELAYS [IN FRAME PERIODS]
FOR PRERECORDED AND LIVE VIDEO ARE ALSO PROVIDED.THESE RESULTS APPLY FOR , 16, 32, 64 MULTIPLEXED STREAMS FOR BOTH
AND
TABLE I
AGGREGATION LEVELS FOR BASIC SMOOTHING SUCH THAT FOR
H.264 SVC (G16-B15)ISLESS THAN FOR H.264/AVC (G16-B3)FOR BOTH
AND .WEPROVIDE ( VALUE FOR LOW COMPLEXITY
SEQUENCE)( VALUE FOR HIGH COMPLEXITY SEQUENCE)
the aggregation level also hold for a heterogeneous mix of
the five video sequences. We organized the H.264/AVC video
streams and the H.264 SVC video streams each into three
quality groups based on average PSNR values: low quality
(32–34 dB), medium quality (35–37 dB), and high quality
(38–40 dB). We conducted multiplexing simulations for each
quality group to determine the minimum link capacities
required to achieve loss probabilities below and
, respectively. In each simulation, we multiplex
streams drawn randomly from the five video sequences
(while equalizing for the different stream lengths so that each
video sequence is selected with approximately equal proba-
bility). The respective estimated values are reported in
Table III for , and in Table IV for .
From the data in Tables III and IV, we conclude that the
above recommendations for the aggregation levels also hold
for the heterogeneous mix of the video streams; furthermore, the
recommendations hold across quality groups and for both
and . The recommended aggregation levels for
approaching the optimal smoothing value within 15\%, are
to for H.264/AVC streams and to for
H.264 SVC streams. This observation confirms that H.264 SVC
streams require higher aggregation levels to approximate the op-
timal smoothing . We also reconfirm that the aggregation
level at which H.264 SVC streams achieve link capacities
below H.264/AVC capacities, is at least and even as high
as . Since these multiplexing experiments with hetero-
geneous video sequences reconfirm the aggregation level rec-
ommendations, we conclude that the different encoder config-
urations, i.e., hierarchical B frames for H.264 SVC (G16-B15)
versus classical B frames for H.264/AVC (G16-B3) are the de-
termining factors in the statistical multiplexing behavior of the
respective video streams.
V. B UFFERED STATISTICAL MULTIPLEXING
Next, we study the buffered statistical multiplexing of video
streams encoded with H.264/AVC (G16-B3), H.264 SVC (G16-
TABLE III
BIT RATES FOR MIXES OF VIDEO STREAMS DRAWN FROM ALL
FIVE VIDEOS FOR DIFFERENT BASIC SMOOTHING LEVELS AND OPTIMAL
SMOOTHING (OPT.S
M.) FOR
TABLE IV
BIT RATES FOR MIXES OF VIDEO STREAMS DRAWN FROM ALL
FIVE VIDEOS FOR DIFFERENT BASIC SMOOTHING LEVELS AND OPTIMAL
SMOOTHING (OPT.SM.) FOR
B15), and MPEG-4 Part 2 (G16-B3). The video traffic is not
smoothed in order to assess the direct impact of the multiplexer
buffer size. The buffer serves the purpose of absorbing some of
the rate variability of the video streams that are multiplexed on
the link. From among the wide range of buffer management and
scheduling policies, see e.g. [73]–[76], we consider the elemen-
tary taildrop policy with first-come-first-served scheduling, to
assess the fundamental impact of the multiplexer buffer. Specif-
ically, with given in (3) denoting the aggregate bit rate [in
bit/s] of the ongoing video streams in frame period ,
denoting the buffered video traffic [in bit] at the end of the pre-
ceding frame period (i.e., at the beginning of frame period
), and noting that traffic is served at bit rate , the amount of
buffered video traffic at the end of frame period is obtained
as
(5)
where denotes the buffer capacity [in bit]. The amount of lost
video bits during frame period is given by
and the expected long run fraction of lost bits gives
the information loss probability, which is required to be less than
.
Fig. 8 depicts simulation results for the five CIF se-
quences. The channel capacity is and .
Curves are presented for buffer sizes 24, 192, and 3840 KB.
(We also examined the buffer sizes 48 and 96 KB, which
Authorized licensed use limited to: Arizona State University. Downloaded on August 31, 2009 at 13:04 from IEEE Xplore. Restrictions apply.

552 IEEE TRANSACTIONS ON BROADCASTING, VOL. 55, NO. 3, SEPTEMBER 2009
Fig. 8. buffered multiplexing simulation results for five long CIF sequences encoded with H.264/AVC (G16-B3), H.264 SVC (G16-B15), and MPEG-4
Part 2 (G16-B3). The channel capacity is and the bit loss probability is . curves are provided for unsmoothed traffic. Curves are
presented for buffer sizes are set to 24, 192, and 3840 KB. The bufferless multiplexing results are included for reference: (a) Silence of the Lambs; (b) Star Wars
IV; (c) Sony Demo; (d) NBC 12 News; (e) Tokyo Olympics.
