You are on page 1of 48

2002:122 CIV

MASTERS THESIS

Admission Control for


Streaming Services in
EGPRS Radio Networks

PER JOHANSSON

MASTER OF SCIENCE PROGRAMME

Department of Computer Science and Electrical Engineering


Division of Signal Processing

2002:122 CIV ISSN: 1402 - 1617 ISRN: LTU - EX - - 02/122 - - SE


Admission Control for Streaming Services
in EGPRS Radio Networks

Per Johansson
Lule University of Technology

Masters Thesis
Performed at Ericsson Erisoft AB
March 2002

Examiner: Supervisor:
Frank Sjberg Tomas Jnsson
Lule University of Technology Ericsson Erisoft AB
Abstract
In third generation cellular systems, new services will be offered. Some
of these suggested services require a guaranteed Quality of Service (QoS).
An example of these services is streaming, which for example requires a
certain guaranteed bit rate and delay. Since the system resources are limited,
it may sometimes be necessary to deny access for new users in order to
preserve the QoS for users already active in the system.
In this thesis it is indicated that admission control is necessary for stream-
ing in order to maintain the quality demands in overload situations. A theo-
retical model of a cellular streaming scenario is developed, and based on this
model a simulator is implemented in Matlab. Different load combinations
of speech and streaming users have been simulated. It is shown that a sys-
tem without admission control may become overloaded (jammed) at higher
loads and the number of satisfied users decreases rapidly.
In order to investigate methods of how to avoid overloaded systems and
unsatisfied users, a number of different admission control algorithms have
been developed and evaluated. These algorithms can be combined for a spe-
cific scenario depending on system parameters and known variables. The
finally proposed algorithm is based on a prediction of the total through-
put available for streaming users, with information from the admitting and
neighboring cells. Based on this, it is possible to allow a certain number
of users in the admitting cell, with a certain probability of keeping all users
satisfied. A load dependent threshold is applied to decrease the blocking
rates at low loads. Speech users have strict priority over streaming users.
This means that timeslots used by streaming users may be stolen, or pre-
empted, and the streaming users session may be terminated
With the proposed algorithm it is shown that admission control is bene-
ficial, and necessary to keep a system stable. The algorithm gives, in terms
of the number of satisfied users, nearly the same performance as without
admission control for low loads, and outstanding performance at high loads.
Acknowledgements
This Thesis would not have been the same without some very important
persons.
First of all of all I would like to thank my supervisor, Tomas Jnsson
at Ericsson Erisoft AB, for great guidance and valuable support during the
Thesis. Thanks also to Peter de Bruin for great knowledge about the subject.
I would also like to thank the whole department for a pleasant time at Eric-
sson Erisoft AB. And of course I would like to thank my examiner, Frank
Sjberg at Lule University of Technology. Finally I would like to thank my
dear girlfriend and my great family for a lot of encouragement when it was
needed.
Contents v

Contents
1 Introduction 1

2 Background 2
2.1 Cellular networks . . . . . . . . . . . . . . . . . . . . . . . 2
2.1.1 Evolution of cellular networks . . . . . . . . . . . . 3
2.1.2 The GSM System . . . . . . . . . . . . . . . . . . . 3
2.2 Circuit Switched Communication vs. Packet Switched Com-
munication . . . . . . . . . . . . . . . . . . . . . . . . . . 4
2.3 GPRS . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 5
2.3.1 GPRS Radio interface . . . . . . . . . . . . . . . . 5
2.3.2 GPRS Architecture . . . . . . . . . . . . . . . . . . 6
2.4 EDGE/EGPRS . . . . . . . . . . . . . . . . . . . . . . . . 6
2.5 Streaming . . . . . . . . . . . . . . . . . . . . . . . . . . . 7
2.6 Quality of Service . . . . . . . . . . . . . . . . . . . . . . . 8
2.7 Admission Control . . . . . . . . . . . . . . . . . . . . . . 9
2.7.1 Local traffic load admission . . . . . . . . . . . . . 9
2.7.2 Global traffic load admission . . . . . . . . . . . . . 10

3 Theoretical Model 11
3.1 User session model . . . . . . . . . . . . . . . . . . . . . . 11
3.1.1 User arrivals . . . . . . . . . . . . . . . . . . . . . 11
3.1.2 Session length . . . . . . . . . . . . . . . . . . . . 11
3.2 Radio channel model . . . . . . . . . . . . . . . . . . . . . 12
3.2.1 Path loss . . . . . . . . . . . . . . . . . . . . . . . 12
3.2.2 Diffraction, Reflection and Scattering . . . . . . . . 13
3.2.3 Co-channel interference . . . . . . . . . . . . . . . 13
3.2.4 Multipath fading . . . . . . . . . . . . . . . . . . . 13
3.2.5 System throughput . . . . . . . . . . . . . . . . . . 13
3.3 System traffic . . . . . . . . . . . . . . . . . . . . . . . . . 15
3.4 Satisfied users . . . . . . . . . . . . . . . . . . . . . . . . . 15
3.5 Model verification . . . . . . . . . . . . . . . . . . . . . . . 16
3.6 Normalized load . . . . . . . . . . . . . . . . . . . . . . . . 20
3.7 Admission control example . . . . . . . . . . . . . . . . . . 20

4 Admission Control Algorithms 22


4.1 Local admission control . . . . . . . . . . . . . . . . . . . . 22
4.2 Weighted semi-global admission control . . . . . . . . . . . 22
4.3 C/I based admission control . . . . . . . . . . . . . . . . . . 23
vi Contents

5 Comparison of AC-Algorithms 26
5.1 Full-global admission control . . . . . . . . . . . . . . . . . 27
5.2 Load dependent threshold . . . . . . . . . . . . . . . . . . . 28
5.3 Dropping . . . . . . . . . . . . . . . . . . . . . . . . . . . 30
5.3.1 Evaluation of admission control combined with drop-
ping . . . . . . . . . . . . . . . . . . . . . . . . . . 30

6 Final Algorithm Proposal 33


6.1 Performance of the proposed algorithm . . . . . . . . . . . 33
6.1.1 Revenue based measurement . . . . . . . . . . . . . 33

7 Conclusions 37
7.1 Further work . . . . . . . . . . . . . . . . . . . . . . . . . 37

Appendix A: Abbreviations 38

Appendix B: Confidence Analysis 39

Appendix C: Maximize the Number of Satisfied Users 40

References 42
Introduction 1

1 Introduction
The third generation cellular systems (3G), for example EGPRS (Enhanced
General Packet Radio Service) [12], will offer a collection of different cus-
tomer services. These services require various amounts of bandwidth, and
they all have a specific desired Quality of Service (QoS) profile to define
their requirements. An upcoming service is streaming, which is a continu-
ous stream of data that can be continuously played back during download,
e.g., music or movie clips.
The bandwidth resources in an EGPRS radio network system are lim-
ited in terms of, for example, the number of timeslots. Also, for a specific
channel the bandwidth varies over time as a function of the radio quality.
The system load (in terms of number of users competing for the system re-
sources) will also vary over time and thereby affect the throughput for each
user. This described relationship makes it necessary to supervise the number
of users that can enter the system with admission control. Increased load and
limited resources may otherwise result in saturation of the system causing a
general decrease of QoS for all users.
An ideal admission control scheme should make it possible to control
and preserve the QoS level demanded by the users, while simultaneously
utilizing radio resources efficiently and achieving a desired relation between
served users and blocked users (users not admitted to the system). This could
also enable an operator to reserve resources for services that have a higher
priority and/or give higher revenue.
The goal of this thesis is to investigate system performance and proper-
ties for a service mix of circuit switched speech and packet switched stream-
ing, with and without some kind of admission control.
2 Background

2 Background
2.1 Cellular networks
The breakthrough for mobile communication came with cellular systems.
Theoretically, base stations are placed in a hexagonal pattern as in Figure 1,
where a three sector site with three base stations covers an area with three
cells, using directional antennas with 120 degrees lobe width. These cells
have one or more carrier frequencies assigned to them, but cells close to each
other can not use the same frequencies since this would cause interference
between co-channel mobiles, producing an insufficient carrier to interfer-
ence ratio (C/I). Therefore, a systematic reuse pattern for the frequencies is
applied, illustrated by the grey cells in Figure 1, which shows a systematic so
called 1 3-reuse where the same frequencies are used in every 3rd cell. The
grey cells all use the same frequencies, which means that they are co-channel
interferers to each other. The reuse distance depends on how often the fre-
quencies are reused. Typically, coverage is good in cellular systems today,
which means that the received signal power is sufficiently strong. The in-
terfering power from neighboring cells with the same frequencies are above
the noise floor. This means that the system is interference limited rather than
noise limited. Of course, there are also channel limited systems, but they are
less interesting for this study.

