Azza Jedidi, Fr
eric Weis
IRISA/INRIA, Rennes, France
Sylvaine Kerboeuf, Marie-Line Alberi Morel
Alcatel-Lucent Bell Labs, France
Mobility, logical discontinuous coverage wireless network, caching, scheduling, scalability, start-up delay,
resource management, layer encoded flow.
The past few years have witnessed the deployment of a wide variety of multimedia applications over wireless
networks. As a consequence, mobile users are demanding fast and efficient connectivity, especially concern-
ing start-up delays and flow quality. Wireless networks aim at providing an everywhere coverage without
prohibitive deployment costs. Thus, the coverage of each access point is extended with detriment to the mean
throughput of the radio cell. The throughput of each zone decreases non-uniformely from the cell center to
the cell edge. The coverage is continuous, but radio conditions met by users are not homogeneous. In fact,
the data rate varies according to user mobility. In the context of streaming applications, intermittent high rate
availability may lead to service disruption, especially when user number’s increases. Thus, it is essential to
design a network architecture able to efficiently deliver video flows to mobile users. In our work, a new equip-
ment, called network cache, is introduced. The latter performs efficient flow caching and scheduling, in order
to guarantee service continuity. The simulation shows that the proposed architecture gives satisfying service
start-up delays and improves system scalability.
data transfers play an important role in cur-
rent multimedia communications, especially thanks
to the increase of available bandwidth and the grow-
ing needs for mobility. As a consequence, users are
demanding fast and efficient ubiquitous connectivity.
They require ¡¡any-time any-where¿¿ services, with-
out delays or disruptions. Therefore, new network in-
frastructures had to be designed in order to guarantee
ubiquitous coverage and to face wireless network con-
straints, in terms of bandwidth, error rates and other
perturbations linked to environment heterogeneity.
In this context, the logical discontinuous coverage
wireless network model has been proposed. This new
model consists of a set of access points, i.e. anten-
nas around which are defined the radio cells. Those
antennas are discontinuously spread on the network
This paper presents results from a collaboration be-
tween INRIA ACES research team and the MAG project
from Alcatel-Lucent Bell Labs
area, thus providing a ¡¡many-time many-where¿¿
service. Actually, the idea of coverage discontinu-
ity brings two major advantages. First, as it implies
the use of a fewer number of access points, the ar-
chitecture deployment will be cheaper. Second, ra-
dio cells disjunction hypothesis simplifies the radio
frequency band management and avoids interference
problems. The mechanisms defined in the frame-
work of discontinuous radio coverage network model
can be easily applied to the context of continuous ra-
dio coverage networks, where bandwidth is highly
varying. High bandwidth areas are similar to cov-
erage areas, and low bandwidth areas are similar to
out of coverage areas. The main idea is to discrim-
inate terminals by their radio conditions in order to
select the data to be sent. In fact, terminals are sup-
plied with data when they cross high bandwidth ar-
eas (transfer areas). And, low bandwidth areas are
dedicated to detecting terminals presence (presence
areas). Our research team has proved through sim-
ulations that avoiding data transfer in low bandwidth
Jedidi A., Weis F., Kerboeuf S. and Alberi Morel M. (2008).
In Proceedings of the International Conference on Wireless Information Networks and Systems, pages 66-73
DOI: 10.5220/0002022200660073
areas enhances the throughput use over the network
and improves system scalability(Luu et al., 2007).
This new logically discontinuous vision of the net-
work makes this model fit to 3
generation net-
works, where coverage is continuous but with im-
portant bandwidth fluctuations, as well as 4
ation networks, where radio coverage could be physi-
cally discontinuous. Even if this model simplifies net-
work deployment, the connectivity intermittence in-
duces important challenges in order to avoid service
disruption. Thus, terminals have to take advantage of
the high bandwidth available when crossing a transfer
area in order to store some part of the multimedia flow
on their own caches, and to consume it when crossing
presence areas.
Our works focus on logical discontinuous coverage
wireless networks, and aim at providing a multitude
of services. In this paper, our main goal consists in
designing a scalable unicast streaming service, while
reducing the start-up delays.
On the one hand, streaming servers are usually lo-
cated in ¡¡classical¿¿ IP networks. Such networks are
submitted to many constraints, like bandwidth varia-
tions. On the other hand, mobile terminals evolve in
wireless logical discontinuous coverage areas, consti-
tuted of several access points. Each access point de-
fines a high bandwidth area in each radio cell. The
main idea we propose in the context of such net-
works, is to design an intermediate equipment that
¡¡catches¿¿ the video flows sent by the content server,
and then efficiently distributes them to mobile termi-
nals in the network. This equipment is called the net-
work cache (see figure 1).