are not included to avoid clutter in the plots.) The bufferless
multiplexing results are depicted for comparison. Analogous to
the minimum channel capacity experiments, we determine the
buffer size that gives near optimal statistical multiplexing re-
sults for H.264 SVC, H.264/AVC, and MPEG-4 Part 2 streams,
whereby we adopt as benchmark for optimal results the
curve for the largest buffer size 3840 KB. Comparisons of
the results in Figs. 1 and 8 indicate that the curve for
3840 KB is very close to the PCBR curve, which gives the
maximum number of streams that can be supported on the
link. We identify the buffer sizes that result in values
that are relatively close to the values for buffer size 3840
KB. The recommended buffer size ranges for each encoder
are summarized in Table V. We determine the buffer ranges
across the five video sequences, with the largest buffer sizes
corresponding to complex sequences. The H.264 SVC streams
require approximately twice the buffer size compared to the
H.264/AVC streams, which in turn require about double the
buffer size required for MPEG-4 Part 2 streams. With the delay
analysis presented in the Appendix, we obtain a delay of 25
frame periods for transmitting unsmoothed live H.264 SVC
video over a transmit path with a single buffer stage with 192
KB compared to 9 frame periods for transmitting H.264/AVC
video over a transmit path with a single 96 KB buffer stage.
We similarly studied the case when ; the corre-
sponding plots are not included due to space constraints. For
, the recommended buffer size ranges are significantly
smaller (approx. half) for each encoder than for .
Authorized licensed use limited to: Arizona State University. Downloaded on August 31, 2009 at 13:04 from IEEE Xplore. Restrictions apply.

VAN DER AUWERA AND REISSLEIN: STATISTICAL MULTIPLEXING OF H.264/AVC AND SVC VIDEO STREAMS 553
Fig. 9. Delay analysis of classical B frame H.264/AVC encoding with GoP structure G16-B3 for no smoothing and for basic smoothing with : (a) no
smoothing, ; (b) basic smoothing with .
Fig. 10. Delay analysis of hierarchical B frame H.264 SVC encoding with GoP structure G16-B15 for no smoothing, and for basic smoothing with : (a) no
smoothing, ; (b) basic smoothing with .
TABLE V
OVERVIEW OF RECOMMENDED BUFFER SIZE RANGES FOR BUFFERED
STATISTICAL MULTIPLEXING WITH
However, the double buffer size relationship between en-
coders remains, as well as the corresponding delay differences.
We conclude that the RD efficiency improvements between
TABLE VI
DELAYS [IN FRAME PERIODS]FOR LIVE H.264/AVC (G16-B3)AND H.264
SVC (G16-B15)S
TREAMS
the encoders comes at the price of increased buffer sizes and
corresponding delays in the buffered statistical multiplexing
scenario.
Authorized licensed use limited to: Arizona State University. Downloaded on August 31, 2009 at 13:04 from IEEE Xplore. Restrictions apply.

554 IEEE TRANSACTIONS ON BROADCASTING, VOL. 55, NO. 3, SEPTEMBER 2009
TABLE VII
DELAYS [IN FRAME PERIODS]FOR PRERECORDED H.264/AVC (G16-B3)AND
H.264 SVC (G16-B15)S
TREAMS.T
HE DELAY WITH OPTIMAL SMOOTHING
WITH (ADDITIONAL)STAR T-UPDELAY ISIDENTICAL TO THE
DELAY FOR UNSMOOTHED TRAFFIC
VI. CONCLUSIONS
We have examined the statistical multiplexing behavior of
H.264 SVC, H.264/AVC, and MPEG-4 Part 2 encoded video
with long video sequences. In particular, we have considered the
bufferless statistical multiplexing of smoothed video streams
and the buffered statistical multiplexing of unsmoothed video
streams. We have found that off-line optimal smoothing ensures
that the RD efficiency gains of H.264 SVC with hierarchical
B frames over H.264/AVC with classical B frames translate
into commensurate gains in the number stream supported with
statistical multiplexing. (Without smoothing, the higher rate
variability of H.264 SVC may actually result in fewer supported
streams than with the less RD efficient H.264/AVC and in some
scenarios even fewer SVC streams than with the even less RD
efficient MPEG-4 Part 2.) We further examined basic smoothing
which averages the sizes of blocks of successive video frames
and is thus simple to implement in on-line fashion and readily
applicable to live video. We characterized the trade-off between
increased delay with increased levels of smoothing (for larger
) and the resulting reduced rate variability and corresponding
increased number of supported streams with statistical multi-
plexing. Specifically, we identified the basic smoothing levels
that ensure that (i) more H.264 SVC than H.264/AVC streams
are supported with statistical multiplexing, and that (ii) the
number of H.264 SVC streams and H.264/AVC streams sup-
ported with basic smoothing closely approaches the number
of streams supported with optimal smoothing. Moreover, we
identified the sizes of the multiplexer buffers that ensure that
the numbers of supported H.264 SVC streams and H.264/AVC
streams approach the theoretical maximum given by the link
capacity divided by the average stream bit rate; we found that
H.264 SVC requires roughly twice the multiplexer buffer of
H.264/AVC, which in turn requires twice the buffer of MPEG-4
Part 2.
There are numerous directions for future research on the
statistical multiplexing of H.264 SVC and H.264/AVC encoded
video. One important direction is examining collaborative
smoothing strategies and active buffer management strategies
considering the frame playout deadlines for H.264/AVC and
H.264 SVC encoded video.