41
   42

  40
43 
39
36 37 38
32    29
33    30
34
   31
  
35

25 26 27 28
20
"! "! #$ #$ 16
21
*) %& *) (' (' 17
22
/0 431 /0 2 1 2 18
23
= :9> = <; <; 19
AB AB
24

13 14 15
9
CDC 6
10
+ .-.-, + 7
11
8765 8
@?@? @?@?
12

4 5
1 FEFE 2 GH GH 3 II JJ II
Figure 1: Hexagonal cellular pattern.

In order to create multichannel systems, the frequency spectra is usually


divided into different sub-bands, referred to as Frequency Division Multiple
Access (FDMA). Each user is then assigned one subcarrier. Digital cellu-
lar systems often use additional Time Division Multiple Access (TDMA)
Background 3

as well, dividing each frequency band into several timeslots. This means
that several users can use the same frequency, and take turns on different
timeslots. Thereby, sharing the carrier frequency in the time domain.
Since there is no time for retransmission of frames during a speech con-
versation, a good method to avoid bit errors is Forward Error Correction
(FEC) [1]. This means that the information is coded with, for example, a
convolutional encoder that adds extra redundant bits to the information bits.
This type of encoders work well for individual bit errors. However, when
transmitting over a radio channel the bit errors often occur in bursts. Bursts
of errors can unfortunately not be handled by a convolutional decoder. But
if the data is interleaved before transmission the burst of errors will be scat-
tered in the deinterleaving and the decoder will only have to deal with a
number of short errors.

2.1.1 Evolution of cellular networks


First generation cellular systems (1G) were launched with start in the early
eighties. Examples are NMT 450 and NMT 900 (Nordic Mobile Telephone
system), TACS (Total Access Communication System) in United Kingdom
and AMPS (Advanced Mobile Phone Service) in the United States and Canada.
All these systems are based on analogue speech transmission using analogue
modulation and FDMA [9].
Second generation cellular systems (2G) were launched in the beginning
of the nineties and these systems employed digital transmission and TDMA.
Examples of 2G systems are GSM (Global System for Mobile Communica-
tion), D-AMPS (Digital-Advanced Mobile Phone Service) in the America
region and PDC (Personal Digital Cellular) in Japan.
Third generation cellular systems (3G) are being launched with start
2001. The purpose of 3G is to offer very high bit rates for multimedia, and
flexible connections for a variety of services to external packet networks.
UMTS and EDGE are standardized by 3rd Generation Partnership Project
(3GPP), which is a co-operation between different standardization bodies
world wide.

2.1.2 The GSM System


One of the second generation cellular systems is GSM, which was launched
in 1991. The spectra in GSM are divided into sub-bands with 200kHz carrier
spacing enabling FDMA. Each specific carrier is divided into eight timeslots

K
enabling TDMA. Binary Gaussian Minimum Shift Keying (GMSK) modu-
lation is used, providing a gross bit rate of 33 9 kbps per timeslot.
A common way to digitize speech is to use an 8 kHz sampling frequency
and 8 bits quantization, resulting in a bit rate of 64 kbps. In GSM this has
been compressed, using a mix of vocoding and waveform coding resulting in
4 Background

a 13kbps coded bitstream [9]. The coded speech will also be burst formatted,

K
channel coded and interleaved before it is sent on a timeslot with a total rate
of 33 9 kbps. A formatted sequence of bits sent during one timeslot is called
a burst.
The hierarchy of the GSM system is shown in Figure 2. Note that SGSN
and GGSN are parts belonging to the packet switched GPRS, see section 2.3,
and not to the circuit switched GSM system. The mobile station (MS)
is the equipment that users use to communicate with the network (that is
the mobile phone). Base Transceiver Station (BTS) is the site with radio
equipment, e.g., antennas and transceivers that communicate with the MS:s.
The BTS is controlled by the Base Station Controller (BSC). The BSC is
in charge of, e.g., handovers, what transmit power to use, and other higher
level system tasks. Mobile services Switching Center (MSC) sets up, su-
pervises and releases calls. It switches calls within the GSM network, to
the Public Switched Telephone Network (PSTN) or to other networks. The
Gateway Mobile Services Switching Center (GMSC) switches all incoming
circuit switched traffic from PSTN to the GSM network.

External
IP Network
(Internet)

GGSN External
IP Network
(Corporate LAN)

PSTN/ISDN
IP-Backbone
Network Other
PLMN

GMSC MSC SGSN

BSC

MS BTS BTS MS

MS MS

Figure 2: GSM/GPRS network.

2.2 Circuit Switched Communication vs. Packet Switched


Communication
In new generations of cellular networks, efforts have been made to provide
well suited interfaces for connection to external Packet Switched (PS) net-
Background 5

works, e.g., Internet, and to offer advanced data services. Therefore, it is


beneficial to use packet switched communication instead of the traditional
Circuit Switched (CS) communication [11].
With circuit switched communication, a connection is set up by allocat-
ing one fixed channel all the way (end to end). The channel is occupied for
the whole session, also during idle periods. Typically, the user is paying
during the whole session time. Circuit switched communication is therefore
suitable for constant bandwidth data flows, or when data is delay sensitive.
With packet switched communication temporary channels are used and
data in packets are forwarded on the first free channel, usually not allocated
end to end. Several users can therefore share one channel in time. Data
is queued and a scheduling algorithm gives access to the users, one at the
time. If users have a bursty behaviour of sending data the channel can used
efficiently. If packet errors occur, it is possible to use Automatic Repeat Re-
quest (ARQ) [2]. This means that packets with errors will be re-transmitted.
With PS communication the user often only pays for the amount of transmit-
ted data, and not for the connection time. Packet switched communication
is suitable when data is sent in bursts or when it is sensitive to errors.

2.3 GPRS
To offer new services with IP interoperability and high bit rates for data,
General Packet Radio Service (GPRS) was developed [8, 11]. GPRS is a
standard from ETSI for packet data in GSM systems. For GPRS, packet
switching is used for data instead of circuit switching, which is used for
speech. With GPRS it is possible to have PS access to external IP-based
networks. Users can have access all the time to external networks, only
consuming the amount of network resources that is actually needed.

2.3.1 GPRS Radio interface

K
GPRS uses a similar radio interface as GSM with GMSK modulation, eight
timeslots, and 33 9 ksymbols s per timeslot. Four different Coding Schemes
(CS) are used for different radio conditions, allowing data rates up to more
than 20 kbps per timeslot. ARQ is used to ensure error free transmission.
GRPS employs a feature called multislot, where a single user can use a
number of timeslots instead of only one. If the system load is low, higher bit
rates are possible by using the larger bandwidth provided by several times-
lots. However, when the total load increases the bit rate will decrease since
the resources must be shared with other users. In order to achieve fair shar-
ing of the channel resources when several users are allocated on the same
timeslot, a scheduling algorithm is used to decide in which order packets
should be forwarded.
6 Background

2.3.2 GPRS Architecture


The GPRS architecture is very similar to GSM. The main difference is that
PS data traffic is used instead of CS, which introduces some new compo-
nents.
The protocol structure of GPRS is illustrated in Figure 3 and is arranged
as follows: Before data packets are transmitted between the MS and the BTS
they are divided into frames at the Logical Link Control (LLC) layer. Each
frame is then divided into Radio Blocks at the Radio Link Control RLC
layer. Each Radio block is interleaved over four GSM bursts. These bursts
are then sent from the MS up to the BTS as in GSM, see Figure 2. Now, the
BTS separates circuit switched calls from packet switched data communica-
tion before it forwards CS calls to the MSC and PS data to the Serving GPRS
Support Node (SGSN). SGSN is a new component needed for GPRS, and
it forwards incoming and outgoing IP packets to the IP backbone network.
From the SGSN the packets can either go to other PLMNs or further to the
Gateway GPRS Support Node (GGSN), another new component needed for
GPRS. The GGSN provides the interface to external IP packet networks.

Header Information Network Layer


Packet

Header Information LLC Layer


Frame

Header Information ... Header Information RLC Layer


Radio Block

Burst Burst Burst Burst Physical Layer

Figure 3: Packet transformation in GPRS.