In a previous version of our work, the network cache
Wired network
« Classical IP architecture »
Bandwidth limitations
Content server
Access Point
Radio zone
Transfer Area
(i.e. highest radio throughput)
Figure 1: Logical discontinuous coverage wireless network.
acted as a temporary storage memory. In fact, it tem-
porarily stored the part of the flow to be delivered to
the mobile terminal, while the latter crossed a pres-
ence area. Then, when the terminal entered a transfer
area, the network cache rapidely delivered the stored
flow to him, taking advantage of the high bandwidth
available. More precisely, the network cache was pro-
vided with a set of buffers. Each buffer was related to
a single streaming request and contained the part of
the flow to be sent to the corresponding mobile termi-
nal(Luu et al., 2007). The use of buffers considerably
enhanced the network scalability. In fact, it allowed
users, crossing transfer areas, to take advantage of the
high bandwidth available, in order to rapidly access
the requested flow. Flow delivery was no longer sub-
ject to the bandwidth limitations at the content server
level, as the flows were already stored in the network
The terminal was also provided with buffers to store
some parts of the flow in advance, when it crosses
transfer area, taking benefit of the high bandwidth
available. The terminal consumes this prefetched data
while crossing out of coverage areas and, thus, avoids
service disruption. The flow playback is smooth (low
jitter), as the terminal cache prevented delay variation
problems (see figure 2).
Moreover, the network cache played a scheduling
Data transfer
Data transfer
Data transfer
Figure 2: Caching mechanisms.
role. It classified flows in priority queues, depend-
ing (1) on terminal positions in the network and (2)
on terminal buffers content. The scheduling policy
favored terminals which were in transfer areas, as
they will rapidly receive their flows thanks to the high
bandwidth available. It also favored terminals whose
buffer level was under a certain threshold, in order to
avoid service disruption.
We proved through simulations that, thanks to the
network cache buffers and the scheduling policy, the
scalability of the network was considerably enhanced
as the service disruption was avoided for a larger ter-
minal density (Luu et al., 2007).
Nevertheless, buffers are temporary storage memo-
ries. So, the flow stored in a buffer is used by only
one user, and could not benefit to users who ask for
the same file later. In this paper, we propose to re-
place the buffers in the network cache with caching
mechanisms. Actually, a copy of any requested video
flow will be kept in the network cache, after the end
of the streaming request. Later requests to the same
video flow will be accessed faster as the flow is al-
ready stored in the cache network, nearer to the ter-
minal than if it was in the content server. The de-
livery is no more subject to ¡¡classical¿¿ IP network
bandwidth constraints (i.e. bandwidth constraints in
the path between the content server and the network
cache). Buffers, in the terminal, will also be replaced
by a cache structure in order to store sufficient parts
of the flow and to avoid service disruption.
Moreover, we propose to bring a more efficient
scheduling mechanism in the network cache, taking
into account some properties of the video flow in ad-
dition to terminal position and cache level. As H.264
flows are more and more used for nowadays mo-
bile video applications, we chosed to consider H.264
video flows for our streaming service.
The remainder of this paper is organized as fol-
lows: first we make a brief study of the main video
flow caching and scheduling mechanisms. Then, we
present the mechanisms we chose in the context of
our network. After that, the targeted architecture is
detailed. Finally, the performances of our solution are
In the context of streaming applications, many
caching algorithms could be adopted in order to de-
sign the network cache. For instance, prefix caching
approach previledges caching the prefix of video
flow in order to fasten the service start (Sen et al.,
1999). Segment-based distance-sensitive caching
approach groups the frames of the media object into
variable-sized segments (the size of segment i is 2
Actually, segments size increases exponentially. As
a consequence, the last segment length is half the
cached portion length (Wu et al., 2001), which is es-
pecially interesting regarding the cache replacement
policy. Indeed the latter will delete segments starting
from the end of the cached flow. So, it will rapidly
discard as big chunks as needed. A more efficient
approach is the layered encoding and transmission
method. Layer coders compress the flow in several
layers. The base layer contains the data representing
the most important elements of the flow. It represents
the video flow in its minimal quality. Thus, it is
the most significant layer. The other layers are
hierarchically added to refine progressively the flow
quality (Kangasharju et al., 2002). Depending on its
capabilities, the terminal will subscribe to a set of
cumulative layers to reconstruct the stream.
