Accelerat ing t he world's research.

Reliable Multimedia Download

Delivery in Cellular Broadcast
Networks
Michael Luby

IEEE Transactions on Broadcasting

Cite this paper Downloaded from Academia.edu 

Get the citation in MLA, APA, or Chicago styles

Related papers Download a PDF Pack of t he best relat ed papers 

Mobile Dat a Broadcast ing over MBMS Tradeoffs in Forward Error Correct ion
Yajun Kou

Mult icast Delivery of File Download Services in Evolved 3G Mobile Net works Wit h HSDPA and MBMS
David Gómez-Barquero

PERFORMANCE ANALYSIS OF AL-FEC RAPT OR CODE OVER 3GPP EMBMS NET WORK
Edit or IJRET
IEEE TRANSACTIONS ON BROADCASTING, VOL. 53, NO. 1, MARCH 2007
Reliable Multimedia Download Delivery in Cellular
Broadcast Networks
Michael Luby, Tiago Gasiba, Thomas Stockhammer, and Mark Watson
Abstract—Raptor codes have been standardized as application
layer forward error correction (FEC) codes for Multimedia Broad-
cast and Multicast Services (MBMS) and Digital Video Broadcast
(DVB) due to their extraordinary advanced FEC protection and
performance. Raptor codes are known to have characteristics very
close to ideal and provide a wide range of operating flexibility and
efficiency unmatched by other codes, and in particular they are an
excellent implementation of fountain codes. Until now, investiga-
tions of the application of Raptor codes to UMTS and EPGRS have
used an overall system model that does not accurately model the
physical channel and user mobility, simplistically assuming inde-
pendent random packet losses at the application layer. We investi-
gate MBMS in UMTS in a much more realistic and complete simu-
lation environment by considering advanced and complete channel
models that simulate the physical channel and user mobility in
a cellular network. We use this realisitic simulation environment
to determine optimal system parameters under different mobility
models, with different bearer parameters, and without and with
selective combining. More specifically, we investigate joint settings
of the Raptor code rate, the Turbo code rate, transmission power,
etc., to find settings which provide reliable download delivery of
files using minimal transmission energy. One of our main results
is that optimal system-wide operating points use low transmission
power and a modest amount of Turbo coding that results in rela-
tively large radio packet loss rates that is compensated for by using
a substantial amount of Raptor coding. These optimal operating
points use far less transmission energy for download delivery of
files than possible operating points without Raptor.
Index Terms—Application layer FEC, code rate, cross-layer op-
timization, FEC, forward error correction, fountain codes, MBMS,
Raptor code, reliable multicast transmission, UMTS.
I. INTRODUCTION
ELIABLE delivery of files of all sizes simultaneously to
R many users over unreliable and bandwidth-limited net-
works is an extremely challenging task. The need to supply
mobile and heterogeneous users with large chunks of data and
multimedia information is more desired and required than ever
in many different application scenarios. Examples include the
delivery of news, software updates, or other popular content to
mobile clients such as handhelds and cars, or the distribution of
location-dependent information in highly populated areas, e.g.
airports, sport stadiums, traffic and weather information, etc.
Manuscript received July 14, 2006; revised December 1, 2006.
The authors are with Digital Fountain, Fremont, CA 94538 USA. (e-mail:
luby@digitalfountain.com; gasiba@digitalfountain.com; stockhammer@digi-
talfountain.com; mark@digitalfountain.com).
Color versions of one or more of the figures in this paper are available online
at http://ieeexplore.ieee.org.
Digital Object Identifier 10.1109/TBC.2007.891703
0018-9316/$25.00 © 2007 IEEE
Authorized licensed use limited to: UNIVERSIDADE DE BRASILIA. Downloaded on August 17, 2009 at 12:08 from IEEE Xplore. Restrictions apply.
235
It has been recognized that broadcast networks must be used
for the efficient distribution of such content. Due to the high
penetration of IP-based applications, IP-based datacast services
are most attractive. The IETF has defined protocols which allow
to deliver files to many users using a broadcast or multicast net-
work. The most popular protocol for download delivery of files
is the File Delivery over Unidirectional Transport (FLUTE) pro-
tocol [1] which is particularly suited for such networks.
3GPP has defined a Multimedia Broadcast and Multicast Ser-
vice (MBMS), which extends the existing architecture by the
introduction of an MBMS bearer service and MBMS User Ser-
vices [2]. The former is provided by the packet-switched do-
main to deliver IP multicast datagrams to multiple receivers
using minimum radio and network resources. The bearer ser-
vices re-uses many existing components of the UMTS system
such as radio access including physical layer coding based on
Turbo codes. The end user is provided by two MBMS user ser-
vices [3], download delivery for reliable multicasting of files
and streaming delivery to deliver real-time multimedia streams.
Thereby, a new architectural component, the Broadcast Mul-
ticast Service Center (BM-SC) provides the MBMS User ser-
vices to the User Equipment (UE). Both application layer de-
livery methods in MBMS mandate that the UE supports Raptor
codes [3]–[5]. Hence, the MBMS system includes forward error
correction on two layers, Turbo codes at the physical layer and
Raptor codes at the application layer.
Commonly, for unicast bearers in UMTS, the Quality-of-Ser-
vice (QoS) is determined and guaranteed by the physical layer
FEC and the residual IP packet loss rates are quite low, usually
at most 1%. This is essential as common IP-based applications
cannot withstand packet losses without sacrificing quality or ef-
ficiency or both. However, with the introduction of broadcast/
multicast services and application layer FEC, the QoS tradeoff
between the two layers is not obvious, especially considering the
time-varying and heterogeneous reception conditions in cellular
mobile broadcast. Therefore, in this work we take a comprehen-
sive approach and investigate different configurations of phys-
ical layer resources such as transmit power and the Turbo code
rate as well as the Raptor code rate such that resource consump-
tion is minimized and user perception is maximized.
This work focuses on download delivery services. We intro-
duce the MBMS Download Delivery method and the MBMS
protocol stack in Section II. Section III gives a brief overview of
Raptor codes as specified in MBMS. In Section IV we introduce
our comprehensive system level simulation approach which in-
cludes models for the mobile radio channel, the user behavior,
the physical layer code, and the user experience. Selected sim-
ulation results are provided in Section V before concluding the
paper in Section VI.
236
Fig. 1. Simplified MBMS system architecture.
II. MULTIMEDIA DELIVERY SERVICES IN MBMS
A. MBMS Architecture
MBMS is a point-to-multipoint service in which data is trans-
mitted from a single source entity to multiple recipients. Trans-
mitting the same data to multiple recipients allows network re-
sources to be shared. The MBMS bearer service offers a Broad-
cast Mode and a Multicast Mode. MBMS architecture should
enable the efficient usage of radio-network and core-network
resources, with an emphasis on radio interface efficiency, i.e.
transmission power efficiency. In the bearer plane, this service
provides delivery of IP Multicast datagrams to UEs. A new func-
tional entity, the Broadcast Multicast Service Center provides a
set of functions for MBMS User Services. The system architec-