2.4 EDGE/EGPRS
The objective of EDGE/EGPRS (Enhanced Data rates for GSM Evolution/
Enhanced GPRS) is to increase user bit rates without making major changes

K
to previously installed GSM networks. Just as for GSM/GPRS, eight times-
lots, 33 9 ksymbols s per timeslot and 200 kHz carrier spacing is used. The
aim is maximum data rates up to about 384 kbps (perceived by a single user
on eight timeslots) [3].
Background 7

EDGE uses 8PSK (Phase Shift Keying) modulation in addition to GMSK.


This means that three bits are sent for every symbol instead of only one sin-

K L M
gle bit. Therefore, three times higher bit rates are theoretically possible.
This results in a gross bit rate of 33 9 3 102 kbps per timeslot at the phys-
ical layer, see Figure 3, achieving the maximum user bit rates stated above.
However, note that 8PSK is much more sensitive to noise and interference.
Therefore, GMSK modulation can still be used when a more robust modu-
lation is needed. One idea is to use 8PSK/GMSK for downlink communi-
cation, where the demand for high bit rates is larger, and use only GMSK
for uplink communication, which results in less complex transmitters in the
terminals.
To optimize the throughput for all radio conditions Link Quality Con-
trol (LQC) is used to dynamically adapt the modulation and code rate with
respect to the current link quality. EGPRS combines two LQC techniques,
Link Adaptation (LA) and Incremental Redundancy (IR).
EGPRS has nine different Modulation and Coding Schemes, MCS-1 to
MCS-9. The first four use traditional GMSK and the last five use the new
8PSK modulation scheme. The MCS:s also have different code rates and
are therefore suitable for different radio conditions. If the conditions are
very good, no coding is necessary and high bit rates can be achieved. Bad
conditions require stronger coding, which results in less payload and lower
bit rate. The technique of adapting the MCS to radio conditions is called
Link Adaptation (LA) and it is aiming at always be on the envelope of the
curves in Figure 4.
EGPRS also has an enhanced version of ARQ called Incremental Redun-
dancy (IR). First, the information is coded with for example a convolutional
encoder. The output from the encoder is punctured, which means that re-
dundant data is separated from the information. The information is then
transmitted uncoded or with a very small number of redundant bits. If the
information is not correctly received, more redundant bits are transmitted.
These redundant bits are combined with the information received earlier,
and a decoding process is performed. If there are still too many errors for
the decoding process to succeed, more redundant bits are transmitted again.
This means that punctured redundant data can incrementally be transmitted
until the information can be decoded successfully. This results in a method
that is throughput optimized, Figure 4 shows increased throughput per-
formance [3, 12]. However, due to more signal processing and memory
handling a larger delay is also introduced in some cases.

2.5 Streaming
An upcoming service in 3G networks is streaming, which is a continuous
stream of data that can be continuously played back during download. The
8 Background

70
Link Adaptation
Incremental Redundancy MCS9

60

50 MCS8

Throughput (kbps per timeslot)


MCS7
40

MCS6

30

20 MCS5

10

0
0 5 10 15 20 25 30 35 40 45 50
C/I (dB)

Figure 4: Comparison of throughput for LA, IR and for some different


MCS.

client must be able to receive data continuously and send it as a steady stream
to the application that processes the data and converts it to, e.g., sound and/or
pictures. With streaming it is not necessary to download the entire file before
displaying it. Some advantages with this is that the entire file does not have
to be stored in memory and the delay from the beginning of download to
start of playback can be relatively small.
Since the bit rate and the delay over the radio channel varies over time
due to channel effects, it is necessary to have a buffer from which data can be
taken when the transfer bit rate is too low for continuous playback without
interrupts. The drawback of buffering is that memory is needed and there
is also a startup delay introduced when the buffer is initially filled. The
delay from the initial buffering is expected not to be too disturbing for the
user, and it is also possible to use ARQ if the buffer is big enough. Minor
packet losses may also be tolerated using error concealment techniques that
minimizes perceived degradation. Examples of streaming services are news
clips, movie clips or birthday greetings.

2.6 Quality of Service


For some of the new services that will be available in 3G there is a need to
be able to guarantee a certain quality of transmit delay, bit rate and/or bit
error rate [7]. Therefore 3GPP has defined the concept of Quality of Service
(QoS). All services have specific needs and has therefore been sorted into
four traffic classes. The four classes are:
Background 9

Conversational: Telephony speech, voice over IP and video


conferencing (real time services).
Streaming: Video/audio streaming (semi real time services).
Interactive: Web browsing, data base retrieval and server access.
Background: Delivery of E-mails, SMS and downloads from
databases.

The main difference between these classes is how delay sensitive the traffic
is. The conversational class has the highest demands for low delay due to the
human perception. Streaming can allow a greater delay because of buffer-
ing. The interactive class expects respond in an acceptable time and the
background class does not have any specific needs at all considering the de-
lay. Conversational and streaming classes require that the data is received in
the same order as it was sent. Interactive and Background classes typically
expect error free transmission, which makes re-transmission necessary.

2.7 Admission Control


All new services that are offered in 3G have different QoS-needs. To main-
tain a certain QoS on system level for different services at different load
situations, admission control (AC) may be needed. The main idea of AC is
to let all users in, until a certain load limit has been reached. When the limit
is reached, new users are denied access, or blocked, to avoid saturation of
the system and degraded quality for the already admitted users. Admission
control can be controlled by one or more variables, e.g., number of channels,
number of users, bandwidth, priority and power consumption of each user.
Therefore, it is possible to trade between QoS and system capacity for differ-
ent services. Below, two simple examples of admission control algorithms
are described [4].

2.7.1 Local traffic load admission


A simple algorithm only considers the traffic load in the admitting cell using
two parameters, c the total number of available channels, and u the number
of active users in the cell. The utilization factor, F is defined as

F M uc K
A new user is admitted if
F N Fthreshold
Here, the threshold Fthreshold determines the trade-off between capacity and
quality in the system.
10 Background

2.7.2 Global traffic load admission


The second algorithm is similar to the first one, but also considers the traffic
in interference areas. Denote the total number of available channels in cell
i by ci . Define the area Ai as the set of dominating co-channel interference

MPO Q Q Q Q Q Q R
cells. Usually the admitting and the nearest cells with the same frequencies,

M
A22 10 11 21 22 23 33 34 in Figure 1. Denote the cells that have cell

O Q Q Q Q Q Q R
i in their interference area by Si , e.g., cell 22 is a part of A10 and S22
A10 A11 A21 A22 A23 A33 A34 .
The utilization factor in one interference area A i , denoted by FAi , is then:

FAi M jAi u j
jAi c j
K
A new call is admitted if

FA j N Fthreshold QTS A j Si K
So, if any utilization factor in all the interference cells is higher then the
threshold value the user is blocked, otherwise admitted.
Theoretical Model 11

3 Theoretical Model
To be able to analyse the problem, a suitable model of the system is needed.
In this thesis the main idea is to stochastically model required streaming
bandwidth and available radio channel bandwidth and compare these. For
the radio channel bandwidth, the pre-emption of channels by speech users
must be taken into consideration.

3.1 User session model


The user session model consists of two parts, user arrivals and session length.
These distributions are necessary in the simulator to achieve a statistically
varying load in the system over the simulation time.

3.1.1 User arrivals


For the model of user arrivals to the system, it is reasonable to assume that
arrivals occur randomly in time. If events occur randomly in the time or
space domain, the number of events at a given time or space is Poisson dis-
tributed [14]. Therefore, users are assumed to arrive to the system according
to a Poisson process with probability density function:

U VWM eX
fX k k
k! Q (1)

where X is the number of arrivals per time unit. The arrival intensity is
chosen to get a certain load in the system. The parameter is set separately
for speech and streaming users.

3.1.2 Session length


The probability of a speech user ending a session is assumed to be the same
at all times. Therefore, the session length is assumed to be exponentially
distributed with probability density function:

U VYM
fX x
1
m
e X Z Q
x m
(2)

where m is the mean value. The same assumptions are made for streaming
users as well. The streaming service uses a buffer with several seconds of
memory and users are not expected to terminate the transfer before the buffer
is full and the playback has started. Therefore, the streaming session time is
divided into two parts. One exponentially distributed part with a mean value
of 10s plus a constant with value 10s where the latter part corresponds to ini-
tial buffering [15]. A speech users session time is exponentially distributed
with a mean value of 40s.
12 Theoretical Model

3.2 Radio channel model


We assume a regular hexagonal cell plan with reuse pattern 1 3, as depicted
in Figure 5. The nine numbered grey co-channel cells use the same fre-
quencies and are assumed to serve both streaming and speech traffic. Only
the first tier of interfering cells are assumed to affect C/I in the studied cell.
A wrap-around technique is used to model cells outside the cell plan. For
example, the next co-channel cell to the right of cell number three is cell
number one again.