In our study, we consider H.264 SVC (scalable
video coding) video flows. Those flows are layer
encoded, which means that they are constituted of
a base layer that represents the flow in a minimal
quality, and some enhancement layers that could be
hierarchically added to the base layer in order to
enhance its quality. In our work, we use a model for
the H.264 SVC video flows with a base layer and two
enhancement layers, as shown in figure 3. A layered
caching approach seems to be the most appropriate
caching strategy in the context of such flow model.
Content server
Access Point
Terminal cache
Enhancement Layer 2
Enhancement Layer 1
Base Layer
Network cache
Layered scheduling
Layered Caching
Layered Transmission
Layered Transmission
Figure 3: Layer encoded flow model.
In this paper, our goal is to deliver H.264 video flow
over a logical discontinuous coverage architecture. To
aim that goal, we had to design a scheduling algo-
rithm that fits our architecture needs and our video
flow properties.
In the context of logical discontinuous coverage net-
works, video flow scheduling may become a very hard
task as it is subject to a multitude of constraints, such
as bandwidth limitation and asymmetry, important er-
ror rates and the high probability of service disruption
due to client mobility.
Multirate Wireless Fair Queuing (MR-FQ) Al-
gorithm. Consists in transmitting the video flow
at different rates depending on the channel condi-
tions(Wang et al., ). This algorithm takes into consid-
eration both time and service fairness. This algorithm
increases the overall system throughput and guaran-
tees fairness and bounded delays.
WINSYS 2008 - International Conference on Wireless Information Networks and Systems
Layered Quality Adaptation Algorithm. Con-
sists in adding and dropping layers of a layer en-
coded video stream to perform smooth flow adapta-
tion(REJAIE and REIBMAN, 2001). The main goal
of the algorithm is to fit available bandwidth changes,
without rapid and disturbing variations in quality. Ac-
tually, congestion control changes the transmission
rate rapidly, over several round-trip time, whereas, hi-
erarchical encoding permits smooth video quality ad-
justment, over long periods of time. The difference
between both timescales is obtained thanks to data
buffering at the receiver, which smoothes the rapid
variations in available bandwidth and allows an al-
most constant number of layers to be played.
Buffer Sensitive Bandwidth Allocation Algorithm.
Recommends that a client maintains sufficient video
frames at the buffer in order to improve the system
adaptability and minimize the impact of overload-
ing(Yuen et al., 2002). The video playback starts only
after reaching a determined buffer level. Based on the
play rate of a video, the system calculates the buffer
playback duration. The main idea of the method is
to allocate the available bandwidth at a base station,
which will serve the concurrent requests in the cell,
based on their playback buffer durations.
Discussions. The layer encoded flow model that we
adopted suggests the layered quality adaptation as
the most appropriate scheduling approach. Moreover,
this approach favors system scalability. Our schedul-
ing algorithm, presented in the next section, is also in-
spired by the Buffer Sensitive Bandwidth Allocation
strategy, as we took into account the terminal cache
level, when the user enters in the radio cell.
The key element for video flows delivery in the de-
signed network is the network cache. This equipment
is mainly made of two modules: the cache and the
The cache is divided into three parts, each of them
corresponds to the memory storage of one layer.
Thus, for each flow, each layer is stored in the cor-
responding part of the cache. And, these parts are
provided with a mapping table in order to match
the memory cells with the corresponding flow identi-
fiers. A similar cache structure is adopted for terminal
The goals of the scheduler are (1) to guarantee the
continuity of the streaming service and (2) to shorten
the start-up delay. To aim that target, it has to effi-
ciently distribute the video flows to the terminals.
In the previous version, the scheduler dynamically
classified the video flows in priority queues, taking
into account the terminal position (i.e. transfer area /
presence area) and its buffer filling level. The stream-
ing service was started when the buffer reached a
¡¡start-up buffer level¿¿ that corresponds to a play-
back duration of 30 seconds, which is a typical size
of a terminal internal buffer for a streaming service.
Now, the scheduler classifies the flow layers sepa-
rately. In fact, it considers the hierarchical order
of layers, in addition to terminal position and cache
level. As a consequence, the streaming service start-
ing is only linked to base layer caching. Since
the equivalent of 30 seconds of playback duration
is cached in the terminal, the video playback starts.