ture is shown in Fig. 1.
The MBMS User Service architecture is based on an MBMS
receiver on the UE side and a BM-SC on the network side.
Reception of an MBMS service is enabled by different phases
such as subscription, joining, data transfer, and leaving. Broad-
cast services do not include joining and leaving phases, but the
data transfer phase is common and identical for both, broad-
cast and multicast. In the remainder of this work we focus on
the data transfer phase exclusively. Furthermore, we concentrate
on MBMS delivery over UTRAN and specifically focus on the
transmission power efficiency.
B. MBMS Protocol Stack
MBMS defines two delivery methods—download and
streaming delivery. MBMS delivery methods make use of
MBMS bearers for content delivery but may also use the asso-
ciated delivery procedures for quality reporting and file repair.
The MBMS protocol stack related to the data delivery phase
for streaming and download delivery is shown in Fig. 2.
Streaming data such as video streams, audio programs or
timed text being encapsulated in RTP are transported over
the streaming delivery network. In this case, application layer
Raptor codes are applied on UDP flows, either individually
or on bundles of streams. The streaming framework provides
significant flexibility in terms of code rates, protection periods,
etc. [5]. Discrete objects such as still images, multimedia
streams encapsulated in file formats, or other binary data are
transported using the FLUTE protocol (RFC 3926 [1]) when
delivering content over MBMS bearers. In both delivery ser-
vices the resulting UDP flows are generally mapped on the
MBMS IP multicast bearers.
The MBMS bearer services reuses most of the legacy
UMTS protocol stack in the packet-switched domain. Only
Authorized licensed use limited to: UNIVERSIDADE DE BRASILIA. Downloaded on August 17, 2009 at 12:08 from IEEE Xplore. Restrictions apply.
IEEE TRANSACTIONS ON BROADCASTING, VOL. 53, NO. 1, MARCH 2007
Fig. 2. MBMS protocol stack.
minor modifications are introduced to support MBMS. The IP
packets are processed in the Packet Data Convergence Protocol
(PDCP) layer where for example header compression might
be applied. In the Radio Link Control (RLC) the resulting
PDCP-PDUs, generally of arbitrary length, are mapped to fixed
length RLC-PDUs. The RLC layer operates in unacknowl-
edged mode due to the unavailable feedback links on the radio
access network. Functions provided at the RLC layer are for
example segmentation and reassembly, concatenation, padding,
sequence numbering, reordering and out-of-sequence and du-
plication detection. The Medium Access Control (MAC) layer
maps and multiplexes the RLC-PDUs to the transport channel
and selects the transport format depending on the instantaneous
source rate. The MAC layer and physical layer appropriately
adapt the RLC-PDU to the expected transmission conditions by
applying, among others, channel coding, power and resource
assignment, and modulation.
C. MBMS Bearer Service Over UMTS
The UMTS bearer provides services with different QoS
which are fundamental to support the MBMS broadcast
mode. Radio bearers are specified among others by the data
throughput, the data transport format, physical layer channel
coding, rate matching, power allocation, and many other
things. MBMS uses the Multimedia Traffic Channel (MTCH),
which enables point-to-multipoint distribution. This channel is
mapped onto the Forward Access Channel (FACH) which is
finally mapped into the Secondary-Common Control Physical
Channel S-CCPCH physical channel [6]. Among others, an
MBMS radio bearer is defined by the transport format size
and number of transport blocks that are to be protected by
physical-layer channel code at every transmission time interval
(TTI). The TTI is transport-channel specific and can be selected
from the set {10 ms, 20 ms, 40 ms, 80 ms} for MBMS. The
higher TTI values are supported by increasing the interleaving
of the channel code and/or using longer codeword sizes of
channel code. This has the advantage of increasing time diver-
sity but at the expense of longer delays.
LUBY et al.: RELIABLE MULTIMEDIA DOWNLOAD DELIVERY IN CELLULAR BROADCAST NETWORKS
Fig. 3. Structure of the UMTS radio frame.
TABLE I
BEARER PHYSICAL LAYER PARAMETERS
Refer again to Fig. 2. After RLC layer processing the re-
sulting RLC/MAC blocks are mapped into the transport blocks
according to the specified radio bearer settings and a 16 bit CRC
is appended. Afterward these blocks might be concatenated and
then further segmented into code blocks such that the maximum
length of a code block does not exceed 5114 bits [7], [8]. This
limitation comes from the restrictions on complexity, memory
and power consumptions of Turbo decoders in handheld de-
vices. After the Turbo coding is applied, the resulting blocks are
concatenated, interleaved and eventually rate matching is per-
formed such that the code rate of the physical layer FEC, , Specifically, if CRC for a transport block fails after Turbo
can be adapted between . The resulting code-
word is mapped to one or more transmission slots which are
finally mapped into radio frames, as shown in Fig. 3. A radio
frame consists of 15 slots, whereby the number of bits per slot
depends on the applied spreading code. Finally, radio frames
are transmitted every 10 ms using a chip-rate of 3.84 Mcps and
QPSK.
For interoperability and testing purpose a number of MBMS
reference radio bearers have been defined as a preferred config-
uration [9] for a specific data rate. The configurations define a
single default radio parameter set including physical layer code
rates, transport format, transmission time interval, symbol rate,
etc. Since our main goal is to investigate the trade-off between
the physical layer FEC and the application layer FEC, only a
subset of the physical layer parameters of the typical MBMS
bearers are adjusted as we vary the application layer parameters.
The Turbo code rate, is not fixed, but kept flexible. Note
that the configurations still comply to the specification. Conse-
quently, we specify an MBMS bearer not by its bitrate at the
RLC layer, but instead by the symbol rate at the physical layer.
The considered bearers with their respective settings are pro-
vided in Table I. The variation of physical layer code rate will
result in varying RLC bit rates: higher code rates offer higher
data rates but also result in higher RLC-PDU loss rates whereas
more physical layer overhead lowers the loss rates, but also the
RLC data rates.
Authorized licensed use limited to: UNIVERSIDADE DE BRASILIA. Downloaded on August 17, 2009 at 12:08 from IEEE Xplore. Restrictions apply.
237
Fig. 4. Selective combining.
At the receiver side, the inverse operations are applied.
decoding, this block is declared as an erasure. All correct data
is delivered to the upper layers. At RLC layer, only correctly
received RLC-SDUs (usually containing an entire IP packet)
are delivered to higher layer. Therefore, any IP packet which
is partly contained in an erroneous transport block, will be lost
and will not be available at the receiver.
To improve receiver performance especially at cell edges,
a mobile terminal can perform selective combining, which is
a simple form of receiver diversity. Applying this, an MBMS
UE listens to more than one Node-B signal simultaneously, as
shown in Fig. 4, and individually decodes the streams. Obvi-
ously, selective combining can only be applied if the MBMS
signal is transmitted from multiple transmission sites using the
same bearer service parameters. The receiver decodes the mul-
tiple signals in the hope that for at least one signal the CRC does
not fail and the correct RLC-PDU can be forwarded to the upper
