PERFORMANCE CONSIDERATIONS ON ADMISSION CONTROL

FOR MULTIMEDIA SERVICES

Brikena Statovci-Halimi and Harmen R. van As

Institute of Broadband Communications, Technical University of Vienna, Favoritenstrasse 9/388, A-1040 Vienna, Austria

Keywords:

Quality of Service, Admission Control.

Abstract:

Admission control represents a convenient mechanism to provide high-quality communication by ensuring

resource availability. This paper gives on overview on different measurement-based admission control algo-

rithms suitable to be applied in multimedia service environments. A new estimator used for the measurement

process is introduced, which dynamically changes the time window used for measurements. The performance

metrics of interest within the performance analysis are made up of average utilization, packet loss and per-

centage of admitted ﬂows.

1 INTRODUCTION

Network management techniques have been of inter-

est to the networking research community for a long

time. They involve both data control and maintain-

ing a general controlled state throughout the network

by providing QoS guarantees. The provision of QoS-

controlled service requires the coordinated use of ad-

mission control, trafﬁc access control, packetschedul-

ing, and buffer management. Other techniques in-

clude ﬂow and congestion control and QoS routing.

Within service management, admission control

provides a mean in fulﬁlling the contracted service

level agreement (SLA) between the user and the net-

work provider. As a preventive congestion avoidance

mechanism, it attempts to make best use of the ﬁnite

link capacity across a network by admitting a new

ﬂow of data into a network without impacting the

guarantees of existing data ﬂows. It can be realized

in several ways. Generally we differentiate between

distributed and centralized approaches. The admis-

sion decision of distributed approaches can be based

on particular parameters of signalling messages (e.g.

peak rate or mean rate) or on measurements.

A particular method for realizing an admission

control decision is by performing measurements upon

arrival of a new admission request. These schemes

usually comprise of two phases - the measurement

phase performed by the estimator and the admission

phase performed by the policy rules. Several ap-

proaches have been investigated to date and were re-

ported in some valuable contributions such as (Jamin

et al., 1997), (Breslau et al., 2000), (Kelly, 2000),

(Casetti et al., 1996).

This paper provides a performance analysis on

measurement-based admission control for multime-

dia services. In Section 2 an overview of several es-

timators and admission policy schemes are presented.

Section 3 introduces a dynamic estimator, as well as

its performance analysis. Finally, Section 4 gives

some concluding remarks.

2 MEASUREMENT-BASED

ADMISSION CONTROL

There are several admission control approaches in the

literature, and no standardized method for the use on

a particular network. Data packet measurement-based

admission control (also referred to as passive MBAC)

is a group of admission control algorithms, which

usually measures the actual trafﬁc load and performs

the AC function using the estimation value based on

the current measured trafﬁc volume. For this pur-

pose the authors of (Jamin et al., 1997) have evalu-

ated different AC algorithms and compared their per-

formance. Further, (Casetti et al., 1996) describes an

adaptive admission control algorithm based on mea-

surements, which is the base for our observations.

A measurement-based admission control algo-

rithm can be divided into a measurement-based es-

timator and an admission policy component. The pol-

384

Statovci-Halimi B. and R. van As H. (2008).

PERFORMANCE CONSIDERATIONS ON ADMISSION CONTROL FOR MULTIMEDIA SERVICES.

In Proceedings of the International Conference on Signal Processing and Multimedia Applications, pages 384-387

DOI: 10.5220/0001938003840387

 SciTePress

icy of an algorithm is the procedure to follow at ﬂow

admission, whereas the role of the estimator is to sup-

ply the information required for the admission deci-

sion based on measurements.

2.1 Policy Algorithms

The measured sum algorithm (MS) uses measurement

to estimate the load of existing trafﬁc. Let µ be the

link bandwidth, α the new ﬂow requesting admission,

and r

the rate requested by ﬂow α. The new ﬂow is

admitted if the following test succeeds:

ν+ r

< cµ (1)

where c is a user-deﬁned utilization target and

0 < c < 1. The measured load of existing trafﬁc is

denoted with

ν. Upon admission of a new ﬂow, the

load estimate is increased using:

′

ν+ r

(2)

A measurement-based approach is doomed to fail

at very high utilization when delay violations become

exceedingly large. It is thus necessary to identify a

utilization target and require that the algorithm strives

to keep link utilization below this level.

The acceptance region algorithms compute an ac-

ceptance region that maximizes the reward of utiliza-

tion against the penalty of packet loss. These algo-

rithms are based on Chernoff bounds. Given link

bandwidth, switch buffer space, a ﬂow’s token bucket

ﬁlter parameters, the ﬂow’s burstiness, and desired

probability of actual load exceeding bound, an accep-

tance region can be computed for a speciﬁc set of ﬂow

types, beyond which no more ﬂow of those particular

types should be accepted.