APPENDIX
DELAY ANALYSIS OF SMOOTHED TRANSMISSION OF H.264
SVC AND H.264/AVC VIDEO
In this Appendix we analyze the end-to-end delay introduced
by the video encoding and decoding in conjunction with the
smoothing of the video frame sizes for network transport. We
initially consider live video and evaluate the time shift between
the capture of a frame at the sender and the display of the frame
at the receiver; we subsequently examine prerecorded video.
Throughout, we normalize time by the frame period (33 ms for
NTSC video). (For all delays reported in units of frame periods,
the corresponding delays in units of seconds are obtained by di-
viding the delay in units of frame periods by the frame rate
in units of frames/second, which is frames/second for
NTSC video.) In general, the time shift between frame cap-
ture and display can be decomposed into the following compo-
nents:
•: Delay introduced due to the dependencies of the
encoded frames, i.e., maximum delay a given captured
frame experiences due to waiting for the capture of subse-
quent frames that are needed for the encoding of the given
captured frame.
•: Delay introduced by the computations needed
for the encoding.
•: Delay introduced by the smoothed transmission.
•: Delay introduced by the computations needed
for the decoding of a frame.
•: Delay introduced by reordering of frames to ensure
uninterrupted display sequence.
The total end-to-end delay is obtained by summing the delay
components
(6)
For each of the following delay analyzes we initially suppose
that the encoding computations and the decoding computations
take one frame period, i.e., , we sub-
sequently consider the cases when computation times become
negligible. Throughout, we suppose that it takes one frame pe-
riod to transmit one (unsmoothed) frame, and frame periods to
transmit a block of smoothed frames, as is consistent with the
evaluation of the aggregate bit rate in (3). We note that the trans-
mission of unsmoothed video is equivalent to basic smoothing
with the aggregation level of one frame, i.e., . We let ,
, denote the number of B frames between successive key
picture (I or P frames).
A. Live Video With Classical B Frames
Fig. 9 illustrates the delay structure for live streaming of
H.264/AVC video encoded with classical B frames for GoP
structure G16-B3, i.e., . The capture time index axis
represents the frame type (I, P, or B) that is used to encode each
captured frame. Each frame is designated by its frame type and
its capture time, e.g., is the frame captured at time index
four and is encoded as a P frame. We suppose that the capture
time itself is infinitesimally short and negligible. On the encode
time axis, the frames are put in encoding order according of the
motion compensated prediction frame dependencies, which are
indicated by the arrows above the capture time axis. The time
shift between the capture and encode axes represents the delay
due to the encoding dependencies . Specifically, we
observe that , since frame needs to wait
for the capture of frame before frame can be encoded.
The time shift between the encode and transport time axes
represents the delay due to encoding computations .
Authorized licensed use limited to: Arizona State University. Downloaded on August 31, 2009 at 13:04 from IEEE Xplore. Restrictions apply.

VAN DER AUWERA AND REISSLEIN: STATISTICAL MULTIPLEXING OF H.264/AVC AND SVC VIDEO STREAMS 555
For example, the frame is encoded in between time indices
and , followed by the encoding of the frames
, , and that depend on and . In general, with
smoothed transmission, all frames of a smoothing block need
to be encoded before transmission of the block can commence,
hence .
Subsequently, the encoded frames are transmitted in encoding
order since the decoder needs the frames in encoding order for
the decoding process to run without introducing unnecessary re-
ordering delays. In the illustrated unsmoothed example, frame
is transmitted between time indices and ,
while for the illustrated smoothed transmission example, the
first block of frames is transmitted between time indices
and . Generally, noting that the decoding can only
start when the entire block of frames is received, we obtain
, which is represented by the time shift between
the transport and decode axes in the illustration in Fig. 9. We do
not consider store-and-forward transmission delays nor propa-
gation, queueing, or processing delays in the transport network;
these delays could be subsumed in in straightforward
fashion. In particular, for buffered multiplexing of unsmoothed
video , as considered in Section V, the transmission
delay in frame periods with a single buffered multiplexing stage
on the transmission path is bounded by one frame period (for
the transmission by the sending host) plus the maximum buffer
delay, namely the buffer capacity [in bit] divided by the bit
rate [in bit/s] normalized with the frame rate F [in frames/s],
i.e., .
Next, the decoder processes frame in between time indices
and in Fig. 9(a); generally, .In
addition, the receiver needs to reorder the decoded frames into
display order to ensure uninterrupted playback. This reordering
introduces one frame period delay, i.e., , since frame
in Fig. 9(a) is not available for display until time instant .
In summary, we obtain for live video with classical B frames
(7)
We remark that we have not included the first I frame in
the data blocks for basic smoothing. Alternatively, this I frame
can be included and the non-overlapping blocks would shift
one frame index to the left without any implications for the
end-to-end delay. The advantage of not including the first I
frame is that the first block already contains a large P frame.
Singling out the first I frame allows for spreading its trans-
mission over multiple frame periods if the I frame is encoded
immediately when it is captured. For example, in Fig. 9(a) the
first I frame can be transmitted over four frame periods, if it is
immediately encoded after time index zero.
We briefly adapt the above delay analysis to scenarios with
negligible encoding and/or decoding computation times as fol-
lows. We focus on scenarios where either or
is an integer. If an arbitrary number of video frames can be en-
coded in negligible time, , then the delay due to
frame encoding dependencies becomes
. To see this, note that two conditions need to be met before
transmitting the first block of encoded frames: (i) the first B
frame needs to await the capture of the successive P frame, i.e.,
for frame periods, and (ii) the first frame to be transmitted in
a smoothing block needs to await the capture of the remaining
frames for the block. If an arbitrary number of video frames
can be decoded in negligible time, , then the
display reordering delay becomes , where
denotes the indicator function which is one if is true,
and zero otherwise.