Moreover, as the scheduling privileges the base layer
transmission, enhancement layers could be delayed or
even omitted in order to guarantee service continu-
ity. In fact, service continuity in such a flow model
is equivalent to base layer streaming continuity. Ac-
tually, the level of terminal cache filling-up necessary
to streaming start-up is defined by making a compro-
mise between (1) a short start-up delay on the one
hand, and (2) caching a part of the flow, sufficient to
cross an out of coverage area without service disrup-
tion, on the other hand. Depending on the network
conditions (i.e. available bandwidth, terminal den-
sity, etc.), enhancement layers could be progressively
added or ommited, in order to ”smoothly” adapt the
quality of the received flows.
Combining those different criteria, we conceived the
algorithm detailed below. BL refers to the base layer.
EL1 represents the first enhancement layer and EL2
the second one. TA means that the terminal is located
in the Transfer Area. CL refers to the critical terminal
cache filling-up level. The priority queues are num-
bered from one to six, classified with a decreasing or-
der of priority. Let us consider a flow requested by a
terminal. The three layers of that flow will have to be
classified by the scheduler. Finally for each layer, the
following algorithm will be iterated.
Algorithm 1 : Scheduling policy in the Network Cache.
if (Layer = BL) & (TerminalCacheLevel CL)
Priority Queue 1
if (Layer = EL1) & (TerminalCacheLevel CL)
& TA then
Priority Queue 2
if (Layer = EL2) & (TerminalCacheLevel
CL) & TA then
Priority Queue 3
if (Layer = BL) & (TerminalCacheLevel
CL) & TA then
Priority Queue 4
if TA then
Priority Queue 5
Priority Queue 6
end if
end if
end if
end if
end if
This algorithm favors service continuity as it guar-
antees, with the highest priority, the delivery of the
base layer to terminals whose cache is under the crit-
ical threshold. Moreover, it clearly appears that it fa-
vors flow delivery to terminals which are in transfer
area. In fact, they take benefit of the high bandwidth
available, wich allow them to store some extra part
of the flow in their cache to avoid service disruption
while crossing presence areas.
Our solution has been evaluated in the framework of
a discrete event modeling based on a Java simula-
tor, which uses DESMO-J (Discrete event simulation
framework in Java) library(Des, ). The simulated sur-
face is 1 km
. There are one network cache and six
Access Points (AP). The server streaming rate is 1
Mb/s ( 256 kb/s for the base layer, 256 kb/s for the
first enhancement layer and 512 kb/s for the second
one). The simulated environment is dense urban en-
vironment like Manhattan. Each AP defines a radio
cell, with a range of 50 m. A set of terminals move
in a Manhattan topology, randomly crossing different
radio cells and out of coverage areas, at a speed of
50 km per hour (vehicular model). Their cache con-
tent allows them staying out of coverage for a certain
duration without service disruption(Luu et al., 2007).
This duration is called the Time Out-of-cover (ToC).
It corresponds to the longest time for which the ser-
vice is satisfied, while the user is out of cover. So, this
duration corresponds to the amount of data that the
terminal must store in its cache to guarantee a con-
tinuous data delivery. As a consequence, if the user
remains out of cover more than the ToC value, sevice
disruptions may appear. In our previous studies, we
evaluated the ToC value for our mobility model and
access point topology to 240 seconds.
The terminal may ask for any video flow in the
streaming server. If this flow is already stored in the
network cache, its layers are scheduled and transmit-
ted directly from the network cache to the terminal.
Else, the three layers are sent from the server, sched-
uled by the network cache and transmitted to the ter-
minal, while being simultaneously stored in the net-
work cache. The terminal can start playing the video
flow since the base layer reaches the start-up thresh-
old, which is equivalent to a playback duration of 30
In order to evaluate our system, it seems important
to point out our main goals. In fact, we need to sup-
port an important number of terminals while guaran-
teeing short start up delays, and service continuity. To
aim at that target, we adopted a layered flow model,
and we took this model into account for the design
of the caching and scheduling mechanisms. The main
principle was to distribute the layers depending on the
terminal position and cache level. We chose to guar-
antee service continuity, even if we had to temporarily
degrade the flow quality, by omitting some enhance-
ment layers.
In the next sections, the performances of the de-
signed system are evaluated, considering the start-up
delay (section 5.1) and the system behavior when the
terminal density increases (section 5.2).
5.1 Start of the Streaming Application
Let us consider a terminal, alone in the network. First,
the terminal requests a video flow which is not al-
ready cached in the network cache. The flow is sent
to the terminal from the streaming server, and simul-
taneously stored in the network cache.
