i

Congestion Control Approach based on Effective Random Early

Detection and Fuzzy Logic

By

Maimuna Abdulfattah Ahmad Khattari

Supervisor

Dr. Gassan Samara

This Thesis was Submitted in Partial Fulfillment of the

Requirements for the Master Degree in Computer Science

Faculty of Graduate Studies

Zarqa University

Zarqa, Jordan

First Semester

December, 2015

ii

خبمعت الضسقبء

إلرار تفىض

, أفض خبمعت الضسقابء تخضاذ و ان ماه سيابلخ / أيشحخا لتم خبابث, أ ممو عبذ الفخبذ احمذ خطبسأوب

المؤي بث, أ الئبث, أ األشخبص عىذ يتبم ح ب الخعتمبث الىبفزة ف الدبمعت.

الخقع:

الخبسن:

Zarqa University

Authorization Statement

I Maimuna Abdulfattah Ahmad Khatari, authorize Zarqa University to supply copies

of my thesis to libraries or establishments or individuals on request, according to the

University regulations.

Signature:

Date:

iii

COMMITTEE DECISION

This is to certify that the thesis entitled (Congestion Control Approach based on

Effective Random Early Detection and Fuzzy Logic) was successfully defended and

approved on -----------------------------.

Examination Committee Members Signature

Dr. ……………………….. (Supervisor) -------------------------

Dr. ……………………….. (Member) -------------------------

Dr. ……………………….. (Member) -------------------------

Dr. ……………………….. (Member) -------------------------

iv

ACKNOWLEDGMENTS

First and foremost, I would like to thank Allah for this blessing to complete this

thesis successfully.

I would like to express my deepest gratitude to my advisor, Dr.Gassan Samara,

for his support and guidance throughout this thesis.

I would also like to thank my family for their continuing love and support,

especially my husband, thank you from the bottom of my heart, Loads of love and

thanks to you all.

v

TABLE OF CONTENTS

Contents Page

List of Tables.............................................................................................................. vii

List of Figures….........................................................................................................viii

List of Acronyms ………………………………………………………………...........x

List of Publications …………………………………………………………….……xii

Abstract in Arabic…...................................................................................................xiii

Abstract in English…..................................................................................................xiv

Chapter 1: Introduction ………………………………………………………………... 1

1.1 Introduction………................................................................................................. 1

1.2 Problem Statement…….......................................................................................... 5

1.3 Objectives................................................................................................................ 5

1.4 Thesis Contributions............................................................................................... 6

1.5 Scope and Limitations............................................................................................. 6

1.6 Thesis Organization..................................................................................................6

Chapter 2: Literature Review and Related Works ………………………………..…. 8

2.1 Introduction............................................................................................................. 8

2.2 Overview of Congestion Factors at the Router Buffer............................................ 9

2.3 Congestion Control at the Router Buffer….......................................................... 12

2.4 Queue-Based Active Queue Management............................................................ 14

2.4.1 Average- Based...................................................................................... 14

2.4.2 Queue- Based………............................................................................. 17

2.4.3 Average & Queue- Based....................................................................... 17

2.4.4 Packet loss- Based…………...…………............................................... 18

2.5 Load-Based Active Queue Management............................................................... 19

2.5.1 Arrival- Based…………………............................................................ 19

2.5.2 Load Factor- Based..................….......................................................... 20

2.6 Hybrid Active Queue Management....................................................................... 21

2.7 Fuzzy Based Active Queue Management............................................................. 22

2.8 ERED……………………………........................................................................ 25

2.9 AQM and Delay………………………………………………………………… 26

vi

2.10 Conclusion……................................................................................................... 32

Chapter 3: Methodology ...........................................................................................33

3.1 Introduction…....................................................................................................... 33

3.2 Methodology……................................................................................................. 33

3.2.1 Initialization........................................................................................... 36

3.2.2 Fuzzification…....................................................................................... 36

3.2.3 Rules Evaluation.................................................................................... 36

3.2.4 Aggregation............................................................................................ 37

3.2.5 Defuzzification....................................................................................... 37

3.3 Conclusion …........................................................................................................ 38

Chapter 4: System Design .........................................................................................39

4.1 Hybrid-based ERED……...................................................................................... 39

4.1.1 Delay Calculation................................................................................... 39

4.1.2 Hybrid-ERED Mechanism..................................................................... 41

4.1.3 Algorithm………………………………............................................... 42

4.2 Fuzzy-based hybrid-ERED….….......................................................................... 48

4.2.1 Initialization...…………………............................................................ 48

4.2.2 Fuzzification…………………............................................................... 50

4.2.3 Rules Evaluation…………….................................................................51

4.2.4 Aggregation…………………................................................................ 53

4.2.5 Defuzzification...……………................................................................ 53

4.2.6 Calculate Final Dropping Probability................................................. 54

4.3 Conclusion ……………………........................................................................... 55

Chapter 5: Results…………………………………………………………………..56

5.1 Introduction…....................................................................................................... 56

5.2 Environment…...................................................................................................... 56

5.3 Parameter settings……….……............................................................................ 58

5.4 Results……........................................................................................................... 59

5.5 Summary………………………........................................................................... 70

Chapter 6: Conclusion and Future Works …………………………………….....71

6.1 Conclusion…......................................................................................................... 71

6.2 Future Work…...................................................................................................... 72