7 cc fedd cc 8
] `_^ ] 9
a \[b a
4 jiji hg jiji 5
o prq o 6
zy xw zy
1 ~}~}  |{ ~}~} |{ 2
nm nm lk lk 3 vu s t s t

Figure 5: Hexagonal cellular pattern for the model.

Every cell is assumed to have two frequencies with eight timeslots each,
which results in a total of sixteen timeslots per cell that can be shared by both
streaming and speech users. Speech users use a whole timeslot during their
session and streaming users may use several timeslots but can share these
with other streaming users. Speech users have strict priority over streaming
users, which means that a speech user entering the system will steal one
timeslot from streaming users if needed, so called pre-emption.
The throughput in a cellular radio network depends on the radio quality
(generally C/I), which depends on several factors. The most important rea-
sons for a varying C/I are different kinds of channel effects, e.g., path loss,
diffraction, scattering, co-channel interference and multipath fading .

3.2.1 Path loss


Path loss or attenuation causes the signal to be weaker further away from
the transmitter. This effect limits the coverage area of the cells but it also
decreases interference from other cells far away. For free space propagation,

X
the path loss is inversely proportional to the square of the distance [1]. In an
urban environment, the propagation is typically proportional to r 3 5 . The
coordinates of the mobiles are not modeled and therefore it is not possible
to model path loss for each mobile.
Theoretical Model 13

3.2.2 Diffraction, Reflection and Scattering


When the radio path does not have free line of sight, the radio waves are
diffracted and reflected around obstacles in the signal path. This effect is also
called shadowing since the signal path is shadowed by obstacles. This will
cause variations in signal strength depending on the relative velocity and the
obstacles between the transmitted and the receiver. The fading frequency is
normally around fractions of Hertz, in other words slow fading. A common
model for slow fading attenuation is the lognormal distribution [13].
Reflections occur when the radio wave is reflected against large objects
compared to the signals wavelength, e.g., buildings and the surface of the
earth.
Scattering occurs when the radio wave is spread out (scattered) in all
directions, when hitting a surface which dimensions are in the order of the
signal wavelength or less.
The combination of radio waves received from reflection and scattering
adds up constructively or destructively at the receiver, creating an interfer-
ence pattern. This causes rapid variations in signal strength and is referred
to as multipath fading or fast fading.

3.2.3 Co-channel interference


All users in the co-channel cells (using the same frequency and timeslot)
will cause interference. The ratio between desired signal and interfering
signals is referred to as carrier to interference ratio (C/I). The received C/I
strongly affects the throughput or speech quality for a mobile. The model
used here assumes that interference from users in adjacent co-channel cells,
e.g., the first tier, are predominant and interferers in other co-channel cells
are disregarded.

3.2.4 Multipath fading


Multipath fading (fast fading) consists of very fast variations compared to
the length of the streaming buffer that has several seconds of memory. The
streaming buffer will therefore act as a low-pass filter on fast fading through-
put effects. Scattering will only cause some ripple on the buffer content as
well, and the users are assumed not to notice anything. Therefore, fast fading
is excluded from the model.

3.2.5 System throughput


To summarize the description above of radio channel effects in the model,
the channel attenuation is assumed to be lognormally distributed (due to the
shadow fading). This means that the attenuation measured in dB will be
14 Theoretical Model

normally distributed. The C/I, with mean and standard deviation , for
each timeslot is assumed to be a stochastic process:

UV
g t N U Q VQ (3)

where and are functions of the assumed total load in the interfering cells
Ltot , which still remains to be defined,

M U V Qf Ltot (4)
M f U L VK tot (5)

In some sense, path loss is implicitly included, since and will be


affected by the path loss.
In Figure 6 the throughput with incremental redundancy and link adap-
tation is plotted as function of C/I. A reasonable approximation is to say that
the throughput (in kbps) is a linear function of C/I (in dB) with slope two.
This proposed mapping from C/I to throughput is shown in Figure 7. The
throughput for timeslot i can therefore be written as:

T Pi M L U V U Q VQ
2 g t N 2 2 (6)
2 M 2 L U L VQ (7)
M 2 L U L VK
tot
2 tot (8)

70
Link Adaptation
Incremental Redundancy MCS9
k=2
60

50 MCS8
Throughput (kbps per timeslot)

MCS7
40

MCS6

30

20 MCS5

10

0
0 5 10 15 20 25 30 35 40 45 50
C/I (dB)

Figure 6: Comparison of throughput for different MCS.

V
The total throughput available for streaming is modeled as the sum of
throughput on all available timeslots for streaming T S st . These are the U
total number of modeled timeslots reduced by the number of timeslots used
for speech. The total throughput per cell for streaming becomes:
Theoretical Model 15

70

60

k=2
50

Throughput (kbps per timeslot)


40

30

20

10

0
0 5 10 15 20 25 30 35
C/I (dB)

Figure 7: Mapping from C/I to throughput.

M T Pi N U 3 Q 3 VQ
T Sst


T Pst (9)

M L U V L T S K
i 1
3 2 Ltot (10)
M 2 L U L VL T S K
st
3 tot st (11)

3.3 System traffic


The source bit rate for streaming users is assumed to be constant at 32 kbps
and the total streaming bit rate demand per cell is therefore:

L Q
32 ust (12)

where ust is number of streaming users in the admitting cell.

3.4 Satisfied users


To be able to compare different admission control algorithms, some kind of
performance measure is needed to tell whether a user is satisfied or not. This
measure can be defined in several ways, which may give different results. In
this evaluation, a streaming user is said to be satisfied if the bandwidth is
large enough for the buffer never to run out of data. The fraction of satisfied
16 Theoretical Model

streaming users could be defined in several ways. Here it is defined by the

U L N L V
equation:
Time 32 ust T Pst c
(13)
Total time
where c is a security factor near one, because a little margin is assumed to be
necessary. In words, the fraction of time when the throughput provided by
the system is higher than the demanded capacity. Actual capacity is the total
throughput for all timeslots available for streaming. The factor c is used to
adjust the estimate of the fraction of satisfied users better. It is also affected
by the buffers sizes that the users are assumed to have. If the actual capacity
is below the desired capacity for only a short moment, the data in the buffers
will decrease but the users will not notice anything. If the actual capacity
goes above the demanded capacity again, the buffer will hopefully increase
to the maximum limit again.

3.5 Model verification


To verify the model it has been simulated in Matlab and compared to similar
scenarios simulated in PETRA (Packet Emulator Tool for Radio network
Applications). PETRA is a radio network simulator used at Ericsson. It is
especially designed for evaluating EGPRS packet data systems, and models
speech and video streaming, as well as WWW traffic. Since the streaming
buffer is assumed be 10s, in PETRA, streaming users are said to be satisfied
if they receive all their packets with a maximum delay of 10s, which means
that no re-buffering is necessary. Some of the parameters used in the PETRA
simulations are presented in Table 1. Multislot capability is the number of
timeslots that one mobile can use.

Frequency reuse pattern 1/3


Frequencies per sector 2
Timeslots per frequency 8
Multislot capability 4
Cells in the system 27

Table 1: Parameters used in PETRA simulations.

Suitable values for the parameters of the C/I distribution, and , were
obtained from PETRA simulations with various streaming and speech user
loads.
Since a streaming users normally require more than one timeslot, a stream-
ing user is assumed to affect the resulting C/I more than a speech user. The
Theoretical Model 17

total load in one cell is therefore defined as:


Ltot M ust L u Q sp (14)
where is a constant and usp and ust are the number of speech and streaming
users per cell, respectively. The constant is applied to match the influence
from speech compared to streaming. In Figure 8, the mean system C/I is
plotted for different loads. For high load, including speech users, a large
number of streaming users are blocked. Therefore, the produced C/I val-
ues are too high in relation to the number of users, since blocked users are
also included in this measure. Only the simulations with normal load are
therefore used to calculate . By shifting the curves with a factor times

K
the number of speech users and calculating the error, compared to the curve
without speech users, using least squares method, is found to be 0 553.
35
No speech users
Two speech users
Four speech users

30
Mean value of C/I (dB)

25

20

15

10
0 1 2 3 4 5 6 7 8 9 10
Average number of streaming users per cell in the interference cells

Figure 8: Mean value of C/I for different numbers of streaming and speech
users in the interference cells, simulated in PETRA.

