A PREDICTIVE MULTI-CHANNEL MBAC TECHNIQUE FOR

ON-LINE VIDEO STREAMING

Pietro Camarda, Cataldo Guaragnella, Domenico Striccoli

D.E.E. − Politecnico di Bari, via E. Orabona, 4, 70125 − Bari, Italy

Keywords: Multimedia Streaming, Measurement Based Admission Control, Bandwidth Prediction.

Abstract: A measurement based admission control predictive technique is introduced for on line streaming systems

exploiting the GOP length demultiplexing of the aggregate bit stream in conjunction with a linear predictive

algorithm. Due to the long latency of the statistical aggregate, the predictive technique is able to predict the

bit rate over about two seconds of time. The prediction information is used in an admission control system

to estimate the bit rate and the margin with respect to the channel capacity in the proposed streaming

system. These measures have been used to estimate the overflow probability in a general aggregate

situation. Tests conducted over real video sequences confirm the feasibility of the proposed technique.

1 INTRODUCTION

Multimedia streaming applications (e.g., audio-video

streaming, Digital Video Broadcasting, UMTS

services, etc.) are rapidly growing in the actual

telecommunication world. They perform multimedia

reproduction while taking into account some user

feedback interactivity. In a multiplexed scenario,

several multimedia flows with a high bit rate

variability, data burstiness, self-similarity and Long

Range Dependence (LRD) properties (Garrett and

Willinger, 1994) are aggregated. Such a kind of

Variable Bit Rate (VBR) traffic makes very difficult

an optimal resource assignment and estimation, in

terms of bandwidth occupation and client buffers, to

properly manage video streaming.

An efficient aggregated VBR data transmission

assumes that multimedia flows are delivered with an

assigned Quality of Service (QoS), that is,

end-to-end delay, data losses, and jitter

specifications. To guarantee QoS requirements for

each of the aggregated flows, Call Admission

Control (CAC) schemes can be implemented, with

the purpose to verify that each flow, sharing link

bandwidth resources with other flows, can be

delivered to destination while respecting QoS

specifications. This task is performed whenever a

new flow requests admission in the network link. If

QoS requirements are respected, the new flow is

admitted, otherwise is rejected.

Admission Control algorithms need the

aggregate traffic to be accurately characterized,

through a priori determined traffic parameters, or

traffic measurements. In the first case,

“parameter-based”, in the second

“measurement-based” admission control algorithms

are implemented. As regards the first approach, the

main flow characteristics are a priori set to estimate

the amount of network resources needed by a flow

aggregation, including the new flow requesting

admission. The second approach is based on

aggregate traffic measurements, to estimate the

current link load.

Nevertheless, bursty compressed multimedia

traffic is very difficult to characterize, due to the

strong unpredictable nature of compressed aggregate

traffic, and relative traffic descriptors could provide

imprecise information for admission control

purposes (Crovella and Bestavros, 1996 and Floyd,

1997). This is particularly true for the

parameter-based approach, where traffic

characteristics are a priori set. This makes an

accurate traffic characterization hard to perform,

since admission control criterion does not properly

take into account dynamics of the aggregate running

into the network link. For bursty sources, this often

translates in a network link underutilization, or a

failed respect of QoS guarantees (Floyd, 1997).

Measurement Based Admission Control (MBAC)

algorithm class generally performs traffic evaluation

and bandwidth estimation through a dynamic

Camarda P., Guaragnella C. and Striccoli D. (2006).

A PREDICTIVE MULTI-CHANNEL MBAC TECHNIQUE FOR ON-LINE VIDEO STREAMING.

In Proceedings of the International Conference on Signal Processing and Multimedia Applications, pages 25-32

DOI: 10.5220/0001568300250032

 SciTePress

observation of traffic behaviour. This approach is

useful to take into account aggregate bandwidth

fluctuations by repeated “on-the-fly” measurements

in past periodic time intervals, and subsequent actual

bandwidth estimation, even if this generally comes

at a cost of a computation complexity overload. This

approach is particularly suitable in an on-line

context, where only part of the aggregate data set is

available for bandwidth estimation and actual load

prediction.