References ………………………………………………………………..…….……73

vii

LIST OF TABLES

Number Table Caption Page

Table 2-2

Evaluation of AQM Methods

24

Table 3-1 Parameter settings 35

Table 4-1 Examples of the Set of Fuzzy Logic Rules 52

Table 4-2 Complete Set of Fuzzy Logic Rules 53

Table 5-2 Empirical weight settings 58

Table 5-3 Empirical FIP settings 59

Table 5-4 Performance comparison with arrival rate 0.93 60

Table 5-5 Performance comparison with arrival rate 0.66 62

Table 5-6 Performance comparison with arrival rate 0.5 63

Table 5-7 Performance comparison with arrival rate 0.33 64

viii

LIST OF FIGURES

Number Figure Caption Page

Figure 1-1 Buffer Queue and RED Parameters 3

Figure 2-1 The Congestion at the Router Buffer 9

Figure 2-2 Bad vs. Good Network Performance Measures 11

Figure 2-3 Congestion Control using AQM 13

Figure 2-4 AQM Methods Classification 13

Figure 2-5

Illustration of Traffic Scenarios in Router Buffer with a

Constant Departure Rate 27

Figure 2-6

Illustration of Packets Arrival/Departure in the Buffer over

Time in Optimal Case 28

Figure 2-7

Illustration of Packets Arrival/Departure in the Buffer over

Time in Nearly Optimal Case 29

Figure 2-8

Illustration of Packets Built-up Queue over Time when Delay

Increase 30

Figure 2-9

Illustration of Packets Overload the Buffer over Time in

Unstable Traffic 31

Figure 3-1 Flowchart of the Hybrid ERED 34

Figure 3-2 Flowchart of Fuzzy Hybrid ERED 35

Figure 4-1 DEsti vs. (ArrEsti /DepEsti) 41

Figure 4-2 Initial Membership Functions for avg 49

Figure 4-3 Initial Membership Functions for q 49

Figure 4-4 Initial Membership Functions for Dp 49

Figure 4-5 Mapping Crisp into Linguistic Term 51

Figure 4-6 Aggregating Output 54

Figure 5-1 (a) Mean Queue Length Comparison 65

Figure 5-2 (b) Number of Packet Dropped Comparison 66

Figure 5-3 (c) Number of Packet Lost Comparison 67

Figure 5-4 (d) Number of Packet Missed Comparison 68

ix

Figure 5-5 (e) Throughput Comparison 69

Figure 5-6 (f) Delay Comparison 70

x

LIST OF ACRONYMS

QoS Quality of Service

TCP Transmission Control Protocol

UDP User Datagram Protocol

Q Queue Length

Avg Average Queue Length

AQM Active Queue Management

Dp Dropping Probability

RED Random Early Detection

ERED Effective Random Early Detection

IETF Internet Engineering Task Force

RFC Request for Comment

FIP Fuzzy Inference Process

W Weight

Qi Queue Size at The Current Time

Avgi-1 Avg Calculated in The Previous Time Slot

PL Packet Loss

pk The Probability the Router Buffer Being Full

Β Packet Departure Probability

T Throughput

P0 Probability of Router Being in The Idle Status

D Delay

Mql Mean Queue Length

DT Drop Tail

xi

FRED Flow Random Early Detection

DSRED Double Slope Random Early Deduction

ARED Adaptive RED

AIMD Additive Increase Multiplicative Decrease

GRED Gentle RED

SRED Stabilized Random Early Drop

MRED Modified RED

AVQ Adaptive Virtual Queue

SAVQ Stable Adaptive Virtual Queue

EAVQ Enhanced Adaptive Virtual Queue

LUBA Link Utilization Based Aqm

SVB Stabilized Virtual Buffer

RAQM Rate AQM

RaQ Robust Active Queue Management

PI Proportional Integral Controller

FL Fuzzy Logic

BL Binary Logic

FEM Fuzzy Explicit Marking

COG Center of Gravity

IDE Integrate Development Environment

xii

List of Publications

Maimuna Khatari, Gassan Samara “Congestion Control Approach based on

Effective Random Early Detection and Fuzzy Logic", MAGNT Research

Report,(ISSN.1444-8939) Vol.3 (8). PP: 180-193, 2015. [Indexed in Thomson ISI].

xiii

نهج التحكم باالختناق اعتمادا على التحمك المبكر العشىائ والمنطك الضباب

إعذاد

ممىنه عبذ الفتاح احمذ خطاري

إشراف

د. غسان سمارة

الملخص

ا ماه عتا ازي العاماا الحاذ مىاب، ال طشة الحضماالصدحبم ف المخضن المخ ضذ مه حأخش فقذان

داسة قبئمات االوخااابس الى ااطت اا مدمعاات معشفاات مااه حااا ان ااذات لخقىاابث داسة الماااسد المخختفاات.األ اام

الحاضم تحازتيشق الخح م ف االصدحبم لدبص الخخ ح ف االصدحبم ح طش عت ف مشحتت مب شة، حبذأ

. حضمتالفقذان الخأخش ف قج مب ش تذت الحذ مه

ال ثشة الخ حضدوب تبالداء الدذ ف ك ف الخاسصمبثفعبلت ال ف الع ائ المب ش احذ خاسصمت

ال حبخاز تعاه االصدحبم الاخح م فا الخقتاا ماه فقاذان الحاضم ماع رلاى ماه احاذ عتاب ا ان ازي الطشقات

المعبش الثاتج. حققت ف ححذذ قم ضم , اضب حاخ م تتصل الحاالعخببس عبما الخبخش ف

اقخشحج ف زي االيشح يشقت امخذاد لخاسصمت فعبلت ال ف الع ائ المب ش حأخز تعه االعخبابس

الىخابئح حباه م المعابش .عبما الخبخش ال حمت ثم حدمع تىب ته االيخذالل الغبمض لتخقتا مه م تت ححذذ ق

: الخابخش ,حازت الحاضم س اياقبيب ان الطشقت المقخشح حعطا وخابئح افضاا فا االداء فقاب لتمقابل الخبلات

الخابخش تى ابت حقتاا% تبالعخمابد عتا احخمبلات صال الحاضم حام 100-10خ بسة الحضم تى ابت قتتج فقذاوب .

كبواج وخبئدا افضاا % امب تبلى بت الداء الىمرج الز خىذ عتا االياخذالل الغابمض لتطشقات المقخشحا 00-00

مه حا الخبخش فقذان الحضم.

xiv

Congestion Control Approach based on Effective Random Early

Detection and Fuzzy Logic

By

Maimuna Abdulfattah Ahmad Khattari

Supervisor

Dr. Gassan Samara

ABSTRACT

Congestion in router buffer increases the delay and packet loss, subsequently,

controlling these factors and reducing them are objectives for various resource

management techniques. Active Queue Management (AQM) is a well known set of

methods for router buffer’s congestion control. AQM methods detect and control

congestion in early stage and start dropping packets early aiming at reducing packet

loss and delay.

Effective Random Early Detection (ERED) method, among many other AQM

methods, gives a good performance in detect and control congestion and preserve

packet loss. However, the ERED’s drawback is in neglecting the delay. Moreover,

ERED faces a real parameterization problem.

This Thesis proposed an extended ERED method that considers the delay in its

process and combines the extended ERED method with a Fuzzy Inference Process

that eases the problem of parameter initialization and parameter dependency. The

results shows that the parametric-based form of the proposed work gives a better

performance results, according to the performance measures, delay, dropping and

xv

packet loss. The loss has been decreased by 10-100% depends on the arrival

probability. Delay has been decreased by 30-60%.The performance of the fuzzy-based

form of the proposed method is better than the parametric-based form and ERED in

terms of delay and packet loss.

1

Chapter 1

Introduction

1.1 Introduction

A lot of data is being transmitted and exchanged over the Internet for various

purposes as the internet technology is rapidly evolved day after day. The huge data

are transmitted over various network topologies are intermediated by links and

routers. Congestion referred to increase the data rate in a way that the network

resources cannot handle it. A host sending data over Transmission Control Protocol

(TCP) avoid congestion by a windowing mechanism which adjust the sending rates

based on the status of the network estimated by the time and amount of the received

acknowledgement. Other protocols, such as User Datagram Protocol (UDP) and

protocols used in wireless and mobile ad-hock networks do not adhere to such

mechanism. Transmitting using such protocols and even with the TCP protocol may

lead to congestion at the routers (Lim, et al., 2011).

A router receives data packets from source(s) and directs them into their

destination. The data being transferred are accommodated temporarily at the router

buffer before it is processed and sent out to their destinations. The number of packets

accommodated in the router buffer is defined as the queue length q, and the average

number of packets over a time window is called the average queue length avg. Both q

and avg are affected by the packet arrival rate, packet departure and the rates of

packets processing at the router. The router buffer has a fixed capacity that allows

limited numbers of packets to be accommodated [(Woodward, 1993), (Chitra, and

Padamavathi, 2010), (Seifaddini, 2013)]. A low packet departure rates compared to

2

high packet arrival cause the packets to stack in the queue and subsequently increase q

and avg accordingly. Increasing the q and avg cause what so called congestion

(Abbasov and Korukoglu, 2009).

Congestion in router buffer comes with the increment in packets queued. If the

number of stacked packets increased, these packets will be subject to notable delay.