Since the curves including speech users are assumed to produce a too
high C/I for high loads, the curve without speech users is used in a least
squares method polynomial fit. The following equations for and , i.e.,
mean and standard deviation of C/I, were obtained:

M 0 K 0209 L L 0 K 829 L L 10 K 2 L L 50 K 1 Q
3 2
(15)
M 0 K 0279 L L 0 K 709 L L 9 K 34 K
tot tot tot
2
tot tot (16)
Based on these functions, the distribution of C/I is obtained from (3).
In Figure 9, the estimated function (15), from total load (Ltot ) to the mean
of C/I and the simulated values are shown. In Figure 10 the corresponding
plot for the standard deviation (16), is presented.
18 Theoretical Model

35
Theoretical mapping
Simulated values

30

Mean value of C/I (dB)


25

20

15

10
2 3 4 5 6 7 8 9 10
Average total load per cell in the interference cells

Figure 9: Mapping from total load to mean value of C/I ().

8.5
Theoretical mapping
Simulated values

7.5
Standard deviation of C/I (dB)

6.5

5.5

5
2 3 4 5 6 7 8 9 10
Average total load per cell in the interference cells

Figure 10: Mapping from total load to standard deviation of C/I ().

With the C/I distribution implemented in the simulator for the theoret-
ical model, several simulations were performed. The fraction of satisfied
users in simulations with varying streaming load without speech users where

K
compared to simulations with identical loads in PETRA. After some small
adjustments of the security factor c to 0 9, simulations with PETRA and
the theoretical model gave similar results. However, for simulations with
a constant load of four speech users included the results did not match so
well. Presumably, that has something to do with the pre-emption model.
The theoretical model has a soft ideal allocation of channels. This means
that all channels are always used and equally shared by all users. Therefore,
Theoretical Model 19

one speech user entering the system will always steal resources from all
other existing streaming users and when the speech user leaves the system
the resources are always immediately returned to streaming users.
PETRA has a more discrete and realistic channel allocation. The speech
users allocate timeslots randomly. This means that sometimes a free timeslot
is used and sometimes a streaming user is pre-empted on one of its timeslots.
If a timeslot is stolen from a streaming user by pre-emption, it will in most
cases be lost for the rest of the streaming user session, even if the speech user
leaves the system earlier. This may also result in that streaming users do not
share the channel resources equally. This difference causes the theoretical
model to be a little bit too optimistic in scenarios with mixed traffic. To

K
compensate for this, the assumed number of timeslots that one speech
user occupies is changed from 1 to 1 9 in the theoretical model. This makes
the results more similar. In Figure 11 the fraction of satisfied streaming
users is plotted as a function of a load of varying streaming load without
speech users and in Figure 12 the corresponding for a load of four speech
users and varying streaming load. The normalized load measure is defined
in section 3.6 below. We can see that all simulations look similar and the
maximum deviation is around 4%, which is considered sufficiently.

100
PETRA
Theoretical model

90

80
Satisfied streaming users (%)

70

60

50

40

30
0.6 0.8 1 1.2 1.4 1.6 1.8 2
Normalized total offered load

Figure 11: Satisfied users for the theoretical model and PETRA for a load
of zero speech users and varying streaming load.
20 Theoretical Model

100
PETRA
Theoretical model
90

80

Satisfied streaming users (%)


70

60

50

40

30

20

10
0.6 0.8 1 1.2 1.4 1.6 1.8 2
Normalized total offered load

Figure 12: Satisfied users for the theoretical model and PETRA for a load
of four speech users and varying streaming load.

3.6 Normalized load


To be able to compare different load scenarios, a relative measure of a sat-
isfied system load is needed. Normalized load is defined as 1 when the
fraction of satisfied streaming users is 90% without admission control (e.g.
a fully loaded system). In order to find the normalization factors, differ-
ent load combinations where the fraction of satisfied users was 90% were
measured. The least square solution of these simulations was

Lnorm M 6uK 5 7uK 0 K


sp st
(17)

Note the difference from the total load measure, Ltot (14), which only con-
siders the total load in the system and not whether the users are satisfied or
not.

3.7 Admission control example


To give an example of why admission control may be necessary, the individ-
ual streaming user throughput for different number of users is modeled, see
Figure 13. The theoretical model is used, but for simplicity, the number of
total timeslots were chosen to eight instead of sixteen. That means that two
streaming users will have an average access to four timeslots each, resulting
in a high possible total throughput. If the number of users increase the num-
ber of available timeslots per user will decrease, and so will the available
throughput. When the number of users goes from three to four per channel
Theoretical Model 21

three satisfied users become four unsatisfied users. Obviously, it would have
been better with three admitted satisfied users and one blocked user instead
of four unsatisfied users.
2 users
100

90

80

70

Throughput per user (kbps) 3 users


60

50

40
Acceptable quality 4 users

30
Unacceptable quality

20

10
0 5 10 15 20 25 30
Time (s)

Figure 13: Throughput per user when the number of streaming users in-
crease.
22 Admission Control Algorithms

4 Admission Control Algorithms


In the upcoming 3G systems there are suggestions of new services, like for
example streaming, that have to be guaranteed a certain quality of service.
The main idea with admission control is to secure the quality for admitted
users. In other words, the goal is to deny access for users that would overload
the system and cause already admitted users to become unsatisfied. If these
users are successfully blocked, a high fraction of satisfied streaming users
will be obtained, possibly with a low blocking rate as well. The goal in this
thesis is to achieve at least 90% satisfied admitted streaming users with as
low blocking rate as possible. This section describes different algorithms
that have been developed and evaluated through simulations.

4.1 Local admission control


The first, and simplest, considered algorithm is based on local traffic admis-
sion control, and very easy to implement in a system. It can be used as a
reference for more complex algorithms. Local traffic load admission was

K
described in section 2.7.1 but in this section a small modification is made
to adjust for the influence of speech users (the factor 1 9 from section 3.5).
The number of free channels available for streaming c st is defined as

cst M ctot 1K 9 L u Q sp (18)

where ctot is the total number of channels and u sp is the number of speech
users. A new user is admitted if
ust
cst
N Fthreshold Q (19)

where Fthreshold is the value of how many streaming users that can be allowed

K
per timeslot. For example, if one streaming user is assumed to need at least
two timeslots, Fthreshold should be 0 5 or less. If a certain fraction of satisfied
users is desired, several simulations with different values of Fthreshold are
performed until an adequate threshold is found.

4.2 Weighted semi-global admission control


An admission control scheme, here defined as weighted semi-global ad-
mission control, considers only the experienced interference due to load in
neighboring cells, and in this case only the first tier, to decide whether a user
can be admitted or not. However, it does not consider if the admitted user
will cause unsatisfied users in the neighboring co-channel cells by increased
interference.
Admission Control Algorithms 23

The average interfering load per cell from the first tier is denoted Ltot ,

Ltot M ust f t 0 K 553 L u Q sp f t (20)

K
where ust f t and usp f t are the average number of streaming and speech users
per cell in the first tier. The constant 0 553 is the C/I coefficient from one
speech user compared to one streaming user, see section 3.5. To be able to
control the effect from surrounding cells compared to the admitting cell, two
constants are added. A user is admitted if the total mean load per cell in the
first tier of co-channel cells and the ratio of streaming users compared to the
available channels is below Fthreshold ,

L
c1 Ltot c2
ust
cst
N Fthreshold Q (21)

where c1 and c2 are used to control the effects from each term. For example,

M M
local admission is a special case of weighted semi-global admission control
with c1 0 and c2 1.
The number of free channels in the co-channel cells is not included since
it is not assumed to affect the users in the admitting cell. Fthreshold is, as
previously described, the maximum number of users per timeslot.