An efficient MBAC scheme should thus provide

an intelligent admission decision, based on network

resource measurements, whose accuracy strongly

affect MBAC efficiency and robustness. MBAC

algorithms exploiting past traffic measurements and

QoS parameters a priori specified decide, based on

actual traffic estimation, the new flow admission or

rejection.

Several different strategies for traffic

measurements have been developed, each of them

strongly influencing the MBAC algorithm

performance. In Jamin et al (1997a) three different

MBAC techniques are compared. In Jamin et al

(1997b) a MBAC algorithm is proposed for a packet

delivery service with bounded delay requirements.

In this algorithm, admission decision is based on a

token bucket filter model, that is able to calculate

worst case delays. The algorithm has been tested in a

wide variety of scenarios, and including also LRD

traffic. The MBAC proposed in Casetti et al (1996)

presents a quite simple traffic estimation process in a

single-link scenario, able to achieve a high

bandwidth utilization, without violating, at the same

time, QoS guarantees, specified as delay bounds. In

Grossglauser and Tse (1999) the impact of some

parameters, like estimation errors, flow dynamics

(departures, arrivals, and bursty traffic time scales)

and system memory, on MBAC robustness are

analyzed. Their reciprocal interactions and impact

on QoS specifications are considered for a unified

framework. In Grossglauser and Tse (2003) a

MBAC algorithm is proposed, based on a time scale

decomposition of a flow aggregate, taking also into

account flow arrivals and/or departures. Bandwidth

fluctuations are decomposed into “fast time scale”

and “slow time scale” contributes, separated by the

“critical time scale” threshold. Aggregate

fluctuations slower than the critical time scale are

exploited for admission control policy while the

faster are used to estimate the necessary amount of

bandwidth needed to absorb high bandwidth

fluctuations in a brief time period.

In this paper a novel MBAC algorithm is

proposed. It is based on a set of aggregate traffic

measurements performed in past time intervals, to

derive statistically an estimation of the aggregate

bandwidth and of the available bandwidth margin,

necessary to absorb instantaneous peak rates, with

the same approach used in Grossglauser and Tse

(2003). Available bandwidth estimation is then

exploited to admit new flows without violating QoS

requirements of the whole aggregation.

The main novelty of the proposed algorithm lies

in the criterion adopted for aggregate bandwidth

estimation, that makes use of linear predictive

techniques. In particular, aggregate traffic

measurements are performed in time intervals

preceding the new flow request of admission. This

data set is exploited to statistically derive a

prediction of the aggregate behaviour in a future

time interval. This prediction is then combined with

the QoS parameter, that is, overflow probability, to

decide the new flow admission. In this context the

overflow probability is the probability that the whole

aggregate, included the new flow, exceeds the

available link bandwidth.

The linear prediction technique dynamically

performs an estimation of the statistical aggregate

characteristics; the algorithm utilized is based on a

“multi-channel” linear prediction, to determine the

predictable characteristics of a set N different

multiplexed video sources.

Predictive

bandwidth

estimation

subsystem

Measurement

Based

Admission

Control

subsystem

Figure 1: the processing scheme adopted in the MBAC

system.

As it will be more clear in the next sections, this

work aims at implementing a numeric filter that

properly describes the predictable portion of

aggregate traffic to quantify the overflow probability

for admission decision.

The paper is structured as follows. In Section 2,

the novel MBAC approach based on a bandwidth

prediction system, is presented and explained. In

Section 2.1 the predictive system is analyzed in

detail. Bandwidth prediction is then used in the

MBAC algorithm, whose principles are explained in

Section 2.2. In Section 2.3 the whole MBAC

algorithm is implemented. Section 3 provides some

significant numerical results. Finally, Section 4

gives some conclusions to the proposed work.