When the number of packets reached the maximum capacity of the router buffer, the

buffer will be overflowed and all the arriving packets are being loss. Delay and Packet

loss are major problems of network and their control and reduction are objectives for

various management techniques. Consequently, a congestion control management

mechanism is operated on the router to response to the congestion when occur

(Naixue, et al., 2010).

Active Queue Management (AQM) is a well known set of methods for router

buffer’s congestion control. AQM methods detect and control congestion in early

stage and start dropping packets early aiming at reducing packet loss and delay. These

methods depend on continuously calculating dropping probability Dp for the

upcoming packets in order to prevent congestion. Examples of the existing AQM

methods are Random Early Detection (RED), Adaptive RED (ARED), Gentle RED

(GRED), Adaptive GRED (AGRED), Dynamic GRED (DGRED), BLUE, dynamic

threshold based BLUE (DT BLUE), GREEN, ERED, Subsidized RED (SubRED) and

others (Alfa, 2010).

RED (Floyd and Jacobson 1993), and many other AQM methods, use avg as

indicator of the queue status. The avg is evaluated by multiple thresholds.

MinThreshold is a position in the buffer, if not exceeded by avg, the buffer can be

considered to be of fair amount of packets in the queue. While, maxThreshold

3

parameter is another position, if reached or exceeded by avg, may be an indication of

congestion. Accordingly, Dp increases when the queued packets reach maxThreshold.

Figure 1.1 illustrates the buffer router and the positions of the parameters used by

RED.

Figure 1.1: Buffer Queue and RED Parameters

RED avoids congestion by dropping packets at early stage with a probability

value Dp. The dropping probability is a function of avg. Subsequently, when the avg

is less than minThreshold, RED drops no packets. All packets are dropped if avg

exceeds maxThreshold. The actual congestion control is implemented when avg is

between minThreshold and maxThreshold, the advantages of RED are as following

(Kwon and Fahmy, 2002):

RED predicts congestion in early stage, by dropping packets.

RED avoids global synchronization (that happened when packet not

dropped randomly) by dropping packets randomly.

RED avoids dropping packets unnecessarily when a short heavy traffic

presented (false congestion) as it uses avg as congestion indicator not q. avg

dose not increased rapidly when false congestion occurs.

4

RED has been successfully adopted by the Internet Engineering Task Force

(IETF) in RFC 2309. However, the problems of RED can be summarized as (Kwon

and Fahmy, 2002):

RED suffers from insensitivity to current queue status, in sudden

congestion, RED slowly adapts and results is packet loss, as it uses avg

instead of q. Notice that considering q will lead to drop packets

unnecessarily when a short heavy traffic presented.

The parameter initialization problem of the RED which affects the

performance of RED.

RED may harm the delay as it does not take into consideration the arrival

and departure rates (traffic-load amount).

The Effective RED (ERED), (Abbasov and Korukoglu, 2009) was proposed to

overcome the insensitivity to current queue size. ERED uses the queue size (q) as

congestion indicator side by side with the avg. Congestion is controlled based on the

following scenarios: When avg is between minThreshold and maxThreshold and q is

greater than minThreshold, ERED drop arrival packets with Dp as calculated in RED.

While, when avg is less than minThreshold and q is less than 1.75* maxThreshold,

Dp is set to a high value. In the rest of the scenarios ERED acts like RED does.

Subsequently, ERED solves the problem of insensitivity to current queue status, in

sudden congestion. However, the ERED’s drawbacks are the parameterization.

Moreover, ERED does not take into consideration the delay in the calculation of Dp.

1.2 Problem Statement:

5

Generally, the sensitivity to current queue size and detection congestion in early

stage, detect of false congestion, delay avoidance, global synchronization avoidance,

avoid parameterization are the factors that should be presented in a congestion

controller. Unfortunately, there is no single congestion control algorithm that takes

into consideration these factors. This problem can be further divided into sub-

problems, these are:

How to implement congestion control method that detects congestion in

early stage, detects false congestion, avoids delay and avoids global

synchronization.

How to solve the problem of parameter initialization and defeat its effects

on the performance of the congestion control method.

1.3 Objectives:

Because ERED has gained the most of the factors listed above for optimal

congestion controller, it will be adapted as the base for achieving the goal of this

thesis that are developing a congestion controller that takes into consideration the

previously mentioned factors for an optimal congestion controlling process. The

objectives of this research are as follows:

To extend the ERED method to consider the Delay in the calculated Dp and

extend the range of the calculated Dp value so that packets are dropped

only if it is necessary to do so, and are dropped rapidly when the risk of

congestion increases.

To combine the ERED algorithm with a Fuzzy Inference Process that shall

solve the problem of parameter initialization.

6

1.4 Thesis Contributions:

The contribution of this thesis:

1. Proposing a congestion control algorithm that takes into consideration

detection congestion in early stage.

2. Detect of false congestion, avoid delay, global synchronization and

parameterization.

1.5 Scope and Limitations:

In computer networks, congestion management and control can be implemented

at several levels in the network; this work focuses on congestion control at the router

buffer. Thus, this work shall consider congestion at a router buffer. Single router

buffer will be used in simulation and testing. The queue for testing and verification is

modeled using the discrete time queue.

1.6 Thesis Organization:

This research is organized into six chapters. Chapter One introduced the principles

behind network congestion control along with the research problem, objectives and

contribution. Chapter Two presents the literature review of the congestion

occurrences effects and influences and the congestion control algorithms. Chapter

Three covers the proposed work. Chapter Four covers the simulation, analysis and

discussion of the proposed algorithm as well as its performance through detailed

experiments. Chapter Five provides results. Finally, Chapter six provides the

conclusions and recommendations for further research.

7

Chapter 2

Literature Review and Related Works

This Chapter discusses the problem of congestion at the router buffer and the

existing methods that were proposed to control this problem and ease its bad influence

on Quality of Service (QoS). The discussion starts by explaining the congestion

problem in the networks, the influence of congestion on the network and the

performance measurements that are effected by congestion. Then, the existing queue-

based congestion control methods are discussed followed by the load-based methods

and then the hybrid-based methods. Then, discussed ERED followed by induction on

the traffic statues that may or may not cause congestion and how congestion

controller response to each status with reference to queue size and delay. The recent

advances in congestion control mechanisms using fuzzy logic are then discussed.

Finally, this chapter is closed by a conclusion.

2.1 Introduction:

Congestion initiated with increment of the amount of packets queued in the

buffer. As the queue is built up, the packets in the queue start to face considerable

delay. When the packets continuously queued in the buffer, the queue gets filled and

no more packets can be served, packet loss occurs. In this case, the buffer is described

as suffered from the problem of congestion, as illustrated in Figure 2.1. Generally,

congestion at the router buffer occurs if the packet departure and packet processing

rates are not sufficient compare to packet arrival rate. Congestion had a bad influence

on the network; this bad influence was the motivation behind the development of

congestion control methods which will also be discussed in this chapter.

8

Figure 2.1: The Congestion at the Router Buffer

2.2 Overview of Congestion Factors at the Router Buffer

Congestion at the router buffer occurs if the packet departure and packet

processing rates are less than packet arrival rate. Arrival rate and departure rate

(together can be referred to as load factor) are the main reason behind congestion.

Queue size, packet loss, throughput and delay are factors influenced by the congestion

(Woodward, 1993).