layer. However, note also that selective combining increases the
complexity and costs of a receiver. Alternative implementations
might only decode the strongest signal. An even more advanced
receiver technique is referred to as soft combining which appro-
priately combines the MBMS signals from different transmis-
sion sites before the channel decoding. However, to provide this
feature to the MBMS receivers, the network requires transmis-
sion sites which strictly synchronize the transmission of MBMS
streams. An early deployment of soft combining in networks is
not foreseen.
238
D. MBMS Download Delivery Service
To deliver a file1 in a broadcast session, FLUTE provides
mechanisms to signal and map the properties of a file to Asyn-
chronous Layered Coding (ALC) [10] protocol such that re-
ceivers can assign these parameters to the received files. The
file is partitioned in one or several so-called source blocks. Each
source block represents a single channel coding word for the
upper layer FEC. Each source block is split into so-called
source symbols, each of length except for the last one, which
can be smaller. Both parameters and are signaled in the ses-
sion setup and are fixed for one session. For each source block,
additional repair symbols can be generated by applying Raptor
encoding which is explained in some detail in Section III. Each
encoding symbol, i.e. source and repair symbols, gets assigned
a unique encoded symbol ID (ESI) which identifies the given
symbol. In particular, if the ESI is less than , then it is a source
symbol, otherwise it is called a repair symbol. Let us denote
the total number of encoding symbols as and define the re-
sulting code rate of the Raptor code as . Either
individual or multiple symbols are mapped to a FLUTE packet
payload. The source block number as well as ESI of the first en-
coded symbol in the packet are signaled in the FLUTE header.
FLUTE packets themselves are then encapsulated in UDP and
then distributed over the IP multicast MBMS bearer.
Receivers collect correctly received FLUTE packets, and
with the information available in the packet header and the
file session setup, the structure of the source block can be
recovered. If no more encoded symbols belonging to the actual
source block are expected to be received, the Raptor decoder,
attempts to recover the source block from all received data.
Obviously, due to heterogeneous receiving conditions, the
set of received encoded symbols typically differs among the
receivers. Nevertheless, with exploitation of the repair data,
most receivers should be able to recover the source block. If all
source blocks belonging to the file are received correctly, the
entire file is recovered. Only if recovery fails, a post-delivery
repair phase might be invoked which is not further discussed.
III. RAPTOR CODES
Raptor codes were introduced in [4], [11], [12] and have re-
cently been standardized by 3GPP [3]. In this section we will
briefly summarize the encoding and decoding algorithms of sys-
tematic Raptor codes as specified in [3, Annex B]. More details
on the notations can be found in [5]. Raptor codes in general
consist of an inner high-rate block code followed by a Luby
transform (LT) encoder [13], [14] with some generation ma-
trix . The encoding and decoding procedures make use of
the concept of a code constraint processor as shown in Fig. 5.
The code constraint processor basically inverts a constraint ma-
trix to obtain intermediate symbols which
serve as the input to an LT encoder. contains
the constraints of the inner code as well of the LT code taking
into account ESIs . The construction of the Raptor
code is such that the first encoded symbols are
equivalent to the source symbols such that the LT encoding
1For simplicity we continue with the notion of a file in the following though
the ALC/LCT concept uses the more general terminology transport object.
Authorized licensed use limited to: UNIVERSIDADE DE BRASILIA. Downloaded on August 17, 2009 at 12:08 from IEEE Xplore. Restrictions apply.
IEEE TRANSACTIONS ON BROADCASTING, VOL. 53, NO. 1, MARCH 2007
Fig. 5. Practical implementation of a Raptor code system.
process can also be used at the decoder to obtain the source
symbols from the intermediate symbols. The code in [3, Annex
B] is constructed such that is invertible for any
.
At the decoder, only if an appropriate set of encoded sym-
bols with is available such that the ma-
trix has full rank, is decoding successful. In
average, the number of necessary encoded symbols, , is only
slightly more than .
The conventional receiver performs Raptor decoding on en-
coded symbols, i.e. symbols that have been received correctly
from the channel. Note that if more than one symbol is con-
tained in a UDP/IP packet, in case of the loss of an IP packet, all
symbols in this packet are lost. At the receiver the correctly re-
ceived symbols are processed by the code constraints processor
that computes the original intermediate symbols. These are then
forwarded to the LT encoder which, due to the systematic Raptor
code, is enable to recover the source symbols and consequently
the original information.
Due to the specific properties of the decoding matrix, namely
the fact that it is a sparse matrix, a very efficient implementa-
tion of the code constraint processor is possible by applying a
so-called decoding schedule as proposed in [15]. A practical im-
plementation of Raptor codes for MBMS [5] results in linear-
time encoding and decoding. Raptor codes are thus very attrac-
tive, since they can be implemented with very low complexity,
have a performance very close to ideal and can encode very
large amounts of data with complexity linearly proportional to
the amount of data processed and repair generated.
IV. MBMS SYSTEM-LEVEL SIMULATION
A. Simulation Setup
The assessment of the performance of MBMS User Service
performance is quite difficult due to the many different pa-
rameters in the system, the long time-scales necessary to be
considered, and the complexity of the involved components.
LUBY et al.: RELIABLE MULTIMEDIA DOWNLOAD DELIVERY IN CELLULAR BROADCAST NETWORKS
Fig. 6. Concept of simulation setup for MBMS download delivery over UMTS.
Therefore, we have developed a comprehensive MBMS sim-
ulator which takes into account realistic effects of MBMS
download delivery over UMTS. The concept of the simulation
approach is shown in Fig. 6.
The simulator is composed of different modules which simu-
late and/or model different components of the entire system. The
simulator is mainly divided in three blocks, the mobile cellular
channel model, physical layer code, and the MBMS User Ser-
vice protocol stack including all protocols and the Raptor coding
and decoding process. In principle, the mobile cellular channel
model generates traces for the carrier, interference and noise
present at the mobile terminal and also the observed orthogo-
nality factor (required to compute self-interference), with a res-
olution of 2 ms and for as many as different users with dif-
ferent random initial position, whereby each trace corresponds
to values captured over 10 minutes. The number of users in our
case is . every TTI is required. This is achieved using a link error pre-
These traces, which are generated for normalized transmit
power and no spreading gain, are then modified to obtain an ef-
fective SINR for each TTI by applying some appropriate com-
bining, referred to as Equivalent SNR Method based on Convex
Metric (ECM), power assignment and spreading code assign-
ment. A resulting SINR is obtained for each TTI, which is con-
verted to a sequence of RLC-PDU loss traces by applying a
suitable table lookup for the physical layer code. The resulting