SIGMAP 2006 - INTERNATIONAL CONFERENCE ON SIGNAL PROCESSING AND MULTIMEDIA

APPLICATIONS

2 THE MBAC PREDICTIVE

APPROACH

Statistical multiplexing of video sources allows an

efficient use of the channel bandwidth for the video

streaming application even if high frame variability

is present (Salehi et al, 1998).

It can be observed that the bit rate of MPEG

coded videos arranged in a composite bit stream

(statistically multiplexed transmission system)

presents a periodicity linked to GOP size: if the GOP

length is N, the multiplexed video stream bit rate

will still present a periodicity of N frames.

This particular case makes the bit rate time series

“cyclostationary” over limited time intervals, and

somehow “predictable”; the N frame-step

stationarity can be exploited to predict the future

behavior of the bit rate over a predictable time

window.

The predictive nature of the video aggregate can

be fruitfully exploited for designing MBAC system

for on-line application in video streaming systems. It

can be subdivided into two main parts:

the predictive system, used to predict the

bandwidth total allocation upon a future time, based

on the bandwidth measures accumulated on a

previous limited observation time;

the admission control system.

Figure 1 pictorially represents the MBAC

system.

2.1 The Predictive Multi-channel

System

Predictive bandwidth estimation is carried out by

standard predictive algorithm, with a multi-channel

approach. The proposed system performs several

predictions of the bit rates for several future time

lags in order to estimate the whole bandwidth

allocation required over a time-finite length

observation window.

As it can be seen by figure 2, a quasi-periodicity is

still present in the aggregate bit stream, and such

characteristic comes from the intrinsic nature of the

MPEG video encoded stream. The GOP quasi

periodicity allows to think the process as a cyclo-

stationary process, so that statistical parameters can

be computed separately on each stationary random

process extracted by subsampling the aggregate bit

rate time series with GOP length time step.

Figure 2: Time evolution and spectrum of a 6 streams CIF

aggregate.

aggregate

ch. 1

ch. 2

ch. 12

Figure 3: the channel splitting system; 12 sub-channels

depend on the 12 GOP length.

Figure 3 represents a 12 (GOP size) channel

demultiplexer of the aggregate bit rate information.

Each channel will undergo a time prediction to

estimate the bandwidth allocation of each channel

video sub-stream. All the measures will contribute to

the prediction of the whole aggregate video

bitstream on a window extended over the predictable

time length of the process at hand. Figure 4 reports a

time average evolution of the process correlation

function (normalized auto-covariance function of

each sub-samples process).

Of course, such analysis can be still valid over a

limited time duration, depending on the correlation

of the process at hand.

The first step to this goal is thus oriented to the

estimation of the time average correlation length of

each sub-channel bandwidth time series. At this aim,

each channel normalized covariance function

(correlation coefficient) has been carried out. As

long as the process we are dealing with is not

properly stationary, several estimates of each

channel correlation functions have been averaged to

estimate the time correlation length.

In the proposed approach, the time correlation length

has been chosen as the time lag where the

correlation function is greater than 0.5. Prediction

A PREDICTIVE MULTI-CHANNEL MBAC TECHNIQUE FOR ON-LINE VIDEO STREAMING

can thus be carried out over time extending not

further than 2 seconds. This means, at our full time

resolution (25 frames/s), 25×2 = 50 frames, so that,

for each channel, only 4 frames can be considered as

the goal of our predictive technique.

a) Sub-channel-1 correlation function

b) detail of the correlation function: samples only 5 samples result

above the 0.5 value.

Figure 4: a) Time averaged correlation function of the

sub-channel 1 process and b) detail of the correlation

length.

The predictive scheme for time prediction of

bandwidth allocation is reported in figure 5. After

demultiplexing the data stream, filtering of each

channel output is carried out. The filtering procedure

is performed to extend process stationarity over

longer time intervals, as subsampling tend to reduce

the correlation between successive measures of

bandwidth allocation (Shanmugan and Breipohl,

1988). This operation, in any case is straightforward

and requires no computation: if a rectangular

impulse response is used in the filter, the low-passed

channel time series is at each time instant the sum of

all the bandwidth measures along the widow length.