Based on different combinations of measured and

declared parameters, four related techniques based

upon Chernoff bounds are presented in (Gibbens and

F.P.Kelly, 1997). The availability and ease of mea-

surement extractions (e.g., per-ﬂow vs. aggregate)

and the need for a priori trafﬁc declarations (e.g., av-

erage rate as well as peak rate) will each affect the

relative practicability of the four approaches, namely:

tangent at peak (ACTP), tangent at arbitrary location,

tangent at slope one, tangent at origin (ACTO). Table

1 illustrates basic features of these four algorithms.

For a better overview let us illustrate the compu-

tation of the effective bandwidth requirement of the

trafﬁc aggregate (all classes added together) for the

tangent at slope one algorithm:

ν = X +

K−1

∑

k=0

(3)

where,

ν is the estimate for trafﬁc load, K is the

number of different ﬂow types, n

is the number of

Table 1: Characteristics of acceptance region schemes.

Acceptance region Measurement Per-class

scheme declaration

Tangent at peak Per-class Peak rate

measurements

Number of connections

per class

Tangent at Per-class Peak rate,

arbitrary location measurements

Number of connections Average rate

per class

Tangent at Aggregate (line) Peak rate

slope one measurements

Tangent at origin Aggregate (line) Peak rate

Tangent at origin measurements

individual ﬂows of a particular type, p

represents the

peak-rate for a particular ﬂow type, X is the measured

aggregate utilization, andC is a scaling factor. For ad-

mission decision this estimated aggregate load should

be smaller or equal to link capacity µ.

The Hoeffding bounds scheme is in fact the com-

putation base for the equivalent bandwidth algorithm

(Guerin et al., 1991). It sets a probability threshold on

the sum of the source transmission rates.

C(ε) = r

ln(1/ε)

∑

(4)

where, C(ε) represents the equivalent bandwidth,

is the average aggregate arrival rate, p

is the peak

rate, ε is the target loss rate and k is the number of

ﬂows. When a new ﬂow α requests admission, the

admission control check is then based on this crite-

rion:

C(ε) + p

≤ µ (5)

2.2 Estimators

In order to be able to maintain a level of service or

guarantee of QoS, the algorithm must have available

an estimate of current resource requirements, typi-

cally bandwidth requirements. Bandwidth estimation

may be based upon predictive trafﬁc models, mea-

surements, or a combination of both. Those based on

measurements are of interest for this study.

As shown in Figure 1 an average load is computed

for every sampling period S with the time window

estimator, where S represents an integer number of

stochastic packet transmission times. At the end of

a measurement window T, which is an integer num-

ber of sampling periods S, the highest average en-

countered within the window is used as the load es-

timate for next window T. Additionally, whenever a

new ﬂow is admitted to the network, the estimate is

increased according to the advertised ﬂow informa-

tion (e.g., peak rate of the requesting ﬂow), and the

window is restarted. The estimate is also increased

PERFORMANCE CONSIDERATIONS ON ADMISSION CONTROL FOR MULTIMEDIA SERVICES

385

T T

Load = max sample

in previous window

New flow

Sample above

estimate

Rate

estimate

time

load

Restart T

Figure 1: Time-window measurement of network load.

immediately if a newly measured average is higher

than the current estimate.

As the name suggests, the point samples measure-

ment mechanism used with the acceptance region al-

gorithm takes an average load sample every sampling

period S.

For exponential averaging, an estimate of the av-

erage arrival rate can be used instead of instantaneous

bandwidth to compute admission decision with the

Hoeffding bounds approach. The average arrival rate

is measured once every sampling period S. The

average arrival rate in then computed using an inﬁ-

nite impulse response function with weight w (e.g.,

0.002):

′

= (1− w)

′

+ wr

(6)

where, r

is the average arrival rate,

ν represents

the measured load of existing trafﬁc, S is the sampling

period, and w is the weight.

2.3 Performance Comparison

Table 2 provides the number of transmitted vs.

dropped packets for different estimator - policy pairs.

This comparison as well as the analysis on output

utilization from (Statovci-Halimi, 2008a), prove the

measurement sum algorithm to reveal the best perfor-

mance at least cost. The conclusion on least cost is

based on the simplicity of this algorithm in compari-

son to the other three evaluated ones. This algorithm

is used together with the time-window estimator.

Table 2: Transmitted and dropped packets.

Admission control Transmitted Dropped

algorithms packets packets

ACTP with PointSample 26217923 150

ACTO with PointSample 25906454 35

HB with ExpAvg 26993894 2038

MS with TimeWindow 26127938 120

3 A DYNAMIC TIME-WINDOW

ESTIMATOR

The idea of our estimator is the dynamic adjustment

of the time window size to the changing trafﬁc re-

quirements. The main principle of this algorithm is

based on the attempt of avoiding the use ﬁxed-length

measurement windows, as trafﬁc characteristics are

usually unknown or can vary. By means of enlarging

or shrinking the measurement window, the adaptation

to the changing trafﬁc conditions can be provided, so

as to obtain a more or less conservative admission

process (Statovci-Halimi, 2008b).