RLC-PDU loss traces for each individual user are then applied
to the generated FLUTE transmission session and for each user
it is evaluated if the transmitted file can be correctly recovered.
The individual components of the simulator as well as the ap-
plied modeling is explained in further detail in the remainder of
this section.
B. Mobile Radio Channel in Multicellular Environment
The mobile radio channel places fundamental limits on the
performance of a wireless communication system. Unlike wired
Authorized licensed use limited to: UNIVERSIDADE DE BRASILIA. Downloaded on August 17, 2009 at 12:08 from IEEE Xplore. Restrictions apply.
239
channels which are more stationary and predictable, radio chan-
nels show extremely varying behavior. In fact, the radio trans-
mission path between the transmitter and the receiver can vary
from a simple line-of-sight to one that is severely obstructed by
buildings, mountains, foliage, etc. Also the speed of the mobile
terminal has a great impact on the received radio signal.
Therefore, in order to evaluate the impact of the mobile radio
in a multicellular environment, a channel model that uses stan-
dard models and techniques has been defined and developed in
3GPP [16], partly based on real measurements. In particular, ef-
fects such as path loss, Doppler spreads, shadowing, antenna
radiation pattern and interference are appropriately dealt with.
This channel model allows to simulate pedestrian and vehic-
ular mobile users within a cell, whereby the main difference
between these users is the speed at which they move within the
cell, but also their power-delay profile. Fig. 7 shows examples
of movements of users in a cell for different speeds and dif-
ferent starting points. The applied movement model is based on
random walk with high correlation. The users do not leave the
area, but bounce at the cell edges. However, possible hand-over
effects are simulated as the signal is not necessarily received
from the base station assigned to the hexagonal cell, but from
the one with the highest SINR. The figure shows the position
of four different users for a time of 10 min. Notice that the ve-
hicular user undergoes a much larger distance due to his speed
of 30 km/h, while the pedestrian users at 3 km/h covers less
distance in the same amount of time. The lower part shows the
signal to interference and noise ratio (SINR) for each of the
users. The SINR is varying due to large scale effects such as at-
tenuation and shadowing on both, useful signals and interferers,
as well as due to short-term effects such as fading. Note also that
for the vehicular user, the SINR shows much stronger variations
than for pedestrian.
1) ECM Method: The generic mobile radio simulator com-
putes the signal to interference and noise ratio every 2 ms which
is for example the TTI in HSDPA. However, the investigated
MBMS bearers use transmission time intervals of up to 80 ms.
Therefore, an appropriate conversion of the effective SINR for
diction method called Equivalent SNR Method based on Convex
Metric (ECM), as defined in [16]. This technique allows to com-
bine several SINR values into a single effective SINR due to the
fact that interleaving over these multiple channel access slots is
performed. The method is based on Shannon’s channel capacity
formula and is processed in the following steps:
1) Compute the channel capacity for every TTI
2)
3) Compute such that
whereby the factor is a correction factor that depends on the
mobile speed, the interleaver, etc. For low mobile speeds and
the almost ideal UMTS interleavers it has been found [16] that
is a good value.
2) Power and Spreading Code Assignment: As already men-
tioned, the traces have been generated for a normalized transmit
power. However, the MBMS bearer might be transmitted with
different power resulting in obviously different SINR values.
An appropriate normalization to investigated transmit power is
240
Fig. 7. Example of PedA and VehA movement within a cell.
applied by appropriately increasing or decreasing the effective
SINR. Furthermore, changing the MBMS radio bearer param-
eters, in particular the spreading factor, leads also to different
SINR values. Since the individual values for the carrier, the in-
terference and the noise are stored in the trace files, it becomes
quite easy to update the carrier power and re-compute the effec-
tive SINR by
whereby represents the carrier power, the interference
power, the noise power, the orthogonality factor and
the spreading factor.
Authorized licensed use limited to: UNIVERSIDADE DE BRASILIA. Downloaded on August 17, 2009 at 12:08 from IEEE Xplore. Restrictions apply.
IEEE TRANSACTIONS ON BROADCASTING, VOL. 53, NO. 1, MARCH 2007
C. Physical Layer FEC Modeling
In general, in system level simulation, the loss probability
of the physical layer code is determined by some table lookup
functions which map the effective SINR to some loss proba-
bility. Based on this loss probability, a random generator de-
cides whether the included RLC/MAC block is decodable or
not. However, for codes with different code rates, this still re-
quires a significant amount of simulations, as each possible code
rate needs to be simulated.
Luckily, Turbo codes as applied in UMTS have the prop-
erty that for a given SINR and a given Turbo code rate ,
the decoder is either almost always able to decode or it almost
always fails. The so-called “water fall” region of long Turbo
codes is rather narrow. The waterfall region for practical Turbo
codes coincides quite well with the computational cutoff rate
in a sense that if the code
rate of the code is below the cutoff rate for this specific SINR,
decoding is successful and otherwise it fails. We, therefore, use
the cutoff rate as a suitable function for the evaluation of the
decodability of the Turbo decoder. Note the above equation is
valid for BPSK transmission as well as for each component in
case of QPSK transmission. Therefore, after for each TTI, the
mobile channel simulator computes the effective SINR. Based
on this value and the applied code rate, the RLC-PDU are either
assumed error-free or are dropped.
D. Application Layer FEC Modeling
Raptor codes have been described in some detail in
Section III. Simulation results [4], [5] show that these codes
have a performance very close to ideal, i.e., the failure proba-
bility of the code is such that in case is only slightly more than
symbols are received, the code can recover the source block.
Although our system level simulation would not be different if
an ideal erasure code would be used, we apply a simple method
to estimate the performance of the code. For the small
inefficiency of the Raptor code can quite well be modeled by
the following equation,
if ,
if ,
whereby denotes the failure probability of the code
with source symbols if symbols have been received. Fig. 8
shows simulation results for selected cases and also a graph of
the formula. It is observed that for different , the equation al-
most perfectly emulates the code performance. While an ideal
fountain code would decode with zero failure probability when
, the failure for Raptor code is still about 85%. However,
the failure probability decreases exponentially with increasing
number of received symbols.
To avoid extensive Raptor decoding for all users, the system-
level simulator implements this equation-based model for the
Raptor code, which allows the simulations to be considerably
faster. In particular, to determine if a Raptor code would decode
a given amount of received symbols, a uniform distributed
random variable is sampled and compared with . If ,
Raptor decoding fails, otherwise it succeeds.
LUBY et al.: RELIABLE MULTIMEDIA DOWNLOAD DELIVERY IN CELLULAR BROADCAST NETWORKS 241