This measure is simply obtainable by tracking the

required transmission buffer filling, as the buffering

at the transmission end cannot be avoided.

The data to be transmitted are simply rearranged

in 12 sub-fluxes and each channel buffer filling is

tracked to be used as the data input of the

multi-channel prediction of the whole aggregate bit

rate.

For each channel time series, a multistep

predictive technique is implemented.

Linear prediction assumes that, stated b

the n-th

sample of a predictable time series in additive white

gaussian noise (AWGN, w

), it can be estimated

from its M previous samples x

n-1

,..., x

n-M

by the

relation (Proakis and Manolakis, 1996):

nknkn

bcxw

−

⋅+

∑

(1)

The equation in (1) can be solved by means of a

minimum mean squared error criterion. To obtain a

compact description of the problem and a direct

solution, matrix notation can be introduced:

⋅

= (2)

where X is the [N×M] data matrix of the signal

time series,

c is the [M×1] linear prediction

coefficient vector and

is the [N×1] vector of the

data to be predicted. Let us remember that N

represents the length, in frames, of the aggregate

periodicity. The number of equations of (2) is much

higher than the number of coefficients of the linear

prediction filter.

The solution in the mean squared error sense can

be obtained by the use of the pseudo-inverse of X,

giving:

(

)

cXXXb

−

⋅⋅

(3)

This allows to compute the coefficients of the

linear FIR predictor filter. This filter thus predicts x

(the elements in b vector) from its M preceding

values [x

n-1

, ..., x

n-M

]. The same procedure can be

established to predict the x

n+1

, x

n+2

and so on in the

future. Going a step further, we can perform a joint

estimate of the next four samples [x

, ..., x

n+3

basing on the observation of a certain number M of

preceding channel bandwidth measures. At this aim,

equation (4) calculates a joint four-step prediction

measure for each channel:

(

)

CXXXB

−

⋅⋅

(4)

SIGMAP 2006 - INTERNATIONAL CONFERENCE ON SIGNAL PROCESSING AND MULTIMEDIA

APPLICATIONS

aggregate

ch. 1

ch. 2

ch. 12

Tx buf 1

Predictive

System

Tx buf 2

Tx buf 12

BW allocation

Buffer filling

Data

Figure 5: the predictive system scheme. The orange lines

indicate the measures used in the predictive system.

In (4) C is a 4 columns vector of M coefficients.

Each column in the vector contains the M

coefficients of a linear prediction filters for the four

successive channel bandwidth measures to be

estimated.

The four bit rate measures predicted by (4) for

the 12 channels are then summed together, thus

obtaining the total number

of bits predicted over

a period T of two seconds. This gives us a predictive

estimation of the total channel bandwidth, obtained

This measure is then used to construct a

predictive MBAC system for on line applications as

described in the next section.

Figure 7 represents the cumulative bandwidth

estimated in the predictable observation window

both for the true aggregate and for the predicted. The

difference is the prediction error.

2.2 The MBAC System

In this work, a look-ahead system, able to take care

of the predictable information about the bit rate time

evolution is proposed, to try to build up a

Measurement Based Admission Control system.

The assumed transmission medium is

characterized by a constant maximum channel

capacity, C.

Both statistics of the prediction error and the

predicted (low passed) composite video stream for

all the cases of K-streams composite video have

been computed.

The MBAC system, basing on the information

about the acquired statistics of the composite video,

should be able to accept or deny the request of a end

user for the addition of a new stream in the video

aggregate. The acceptance of the user request should

not compromise the performances of the aggregate

video stream of any user, so that a measurement

based approach should be able to guarantee, within a

given probability, that the inclusion of a new stream

in the aggregate doesn’t lower too severely the

whole system performances.