B. Live Video With Hierarchical B Frames
We consider hierarchical B frames with a dyadic structure,
i.e., B frames between key pictures for some integer
. We do not consider low-delay or constrained delay B
frame prediction structures [2].
Reasoning as above, along the illustration in Fig. 10, we find
that the delay components , , , and
are identical to the above case of classical B frames.
Note however the hierarchical B frame dependency structure,
which is indicated with arrows above the capture time axis, and
the encoding order of the frames on the encode time axis, which
results in minimal reordering delay for the display process [77].
Importantly, we note that due to the hierarchical dependencies
between B frames, the reordering delay for achieving the display
sequence depends on the number of temporal levels, i.e.,
. In summary,
(8)
In Table VI, we summarize the delays for the H.264 SVC
(G16-B15) and H.264/AVC (G16-B3) streams considered in this
study. The end-to-end delays for the H.264 SVC traffic are 15
frame periods larger than for H.264/AVC, which is attributable
to the hierarchical B frame prediction structure. In particular,
with the G16-B15 hierarchical B prediction structure, which re-
sults in improved RD performance, the encoder has to wait until
the frame with time index 16 is captured before it can encode
this frame as an I frame and start encoding all 15 preceding hi-
erarchical B frames. In addition, the reordering delay increases
to four frame periods with the considered RD efficient hierar-
chical B frame structure.
For scenarios with , as well as either or
an integer, we adapt the preceding analysis as fol-
lows. With negligible encoding time, , the en-
coding dependency delay becomes ,
similar to the case of classical B frames. For negligible decoding
time, , the smoothed transmission and display re-
ordering delay become together
(9)
Authorized licensed use limited to: Arizona State University. Downloaded on August 31, 2009 at 13:04 from IEEE Xplore. Restrictions apply.

556 IEEE TRANSACTIONS ON BROADCASTING, VOL. 55, NO. 3, SEPTEMBER 2009
C. Prerecorded Video
For prerecorded video, all frames are preencoded, leaving
only the smoothed transport, decoding, and display reordering
delays, i.e., and
(10)
Effectively, for prerecorded video, the time index is
shifted to the beginning of the frame sequence on the transport
axis in the illustrations in Figs. 9 and 10, e.g., to in Fig. 9.
Specifically, we obtain for classical B frames
(11)
and for hierarchical B frames
(12)
Table VII gives the delays for prerecorded H.264/AVC
(G16-B3) and H.264 SVC (G16-B15) streams. The end-to-end
delays for the H.264 SVC traffic are three frame periods larger
than for H.264/AVC, which is a smaller difference than for live
video in Table VI.
The delays for optimal smoothing of prerecorded video with
the (additional) startup delay of frame periods (defined in [10])
are obtained by replacing by in (11) and (12). This is
because optimal smoothing is designed to deliver the first frame
within frame periods to the decoder; and then ensure that
for each subsequent frame period the next frame is available for
decoding. For the examples in Table VII, the delay for opti-
mally smoothed prerecorded traffic is one to fifteen frame pe-
riods smaller than for basic smoothed prerecorded traffic; op-
timal smoothing is however much more computationally de-
manding than basic smoothing.
ACKNOWLEDGMENT
We are grateful to Prof. Lina Karam, Arizona State Univer-
sity, for insightful discussions on video coding, and to Pras-
anth T. David for developing the automated scheduler of the
encoding jobs.
REFERENCES
[1] H.-C. Huang, W.-H. Peng, T. Chiang, and H.-M. Hang, “Advances in
the scalable amendment of H.264/AVC,” IEEE Communications Mag-
azine, vol. 45, no. 1, pp. 68–76, Jan. 2007.
[2] H. Schwarz, D. Marpe, and T. Wiegand, “Overview of the scalable
video coding extension of the H.264/AVC standard,” IEEE Trans. Cir-
cuits and Systems for Video Technology, vol. 17, no. 9, pp. 1103–1120,
Sep. 2007.
[3] M. Wien, H. Schwarz, and T. Oelbaum, “Performance analysis of
SVC,” IEEE Trans. Circuits and Systems for Video Technology, vol.
17, no. 9, pp. 1194–1203, Sep. 2007.
[4] D. Marpe, T. Wiegand, and G. Sullivan, “The H.264/MPEG–4 ad-
vanced video coding standard and its applications,” IEEE Communi-
cations Magazine, vol. 44, no. 8, pp. 134–143, Aug. 2006.
[5] G. Van der Auwera, P. T. David, and M. Reisslein, “Traffic charac-
teristics of H.264/AVC variable bit rate video,” IEEE Communications
Magazine, vol. 46, no. 11, pp. 164–174, Nov. 2008.
[6] T. Lakshman, A. Ortega, and A. Reibman, “VBR video: Tradeoffs and
potentials,” Proceedings of the IEEE, vol. 86, no. 5, pp. 952–973, May
1998.
[7] A. R. Reibman and M. T. Sun, Compressed Video over Networks.
New York: Marcel Dekker, 2000.