Fig. 8. Raptor source block loss rate for different source block size k , different
0
received overhead m k , and symbol loss rate 40%.

V. SIMULATION RESULTS FOR DOWNLOAD DELIVERY

The simulation concept and details presented in Section IV
allow us to simulate the performance of download delivery ser- Fig. 9. Simulation results for a 240 kbps bearer, Vehicular A mobility model,
vices in MBMS. In order to evaluate system performance and without selective combining.
specifically the trade-off between application layer FEC and
physical layer FEC extensive system level simulations are per-
formed. The basic scenario is that a file of a certain size with
static contents is transmitted in broadcast mode to a large group
of users which are randomly placed in the serving cell.

A. Simulation Parameters and Evaluation Criteria

Due to the large number of possible system and simulation
parameters, a reasonable parameter selection is required. Impor-
tantly, many more simulations than shown in this section have
been carried out, and the results represent a reasonable selec-
tion. We have chosen to simulate the distribution of a 512 KB
file which might represent a short multimedia clip, a still image
or a reasonably sized ring tone. For the finally presented re-
sults, a bearer supporting 240 kbps at the physical layer has been
chosen. Simulations are run for users whereby their
starting position is randomly and uniformly distributed over the
cell area. These users are simulated for vehicular and pedestrian
mobility and propagation model. We also compare receiver per- Fig. 10. Simulation results for a 240 kbps bearer, Vehicular A mobility model,
formance with and without selective combining. with selective combining.
In the assessment of different system configurations, basi-
cally two aspects are of major interest, user perception of the
multimedia delivery as well the resources consumed on the methods to complete their reception. Therefore, we evaluate the
physical layer. Resources are quite suitably measured by the necessary system resources in terms of the required transmis-
duration of time for the distribution of the file during which sion energy to satisfy at least a certain percentage of the user
system resources are occupied. Secondly, the power assigned population in the MBMS service area. The following results
to the distribution of the file is essential. To simplify the pre- show the necessary transmission energy to support 95% of the
sentation, we use the product of the two values, the assigned user population for different system parameter configurations.
transmit power times “on air time” which basically represents Very similar in spirit results are obtained to support for example
the necessary transmission energy, , to distribute the file. 90% or 99% of the user population. We investigate different
In terms of user perception, file download delivery is in some settings of application layer FEC rates and physical layer code
sense only binary, namely it is evaluated if the file is correctly rates. For intuitive interpretation of the results, we present the
received or not. Other aspects such as the experienced “down- RLC-PDU loss rate of the worst supported user and the neces-
load time”, i.e., how long it takes to receive the file after the sary transmission energy to support this user.
joining has happened, are not explicitly evaluated, but could be
taken into account for certain applications. The system is likely B. Selected Simulation Results
to be operated such that a certain percentage of users are satis- Figs. 9, 10, 11 and 12 show the necessary transmission energy
fied, i.e. have received the file. Remaining users are not excluded over the resulting RLC-PDU loss rate, all for 240 kbps bearers
in the MBMS download service, but must rely on post-delivery and all for different power assignments for the MBMS service,