Figure 6: the scheme used in each channel prediction

system. The red time samples are the goal of the predictive

system.

Figure 7: Predicted Vs true time series of the whole video

aggregate in the prediction window (about 2 seconds).

An efficient MBAC system should be able to use

as much as possible the available channel capacity,

so that the higher the number of videos in the

aggregate bit streams, the higher the obtained

performances of the streaming system.

Usually MBAC techniques use few statistic

parameters such as the peak bit rate, the average bit

rate and the standard deviation bit rate to define a

procedure for the admission, so that a low bandwidth

efficiency can be often experienced to be able to

guarantee QoS.

In this work we deal with on line streaming. Not

all the video stream is available, so that statistics for

the generic N-video aggregate case are used to

define the MBAC admission, in conjunction with the

acquired statistics on the past observation of the

process at hand.

The definition of the MBAC system needs a

conservative margin with respect to C, to manage all

the unpredictable bit rate requirements in the K-

streams composite video. Such a margin should be

neither too strict, to avoid too frequent overflow of

A PREDICTIVE MULTI-CHANNEL MBAC TECHNIQUE FOR ON-LINE VIDEO STREAMING

the available channel capacity C, nor too mild, to

avoid low bandwidth efficiency of the on line

streaming system.

The definition of a suitable margin with respect

to the channel capacity C of the system is carried out

basing on the prediction statistics for several N-

aggregate streams, with different values of N.

The channel capacity overflow probability event

can depend on two possible causes:

− overflow coming from a peak in the prediction

error;

− overflow coming from the peak of the required

bandwidth to deliver the composite video

stream.

The probability of the union of these events can

be approximated by the sum of the two probabilities.

As we are searching for a conservative MBAC

algorithm, we can assume valid the following

relation:

qsi

PPP≈+

(5)

where P

is the required overflow probability, P

is the overflow contribute coming from the signal

component and P

is the overflow probability

coming from the prediction error.

As a further requirement for the MBAC system,

we assume the P

probability be small with respect to

=× (6)

with α a coefficient to be defined, so that from

(3.1):

(1 )

=+ ⋅ (7)

In our experiments, we assume that

P be small if

compared with the probability of the overflow event

due to the signal specificity, so that we can choose

an order of magnitude smaller than

P ; thus, we

suppose

0.1

= .

With such an assumption, we can draw the

graphics for the P

as a function of the bit rate for

several orders K of the composite video stream.

By drawing the polynomial fitting curves of the

bit rate as function of the P

, we can compute the bit

rate of the composite video signal as a function of

the overflow probability P

2.3 The MBAC Implementation

The overflow probability can be used to define a

MBAC system; the global bit rate of the K-streams

composite video has to be considered; to take care of

the two components of the bit rate, the predicted bit

rate is computed as the sum of two components: the

estimated bit rate in the prediction window and a

margin with respect to the channel capacity, to avoid

the frequent occurrence of the overflow event.

The predictable bit rate is estimated as

previously described, while the margin depends on

the number of aggregated streams and on the value

of the predicted bit rate.

Two different measures have been carried out,

one over the predicted signal and the other over the

prediction error.

Let

(b) be the probability density function (pdf)

of the prediction error with respect to the bit rate,

and

(b) its corresponding cumulative distribution,

the quantity

(

)

(

)

b-FbF

ncn

represents the

probability that the event “the difference between

estimated bit rate and the true one is greater than b”.

This probability will be used in the proposed MBAC

algorithm.

To accept a new streaming client, the available

composite video bit rate must satisfy the relation:

(

)

(

)

BR N P Br P C

+Δ ⋅ ≤

(8)

where

BR and Br

are evaluated by (9). BR is a

function of the overflow probability Ps, and

represents the composite video stream bit rate in the

hypothesis the new stream is accepted and

multiplexed together with the given K-streams

composite video, to originate a (K+1)-streams

composite video, and

is the corresponding

margin with respect to the channel capacity C to

absorb the peaks of the BR due to the

unpredictability of the phenomenon.