At start of our streaming application, the network
cache is empty. Here, our main goals are (1) to pro-
vide as quickly as possible the amount of data re-
quired to start the streaming application (i.e. terminal
cache level = 30 s * Service Data Rate) and (2) to min-
WINSYS 2008 - International Conference on Wireless Information Networks and Systems
Figure 4: Network cache filling.
imize service disruptions as soon as the application
has begun to consume data in the terminal (i.e. termi-
nal cache level = ToC * Service data rate). Therefore,
we introduced a fast start-up period during which the
network cache is able to retrieve the beginning of the
stream twice as fast as the normal service rate (Fast
start-up rate = 2 Mb/s, normal service rate = 1 Mb/s).
As the terminal is provided with a doubled data rate,
the time necessary to start the streaming application
decreases from 30 s to 15 s (Luu et al., 2007). This
fast start-up phase continues after the service start. So
the terminal continues storing data in its cache while
consuminng. After this fast start up period, the flow
starts to be delivered at the normal application data
rate (1 Mb/s).
Figure 4 illustrates the network cache content dur-
ing the experiment. In this figure, the whole flow (the
total of the three layers) is represented. At 240 sec-
onds, the figure illustrates that application data rate
decreases after the fast start-up phase (the graph’s
slope decreases). After that (at 250 seconds), the ter-
minal decides to stop this first streaming request. At
that moment, the terminal did not end the whole flow
streaming. We may notice that the flow is not to-
tally stored in the network cache at that moment (see
figure 4). The terminal cache empties at the end of
the streaming request. Nevertheless, in the network
cache, the part of the flow stored is kept there. Now
(at 250 seconds), the terminal asks for the same flow,
a second time. The second streaming request starts. A
part of the flow is already cached in the network cache
as shown in figure 4. This figure shows that, during
this second streaming request, the part of the flow, re-
maining at the content server level, is still delivered to
the network cache. At 400 seconds, the whole flow is
stored in the network cache.
Figure 5 represents the evolution of the terminal
cache content. In this figure, the flow layers are repre-
Figure 5: Terminal cache filling.
sented seperately. During the first streaming request,
figure 5 shows that the terminal cache starts filling un-
til its content reaches the playback start level (at 15
seconds). Then, the terminal starts playing the video
flow at a 1 Mb/s rate, while it continues receiving the
flow from the network cache at a 2 Mb/s rate. Since
240 seconds, the fast start-up period lasts and the ap-
plication rate become equal to 1 Mb/s. The mobile
terminal consumes its flow at the same rate it receives
it, that is why we observe a plateau during this period.
At 250 seconds, the terminal ends the first streaming
request and empties its cache. At 250 seconds, the ter-
minal asks for streaming the same video flow a second
time. The second streaming request starts. Figure 5
shows that the base layer, and the first enhancement
layer are sent at 256 kb/s, while enhancement layer
2 is transmitted at 512 kb/s. The second enhance-
ment layer is thus received more rapidely than both
the base layer and the first enhancement layer. The
figure also shows that the layers are scheduled in the
defined hierarchical order, which highlights the role
of the scheduler.
Now, the terminal cache fills-up faster, as the flow
is stored at the network cache level. As a conse-
quence, the start-up threshold is reached more rapidly
(The start-up threshold is the terminal cache level
which corresponds to a playback duration of 30 sec-
onds). Actually, the start-up threshold is reached in
about 2 seconds in the second request while it is
reached in about 15 seconds for the first one. This
demonstrates the network cache impact during ser-
vice starting.
We may notice that, at the beginning of the first
streaming request, the network cache was empty, but
when the second request starts, a part of the requested
flow is already cached at the network cache. As a
consequence, this part of the flow is very rapidely de-
livered to the terminal thanks to the high bandwidth
available. Actually, as the flows are located in the net-
work cache, they are not submitted to the bandwidth
Figure 6: Terminal cache filling.
constraints at the video server level.
During the starting of a streaming request, some
images might be played while only one or two layers
are received. If only the base layer is received, the im-
age is in base quality. If base layer and enhancement
layer 1 are received, the quality is good. Finally, if all
the layers are received, the quality is excellent. Table
1 details the percentage of images received, in each
quality, by the considered terminal.