Authorized licensed use limited to: UNIVERSIDADE DE BRASILIA. Downloaded on August 17, 2009 at 12:08 from IEEE Xplore. Restrictions apply.
242
Fig. 11. Simulation results for a 240 kbps bearer, pedestrian A mobility model,
without selective combining.
Fig. 12. Simulation results for a 240 kbps bearer, pedestrian A mobility model,
with selective combining.
for all users being vehicular users and all users being pedestrian
users and in both cases with and without selective combining
at the receiver. The curves are generated by applying different
Turbo code rates and applying as much application layer
FEC as necessary so that 95% of the users are satisfied. The
curves generally terminate on the left due to the restriction on
the Turbo code rate of 1/3; the leftmost point corresponds to
lowest RLC-PDU loss rates and therefore to lower Turbo code
rate, while the rightmost point of the curve corresponds to higher
RLC-PDU loss rate and therefore to higher Turbo code rate.
From the simulation results it is immediately observed that
there are some optimum system configuration at rather high
RLC-PDU losses which is in general not located at the lowest
Turbo code rate 1/3. For example, in Fig. 9, if the system al-
locates 4 W of transmit power for MBMS service, the optimal
physical layer code rate, achieving minimal required transmis-
sion energy, is such that the average RLC-PDU loss rate for the
Authorized licensed use limited to: UNIVERSIDADE DE BRASILIA. Downloaded on August 17, 2009 at 12:08 from IEEE Xplore. Restrictions apply.
IEEE TRANSACTIONS ON BROADCASTING, VOL. 53, NO. 1, MARCH 2007
worst supported user is about 40%. If stronger physical layer
channel coding is applied, the RLC-PDU loss rate decreases.
However the throughput at the RLC layer also decreases as al-
ready elaborated previously. This leads to an increased down-
load delivery time and consequently to higher required trans-
mission energy.
If a Turbo code rate of 1/3 is chosen, the required transmis-
sion energy for successful delivery is about 60% higher than in
case of the optimum configuration. However, if the Turbo code
rate is set too high then the increasingly higher bitrates cannot
compensate the increasing RLC-PDU loss rates. This also leads
to increased download delivery time and consequently higher
required transmission energy. These results suggest that using a
relatively low Raptor code rate and working with a rather high
RLC-PDU loss rate for the worst supported user is overall very
beneficial for the system resources and reduces the overall re-
quired transmission energy for the file distribution.
Another interesting observation is that transmission with
lower transmission power is considerably advantageous. In all
the presented results, transmission with 0.5 W always results
in the minimal required transmission energy. Although even
lower transmit power might provide even better performance,
eventually other effects such as very long on air times and shut
down of terminals would be counterproductive.
Selective combining, if applicable, has a strong impact on the
required transmission energy. Performing selective combining
improves the system capacity significantly. In Fig. 10, for ex-
ample, the minimum required transmission energy for 0.5 W
less than half the transmission energy required for the corre-
sponding case without selective combining. However, the low
penetration of mobiles applying selective combining and re-
stricted MBMS service area requires system designs without se-
lective combining. The RLC-PDU loss rates for minimal trans-
mission energy with selective combining are lower mainly due
to lower download time, not due to the use of a different Turbo
code rate. Therefore, receivers with and without selective com-
bining can quite well coexist and should be operated with sim-
ilar system parameters. Note, however, that the optimized loss
rates with selective combining are still in the range of 20% to
30%.
When comparing vehicular and the pedestrian mobility sce-
narios, we conclude that less transmission energy is required
to deliver a file if the users are moving at higher speeds, i.e.
use vehicular model. This can be explained due to the fact that,
for higher mobility, there is also higher diversity gains. When a
pedestrian user is in a deep-fade, it will remain in this situation
for longer time than a vehicular user, which moves faster.
Fig. 13 and Fig. 14 show the Raptor code rate for a given
Turbo code rate to support 95% of the users for different trans-
mission powers, without selective combining in Fig. 13 and with
selective combining in Fig. 14. Each line corresponding to a
fixed transmission power is labeled at three points, where the
leftmost point corresponds to the configuration with Turbo code
rate 1/3, the middle point corresponds to the configuration that
minimizes transmission energy for that transmission power, and
the rightmost point corresponds to the configuration with Turbo
code rate 0.95. The percentage label on each point shows the
RLC-PDU loss rate for the worst case supported user for that
LUBY et al.: RELIABLE MULTIMEDIA DOWNLOAD DELIVERY IN CELLULAR BROADCAST NETWORKS 243

TABLE II
OPTIMAL SYSTEM OPERATING POINT FOR F ,u = 512 KB = 95%
, NON
SELECTIVE COMBINING, VEHICULAR A MOBILITY MODEL

TABLE III
OPTIMAL SYSTEM OPERATING POINT FOR F ,u = 512 KB = 95%
, NON
SELECTIVE COMBINING, PEDESTRIAN A MOBILITY MODEL

Fig. 13. Trade-off between physical layer Turbo code and application layer
Raptor code for a 240 kbps bearer and vehicular users with no selective com-
bining, whereby the percentage of satisfied user is 95%. Dots show the optimal
system operating point.

TABLE IV
OPTIMAL SYSTEM OPERATING POINT FOR F ,u = 512 KB = 95%,
SELECTIVE COMBINING, VEHICULAR A MOBILITY MODEL

TABLE V
OPTIMAL SYSTEM OPERATING POINT FOR F ,u = 512 KB = 95%,
SELECTIVE COMBINING, PEDESTRIAN A MOBILITY MODEL

Fig. 14. Trade-off between physical layer Turbo code and application layer
Raptor code for a 240 kbps bearer and vehicular users with selective combining,
whereby the percentage of satisfied user is 95%. Dots show the optimal system
operating point.

configuration. In general, the higher the Turbo code rate, the

lower the Raptor code rate and vice-versa, i.e. as expected weak-
ening the physical layer FEC code requires using a stronger ap-
plication layer FEC code. Interestingly, without selective com-
bining the system configuration that minimizes transmission en-
ergy uses about the same low code rate for the Raptor code and
for the Turbo code and the RLC-PDU loss rate is 60%. For the
case with selective combining, the Turbo code rate for the con-
figuration that minimizes transmission energy is lower than the
Raptor code rate, but still the Raptor code rate is around 0.65
and the RLC-PDU loss rate is 30%.
Tables II, III, IV and V show in more detail the optimal system
operating points for download delivery with minimal transmis-
sion energy to support 95% of the users for the vehicular and
pedestrian models, with and without selective combining. In
these tables a comparison is also made with the conventional
system operating point, whereby the physical layer Turbo code
operates at rate 1/3. It is clear that huge reductions in transmis-
sion energy can be achieved by properly selecting the system
operating point, especially in the case of high transmit power.
These reductions are achieved by relaxing the physical layer
code and applying stronger application layer FEC. Also note
that the optimal system operating point is always achieved for
0.5 W, i.e. very small transmit power. It is also observed that se-
lective combining has a strong influence on the system perfor-
mance, however the application layer FEC is still highly bene-
ficial.