[8] D. Wu, Y. Hou, W. Zhu, Y.-Q. Zhang, and J. Peha, “Streaming video
over the internet: Approaches and directions,”IEEE Trans. Circuits and
Systems for Video Technology, vol. 11, no. 3, pp. 282–300, Mar. 2001.
[9] G. Van der Auwera, P. T. David, and M. Reisslein, “Traffic and
quality characterization of single-layer video streams encoded with
the H.264/MPEG–4 Advanced Video Coding standard and Scalable
Video Coding extension,” IEEE Trans. Broadcasting, vol. 54, no. 3,
pp. 698–718, Sep. 2008.
[10] J. Salehi, Z.-L. Zhang, J. Kurose, and D. Towsley, “Supporting stored
video: Reducing rate variability and end–to–end resource requirements
through optimal smoothing,” IEEE/ACM Trans. Networking, vol. 6, no.
4, pp. 397–410, Aug. 1998.
[11] A. R. Reibman and A. W. Berger, “Traffic descriptors for VBR video
teleconferencing over ATM networks,” IEEE/ACM Trans. Networking,
vol. 3, no. 3, pp. 329–339, Jun. 1995.
[12] G. Van der Auwera, M. Reisslein, and L. J. Karam, “Video texture
and motion based modeling of rate variability-distortion (VD) curves,”
IEEE Trans. Broadcasting, vol. 53, no. 3, pp. 637–648, Sep. 2007.
[13] A. Ortega and K. Ramachandran, “Rate-distortion methods for image
and video compression,” IEEE Signal Processing Magazine, vol. 15,
no. 6, pp. 23–50, Nov. 1998.
[14] P. Seeling and M. Reisslein, “The rate variability-distortion (VD)
curve of encoded video and its impact on statistical multiplexing,”
IEEE Trans. Broadcasting, vol. 51, no. 4, pp. 473–492, Dec. 2005.
[15] A. Alheraish, S. Alshebeili, and T. Alamri, “A GACS modeling ap-
proach for MPEG broadcast video,” IEEE Trans. Broadcasting, vol.
50, no. 2, pp. 132–141, Jun. 2004.
[16] N. Ansari, H. Liu, Y. Q. Shi, and H. Zhao, “On modeling MPEG video
traffics,” IEEE Trans. Broadcasting, vol. 48, no. 4, pp. 337–347, Dec.
2002.
[17] D. P. Heyman and T. V. Lakshman, “Source models for VBR broadcast
video traffic,” IEEE/ACM Trans. Networking, vol. 4, no. 1, pp. 40–48,
Jan. 1996.
[18] X.-D. Huang, Y.-H. Zhou, and R.-F. Zhang, “A multiscale model for
MPEG-4 varied bit rate video traffic,” IEEE Trans. Broadcasting, vol.
50, no. 3, pp. 323–334, Sep. 2004.
[19] M. M. Krunz and A. M. Makowski, “Modeling video traffic using
input processes: A compromise between Markovian and
LRD models,” IEEE Journal on Selected Areas in Communications,
vol. 16, pp. 733–748, Jun. 1998.
[20] D. Marpe, T. Wiegand, and S. Gordon, “H.264/MPEG-4 AVC Fidelity
Range Extensions: Tools, profiles, performance, and application
areas,” in Proc. IEEE Int. Conf. on Image Proc. (ICIP), Sep. 2005, pp.
593–596.
[21] A. Undheim, Y. Lin, and P. Emstad, “Characterization of slice-based
H.264/AVC encoded video traffic,” in Proceedings of Fourth European
Conference on Universal Multiservice Networks (ECUMN), Feb. 2007,
pp. 263–272.
[22] H.-H. Juan, H.-C. Huang, C. Huang, and T. Chiang, “Scalable video
streaming over mobile WiMAX,” in Proceedings of IEEE Int. Sympo-
sium on Circuits and Systems (ISCAS), May 2007, pp. 3463–3466.
[23] P. Li, W. Lin, S. Rahardja, X. Lin, X. Yang, and Z. Li, “Geometrically
determining the leaky bucket parameters for video streaming over con-
stant bit-rate channels,” Signal Processing: Image Communication, vol.
20, no. 2, pp. 193–204, Feb. 2005.
[24] D. T. Nguyen and J. Ostermann, “Congestion control for scalable
video streaming using the scalability extension of H.264/AVC,” IEEE
Journal of Selected Topics in Signal Processing, vol. 1, no. 2, pp.
246–253, Aug. 2007.
[25] T. Ozcelebi, A. Tekalp, and M. Civanlar, “Delay-distortion optimiza-
tion for content-adaptive video streaming,” IEEE Trans. Multimedia,
vol. 9, no. 4, pp. 826–836, Jun. 2007.
[26] M. van der Schaar, Y. Andreopoulos, and Z. Hu, “Optimized scalable
video streaming over IEEE 802.11a/e HCCA wireless networks under
delay constraints,” IEEE Trans. Mobile Computing, vol. 5, no. 6, pp.
755–768, Jun. 2006.
[27] T. Schierl, K. Ganger, C. Hellge, T. Wiegand, and T. Stockhammer,
“SVC-based multisource streaming for robust video transmission in
mobile ad hoc networks,” IEEE Wireless Communications, vol. 13, no.
5, pp. 96–103, Oct. 2006.