4.3 C/I based admission control


As an attempt to base AC on a more realistic measure, a C/I based admis-
sion control was developed based on the findings gained from the theoretical
model. The radio quality could either be based on measured or predicted val-
ues of the current C/I. Below, a description of a prediction based method is
described.
From the total load in the first tier of interfering cells, the mean and
standard deviation predictions for the throughput in the admitting cell are
calculated. Based on this, the distribution of the total bit rate available for
streaming is calculated. This measure makes it possible to predict how many
streaming users that can be added while preserving a certain desired fraction
of satisfied users. The algorithm step by step:

Mean and standard deviation of C/I for each channel is calculated ac-
cording to (15) and (16)

ts M 0 K 021 L L 0 K 83 L L 10 K 2 L L 50 K 1 Q
3 2
(22)
M 0 K 028 L L 0 K 71 L L 9 K 38 Q
tot tot tot
ts 2
tot tot (23)

where Ltot is total mean load per cell in the first tier of interfering
co-channel cells.
24 Admission Control Algorithms

Mean and standard deviation for the system throughput margin, i.e.,
the total channel including the constant c minus the demanded chan-
nel, is calculated
margin M K L L L bitrate L u Q
0 9 2 ts cst (24)
M 0 K 9 L 2 L L c Q
st st
margin ts st (25)
where bitratest is the required bit rate for the streaming users and c st


is the number of channels available for the streaming users.
The variable with mean and standard deviation, margin and margin , is
assumed to be the normally distributed stochastic variable . Figure 14
shows that denotes the probability that is negative, i.e., that a user
is unsatisfied:

U N 0
P V M P margin
margin
N 0 margin
margin
M (26)

0.4

0.35

0.3

0.25

0.2

0.15

0.1

0.05

0
3 2 1 0 1 2 3 4 5 6 7


margin

Figure 14: Probability density function of .


For a certain , is collected from a table of quantiles of the nor-
mal distribution. Then a ratio of margin and margin can be calculated
according to
margin
margin

(27) K
This implies that a user can be safely admitted if

ust N 1 K 8 L L bitrate
c L
ts st
st
margin
Q (28)
Admission Control Algorithms 25

where the threshold parameter is chosen according to the desired


M K
fraction of satisfied users. For example, 1 28 should give a prob-
ability of at least 90% of satisfied users.
26 Comparison of AC-Algorithms

5 Comparison of AC-Algorithms
The proposed algorithms are evaluated in this section, and some compar-
isons are made. Further improvements are also suggested and evaluated in a
corresponding manner.
To be able to compare the algorithms, their thresholds were adjusted to
achieve 90% satisfied admitted users for a reasonably high load.

K
Fthreshold for both local and weighted semi-global admission was cho-
sen to 0 61 [users/TS].
Different combinations of the constants, which control the effects from

M K M K
the admitting and the co-channel cells for semi-global admission, c 1
and c2 were evaluated. For values around c 1 0 02 and c2 0 79, the
best performance was achieved.
K
for 90% reads 1 28 from a table with quantiles of the normal dis-
K
K
tribution, but needed to be adjusted to 1 4 to achieve 90% in the simu-
lation. 1 4 corresponds to approximately 92% in the table.
In Figure 15 we can see the fraction of satisfied users for the three algorithms
with an average of four speech users per cell and varying streaming load.
Without admission control, the number of satisfied users decreases rapidly,
which is not acceptable. It should be noted that the ideal case would yield
50% blocking for a normalized load of 2 (100% overload). But since no
speech users are blocked, the total blocking rates are actually lower than
50% and better than the ideal case without AC. This is possible since
resources can be more efficiently utilized with AC.
C/I based AC has the lowest blocking rate for all loads, but the fraction
of satisfied users is slightly lower for lower loads. Local admission has the
highest blocking rate, but also the highest fraction of satisfied users, and the
weighted semi-global admission results end up between the other two. A
general confidence analysis of simulation results is performed in Appendix
B.
Results from simulations in PETRA with the same algorithms and iden-
tical traffic loads are shown in figure 16. The blocking curves almost per-
fectly match the curves from the theoretical model. The order of fractions
of satisfied users curves is the same but they are generally a bit higher than
the theoretical model. The theoretical model is fitted to PETRA simulations
over the interval from about 30% up to 100% of satisfied users. There-
fore, the values do not match perfectly at high fractions of satisfied users
with admission control. However, the algorithms seem to work equally in
both simulators and the fraction of satisfied users is easily adjusted with the
threshold values, if necessary. Since the C/I based algorithm has the lowest
blocking rate and acceptable fraction of satisfied users it is considered to be
the best algorithm of these three.
Comparison of AC-Algorithms 27

100

90

80 No admission
Local
Weighted semiglobal
70 C/I based

Satisfied and blocked users (%)


60

50

40

30

20

10

0
0.6 0.8 1 1.2 1.4 1.6 1.8 2
Normalized total offered load

Figure 15: Performance for three different AC algorithms in theoretical


model.
100

90
No admission
Local
80 Weighted semiglobal
C/I based

70
Satisfied and blocked users (%)

60

50

40

30

20

10

0
0.6 0.8 1 1.2 1.4 1.6 1.8 2
Normalized total offered load

Figure 16: Performance for three different AC algorithms in PETRA.

5.1 Full-global admission control


A full-global admission control, in contrast to the previously defined weighted
semi-global admission control, also considers if the increased interference
from the new user causes unsatisfied users in neighboring co-channel cells.
If a user is admitted to the admitting cell, users in the neighboring cells might
become unsatisfied because of the increased interference caused by the ad-
mitted user. Therefore, it might be a good idea to use full-global admission
control, where all the neighboring cells are checked [5]. If the increased
interference in the neighboring cells causes degraded quality and unsatis-
28 Comparison of AC-Algorithms

fied users, the user is not admitted even if there is enough bandwidth in the
admitting cell. In Figure 17, C/I based admission control is compared to
full-global C/I based admission control. The thresholds have been adjusted
to achieve equal fractions of satisfied users. It can be seen from these results
that the blocking rate is higher for full-global C/I based admission control.
Most likely this depends on pre-emption by speech users, because if one cell
is overloaded due to pre-empting speech users it will result in that neighbor-
ing cells can not admit any users until some users have left the overloaded
cell. The solution to this problem is to check if a cell is overloaded and if
that is the case, ignore that cell because these users are probably unsatisfied
already, regardless of a new user is admitted to a neighboring cell or not.
This makes the algorithm more complex, the limit where to ignore an
overloaded cell is not clear. If speech users steal a few timeslots, the cell
does not necessary have to be overloaded, but it might be if another speech
user enters the system. A few different ways to ignore these overloaded
cells have been evaluated but it seemed difficult to improve the algorithm
with some kind of full-global admission control.

100

90

C/I based
80
Full global C/I based

70
Satisfied and blocked users (%)

60

50

40

30

20

10

0
0.6 0.8 1 1.2 1.4 1.6 1.8 2
Normalized total offered load

Figure 17: Comparison of C/I based and full-global C/I based admission
control.

5.2 Load dependent threshold


A negative aspect of admission control is that users are blocked before the
system is fully loaded, which will result in lost revenues for the operator. To
lower the blocking rates at low loads, the admission threshold can be set as a
function of the average load per cell in the first tier, including the admitting
cell. A suggested mapping from normalized average load to threshold value
Comparison of AC-Algorithms 29

is shown in Figure 18. When the system load is low, the admission control
is almost turned off and vice versa. The normalized average load is filtered
with a moving average filter over 10s, to get a more stable behaviour.
In Figure 19, C/I based admission is compared to C/I based admission
with a load dependent threshold. We can see that the blocking rate at low
loads is lower with load dependent threshold , but the fraction of satisfied
users is also a little bit lower, which is a natural effect.
3
Load dependent threshold
Fixed threshold

1
Threshold value

4
0.9 0.95 1 1.05 1.1 1.15 1.2 1.25 1.3 1.35 1.4
Normalized total load in the system

M K
Figure 18: Mapping from normalized load to threshold value. The fixed
threshold corresponds to 1 4.

100

90

80 C/I based
C/I based with load dependent threshold

70
Satisfied and Blocking (%)

60

50

40

30

20

10

0
0.6 0.8 1 1.2 1.4 1.6 1.8 2
Normalized total offered load

Figure 19: Comparison of C/I based admission control with and without
load dependent threshold.
30 Comparison of AC-Algorithms

5.3 Dropping
Some users may become unsatisfied if one or several speech users enter a
cell, since the channel resources available for streaming will decrease dras-
tically, as indicated in Figure 20. Streaming users that are pre-empted on
one or more timeslots will probably be unsatisfied, while other streaming
users will remain unaffected. If nothing is done about this sudden decrease
of channel resources, streaming buffers will run empty and several users will
be unsatisfied.
It seems reasonable to try to minimize the consequences from the sudden
decrease of channel resources. One suggestion is to drop some streaming
users, i.e., actively terminate their session, which will result in better per-
formance for the remaining streaming users. With this method, the fraction
of satisfied users will increase, at the cost of some dropped users. However,
the users that get dropped were probably unsatisfied to begin with, so the
difference for them might not be that big.