5.2 Scalability of the Solution
If a terminal, alone in the network, requests a video
flow already stored in the network cache, the three
layers are almost immediately received in the termi-
nal cache. As terminal density increases, reception
becomes slower, because the available bandwidth is
shared between users. Moreover, the flow quality may
need to be degraded (by omitting enhancement layers)
in order to send base layers to all users without service
disruption. Figure 6 represents the filling-up of the
cache of one terminal, for different values of termi-
nal densities. The figure shows that the enhancement
layers reception is delayed as the number of terminals
increases. The idea is that the base layer of the re-
quested flows must be sent to all terminals in order
to maintain service continuity. Enhancement layers
could be delayed or even omitted in order to avoid ser-
vice discontinuity. Moreover, acknowledgements sent
by terminals allow the network cache to avoid sending
enhancement layers of images already played at the
client side with base quality. Table 2 illutrates how
the system progressively degrades the received flows
quality in order to maintain the streaming service con-
tinuity. Actually, the table details the evolution of the
percentage of images received in base quality, in good
quality and in excellent quality by one considered ter-
minal, while terminal density increases.
Table 1: Received images quality.
Received images quality
Base quality 1,6 %
Good quality 3,9 %
Excellent quality 94,4 %
Table 2: Received images quality.
Received images quality
terminals density/Km
10 15 20 25 30 40
Base quality 4,03 % 7,15 % 2,84 % 6,99 % 19,76 % 39,86 %
Good quality 3,15 % 4,26 % 6,54 % 15,04 % 27,09 % 42,68 %
Excellent quality 92,8 % 88,57 % 90,61 % 77,95 % 53,14 % 17,45 %
We work on extending logical discontinuous cover-
age networks to a multitude of services. This paper
adresses the streaming service. We actually tried to
deliver unicast video flows to a multitude of mobile
terminals, in a logical discontinuous coverage net-
work. The main goal was to guarantee a good qual-
ity of service. We especially focused on guarantee-
ing short start-up delays and system scalability. To
achieve this goal, appropriate cache and scheduling
modules have been designed. Those mechanisms took
benefit from the video flow layered model.
The network cache plays a crucial role. As it is
nearer to the terminals than the origin server, start-up
delays are considerably reduced. Actually, flow de-
livery is no more submitted to bandwith limitations
at content server side. Terminal caches allow flow
storing in order to avoid service disruption while be-
ing in out of coverage areas, or low throughput areas.
They also allow smooth video playback. The layered
flow scheduling guarantees system scalability, thanks
to the possibility of quality degradation, in order to
maintain the service without any disruption.
Simulation showed that our goals are achieved : Ter-
minals receive their flows with satisfying start-up de-
lays, and service disruptions are avoided, thanks to the
caching and sheduling approach adopted. The short
periods of flow degradation allowed service continu-
ity, while maintaining a good quality of experience
(as the base layer flow contains the main elements of
the stream).
Our futur work will concentrate on testing new mo-
bility profiles, like the pedestrian model and on ex-
tending the evaluation of scalability performances, by
simulating different video flows, available at different
WINSYS 2008 - International Conference on Wireless Information Networks and Systems
A framework for discrete-event modelling and simulation.
Kangasharju, J., Hartanto, F., Reisslein, M., and Ross,
K. W. (2002). Distributing layered encoded video
through caches. IEEE Transactions on Computers,
51(6), pp. 622-636.).
Luu, A., Morel, M.-L. A., Kerboeuf, S., M
enard, R., Tlais,
M., and Weis, F. (2007). Improving Wireless Network
Capacity by Introducing Logical Discontinuous Cov-
erage. Report).
REJAIE, R. and REIBMAN, A. (2001). Design issues
for layered quality adaptive internet video playback.
In Workshop on Digital Communications (Taormina,
Sen, S., Rexford, J., and Towsley, D. (1999). Proxy pre-
fix caching for multimedia streams. In IEEE INFO-
COM 99, New York, NY.
Wang, Y.-C., Tseng, Y.-C., and Chen, W.-T. MR-FQ:
A Fair Scheduling Algorithm for Wireless Networks
with Variable Transmission Rates. Simulation: Trans-
actions of The Society for Modeling and Simulation
International, accepted, 2005. (SCI)).
Wu, K. L., Yu, P. S., , and Wolf, J. L. (2001). Segment-
based proxy caching of multimedia streams. In World
Wide Web Conference(WWW10), Hong Kong.
Yuen, J., Lam, K.-Y., and Chan, E. (2002). A fair and adap-
tive scheduling protocol for video stream transmission
in mobile environment. In IEEE International Con-
ference on Multimedia and Expo, Lausanne, Switzer-