[28] A. Puri, X. Chen, and A. Luthra, “Video coding using the
H.264/MPEG-4 AVC compression standard,” Journal of Visual
Communication and Image Representation, vol. 19, no. 9, pp.
793–849, Oct. 2004.
[29] J. Ostermann, J. Bormans, P. List, D. Marpe, M. Narroschke, F. Pereira,
T. Stockhammer, and T. Wedi, “Video coding with H.264/AVC: Tools,
performance and complexity,” IEEE Circuits and Systems Magazine,
vol. 4, no. 1, pp. 7–28, First Quarter, 2004.
Authorized licensed use limited to: Arizona State University. Downloaded on August 31, 2009 at 13:04 from IEEE Xplore. Restrictions apply.

VAN DER AUWERA AND REISSLEIN: STATISTICAL MULTIPLEXING OF H.264/AVC AND SVC VIDEO STREAMS 557
[30] G. Sullivan, P. Topiwala, and A. Luthra, “The H.264/AVC advanced
video coding standard: Overview and introduction to the fidelity range
extensions,” in Proc. of SPIE 5558, Conference on Applications of
Digital Image Processing XXVII, Special Session on Advances in
New Emerging Standard: H.264/AVC I, Denver, CO, Aug. 2004, pp.
454–474.
[31] Information Technology–Generic Coding of Audio-Visual Ob-
jects–Part 2: Visual, Final Proposed Draft Amendment 1, ISO/IEC
JTC 1/SC 29/WG 11 N2802, , Geneva, Jul. 1999.
[32] P. Seeling, M. Reisslein, and B. Kulapala, “Network performance eval-
uation with frame size and quality traces of single-layer and two-layer
video: A tutorial,” IEEE Communications Surveys and Tutorials vol. 6,
no. 3, pp. 58–78, Third Quarter, 2004 [Online]. Available: http://trace.
eas.asu.edu, video traces available at
[33] S. Bakiras and V. O. K. Li, “Maximizing the number of users in an
interactive video-on-demand system,” IEEE Trans. Broadcasting, vol.
48, no. 4, pp. 281–292, Dec. 2002.
[34] P. Koutsakis and M. Paterakis, “Policing mechanisms for the transmis-
sion of videoconference traffic from MPEG-4 and H.263 video coders
in wireless ATM networks,” IEEE Trans. Vehicular Technology, vol.
53, no. 5, pp. 1525–1530, 2004.
[35] B. Nikolaus, J. Ott, C. Borrmann, and U. Borrmann, “Generalized
greedy broadcasting for efficient media-on-demand transmissions,”
IEEE Trans. Broadcasting, vol. 51, no. 3, pp. 354–359, 2005.
[36] J. Roberts, “Internet traffic, QoS, and pricing,” Proceedings of the
IEEE, vol. 92, no. 9, pp. 1389–1399, 2004.
[37] Y. Xu and R. Guerin, “Individual QoS versus aggregate QoS: A loss
performance study,” IEEE/ACM Trans. Networking, vol. 13, no. 2, pp.
370–383, 2005.
[38] X.-D. Huang, Y.-H. Zhou, and R.-F. Zhang, “A multiscale model for
MPEG-4 varied bit rate video traffic,” IEEE Trans. Broadcasting, vol.
50, no. 3, pp. 323–334, Sep. 2004.
[39] C. H. Liew, C. K. Kodikara, and A. M. Kondoz, “MPEG-encoded vari-
able bit-rate video traffic modelling,” IEE Proceedings Communica-
tions, vol. 152, no. 5, pp. 749–756, Oct. 2005.
[40] U. K. Sarkar, S. Ramakrishnan, and D. Sarkar, “Modeling full-length
video using Markov-modulated gamma-based framework,” IEEE/ACM
Trans. Networking, vol. 11, no. 4, pp. 638–649, Aug. 2003.
[41] U. K. Sarkar, S. Ramakrishnan, and D. Sarkar, “Study of long duration
MPEG-trace segmentation methods for developing frame size based
traffic models,” Computer Networks, vol. 44, no. 2, pp. 177–188, 2004.
[42] M. Dai and D. Loguinov, “Analysis and modeling of MPEG-4 and
H.264 multi-layer video traffic,” in Proc. of IEEE INFOCOM, Miami,
FL, Mar. 2005, pp. 2257–2267.
[43] D. Fiems, V. Inghelbrecht, B. Steyaert, and H. Bruneel, “Markovian
characterization of H.264/SVC scalable video,” in Proceedings of 15th
Int. Conference on Analytical and Stochastic Modeling Techniques and
Applications (ASMTA), Jun. 2008, Lecture Notes in Computer Science
5055, pp. 1–15.
[44] S. Kempken and W. Luther, “Modeling of H.264 high definition video
traffic using discrete-time semi-Markov processes,” in Proceedings of
20th Int. Teletraffic Congress (ITC), Jun. 2007, Lecture Notes in Com-
puter Science 4516, pp. 42–53.
[45] C. Bewick, R. Pereira, and M. Merabti, “Network constrained
smoothing: Enhanced multiplexing of MPEG-4 video,” in Pro-
ceedings of IEEE International Symposium on Computers and
Communications, Taormina, Italy, Jul. 2002, pp. 114–119.