Arrival rate is the average number of packets received by a single node (router) in

a time window (multiple time slots) and can be determined by monitoring the packets

arrived at the node. Departure rate is the average number of packets departed from a

node in a time window. Load factor is the ratio between arrival rates to the departure

rate as given in Equation 2.1.

⁄

(2.1)

Queue size (q) is the number of queued packets in the router buffer. Mean queue

length (mql) is the average number of packets queued in the router buffer over time.

Average queue length (avg) is an average value of the q at current and earlier time

Packet Loss

High Packet Arrival Rate Low Packet Departure Rate

The Router Buffer

9

calculated as low-pass filter in a time window as given in equation 2.2 (Woodward,

1993).

( )

(2.2)

Where w is a predefined weight, is the queue size at the current time, and

, is the avg calculated in the previous time slot.

Packet loss (PL) is the failure of one or more transmitted packets to arrive at their

destination node. PL is calculated as given in Equation 2.3 (Woodward, 1993).

( )

(2.3)

Where represents the probability the router buffer being full and represents

the packet departure probability. Subsequently, PL is always less than and the

maximum value of PL is that equal to . As such, if the node is always full (e.g.:

=1.0) and no packets are departed (e.g.: =0.0), the PL value will be equal to 1.

Throughput (T) is the probability of packets successfully passes through the

router buffer per time slot, T is calculated as given in Equation 2.4 (Woodward,

1993).

( )

(2.4)

Where , represents the probability of router being in the idle status (no packets

in the buffer) and , represents the packet departure probability. Subsequently, T is

always less than and the maximum value of T is that equals to .

10

Delay (D) is the waiting time for each packet at the router buffer. D is calculated

as given in Equation 2.5 (Woodward, 1993).

(2.5)

As such, these factors and measures, discussed above, can be used as indicators

for early congestion prediction. According to these measures, congestion occurs when

the load factor is above 100%.In this case the arrival rate is higher than the departure

rate and when q or avg are high, which cause packets to queue sequentially in the

router buffer. As such, this will increase PL, D and mql and decrease T. However,

these changes considered as problem for network performance. The aims of the

congestion control process are to predict, detect and prevent congestion in early stage

to decrease PL, D and mql, and increase T, Figure 2.2.

Figure 2.2: Bad vs. Good Network Performance Measures

Packet Loss

Delay

Mean Queue

length

Throughput

Decrease

Decrease

Decrease

Increase

Increase

Increase

Increase

Decrease

Measures Good Network

Performance

Bad Network

Performance

11

2.3 Congestion Control at the Router Buffer

The earliest method for congestion control is a drop tail (DT). DT is simple

congestion control method that is based on predefined queuing threshold. Using DT,

when the number of packets queued in the buffer reach the predefined threshold, all

the arrival packets are dropped accordingly. DT threshold might be set to the buffer

capacity in such a case, no congestion control is conducted and all arrival packets

after the threshold reach is getting lost. On the other hand, when DT threshold is set to

the minimum, huge amount of packets are dropped unnecessary. Subsequently, DT

threshold was mostly set to be above the half of the buffer capacity, given that, in both

cases of normal heavy traffic and sudden congestion status, DT fails to maintain

network performance and keep dropping packets in flat rate, cause unnecessary

dropping in normal traffic and huge amount of packet loss in congestion status.

Generally, DT handles congestion at router buffers in a late stage which consider as

the main drawback of DT (Lin and Morris, 1997).

Active Queue Management (AQM) was proposed to overcome the limitation of

DT. AQM methods maximize the network performance as it does not drop packets

unnecessarily in normal traffic and detect and control congestion in early stage and

start dropping packets early to reduce packet loss and delay. AQM is operated by

gathering information about the performance measures. Then, AQM calculates a

dropping probability (Dp) with each arrival packet. Different method calculates Dp in

different ways, either by estimating the load factor (load-based AQM) or based on the

calculated performance measures (queue-based AQM) or both (hybrid-based AQM).

Based on the calculated Dp, AQM methods decide whether to drop or to allow the

12

packet to be queued in order to maximize the network performance (Jamali, etal,

2014), as illustrated in Figure 2.3.

Figure 2.3: Congestion Control using AQM

A lot of AQM methods were developed that calculates the Dp in different ways,

these methods can be classified into queue-based, load-based, hybrid based and fuzzy

based, this classification is illustrated in Figure 2.4.

Figure 2.4: AQM Methods Classification

Generally, all AQM methods, as an active congestion control have taken into

consideration one or more of the following factors:

1. Avoid global synchronization by dropping packets randomly.

2. Minimize packet loss by predicting congestion in early stage.

Packet Dropping

Measures Calculation

Dp Calculation Packet Queuing

The Router Buffer

Queue-based Hybrid-based

AVG-based Arrival-based

Load-based Fuzzy-based

AQM

Q-based

AVG & Q-based

PL-based Load Factor-based

13

3. Minimize packet loss in response to sudden congestion, by taken into

consideration the sensitivity to current queue status.

4. Avoid unnecessarily packet dropping when a short heavy traffic presented

(false congestion) by taken into consideration the sensitivity to average queue

length over a sufficient time window.

5. Minimize delay by taken into consideration the arrival and departure rates

(traffic-load amount).

6. Avoid parameter initialization problem which affects the overall performance.

The discussion of the AQM methods categories and sub-categories in the

following sections will focus on these congestion control factors and other

requirements and other criteria as coming along.

2.4 Queue-Based Active Queue Management

Queue-based methods use avg, q or PL as congestion indicator. The measures,

avg, q and PL, are calculated continuously. These measures are then used for

calculating Dp with each arrival packet; the calculated value is then used to decide on

dropping packets.

2.4.1 Average- Based

Several AQM methods have used avg as congestion indicator, such as RED

(Floyd and Jacobson, 1993), FRED (Lin and Morris, 1997), DSRED (Floyd and

Jacobson, 1993) (Zheng and Atiquzzaman, 2000), ARED (Floyd, et al., 2001) and

GRED (Floyd, 2000). RED the core of most AQM methods, queue-based active

queue management method uses avg as congestion indicators. The avg status

compares two thresholds values decide the value of Dp and the amount of packets to

14

be dropped. In RED, when avg is less than minThreshold, RED drops no packets

(e.g.: Dp is set to 0). While, all packets are dropped if avg exceeds maxThreshold

(e.g.: Dp is set to 1). The actual congestion control is implemented when avg is

between min and max threshold (e.g.: Dp is set to the value in the range [0-1]). RED

calculates Dp using linear function of avg and maxp and the thresholds. In terms of

advantages and disadvantages, RED avoid global synchronization by dropping

packets randomly, predict congestion in early stage and avoid dropping packets

unnecessarily when a short heavy traffic presented (false congestion) as it is based on

average queue length over a sufficient time window. However, RED does not

response to sudden congestion, as it does not take into consideration the current queue

status, RED also may harm delay as it does not take into consideration the arrival and

departure rates (traffic-load amount) and finally, RED required a lot of parameter

initialization which effects the overall performance.

Flow Random Early Detection (FRED) (Lin and Morris, 1997), extends RED by

promoting fair buffer allocation between flows. Subsequently FRED does not

dropping packets fully randomly. Instead it drops packets that are received from the

source that exceeded its allocated share and keeps the packets from sources that do

not exceed its allocated share. FRED use the same technique as RED, including the

thresholds, parameters and Dp function, the difference is that FRED adds a step to

drop packets from exceeded sources independent from the calculate Dp.