The algorithm comprises two phases. The ﬁrst

phase represents the measurement procedure. In or-

der to increase link utilization, the length of the mea-

surement window is here continually shrunk with a

factor f

small

, until the amount of trafﬁc generated by

accepted requirements reaches a trigger value, which

is smaller than the output link capacity. As a reaction,

the algorithm then enlarges the measurement window

with a factor f

large

until the measured rate drops be-

low the trigger, at which point the window can be

shrunk again. This process changes according to the

variable trafﬁc conditions. The second phase of the

algorithm the outputs of the ﬁrst phase in order to ad-

just the trigger value according to trafﬁc ﬂuctuations.

3.1 Simulation Results

For evaluation purposes network simulations are per-

formed with ns-2, using three source models. Ac-

cording to conclusions of (Statovci-Halimi, 2008a),

our dynamic estimator is used in combination with

the measured sum algorithm.

0 1 2 3 4 5 6 7

Utilization [%]

Time Window

Figure 2: Dependency of utilization from time window.

The simulation environment comprises a two-

node topology, and the total number of ﬂows is 9986.

The voice over IP trafﬁc uses an exponential ON/OFF

SIGMAP 2008 - International Conference on Signal Processing and Multimedia Applications

386

source with transmission peak rate of 64kbit/s, packet

size of 200 and idle time of 325ms. Video trafﬁc is

simulated by and exponential ON/OFF source with an

inter arrival time of 0.1s, exponential holding time of

100s, and video holding time 300s. Background traf-

ﬁc is also applied to the network.

8150 8200 8250 8300 8350 8400 8450 8500 8550

Percentage of admitted flows [%]

Number of admitted flows

Time Window

T = 6

T = 4

T = 1

T = 3

T = 5

T = 2

Figure 3: Percentage of admitted ﬂows.

The performance of the algorithm is evaluated by

measuring the actual link utilization and drop rate.

These metrics are measured starting after an initial

warm-up period of 1600s. Figure 2 clearly illustrates

the impact that the time window length has on the av-

erage utilization. A larger time window causes a de-

crease of the average utilization.

Figures 3 and 4 prove together an interesting

property of the algorithm. For time window T = 4,

loss rate is equal to zero, whereas for a time window

equal to the sample time S, i.e. T = 1, the loss rate

is very high, as to much resources are spent for often

measurement and decision process.

-20

100

120

0 1000 2000 3000 4000

Loss rate [x10-6]

Number of packet drops

Time Window

T = 6

T = 4

T = 1

T = 3

T = 5

T = 2

Figure 4: Loss rate for different T lengths.

4 CONCLUSIONS

This paper introduces a dynamic time windowestima-

tor, which is based on load measurements and adjust-

ment of the time window size. In general, a smaller

measurement window T yields a higher utilization at

higher loss rate and a larger T keeps more reliable loss

rates at the expense of utilization level. The source

burstiness also gives differences in this context. Ex-

treme trafﬁc ﬂuctuation is more difﬁcult to handle un-

der tight guarantees, so the tradeoff between accuracy

and automate conﬁguration is a relevant factor when

deciding upon the implemented approach.

ACKNOWLEDGEMENTS

The work described in this paper was carried out with

the partial support of the BONE-project (”Building

the Future Optical Network in Europe”), a Network

of Excellence funded by the European Commission

through the 7th ICT-Framework Programme.

REFERENCES

Breslau, L., Knightly, E., Shenker, S., Stoica, I., and Zhang,

H. (2000). Endpoint Admission Control: Architec-

tural Issues and Performance. Stockholm, Sweden.

Casetti, C., Kurose, J., and Towsley, D. (1996). An Adaptive

Algorithm for Measurement–Based Admission Con-

trol in Integrated Services Packet Networks. In 5th

Int. Workshop on Protocols for High Speed Networks,

pages 13–28, Antipolis, France.

Gibbens, R. and F.P.Kelly (1997). Measurement–based

Connection Admission Control. In 15th International

Teletrafﬁc Congress (ITC15), volume 2b, pages 879 –

888, Washington DC, USA.

Guerin, R., Ahmadi, H., and Naghshineh, M. (1991).

Equivalent capacity and its application to nadwidth

allocation in high-speed networks. IEEE Journal in

Selected Areas in Communications, 9(3):378–387.

Jamin, S., Danzig, P., Shenker, S., and Zhang, L. (1997).

A Measurement–Based Admission Control Algorithm

for Integrated Services Packet Networks. IEEE/ACM

Transactions on Networking, 5(1):56–70.

Kelly, F. (2000). Distributed Admission Control. IEEE

Journal on Selected Areas in Communications,

18(12):2617 – 2628.

Statovci-Halimi, B. (2008a). Performance Comparison of

Measurement-Based Admission Control. In 13th Eu-

ropean Conf. on Network and Optical Communica-

tions, Krems, Austria.

Statovci-Halimi, B. (2008b). Support of IP Multi-Services

Through Admission Control. In First ITU-T Kalei-

doscope Academis Conference, pages 407 – 414,

Geneva, Switzerland.

PERFORMANCE CONSIDERATIONS ON ADMISSION CONTROL FOR MULTIMEDIA SERVICES

387