Authorized licensed use limited to: UNIVERSIDADE DE BRASILIA. Downloaded on August 17, 2009 at 12:08 from IEEE Xplore. Restrictions apply.
244
Fig. 15. Required transmit power that satisfies 95% users in the cell for a given
fixed broadcast time, with file size 512 KB, 240 kbps bearer and using Vehicular
A mobility model.
C. Selected Simulation Results for Time-Constrained
Broadcast
In this section we consider the scenario where the broadcast
of the file takes place over a limited amount of time. Fig. 15
shows simulation results for time-constrained broadcast of a
512 KB file over a 240 kbps bearer, whereby the users follow a
Vehicular A mobility model. In particular, the figure shows the
required transmit power to satisfy 95% of users in the cell as a
function of the broadcast time. The two curves differentiate be-
tween the two cases where 1) the code rate of the physical layer
Turbo code is optimaly adapted for delivery performance (i.e.
minimal delivery time), 2) the physical layer Turbo code is oper-
ated with standard UMTS parameter, i.e. code rate 1/3. A curve
gap in the horizontal axis indicates a reduction in delivery time,
while a vertical gap indicates a reduction in transmission power.
It can be observed that for shorter broadcast times, hence higher
transmit power, a delivery time reduction of about 70 sec can be
achieved. If we fix the broadcast time, huge reductions in trans-
mission power can be achieved. For example, for a broadcast
time of 75 sec, a significant power reduction of 1/6 is achieved
by properly selecting the physical layer Turbo code rate.
Fig. 16 shows the required transmission energy to deliver a
512 KB file as a function of the supported throughput at RLC
layer for the same channel and mobility models. A curve gap
in the horizontal represents a gain in RLC throughout, while a
vertical gap represents a reduction in required transmission en-
ergy. It can be observed that, in general, higher transmission
energy is required in order to support higher throughput. How-
ever, we also notice that, using the standard UMTS settings for
the Turbo code, the supported throughput does not go beyond
64 kbps. This happens because the RLC throughput is limited
by the Turbo code rate. By relaxing the Turbo code at phys-
ical layer, we can observe the benefits of using the Raptor at
application layer since, not only are higher RLC throughputs
achieved, they are also achieved using less transmission energy.
Therefore, an important and immediate conclusion is that, by
Authorized licensed use limited to: UNIVERSIDADE DE BRASILIA. Downloaded on August 17, 2009 at 12:08 from IEEE Xplore. Restrictions apply.
IEEE TRANSACTIONS ON BROADCASTING, VOL. 53, NO. 1, MARCH 2007
Fig. 16. Required transmission energy to deliver a 512 KB file versus the sup-
ported bit-rate over a bearer type of 240 kbps, whereby 95% of users are satisfied
and follow a Vehicular A mobility model.
properly choosing the physical layer Turbo code rate, some RLC
throughput values are achieved that can not be achieved (inde-
pendent of how much transmission energy is spent) by using
standard UMTS parameter settings.
VI. CONCLUSIONS
In this paper we have investigated MBMS download delivery
services in UMTS systems considering a comprehensive anal-
ysis by applying a detailed and complex channel model and
simulation setup. A significant part of MBMS is application
layer FEC based on Raptor codes, which have been standardized
for MBMS for the broadcast delivery of multimedia content.
The benefits of using Raptor codes to provide mobile broad-
cast services are manifold, including their almost ideal ability
to recover from losses, their low computational complexity and
their fountain code flexibility. Even though Raptor codes have
been analyzed in detail in the MBMS standardization efforts, the
concrete benefits of using application layer Raptor codes in an
overall mobile system design have not been fully understood or
appreciated. Therefore, we have applied an accurate and com-
prehensive simulation model which takes into account the ef-
fects of different layers in the protocol stack and also evaluates
the service for the two most important metrics, user experience
and radio resource consumption.
Of specific interest in the evaluation is the tradeoff between
resources used in the physical layer and resources used in the
application layer. The results of our work clearly indicate that
a tradeoff and thorough balancing of the overheads is neces-
sary. In contrast to some beliefs and conjectures that all prob-
lems can be solved at the physical layer, our results clearly show
that only a well designed system that takes into account the ben-
efits of application layer FEC can optimize system resources
and user perception. In particular it was shown that a well-de-
signed system should allow relatively high loss rates at the phys-
ical layer and use significant amounts of application layer FEC
to overcome these losses. More specifically, allowing higher
RLC-PDU losses and using Raptor codes to overcome these
LUBY et al.: RELIABLE MULTIMEDIA DOWNLOAD DELIVERY IN CELLULAR BROADCAST NETWORKS
losses works in favor of optimizing the overall system efficiency
and reduces the overall transmission energy to deliver a file to a
number of users within a cell. Packet losses are not per se your
foe, they can quite well be your friend.
The principle findings have been verified for different system
parameter settings such as different power assignments, dif-
ferent bit rates, different mobility models, as well as advanced
receiver techniques such as selective combining. Selective com-
bining can improve the system performance quite a bit, but the
optimized physical layer settings are not dramatically different
from the case without selective combining.
Not only does MBMS include application layer FEC with
Raptor codes, but also DVB has also recommended the inclu-
sion of Raptor decoders in their terminals for IP-datacast ser-
vices [17]. Other mobile broadcast standardization bodies are
also currently investigating the inclusion of application layer
FEC in their systems. In general, solutions allowing further in-
creases in the reliability of such systems which have minimum
impact on the protocol stack are highly desirable.
Ongoing and future work addresses the evaluation of
streaming delivery services in 3G broadcast environments.
Similar to MBMS download services, MBMS streaming ser-
vices can also uses Raptor codes, and all MBMS compliant
receivers must support Raptor decoding for both download
and streaming services. However, a different system design
is required for streaming services than for download services
due to stringent delay constraints. Thus, evaluations analogous
to those presented in this work are currently underway for
the MBMS streaming delivery service in order to obtain an
understanding of good tradeoffs between the assignment of
physical layer resources and application layer FEC.
ACKNOWLEDGMENT
The authors would like to thank the staff of Nomor Research