Double Slope Random Early deduction (DSRED) (Zheng and Atiquzzaman,

2000) extends RED with the aim at reducing the packet dropping in non-congestion

status. DSRED uses avg as congestion indicator and a combination of two Dp

calculation functions. These functions are linear which operated in loops, one of them

severely reacts with the increment of avg and the other is non-severed. By adjusting

15

the control parameter, the DSRED will response to the increment of avg by high drop

rate first followed by a low drop rate, or vice versa.

Adaptive RED (ARED) (Floyd and Shenker, 2001), uses avg as similar to RED

and calculates Dp by additive increase multiplicative decrease (AIMD) function. This

approach focus on setting an optimal value for a parameter maxp, by which the

dropping probability increases rapidly when avg increases to avoid buffer overflowed.

The AIMD has proven to give better response to sudden congestion compares to

RED.

Gentle RED (GRED) (Floyd, 2000) also uses avg, and is built as extends to

ARED by optimizing the parameters for maxp. AGRED adds a new position at the

router buffer called doubleMaxThreshold that is a double value of the maxThreshold.

By doing so, AGRED tries to react differently when the number of packets exceeds

maxThreshold in order to avoid buffer overflowing but at the same time do not drop

many packets unnecessarily.

Overall, avg-based AQM avoid global synchronization by dropping packets

randomly, predict congestion in early stage and avoid dropping packets unnecessarily.

On the other hand, these methods do not response to sudden congestion, may harm

delay and required massive parameter initialization. In terms of results, RED, FRED

and SRED achieve mostly similar results; DSRED always has lower queuing delay,

smaller queue size and lower packet drops than the RED, SRED and FRED. ARED,

GRED and AGRED have better results compare to DSRED (Morris, 2000).

In summary, avg-based AQM methods try to set optimal values for Dp to avoid

drop many packets unnecessarily. Other methods were also built on top of RED with

the overall aim to enable stabilization of the avg under all circumstances, e.g.: heavy,

16

light and no congestion in order to enhance network performance. However, it was

notably the inflexibility of these methods with the status of avg, which leads to either

drop packets unnecessarily or more packets store in buffer which lead to overflow and

delay.

2.4.2 Queue- Based

Few AQM methods have used q as congestion indicator, such as Stabilized

Random Early Drop (SRED) (Ott, et al., 1999) uses an estimated number of active

connections or flows with q as congestion indicators and operates similar to RED and

FRED. Subsequently, SRED calculates DP based on q. SRED, similar to RED avoid

global synchronization by dropping packets randomly, predict congestion in early

stage and response to sudden congestion as it is based on q not avg. On the other

hand, SRED may drop packets unnecessarily, when false congestion occurs, may

harm delay and required massive parameter initialization. Some other methods that

use q as congestion indicator were developed ((Fakharian and Abbasi, 2015), (Jamali,

et al, 2014), (Kadhum and Manickam, 2014)).

Overall, as an extension to avg-based method, q-based AQM avoid global

synchronization by dropping packets randomly, predict congestion in early stage.

However, q-based method may drop packets unnecessarily. On the other hand, q-

based methods response to sudden congestion more effectively compares to avg-

based.

2.4.3 Average & Queue- Based

The Effective RED (ERED) (Abbasov and Korukoglu, 2009), is a unique AQM

method, that was proposed to overcome the limitations of q-based and avg-based

17

methods. ERED uses q and avg as congestion indicators. Congestion is controlled

based on the following scenarios: When avg is between minThreshold and

maxThreshold and q is greater than minThreshold, ERED drop arrival packets with

Dp as calculated in RED. While, when avg is less than minThreshold and q is less

than 1.75* maxThreshold, Dp is set to a high value. In the rest of the scenarios ERED

acts like RED does.

In terms of advantages and disadvantages, ERED avoids global synchronization

by dropping packets randomly, predicts congestion in early stage, and avoids

dropping packets unnecessarily when a short heavy traffic presented (false

congestion) as it is based on average queue length over a time and responses to

sudden congestion, as it does take into consideration the current queue status.

However, ERED harms delay as it does not take into consideration the arrival and

departure rates (traffic-load amount) and ERED required massive parameter

initialization which affects the overall performance.

2.4.4 Packet loss - Based

Other AQM methods have used PL as congestion indicator, such as BLUE (Feng,

et al., 1999) and MRED (Koo, et al., 2001). BLUE (Feng, et al., 1999) uses PL and

link idle as indicators of congestion and Dp is calculated in linear function. If the

queue is continually dropping packets due to buffer overflow, BLUE increase Dp,

while Dp is decreased if the queue becomes empty (e.g.: link is idle). Two parameters

are used to adjust Dp based on the congestion status, the freeze_time determines the

minimum interval between two successive updates. The value of freeze_time should

be randomized in order to avoid global synchronization. The other parameters are

used to determine the amount by which Dp is increased/decreased.

18

Modified RED (MRED) (Koo, et al., 2001) uses estimation of packet loss and

link utilization as congestion indicators. MRED is estimate average queue length

using these indicators and implements dropping as in RED.

In terms of advantages and disadvantages, PL-based AQM methods avoid global

synchronization by dropping packets randomly, avoids dropping packets

unnecessarily when a short heavy traffic presented (false congestion) and enhance

delay measure compare to avg-based and q-based. However, PL-based AQM fail

sometimes to predicts congestion in early stage and does not response to sudden

congestion and finally, PL-based AQM required massive parameter initialization.

Generally, Queue-based AQM regardless of its sub-type focus on using the status

of the queue to determine the Dp that is used for congestion control. Different

methods use different indicators, parameters and functions. Overall ERED have

shown robustness in much of the congestion control factors mentioned above.

2.5 Load-Based Active Queue Management

Load-based methods use arrival rate or load factor as congestion indicator.

Similar to queue-based methods avoid congestion by calculating Dp with each arrival

packet, and use the calculated value to decide on dropping packets at early stage to

avoid congestion.

2.5.1 Arrival- Based

Adaptive Virtual Queue (AVQ) (Kunniyur and Srikant, 2003) used link load as

congestion indicator. AVQ maintains a virtual queue whose capacity is less than the

actual capacity of the link. The number of packet queued in the virtual queue

synchronized with the number in the real one. Packets in the real queue are dropped

19

when the virtual buffer overflowed. The dropping increased when the utilization of

virtual capacity is increased and all packets are dropped when the transmitting rate

exceed the virtual queue capacity.

AVQ is responsive to changes in network load (Congestion) and is able to

maintain a small queue length even when network load keeps increasing (Kunniyur

and Srikant, 2003). Stable Adaptive Virtual Queue (SAVQ) (Long, et al., 2005)

extends AVQ. Enhanced Adaptive Virtual Queue (EAVQ) algorithm (Yanping, et al.,

2007) uses an enhance estimation of the arrival rate at the network.

In terms of advantages and disadvantages, arrival-based AQM methods avoid

global synchronization by dropping packets randomly, predict congestion in early

stage, response to sudden congestion and improve delay. However, arrival-based

AQM drop packets unnecessarily when a short heavy traffic presented (false

congestion) and required massive parameter initialization.