320

300

280
Throughput (kbps)

260 Satisfied Unsatisfied

240

220

200
0 20 40 60 80 100 120 140 160 180 200
time (s)

Figure 20: Channel decreasing at time 100 when one speech user enter the
system.

5.3.1 Evaluation of admission control combined with dropping


The performance of admission control combined with dropping was evalu-
ated through simulations.
C/I based admission control was still used but now combined with drop-
ping. The dropping algorithm works almost the same as the admission con-
trol algorithm. The threshold for dropping is, however, changed so that drop-
ping occurs when the throughput of the system is lower than the bandwidth
Comparison of AC-Algorithms 31

demanded by the users. If this happens, one user is dropped randomly. How-
ever, improvements can be made if the user with the lowest C/I, combined
with the shortest time in the system, and worst affected by pre-emption,
could be found and dropped instead. This is possible but complicated and
has not been studied here.
The threshold , see (28), for the ordinary C/I based admission control
M K
without dropping was still set to adm 1 4. The thresholds for the com-
M K
bined algorithm was set to adm 1 1 for admission and dr 1 1 for M K
dropping. The threshold for dropping was chosen to get a few percentage of
dropped users. Dropping occurs when the channel bandwidth is below the
demanded bandwidth with a probability of just below 90%. The admission
control threshold was chosen to achieve equal fractions of satisfied users for
admission control with and without dropping. In Figure 21, C/I based ad-
mission control is compared to C/I based admission control combined with
C/I based dropping. We can see that the fraction of satisfied users is al-
most the same but the blocking rate is much lower with dropping. Even the
sum of dropping and blocking is a little bit lower than the blocking for only
admission control.
However, the big advantage with dropping appears when higher fractions
of satisfied users are required. In Figure 22 the thresholds have been set a
little bit more aggressive to achieve around 95% of satisfied users, which
yields higher blocking rates. We can clearly see that by combining two
different ways of denying service to streaming users, blocking and dropping,
the total amount of denied users decrease.
100

90

Satisfied, AC
80 Satisfied, AC + Dropping
Blocking, AC
Blocking + Dropping, AC + Dropping
70 Dropping, AC + Dropping
Blocking, AC + Dropping

60

50

40

30

20

10

0
0.6 0.8 1 1.2 1.4 1.6 1.8 2
Normalized total offered load

Figure 21: Comparison between admission control with and without drop-
ping.
32 Comparison of AC-Algorithms

100

90
Satisfied, AC
Satisfied, AC + Dropping
80 Blocking, AC
Blocking + Dropping, AC + Dropping
Dropping, AC + Dropping
70 Blocking, AC + Dropping

60

50

40

30

20

10

0
0.6 0.8 1 1.2 1.4 1.6 1.8 2
Normalized total offered load

Figure 22: Comparison between admission control with and without drop-
ping with higher thresholds.
Final Algorithm Proposal 33

6 Final Algorithm Proposal


Based on the results from section 5, the final algorithm is built on C/I based
admission control (section 4.3), but the same ideas are applicable to other
algorithms as well. To achieve good performance at low loads a load de-
pendent threshold (section 5.2) is applied. Dropping (section 5.3) is also
included in the final algorithm because of two main reasons; it can easily
bring up the number of satisfied users, and it makes the number of satisfied
users more stable for different load combinations.

6.1 Performance of the proposed algorithm


A comparison of a system without admission control, with local admission,
and with the final algorithm, is shown in Figure 23. The thresholds that are
used are shown in Table 2. The load dependent threshold is set as in
Figure 18.
The load is a combination of a fixed number of speech users and a vary-
ing number of streaming users per cell. We can see that the final algorithm
has much lower blocking and dropping rate for low loads but also a little bit
lower fraction of satisfied users than local admission control. The dropping
rate is typically 2-3% over the whole interval. The final algorithm flattens
out at 90% satisfied users, but the local admission algorithm will continue to
fall below 90%.
The corresponding results from PETRA are shown in Figure 24, for the
same load combinations. For some reason, PETRA gives dropping rates of
about 4-5% instead of about 2-3% for the theoretical model. This gives a
little bit higher total blocking plus dropping rate but this can be adjusted
with different thresholds in PETRA. The fraction of satisfied users is also
higher in PETRA, which can be adjusted as well.

Fthresholdlocal
0 61
adm
09
dr
15

Table 2: Thresholds used in the final comparison

6.1.1 Revenue based measurement


Most probably, operators would like to maximize the number of satisfied
users in his system, because these users will bring in the long-term revenue.
This makes sense, because then the system resources are assigned to satisfied
users and not to unsatisfied users, or being left as idling reserve resources.
34 Final Algorithm Proposal

a) b)
100 100
Satisfied and Blocking + Dropping (%)

Satisfied and Blocking + Dropping (%)


80 80

60 60
No admission
No admission Local
40 Local 40 Final
Final

20 20

0 0
0.5 1 1.5 2 0.5 1 1.5 2
Normalized total offered load Normalized total offered load

Figure 23: Comparison of the final algorithm, local admission and no ad-
mission for a load of a) two speech users, b) four speech users
and varying streaming load.

a) b)
100 100
Satisfied and Blocking + Dropping (%)

Satisfied and Blocking + Dropping (%)

80 80

60 60
No admission
No admission Local
40 Local 40 Final
Final

20 20

0 0
0.5 1 1.5 2 0.5 1 1.5 2
Normalized total offered load Normalized total offered load

Figure 24: Comparison in PETRA of the final algorithm, local admission


and no admission for a load of a) two speech users, b) four
speech users and varying streaming load.

The total number of users in the system is the total number of users
excluding blocked or dropped users. To get the normalized total satisfied
load, the number of users in the system should be multiplied by the fraction
of satisfied users, and normalized as in (17). Note that speech users are also
included in this measure, which all are assumed to be satisfied.
Final Algorithm Proposal 35

Figure 25 shows the normalized total satisfied load, for a fixed number of
speech users and a varying number of streaming users per cell. We can see
that the scenario without admission control has a peak shortly after the sys-
tem reaches full load, but at this peak the fraction of satisfied users is much
less than 90%. This phenomenon that the normalized total satisfied load
increases despite the decreasing fraction of satisfied users, arises because
the total number of users in the system increases and system resources are
being used more efficiently, see Appendix C for further explanation. How-
ever, after this point the system becomes jammed and users quickly become
unsatisfied.
Figure 25 also shows that admission control has its drawback of block-
ing users before the system is fully loaded, resulting in worse performance
for lower loads. That is why the load dependent threshold works quite well
when it almost turns off the admission control at low loads. The final al-
gorithm is, however, not far behind the system without AC for low loads.
At high loads, admission control keeps the system stable and preserves the
quality. We can clearly see in the figures that the normalized total satisfied
load slowly increases at high loads, and does not drop like the system with-
out admission control. Note that the system without AC does not satisfy
the criteria of at least 90% of satisfied users above normalized load 1, see
Figure 23. Since speech users are included in this measure, the fraction of
satisfied users seems to be a little bit higher than 90% at normalized load 1
in Figure 25.
In Figure 26, the corresponding results from PETRA are presented. We
can see that the levels might be a little bit different but the final algorithm
seems to be overall best algorithm.
How is it possible to have at least 90% of satisfied users and a higher
normalized satisfied load than 1? Without admission control the mean load
must be quite low, otherwise a lot of users will become unsatisfied when the
total throughput has a dip. However, with admission control the number of
users are decreased in those cases and increased when the throughput has a
top. This is why the mean load can be higher than 1, while still having 90%
satisfied users in the system.
36 Final Algorithm Proposal

a) b)
1.2 1.2
No admission No admission
Local Local
Final Final
Normalized total satisfied load

Normalized total satisfied load


1.1 1.1

1 1

0.9 0.9

0.8 0.8

0.7 0.7

0.6 0.6
0.5 1 1.5 2 0.5 1 1.5 2
Normalized total offered load Normalized total offered load

Figure 25: Normalized total satisfied load for a load of a) two speech users,
b) four speech users and varying streaming load.

a) b)
1.2 1.2
No admission No admission
Local Local
Final Final
Normalized total satisfied load

Normalized total satisfied load

1.1 1.1

1 1

0.9 0.9

0.8 0.8

0.7 0.7

0.6 0.6
0.5 1 1.5 2 0.5 1 1.5 2
Normalized total offered load Normalized total offered load

Figure 26: Normalized total satisfied load in PETRA for a load of a) two
speech users, b) four speech users and varying streaming load.
Conclusions 37

7 Conclusions
It has been shown that employing admission control makes it possible to
offer a stable system with approximately 90% of satisfied users, even for
excessive loads.
Several algorithms have been developed and evaluated. These algo-
rithms can be combined to match different scenarios or systems well. For
example, if a high fraction of satisfied users is desired (more than 90%),
dropping is preferably used to avoid very high blocking rates.
Full global algorithms, where neighboring cells are also considered, tend
to give high blocking rates or be rather complex.
The final proposed AC algorithm is based on predicted radio quality
(C/I) and combined with C/I based dropping and a load dependent admis-
sion threshold. The load dependent threshold results in lower blocking rates
at low loads. The algorithm performs nearly as good as the system without
admission control at low loads and at high loads it is outstanding compared
to a system without admission control.
Note that the measure satisfied users can be defined in several different
ways, which may give different results.