[46] H.-C. Chao, C. L. Hung, and T. G. Tsuei, “ECVBA traffic-smoothing
scheme for VBR media streams,” International Journal of Network
Management, vol. 12, pp. 179–185, 2002.
[47] W.-C. Feng and J. Rexford, “Performance evaluation of smoothing al-
gorithms for transmitting prerecorded variable-bit-rate video,” IEEE
Trans. Multimedia, vol. 1, no. 3, pp. 302–312, Sep. 1999.
[48] T. Gan, K.-K. Ma, and L. Zhang, “Dual-plan bandwidth smoothing
for layer-encoded video,” IEEE Trans. Multimedia, vol. 7, no. 2, pp.
379–392, Apr. 2005.
[49] C.-D. Iskander and R. T. Mathiopoulos, “Online smoothing of VBR
H.263 video for the CDMA2000 and IS-95B uplinks,” IEEE Trans.
Multimedia, vol. 6, no. 4, pp. 647–658, Aug. 2004.
[50] M. Krunz, W. Zhao, and I. Matta, “Scheduling and bandwidth allo-
cation for distribution of archived video in VoD systems,” Journal of
Telecommunication Systems, Special Issue on Multimedia, vol. 9, no.
3/4, pp. 335–355, Sep. 1998.
[51] M. Krunz, “Bandwidth allocation strategies for transporting vari-
able–bit–rate video traffic,” IEEE Communications Magazine, vol. 37,
no. 1, pp. 40–46, Jan. 1999.
[52] H. Lai, J. Y. Lee, and L.-K. Chen, “A monotonic-decreasing rate sched-
uler for variable-bit-rate video streaming,” IEEE Trans. Circuits and
Systems for Video Technology, vol. 15, no. 2, pp. 221–231, Feb. 2005.
[53] A. Solleti and K. J. Christensen, “Efficient transmission of stored
video for improved management of network bandwidth,” International
Journal of Network Management, vol. 10, pp. 277–288, 2000.
[54] B. Vandalore, W.-C. Feng, R. Jain, and S. Fahmy, “A survey of applica-
tion layer techniques for adaptive streaming of multimedia,” Real-Time
Imaging Journal, vol. 7, no. 3, pp. 221–235, 2001.
[55] D. Ye, J. Barker, Z. Xiong, and W. Zhu, “Wavelet-based VBR video
traffic smoothing,” IEEE Trans. Multimedia, vol. 6, no. 4, pp. 611–623,
Aug. 2004.
[56] Z. Zhang, J. Kurose, J. Salehi, and D. Towsley, “Smoothing, statistical
multiplexing and call admission control for stored video,” IEEE
Journal on Selected Areas in Communications, vol. 13, no. 6, pp.
1148–1166, Aug. 1997.
[57] Z. Antoniou and I. Stavrakakis, “An efficient deadline-credit-based
transport scheme for prerecorded semisoft continuous media applica-
tions,” IEEE/ACM Trans. Networking, vol. 10, no. 5, pp. 630–643, Oct.
2002.
[58] J. C. H. Yuen, E. Chan, and K.-Y. Lam, “Real time video frames allo-
cation in mobile networks using cooperative pre-fetching,” Multimedia
Tools and Applications, vol. 32, no. 3, pp. 329–352, Mar. 2007.
[59] Y.-W. Leung and T. K. C. Chan, “Design of an interactive video-on-
demand system,” IEEE Trans. Multimedia, vol. 5, no. 1, pp. 130–140,
Mar. 2003.
[60] F. Li and I. Nikolaidis, “Trace-adaptive fragmentation for periodic
broadcast of VBR video,” in Proceedings of 9th International Work-
shop on Network and Operating Systems Support for Digital Audio
and Video (NOSSDAV), Basking Ridge, NJ, Jun. 1999, pp. 253–264.
[61] C.-S. Lin, M.-Y. Wu, and W. Shu, “Transmitting variable-bit-rate
videos on clustered VOD systems,” in Proceedings of IEEE Interna-
tional Conference on Multimedia and Expo (ICME), New York, Jul.
2000.
[62] S. Oh, Y. Huh, B. Kulapala, G. Konjevod, A. W. Richa, and M.
Reisslein, “A modular algorithm-theoretic framework for the fair and
efficient collaborative prefetching of continuous media,” IEEE Trans.
Broadcasting, vol. 51, no. 2, pp. 200–215, Jun. 2005.
[63] S. Oh, B. Kulapala, A. W. Richa, and M. Reisslein, “Continuous-time
collaborative prefetching of continuous media,” IEEE Trans. Broad-
casting, vol. 54, no. 1, pp. 36–52, Mar. 2008.
[64] M. Reisslein and K. W. Ross, “High–performance prefetching proto-
cols for VBR prerecorded video,” IEEE Network, vol. 12, no. 6, pp.
46–55, Nov./Dec. 1998.
[65] S. Racz, T. Jakabfy, J. Farkas, and C. Antal, “Connection admission
control for flow level QoS in bufferless models,” in Proc. IEEE IN-
FOCOM, 2005, pp. 1273–1282.
[66] M. Reisslein and K. W. Ross, “Call admission for prerecorded sources
with packet loss,” IEEE Journal on Selected Areas in Communications,
vol. 15, no. 6, pp. 1167–1180, Aug. 1997.