2.5.2 Load Factor- Based

Link Utilization Based AQM (LUBA) (Agarwal , et al., 2004) use estimation of

the load factor as congestion indicator. If the overload factor U=λ/μ (where λ is the

totla arrival rate at the router and μ is the outgoing link capacity of the router), is

below the target link utilization, the router is non-congested and packets are not

marked or dropped. When it is greater, all arriving packets are dropped, the router is

congestion.

In terms of advantages and disadvantages, load-based AQM methods avoid global

synchronization by dropping packets randomly, predict congestion in early stage,

response to sudden congestion and improve delay. However, load-based AQM drop

20

packets unnecessarily when a short heavy traffic presented (false congestion) and

required massive parameter initialization.

Overall, considering the load in the AQM enhances the delay. However, there is

tradeoff between queuing delay and queue utilization. Intuitively, larger buffers lead

to higher utilizations of the link, but they also result in larger queuing delays.

2.6 Hybrid Active Queue Management

As mentioned previously, both queue-based and load-based through its sub-types

have advantages and disadvantages. In the attempt to maximize the benefits of both

types, hybrid AQM methods were proposed.

Stabilized Virtual Buffer (SVB) (Deng, et al., 2002) considers arrival rate and q

as congestion indicators and calculates Dp in linear equation with the aim at

stabilizing congestion control factors at router buffer.

Rate AQM (RAQM) (Chonggang , et al., 2005) uses the arrival rate and q as

congestion indicators. RAQM works in two modes that each calculates Dp in different

way. In the load dependent mode, only the total arrival rate is used to regulate the

output rate. In queue dependent mode, it also uses the q to further adjust the packet

drop probability and to regulate the queue length to the expected value.

Robust Active Queue Management (RaQ) (Sun and Zukerman, 2007) uses the

arrival rate and q as congestion indicators to calculate Dp. RaQ used dual inner loops

controller. A loop uses the arrival rate and the other loop uses q. Other methods, such

as Proportional Integral Controller (PI) (Hollot, et al., 2001), uses as congestion

indicators, q and traffic load. Yellow (Chengnian , et al., 2005) use the arrival rate and

q as congestion indicator.

21

In terms of advantages and disadvantages, hybrid-based AQM methods avoid

global synchronization by dropping packets randomly, response to sudden congestion

and improve delay. However, hybrid-based AQM methods may drop packets

unnecessarily when a short heavy traffic presented (false congestion) and they

required massive parameter initialization.

2.7 Fuzzy Based Active Queue Management

Fuzzy Logic (FL) has been used recently to solve the parameterization problem in

AQM. Several methods, such as Fuzzy BLUE method ((Feng , et al., 1999), (Wu-

chang , et al., 2002)) and Fuzzy RED method ((Chrysostomou , et al., 2003),(

Chrysostomou, , et al., 2003)), has been developed based on existing non-fuzzy AQM

methods.

FL is a field of study that concerns with the uncertainty and inexact modeling of a

problem [(Negnevitsky, 2009), (Abdel-Jaber, , et al., 2008),( Pedrycz and Vasilakos,

2000) ,( Ingoley ,2012), (Loukas, 2000))) . Opposite to binary logic (BL) (Zadeh,

1975), which depends on two values (0, 1), FL associates a fuzzy value, for the desire

output variables, between [0, 1].

With AQM methods, FL is used to calculate the packet dropping probability.

Fuzzy Explicit Marking (FEM) [(Chrysostomou, et al., 2003), (Negnevitsky, 2005),

(Hiok and Qiu, 2000)] is AQM method that is extending RED by replacing the

utilized thresholds in RED with a fuzzy inference process (FIP). FEM calculates Dp

as an output linguistic variable based on two input linguistic variables, that are the

changing rate in the buffer (CRB) and current buffer length (CBL) as parallel to

using avg in RED ((Negnevitsky, 2009) , (Chrysostomou, , et al., 2003)). REDD1

(Abdel-Jaber, et al., 2008), fuzzy based method, that was built on top of RED method.

22

The REDD1 relies on two input linguistic variables, which are avg and PL, to

generate one output linguistic variable, Dp.

Other fuzzy-based AQM methods were built on other than RED, such as

Adaptive Fuzzy-based AQM AFRED (Wang, et al., 2003) which used Gentle RED at

its core. AFRED uses as its input variable the avg and the output is Dp. Fuzzy Logic

Controller Random Early Marking (FUZREM) (Changbiao and Fengfeng, 2008)

implements a FI for REM (Yongju, et al., 2009). DEEP BLUE (Masoumzadeh, et al.,

2009) consists of combining the BLUE AQM algorithm with the FIP.

Fuzzy Control RED (FCRED) (Sun, et al., 2007) consists of a fuzzy controller

adjusting the maxp parameter of the RED algorithm based on the input variable avg.

RED-based fuzzy controller (Moghaddam, 2010) is also used to adjust the value of

maxp of RED. There are many other methods that use FIP to extend existing AQM

methods [(Negnevitsky, 2009), (Yaghmaee and AminToosi, 2003) ,( Zadeh ,1965)

,(Klir, 1995)].

Overall, these methods have solved the parameter initialization problem that

presented in other AQM methods. The drawbacks of these methods are inherited from

its core methods.

Table 2.2 summarizes the evaluation of the developed methods based on the

congestion control factors.

23

Table 2.2: Evaluation of AQM Methods

Ind

icato

r

Av

oid

Glo

bal

Sy

nch

ron

iza

tio

n

Pre

dic

t C

on

ges

tio

n

in E

arl

y s

tag

e

Res

po

nse

to

su

dd

en

Co

ng

esti

on

Av

oid

un

nec

essa

ry

pa

cket

dro

pp

ing

No

del

ay

Av

oid

Pa

ram

eter

iza

tio

n

RED Avg Yes Yes No Yes No No

FRED Avg Yes Yes No Yes No No

DSRED Avg Yes Yes No Yes No No

ARED Avg Yes Yes No Yes No No

GRED Avg Yes Yes No Yes No No

SRED Q Yes Yes Yes No No No

ERED Avg & Q Yes Yes Yes Yes No No

BLUE PL Yes No No Yes Yes No

MRED PL Yes No No Yes Yes No

AVQ Arrival rate Yes Yes Yes No Yes No

SAVQ Arrival rate Yes Yes Yes No Yes No

EAVQ Arrival rate Yes Yes Yes No Yes No

LUBA Load rate Yes Yes Yes No Yes No

SVB Arrival rate & Q Yes No Yes No Yes No

RAQM Arrival rate & Q Yes No Yes No Yes No

RaQ Arrival rate & Q Yes No Yes No Yes No

PI Load rate & Q Yes No Yes No Yes No

Yellow Arrival rate & Q Yes No Yes No Yes No

FEM Avg Yes Yes No Yes No Yes

24

REDD1 Avg & PL Yes Yes No Yes No Yes

AFRED Avg Yes Yes Yes No Yes Yes

FUZREM Arrival rate & Q Yes No Yes No Yes Yes

DEEP BLUE PL Yes No No Yes Yes Yes

FCRED Avg Yes Yes No Yes No Yes

RED-FC Q Yes Yes No Yes No Yes

2.8 ERED

AQM methods for congestion control at the router buffer have been developed for

the same purpose that is achieving the best QoS; however each method uses different

process with various considerations. ERED, has used as the technique for calculation

Dp as the one used in RED and enhance it. ERED has gained more advantages,

compare to RED, as it considers one more measure than RED that is the queue size q.