7.1 Further work


The fraction of satisfied users seems to depend a little on the load
combination, session time and the bit rate and could need further in-
vestigation
The fraction of satisfied speech users should perhaps be investigated,
since it is an important measure.
More services could be added in order to further investigate effects of
different service mixes, e.g., WWW-traffic.
Admission control of speech users could be a good tool to achieve
good performance. This might lower the blocking rates for the stream-
ing users since the preemption is more under control, but has been
found not to be efficient in previous studies.
C/I for the users could be predicted or measured in a better way, for
example by including system measurement data.
Setup-times for speech might be used to give a forewarning and under
that time no streaming users may be admitted.
More cells than the first tier of interference cells could be considered
with different weights.
38 Appendix

Appendix A: Abbreviations
3G 3rd Generation
3GPP 3rd Generation Partnership Project
AC Admission Control
ARQ Automatic Repeat Request
BSC Base Station Controller
BTS Base Transceiver Station
C/I Carrier to Interference ratio
CS Circuit Switched
EDGE Enhanced Data rates for GSM Evolution
EGPRS Enhanced GPRS
FDMA Frequency Division Multiple Access
FEC Forward Error Correction
GGSN Gateway GPRS Support Node
GMSC Gateway Mobile Services Switching Center
GMSK Gaussian Minimum Shift Keying
GPRS General Packet Radio Service
GSM Global System for Mobile Communication
IP Internet Protocol
IR Incremental Redundancy
LA Link Adaptation
LLC Logical Link Control
LQC Link Quality Control
MCS Modulation and Coding Scheme
MS Mobile Station
MSC Mobile services Switching Center
PS Packet Switched
PSK Phase Shift Keying
PSTN Public Switched Telephone Network
QoS Quality of Service
RLC Radio Link Control
SGSN Serving GPRS Support Node
TDMA Time Division Multiple Access
TS Time Slot
UMTS Universal Mobile Telecommunications System
Appendix 39

Appendix B: Confidence Analysis


To verify the accuracy of the simulations, the same load was simulated ten
times with different seeds for different simulation times. In Figure 27 and 28

L K
we can see that the deviation decreases when the simulation time increases.
For 4 104 s the standard deviation for satisfied users is 0 08 and 0.27 for

L
blocked users. In other words, to achieve enough accuracy the simulation
time should be chosen to an average of ten simulations with 4 10 4 s each.
This simulation time or longer has been used in all simulations.

94.1
Mean
Different simulations

94

93.9
Satisfied users (%)

93.8

93.7

93.6

93.5
0.5 1 1.5 2 2.5 3 3.5 4 4.5
Simulation time (s) x 10
4

Figure 27: Deviation for satisfied users for different simulation times.

41.8
Mean
Different simulations
41.6

41.4

41.2
Blocked users (%)

41

40.8

40.6

40.4

40.2

40
0.5 1 1.5 2 2.5 3 3.5 4 4.5
Simulation time (s) x 10
4

Figure 28: Deviation for blocked users for different simulation times.
40 Appendix

Appendix C: Maximize the Number of Satis-


fied Users
To be able to offer a certain amount of satisfied users, the system load can
not be too high. Therefore, all the system resources are not in use all the
time. If the number of satisfied users in the system (paying customers) shall
be maximized, the system load must be high at the cost of more unsatisfied
users.
This fact can be visualized through two extreme examples. Assume that
the channel is normally distributed, see Figure 29. If the system is loaded
with only one user, this user will most probably be satisfied, but system will
not be used efficiently. The other extreme case is that the system is loaded
with too many users. In this case all users will probably be unsatisfied, which
results in zero satisfied users in the system. Based on these two extreme
cases it can be seen that there must exist a point in between resulting in an
optimum number of satisfied users. The number of satisfied users in the

M L Q
system is assumed to be
usatis f ied u (29)
where u is the number of users in the system and is the probability that the
users are satisfied. In Figure 30 usatis f ied is plotted as a function of u and
K
where we can see that for an average load of 6 25 users per cell, u satis f ied is
maximized but for that point is only 70%. at the optimum value depends
on the throughput characteristics of the channel, but the value is in the most
cases below 90%. Therefore, some trade-offs need to be performed. If the
system is near the optimal value, the system will be used more efficiently
with a bigger average load, but with a lower fraction of satisfied users. The
unsatisfied users will probably get fed up after a while and perhaps choose
another operator. If system load is low, the fraction of satisfied users will
be high, but system resources will be used inefficiently. To get a system
that will satisfy the customers and the operator in the long run, it must be
designed with consideration to a fraction about 90% of satisfied users and
therefore is admission control most probably has to be used.
Appendix 41

3
x 10
4

3.5

2.5

1.5

0.5

0
0 100 200 300 400 500 600
Total throughput of the system (kbps)

Figure 29: Probability density function of the total throughput.

4.5

3.5
Number of satisfied users in the system

2.5

1.5

0.5

0
0 2 4 6 8 10 12 14 16 18 20
Number of users in the system

Figure 30: Number of satisfied users as a function of total number of users


in the system.
42 References

References
[1] Jerry D. Gibson. The Mobile Communications Handbook. CRC Press,
1996.
[2] Robert G. Winch. Telecommunication Transmission Systems. Second
edition. McGraw-Hill, 1998.
[3] A. Furuskr, J. Nslund, H. Olofsson. EDGE Enhanced Data Rates for
GSM and TDMA/136 Evolution. Ericsson Review Vol.76(1999):1, pp.
28-37.
[4] Magnus Karlsson. Admission Control for Packet Data in GSM. Mas-
ters thesis, Royal Institute of Technology, 1998.
[5] Bogdan Timus, Jonas Pettersson. Uplink Admission Control for Con-
versational Services Using Information From Many Cells. Proceed-
ings in IEEE Vehicular Technology Conference, Rhodes, Greece, May
2001.
[6] Tomas Jnsson. A Study of Radio Resource Cost for Streaming Ser-
vices. Masters thesis, Lule University of Technology, 2000.
[7] 3rd Generation Partnership Project: Technical Specification Group
Services and System Aspects; QoS Concept and Architecture. 3G TS
23.107.
[8] H. Granbohm, J. Wiklund. GPRS-General Packet Radio Service. Eric-
sson Review Vol.76(1999):2, pp. 82-88.
[9] APIS Training & Seminars. GSM System Overview. 1999.
[10] P. Beming, M. Frodigh. Admission Control in Frequency Hopping
GSM Systems. In Proceedings in IEEE Vehicular Technology Confer-
ence, VTC-96.
[11] Ericsson. GPRS System Survey. EN/LZT 123 5374. Student Text. 1999.
[12] C. Lindheimer, S. Mazur, J. Moln, M. Waleij. Third-generation
TDMA. Ericsson Review Vol.76(2000):2, pp. 68-79.
[13] Michael Andersin. Power Control and Admission Control in Cellular
Radio Systems. Ph.D. Thesis, Royal Institute of Technology, 1996.
[14] Gunnar Blom. Sannolikhetsteori och statistikteori med tillmpningar
(Bok C). Tredje upplagan, Studentlitteratur, Lund 1980.
[15] Christer Johansson, Henrik Nyberg, Birgitta Olin. A Streaming Video
Traffic Model for the Mobile Access Network. Proceedings in IEEE Ve-
hicular Technology Conference, Atlantic city, New Jersey, USA, Oc-
tober 2001.

You might also like