[67] J. Roberts, U. Mocci, and J. Virtamo, Broadband Network Traffic: Per-
formance Evaluation and Design of Broadband Multiservice Networks,
Final Report of Action COST 242. New York: Springer Verlag, 1996,
vol. 1155, Lecture Notes in Computer Science.
[68] A. Elwalid, D. Heyman, T. Lakshman, D. Mitra, and A. Weiss, “Fun-
damental bounds and approximations for ATM multiplexers with ap-
plications to video teleconferencing,” IEEE Journal on Selected Areas
in Communications, vol. 13, no. 6, pp. 1004–1016, Aug. 1995.
[69] F. Y.-S. Lin, “Optimal real-time admission control algorithms for the
video-on-demand (VOD) service,” IEEE Trans. Broadcasting, vol. 44,
no. 4, pp. 402–408, Dec. 1998.
[70] N. Shroff and M. Schwartz, “Improved loss calculations at an ATM
multiplexer,” IEEE/ACM Trans. Networking, vol. 6, no. 4, pp. 411–421,
Aug. 1998.
[71] J. McManus and K. Ross, “Video-on-demand over ATM: Constant-
rate transmission and transport,” IEEE Journal on Selected Areas in
Communications, vol. 14, no. 6, pp. 1087–1098, Aug. 1996.
[72] S. Sen, J. L. Rexford, J. K. Dey, J. F. Kurose, and D. F. Towsley, “On-
line smoothing of variable-bit-rate streaming video,” IEEE Trans. Mul-
timedia, vol. 2, no. 1, pp. 37–48, Mar. 2000.
[73] Y. Bai and M. Ito, “Application-aware buffer management: New met-
rics and techniques,” IEEE Trans. Broadcasting, vol. 51, no. 1, pp.
114–121, Mar. 2005.
[74] Y. Huang, R. Guerin, and P. Gupta, “Supporting excess real-time traffic
with active drop queue,” IEEE Trans. Networking, vol. 14, no. 5, pp.
965–977, Oct. 2006.
[75] G.-M. Muntean, P. Perry, and L. Murphy, “A new adaptive multi-
media streaming system for all-IP multi-service networks,” IEEE
Trans. Broadcasting, vol. 50, no. 1, pp. 1–10, Mar. 2004.
[76] S. Ryu, C. Rump, and C. Qiao, “Advances in internet congestion con-
trol,” IEEE Communications Surveys and Tutorials, vol. 5, no. 1, pp.
28–39, 2003.
Authorized licensed use limited to: Arizona State University. Downloaded on August 31, 2009 at 13:04 from IEEE Xplore. Restrictions apply.

558 IEEE TRANSACTIONS ON BROADCASTING, VOL. 55, NO. 3, SEPTEMBER 2009
[77] H. Schwarz, D. Marpe, and T. Wiegand, “Analysis of hierarchical B
pictures and MCTF,” in IEEE Int. Conf. Multimedia and Expo (ICME),
Toronto, Canada, Jul. 2006, pp. 1929–1932.
Geert Van der Auwera received the Ph.D. degree
in Electrical Engineering from Arizona State Univer-
sity, Tempe, USA, in 2007, and the Belgian MSEE
degree from Vrije Universiteit Brussel (VUB), Brus-
sels, Belgium, in 1997. Presently, he is Staff Research
Engineer with Samsung Electronics in Irvine, CA.
His research interests are video coding, video traffic
and quality characterization, video streaming mech-
anisms and protocols. Until the end of 2004, he was
Scientific Advisor with IWT-Flanders, the Institute
for the Promotion of Innovation by Science and Tech-
nology in Flanders, Belgium. In 2000, he joined IWT-Flanders after researching
wavelet video coding at IMEC’s Electronics and Information Processing De-
partment (VUB-ETRO) in Brussels, Belgium. In 1998, his thesis on motion es-
timation in the wavelet domain received the Barco and IBM prizes by the Fund
for Scientific Research of Flanders, Belgium.
Martin Reisslein is an Associate Professor in the De-
partment of Electrical Engineering at Arizona State
University (ASU), Tempe. He received the Dipl.-Ing.
(FH) degree from the Fachhochschule Dieburg, Ger-
many, in 1994, and the M.S.E. degree from the Uni-
versity of Pennsylvania, Philadelphia, in 1996. Both
in electrical engineering. He received his Ph.D. in
systems engineering from the University of Pennsyl-
vania in 1998. During the academic year 1994-1995
he visited the University of Pennsylvania as a Ful-
bright scholar. From July 1998 through October 2000
he was a scientist with the German National Research Center for Information
Technology (GMD FOKUS), Berlin and lecturer at the Technical University
Berlin. From October 2000 through August 2005 he was an Assistant Professor
at ASU.
He served as editor-in-chief of the IEEE Communications Surveys and Tuto-
rials from January 2003 through February 2007 and has served on the Technical
Program Committees of IEEE Infocom and numerous other networking con-
ferences. He maintains an extensive library of video traces for network perfor-
mance evaluation, including frame size traces of MPEG-4 and H.264 encoded
video, at http://trace.eas.asu.edu. His research interests are in the areas of In-
ternet Quality of Service, video traffic characterization, wireless networking,
optical networking, and engineering education.
Authorized licensed use limited to: Arizona State University. Downloaded on August 31, 2009 at 13:04 from IEEE Xplore. Restrictions apply.