Subsequently, ERED has gained most of the advantages that a congestion controller

may have. However, ERED has shortages in terms of delay and parameterization

which need to be addressed. Accordingly, ERED will be extended to consider the

delay measure to improve its performance in terms of the delay.

The proposed work extends the ERED algorithms by injecting the measure of

delay into its main algorithm. By doing this, the ERED algorithm is upgraded from

being queue-based into a hybrid-based, as the utilized delay will be estimated based

on the load factor. Then, the extended ERED is wrapped by fussy inference process

(FIP) to solve the parameter initialization problem.

2.9 AQM and Delay

25

In computer networks, the undesirable increase of delay (D) affect on the increase

of the queue size mql. The desirable increase in Throughput (T) is also affect on the

increase of mql. Reducing D through reducing mql, by dropping packets, will increase

T, which is not desirable. As such, no AQM method will be better than the other when

the traffic load is normal or slightly heavy. When the load is low, no sacrifice is

necessary as D, T and mql. However, the differences of the AQM methods in terms of

their achievement on D and T are stated clearly in the congestion and pre-congestion

states. This is because the decision of dropping or allowing new arrival packets to

queue will affect these measures simultaneously in positive or negative manner.

Figure 2.5 illustrates the network statuses with different packet arrival rate, that

creates or does not create congestion in the buffer, compares to optimal rate (e.g.: rate

that can be handle efficiently by the network), with assumption that departure rate are

constant.

#Pac

kets

Arr

ival

#Time

Optimal Line

#Pac

kets

Arr

ival

#Time

Optimal Line Delay

#Pac

kets

Arr

ival

#Time

Optimal Line

Delay

#Pac

kets

Arr

ival

#Time

Optimal Line

Delay

#Pac

kets

Arr

ival

#Time

Optimal Line

Delay

(d) Overload

(a) Optimal (b) Nearly Optimal (c) Built-up Queue

Less than

optimal

26

Figure 2.5: Illustration of Traffic Scenarios in Router Buffer with a Constant

Departure Rate

Figure 2.5(a) illustrates arrival packet is represented by a straight line. The

optimal line, a non-real demonstration shows that the numbers of packets pass by the

router are increased overtime in linear manner, which means that the packets arrival

rate is constant. The optimal line demonstrates the scenario where no packet loss

occurs, no dropping is required and no delay is faced. Subsequently, any load-based

or queue-based will efficiently operates in such a case. Figure 2.6 illustrates a

demonstration of the packets arrival/departure to the queue over time in the optimal

case.

Figure 2.6: Illustration of Packets Arrival/Departure in the Buffer over Time in

Optimal Case

Congestion controller is required when the rate of packet arrival is not at the

optimal rate as shown in Figure 2.5(b). For example, the increase of the packet arrival

rate, with the assumption the departure is constant, will lead to accumulate packets in

Buffer in out

Demonstration

Evolving Process

Tim

e

....

.

27

the buffer and subsequently, packets will start facing delay. However, if this

increment followed by decrement to a value significantly below the optimal rate, the

buffer would build up for a short period only and will get back into its optimal state.

For example, say that the optimal rate leads to have 4 packets queued at any time in

the buffer to avoid any delay and packet loss. Increasing the arrival rate, will lead to

have more packets queued in the buffer, says 6 packets in the queue. Then, having

arrival rate below optimal will lead to have 4 or less packets queued in the buffer

continuously in the following time slots. Early dropping of packets in such a case is

just a waist, as the delay and queue size will increase temporary for a limited time

only and no packet loss will occur. Any load-based or queue-based will efficiently

response to such a case. As the given scenario has a limited and short bad affect and

AQM method does not drop packets in such a case. Figure 2.7 illustrates a

demonstration of the packets arrival/departure to the queue over time in the case.

Buffer in out

Demonstration

Evolving Process

Tim

e

....

.

28

Figure 2.7: Illustration of Packets Arrival/Departure in the Buffer over Time in

Nearly Optimal Case

In another case, when the arrival rate increases significantly and packets are

accumulated in the buffer, queue is built up and the accumulated packets start facing

delay. Then after awhile the rate maybe decreases and got back into the optimal rate,

but the buffer would be already built up and all the subsequent packets will face

similar delay. For example, say that the optimal rate leads to have 4 packets queued at

any time in the buffer to avoid any delay. Increasing the arrival rate for a short while,

will lead to have more packets queued in the buffer, says 6 packets in the queue. Even

if the traffic load got back to the optimal rate, 6 packets queued in the buffer

continuously in the following time slots. Having 6 packets will lead to have a packet

delayed in each time slot. An option in this case is to sacrifice, by dropping, all the

packets that rose above the optimal rate to avoid the delay of all subsequent packets

that reached within the optimal rate. Notice that, the number of packet queued might

not be large, but it is still causes delay. Figure 2.5(c) illustrates such a case. Generally,

drop some packets in this case, will preserve the delay. Only load-based technique

that consider the delay will implements dropping in this case if the queue size is

considered to be small. Figure 2.8 illustrates a demonstration of the packets

arrival/departure to the queue over time in the case.

29

Figure 2.8: Illustration of Packets Built-up Queue over Time when Delay

Increase

In another case if the traffic increased, initially then followed by another increase

and other increases. Avoid dropping in early stage will harms delay and huge packet

loss will occur. In this case, the status can be view clearly by measuring the number of

packets queued without looking into delay, as the queue size increases there is a need

to drop packets. Figure 2.5(d) illustrates such a case. Early dropping of packets is

necessary, as the delay and dropping rates will both be influenced negatively

otherwise. Both load-based and queue-based will finally lead to packet dropping but

the fast response will be initiated from queue- based. Figure 2.9 illustrates a

demonstration of the packets arrival/departure to the queue over time in the case.

Buffer in out

Demonstration

Evolving Process Ti

me

....

.

30

Figure 2.9: Illustration of Packets Overload the Buffer over Time in Unstable

Traffic

Overall, there are some cases wherein the queue size should be used to prevent

congestion and some other times delay should be used to enhance the performance

measure. Subsequently, there is a need to decide whether to drop packets based on

queue and delay states. It is to be noted that there is no optimal rate as similar to the

one drawn in Figure 2.5, in real implementation, subsequently the discussed scenarios

do not applied literally. Moreover, similar to all these scenarios might be happened

sequentially in a short while.

2.10 Conclusion

Buffer in out

Demonstration

Evolving Process Ti

me

....

.

31

Overall, a lot of AQM methods were developed that calculates the Dp in different

ways, these methods can be classified into queue-based, load-based, hybrid based and

fuzzy based. Many of these algorithms were proposed to address some problem in the

core AQM method that called RED.

Overall, ERED has gain most of the factors listed above for optimal congestion

controller. However, ERED has shortage in terms of delay and parameterization

which need to be addressed.

32

Chapter 3

Methodology

This Chapter presents the methodology that will be followed in order to develop

the proposed congestion controller to achieve the desired goals. A brief introduction is

given in section 3.1 followed by the methodology in section 3.2 and the conclusion in

the last.

3.1 Introduction

To start with, the methodology is designed with the aim to clarify the work phases

in clear steps. The desired work is mainly divided into two steps: The first step aims

to achieve the first goal which combining the delay factor into a congestion control

process. The second is to wrap the process with a Fuzzy Inference Process (FIP). Both

of these followed an identified methodology as will be given in next section.