Whenever

BR Br C

Δ< the new client will be

considered as admissible to the service: the

probability of overflow will be estimated and sent to

the client that, considering such information, might

decide whether to accept the service or deny it.

The equality can be assumed in (8) as a limit

condition to determine the searched

value.

For each K, and defined α in (6), the

complementary curves of the probability as a

function of the bit rate can be approximated by

polynomial functions in the variable P

; we compute

a single polynomial function for the unpredictability

of the bit rate, ΔBr, and a family of curves for the N-

streams cases of composite video streams.

Thus we have:

SIGMAP 2006 - INTERNATIONAL CONFERENCE ON SIGNAL PROCESSING AND MULTIMEDIA

APPLICATIONS

() ( )

() ()

isns

sns

rP Br P c P

BR N P a K P

⎧

Δ=Δ⋅=⋅

⎪

⎨

⎪

=⋅

⎪

⎩

∑

(9)

Assuming to be in proximity of the channel

capacity, with a K-streams composite video, the

algorithm steps consist in:

1. Compute the probability Pso due to the (K+1)

case finding the zero of the polynomial

equation BR(Pso) = C:

()

1, ( 1)

on so

CBRK P aK P

=+= +⋅

∑

2. Compute the margin with respect to the

channel capacity for the (K+1)-streams

composite video, Br(Pso):

()

1, ( 1)

on so

Br K P c K P

Δ+⋅= +⋅

∑

3. Compute the new maximum bit rate to be

assumed for the smooth signal, BR1:

()

1 S

BR C Br P

=−Δ ⋅

4. Compute the minimum value assumed at the

new computed bit rate for the (K+1) case,

min

, as the solution of the polynomial

equation BR

=BR(K+1,P

min

1min min

(1,) (1)

RBRK P aK P

=+= +⋅

∑

5. Compute the estimated overflow probability

as:

(1 )

=+ ⋅

3 EXPERIMENTAL RESULTS

The experimental results we present all refer to real

data of composite videos obtained by statistically

multiplexing MPEG encoded YUV-CIF videos, with

GOP of 12 frames and 30 fps. All the created

composite video streams have been obtained

randomly mixing sources with different statistics,

and composite videos are limited to K=20 sources

composite video. Figure 8 refers to the obtained

complementary cumulative distribution function for

the cases of K multiplexed sources, with K in the set

[1,20].

Curves in figure 8 and 9, representing the

complementary cumulative distribution functions for

the aggregate predicted bit rate and errors

respectively, have been polynomial fitted to give

analytical expression in (9).

Two sets of K-streams composite videos time

series bit rates have been produced: a training one,

used to obtain curves in figures 8 and 9, and a test

set, used to obtain experimental results.

The test of the proposed algorithm has been

carried out on the test set of K-streams composite

videos to estimate the expected overflow event.

A capacity C communication channel has been

assumed, and a number of video, K, in the

composite video stream has been selected in order

be close to the channel capacity and give a

negligible overflow probability. A request of service

by a (K+1)-th client has been assumed, and the

overflow probability for that client has been

estimated as previously described.

Figure 8: 1–F

(BR): complementary cumulative

probability functions of the bit rates for several orders of

composite videos.

Repeated experiments in the same conditions

have been carried out. The experimental measure of

the overflow probability for the test cases have been

computed as the fraction between the overflow

experienced time duration and the total duration of

the (K+1)-streams composite videos.

The obtained overflow probability is mostly well

predicted by the proposed algorithm.

Figure 10 presents the histogram of the obtained

probability values for 100 cases of (K+1)-streams

composite videos in the test set. The fraction of test

cases showing an experienced overflow probability

greater than the estimated one results lower than

15% in our simulations; also, the difference in

probability value is never much higher than the

estimated one.