for their support in the generation of this work, specifically
Eiko Seidel for providing useful, constructive, and insightful
comments on the manuscript. In addition, the help of Chris-
tian Buchner and Junaid Afzal in setting the complex simulation
was much appreciated. The authors would also like to thank the
anonymous referees for their valuable and insightful comments.
REFERENCES
[1] T. Paila, M. Luby, R. Lehtonen, V. Roca, and R. Walsh, FLUTE—File
Delivery Over Unidirectional Transport IETF, RFC3926, Oct. 2004,
Tech. Rep..
[2] Technical Specification Group Services and System Aspects; Mul-
timedia Broadcast/Multicast Service; Architecture and functional
description, , 3GPP TS 23.246 V6.9.0, December 2005.
[3] Technical Specification Group Services and System Aspects; Multi-
media Broadcast/Multicast Service; Protocols and Codecs, , 3GPP TS
26.346 V6.1.0, June 2005.
[4] A. Shokrollahi, Raptor Codes Digital Fountain, Tech. Rep.
DR2003-06-001, Jun. 2003.
[5] M. Luby, M. Watson, T. Gasiba, T. Stockhammer, and W. Xu, “Raptor
codes for reliable download delivery in wireless broadcast systems,” in
CCNC 2006, Las Vegas, NV, Jan. 2006, unpublished.
Authorized licensed use limited to: UNIVERSIDADE DE BRASILIA. Downloaded on August 17, 2009 at 12:08 from IEEE Xplore. Restrictions apply.
245
[6] Technical Specification Group Radio Access Network; Introduction of
the Multimedia Broadcast Multicast Service (MBMS) in the Radio Ac-
cess Network (RAN), , 3GPP TS 25.346 V7.0.0, March 2006.
[7] Technical Specification Group Radio Access Network; Multiplexing
and Channel Coding (FDD), , 3GPP TS 25.212 V7.0.0, March 2006.
[8] Technical Specification Group Radio Access Network; Multiplexing
and channel coding (TDD), , 3GPP TS 25.222 V7.0.0, March 2006.
[9] Technical Specification Group Radio Access Network; Typical ex-
amples of Radio Access Bearers (RABs) and Radio Bearers (RBs)
supported by Universal Terrestrial Radio Access (UTRA), , 3GPP TS
25.993 V6.13.0, March 2006.
[10] M. Luby, J. Gemmell, L. Vicisano, L. Rizzo, M. Handley, and J.
Crowcroft, Asynchronuous Layered Coding (ALC) Protocol Instanti-
ation IETF, RFC3451, Dec. 2002, Tech. Rep..
[11] A. Shokrollahi, “Raptor codes,” IEEE Trans. Information Theory, vol.
52, no. 6, pp. 2251–2567, June 2006.
[12] A. Shokrollahi, S. Lassen, and M. Luby, “Multi-Stage Code Generator
and Decoder for Communication Systems,” U.S. Patent 7 068 729, June
27, 2006.
[13] M. Luby, “LT codes,” in Proc. 43rd Annual IEEE Symposium on Foun-
dations of Computer Science, 2002.
[14] ——, “Information Additive Code Generator and Decoder for Com-
munications Systems,” U.S. Patent 6 307 487, Oct. 23, 2001.
[15] 3GPP TSG-SA WG4 S4-AHP238, in Specification Text for System-
atic Raptor Forward Error Correction, Sophia Antipolis, France, Apr.
2005, PSM SWG.
[16] 3GPP TSG-SA WG4 S4-AHP238, in Specification Text for Systematic
Raptor Forward Error Correction, Seoul, South Korea, Oct. 2003, PSM
SWG.
[17] ETSI DVB TM-CBMS1167, IP datacast over DVB-H: Content de-
livery protocols Draft Technical Specification, Sept. 2005 [Online].
Available: http://www.dvb.org
Michael Luby cofounded Digital Fountain, Inc. in
1998, where he currently holds the position of Chief
Technology Officer. Michael is a world-renowned
Theoretical Computer Scientist, and has made break-
through research contributions in the areas of coding
theory, randomized algorithm design and analysis,
transport protocols and cryptography. He has led the
development of several transport standards within
the Internet Engineering Task Force (IETF). After
receiving his Ph.D. in Theoretical Computer Science
from UC Berkeley in 1983 he was a Professor in
Computer Science at the University of Toronto. In 1988 Michael joined the
International Computer Science Institute in Berkeley to become the Leader
of the Theory Group and concurrently an Adjunct Professor at UC Berkeley.
Michael is a recipient of the 2002 Information Theory Society Paper Award
for some of his coding theory research, the 2003 SIAM Outstanding Paper
Prize for some of his cryptography research and the 2007 IEEE Eric E. Sumner
Award for bridging mathematics, internet design and mobile broadcasting as
well as successful standardization.
Tiago Gasiba was born in Oporto, Portugal. He
received his M.Sc. degree in telecommunication en-
gineering from the Technical University of Munich
(TUM) Germany in 2004, and his Eng. degree in
electrical engineering and computer science from
the Faculdade de Engenharia da Universidade do
Porto in 2002.
He is currently working for Digital Fountain and
NoMoR Research GmbH and working towards his
PhD degree under the supervision of Prof. Hagenauer
and Prof. Shokrollahi. In 2005 he was a visiting re-
searcher at the Laboratoire d’Algorithmique et Laboratoire de Mathematiques
Algorithmique (Algo+Lma) in Lausanne, Switzerland. His current research in-
terest include forward error correction codes in particular fountain codes, wire-
less communications networks and video and data broadcast.
246
Thomas Stockhammer has been working at the
Munich University of Technology, Germany, and
was a Visiting Researcher at Rensselear Polytechnic
Institute (RPI), Troy, NY and at the University of
San Diego, California (UCSD). He has published
more than 70 conference and journal papers, is a
Member of different program committees, and holds
several patents. He regularly participates in and
contributes to different standardization activities,
for example, JVT, IETF, 3GPP, and DVB, and has
coauthored more than 100 technical contributions.
He is the Acting Chairman of the video ad hoc group of 3GPP SA4. He is
also the cofounder and CTO of Novel Mobile Radio (NoMoR) Research, a
company working on the simulation and emulation of future mobile networks.
Since 2004, he has been working as a research and development Consultant for
Siemens Mobile Devices, now BenQ mobile in Munich, Germany. His research
interests include video transmission, cross-layer and system design, forward
error correction, content delivery protocols, rate-distortion optimization,
information theory, and mobile communications.
Authorized licensed use limited to: UNIVERSIDADE DE BRASILIA. Downloaded on August 17, 2009 at 12:08 from IEEE Xplore. Restrictions apply.
IEEE TRANSACTIONS ON BROADCASTING, VOL. 53, NO. 1, MARCH 2007
Mark Watson is a Director of Research at Digital
Fountain Inc, focusing on erasure coding appli-
cations such as mobile broadcast and IPTV as
well, development of new coding technologies and
promotion of Digital Fountain codes within industry
standards. Before joining Digital Fountain in 2004
he worked for Nortel Networks on product strategy
and standardization in various areas including Voice
over IP, Next Generation Networks and 3G cellular
networks, including significant contributions to the
Mobile Broadcast Multicast Service (MBMS) stan-
dards in 3GPP. Mark received an MSc in Mathematics from King’s College,
London and a BA(Hons) in Mathematics from the University of Cambridge.