A PREDICTIVE MULTI-CHANNEL MBAC TECHNIQUE FOR ON-LINE VIDEO STREAMING

Figure 9: 1–F

(BR): cumulative complementary

probability functions of the prediction error

(unpredictability) of the bit rates for several orders of

composite videos.

4 CONCLUSIONS

In this paper, a new MBAC algorithm for VBR

multimedia streaming has been proposed. It makes

use of an innovative bandwidth prediction technique

to more efficiently admit new flows in a network

link with aggregate traffic already running, without

violating the QoS requirements of the whole

aggregation.

Figure 10: Histogram of tested sequences: the estimated

probability (signaled by the red arrow) is lower than the

experienced one only for the 15% of the total test cases.

The proposed bandwidth prediction technique

exploits the particular, quasi-stationary nature of

aggregate traffic to predict bandwidth utilization in a

relatively short time period. This approach is

particularly useful to take into account aggregate

dynamics quickly varying in time. Furthermore, this

straightforward predictive technique can be fruitfully

implemented in advanced video smoothing systems,

where aggregate video transmission consists of

Constant Bit Rate (CBR) pieces. The proposed

predictive technique can thus be exploited to

establish the maximum CBR size allowed for video

smoothing in the next two seconds.

The predictive bandwidth estimation technique is

the core of the MBAC algorithm. The predicted

bandwidth information is utilized to perform a new

flow admission, that respects the given aggregate

overflow probability (QoS parameter).

Performed simulation show the effectiveness of

the proposed MBAC for a wide number of

simulation scenarios. Experimental results show that

the observed overflow probability almost never

exceeds the correspondent QoS parameter, testifying

a good performance of the proposed algorithm and

its ease of use in several practical scenarios.

REFERENCES

Garrett,M.W., Willinger,W., 1994. Analysis, Modeling

and Generation of Self-Similar VBR Video Traffic.

ACM Computer Communications Review. Vol.24,

N.4, pp.269-280.

Crovella,M., Bestavros,A., 1996. Self_Similarity inWorld

Wide Web Traffic: Evidence and Possible Causes. In

ACM Sigmetrics ’96.

Floyd,S., 1996. Comments on Measurement–based

Admission Control for Controlled Load Service.

Technical report. Lawrence Berkeley Laboratory.

Jamin,S., Shenker,S., Danzing,P., 1997. Comparison of

Measurement-based Admission Control Algorithms

for Controlled Load service. In INFOCOM ’97.

Jamin,S., Shenker,S., Danzing,P., Zang,L., 1997. A

Measurement–based Admission Control for Integrated

Service Packet Networks. ACM/IEEE Transaction on

Networking, N.5, pp.56-70.

Casetti,C., Kurose, J., Towsley, D., 1996. Adaptive

Algorithm for MBAC in Integrated Services Packet

Networks. In Protocols High Speed Networks

Workshop.

Grossglauser,M., Tse,D., 1999. A Framework for Robust

Measurement Based Admission Control. IEEE ACM

Transaction on Networking, Vol .7, N.3, pp.293-309.

Grossglauser,M., Tse,D., 2003. A Time-Scale

Decomposition Approach to Measurement-Based

Admission Control. IEEE/ACM Transaction on

Networking, Vol.11, N.4, pp.550-563.

Salehi, J.D., Zhang, Z.L., Kurose, J., Towsley, D., 1998.

Supporting Stored Vieo: Reducing Rate Variability

and End-to-End Resource Requirements through

Optimal Smoothing. IEEE/ACM Transactions on

Networking, Vol.6, N.4, pp.397-410.

Shanmugan, K. S., Breipohl, A.M., 1988. Random Signals

– Detection, Estimation and Data Analysis. John

Wiley and Sons. New York.

Proakis, J.G., Manolakis, D.G., 1996. Digital Signal

Processing – Principles, Algorithms and Aplications.

Prentice Hall, Third Edition.

SIGMAP 2006 - INTERNATIONAL CONFERENCE ON SIGNAL PROCESSING AND MULTIMEDIA

APPLICATIONS