3.2 Methodology

The methodology adopted to achieve the desired goals is conducted as follows:

The parameters are initialized first. Then, a packet generated and sent to the queue.

The packet as generated and sent, maybe lost if the queue is full, or it might be

dropped or queued as decided based on calculated the value of drop probability Dp for

each packet arrival by using the proposed method parametric and FIP-based, where

Packet arrival is simulated based on a pre-determined probability value called

“Alpha” that is in the range between 0-1. A random number generation is used to

generate a random number to be compared with alpha. If the number generated below

alpha then the packet is arrived to the buffer, else no packet is arrived.

33

The properties of random number besides it is random, it also generates a

distributed number. When Alpha is 0.5 then for a 100 slots, 50 slots will allow packet

arrival. Packet departure is simulated based on a pre-determined value called “Beta”

and is happened in a similar way as the packet arrival and based on random number

generation.

Congestion Control checks the statuses of the queue calculates the average queue

length and decides whether to drop the packet or to queue it. In the same time slot, a

packet maybe departed. This process is repeated for each time slot. This process is

illustrated in Figure 3.1.

Figure 3.1: Flowchart of the Hybrid ERED

Parameter Initialization

Finish No

Results Collection

Yes

AQM Method Calculate Dp based on Equation (extends ERED by including delay consideration in the implemented)

Packet Arrival

Queue Status Checking Drop/ Loss /Queue-in

Packet Departure

34

The parameters used in the experiments for are given in Table 5.1 as follows:

Table 3.1: Parameter settings

Parameter Initial Value

Probability of packet arrival 0.33, 0.5, 0.66, 0.93

Probability of packet departure 0.5

Number of slots 2,000,000

Warm-up 800,000

Router buffer capacity 20

Queue Weight 0.002

maxp 0.1

minth 3

maxth 9

The process of implementing Fuzzy-based congestion controller is given in

Figure 3.2.

Figure 3.2: Flowchart of Fuzzy Hybrid ERED

Create Fuzzy Sets Input

& Output Variables Create Membership

Functions Create Rule Set

Fuzzy Sets Membership Functions Rule Set

Crisp Input Fuzzification Rules Evaluation

Aggregation Defuzzification

Initialization

35

3.2.1 Initialization

The fuzzy sets should be determined in this step; the set may have the form

{"low", "medium", "high"}. Also, the membership functions are drawn in this step;

the membership function describes the slop of each linguistic term in the fuzzy set in

the space of the crisp input variable. The membership function which can be

formulated for the input and output linguistic variables used well-established

approaches of trapezoidal (Klir, 1995) or triangular (Negnevitsky, 2009). The

initialization of these sets and functions are subject to trial and error approach. Thus,

the initialization is altered and modified after the initial experiments are conducted. In

another word, the sets and membership functions should be trained to determine the

boundary and the slop of each linguistic term.

The rules are also determined, the rules utilized to obtain the linguistic terms for

the output variable from the linguistic terms of the input variables. The fuzzy

linguistic rules are created by error and trial methods.

3.2.2 Fuzzification

Based on the sets and memberships, in the Fuzzification step, the input crisp

values of the input linguist variables are converted into their equivalent linguistic

terms in the fuzzy set. Each crisp value can be converted into one or multiple

linguistic terms in its corresponding fuzzy set.

3.2.3 Rules Evaluation

Fuzzy rules, part of the fuzzy logic theory describe the connection between the

inputs and outputs variables. A fuzzy rule is a conditional statement of the form "IF x

is X AND y is Y THEN z is Z", where "x", "y" and “z” are the linguistic variables and

36

“X”, “Y” and “Z” are their representative linguist terms (Bazaz, et al., 2013). Thus, in

this step, the output of the fuzzification process is injected as input to the rule

evaluation and the output is a linguistic term(s) for the output variable.

3.2.4 Aggregation

The next step is to aggregate the output of all rules applied. If the inputs can be

anticipated on multiple rules, then multiple linguistic terms for the output variable

will be obtained. Each of these will also have a probability value which is obtained in

this step. This step is start with obtain a probability value for each output linguistic

term in which the input can be anticipated. The probability value of the output is

obtained as the maximum value of the probability of the input linguistic terms

anticipated.

Moreover, if the same linguistic term is appeared multiple times then the

maximum value for this term is considered.

3.2.5 Defuzzification:

Defuzzification, the final step generates a single crisp value for the output

linguistic variable. One of the popular defuzzification techniques is the center of

gravity (COG) method, which aims to find out the point located on the center of the

aggregate fuzzy set for each output linguistic. Given that multiple output linguistic

terms is obtained as multiple rules are anticipated, each with a probability value, the

final output crisp result is obtained mostly by the center of gravity.

3.3 Conclusion

37

Generally, original ERED and its ancestor method, RED, harms delay and

number of packet queued as it does not take into consideration the rates of traffic and

its transformation over time. Generally, the queue-size measure is helpful to avoid

congestion and thus, packet loss. However, delay is useful to avoid long packet

queuing time and dropping packets unnecessary especially, when the load decreases

significantly. Hybrid-ERED method is proposed by extended ERED method to

consider the delay measure. Fuzzy hybrid-ERED wrap the proposed method with

fuzzy inference process, to solve the parameter initialization problem.

38

Chapter 4

The System Design

This Chapter presents the proposed method for congestion control. The proposed

method is built on top of ERED method and is developed in two stages: The first

stage extends ERED by including delay consideration in the implemented congestion

controller mechanism (Hybrid-based ERED). The second stage further extends the

method resulted in the first stage by including fuzzy inference process to solve the

parameter initialization problem (Fuzzy-based Hybrid ERED). This chapter starts

with discussed the hybrid-ERED then the fuzzy hybrid ERED. Finally, a conclusion is

presented at the end of the chapter.

4.1 Hybrid-based ERED:

The proposed hybrid-ERED uses three congestion indicators, these are: q, avg

and DEsti. DEsti will be calculated through the arrival rate and departure rate. Dp will

be calculated to be increased/decreased linearly with the increment/decrement of the

value of the congestion indicators. In the hybrid-ERED, a few scenarios are

considered that is drawn based on the status of the indicators according to some

thresholds.

4.1.1 Delay Calculation:

Network delay generally is imposed of different types of delays, these are:

processing delay, transmission delay, propagation delay and queuing delay, the

subject of the proposed method. The queuing delay considered with the congestion

control can be calculated according to various laws. A derivation of the Little's law

39

(Allen, 1990), which has a fair estimation of the time that packet spends in the queue.

As such, estimated delay DEsti, is calculated as given in Equation 4.1.

( ) ( )

(4.1)

The estimated delay is proportional to the increment of the arrival rate and

inverse proportional to the departure rate. The estimated arrival rate is low-pass filter

of the average arrival rates over previous and current packets arrival events. The

arrival estimate is calculated as given in equation 4.2.

( ) ( )

(4.2)

Where p is the time value, is the value calculated previously for the arrival

rate, Arrp is the number of packets arrived in current time and warr is the weight

value. The estimated departure rate is calculated similarly as given in Equation 4.3.

( ) ( )

(4.3)

The low pass filter gives a more value to the older value to contribute to the

calculated output. This is implemented by given the weights parameters and

values below 0.5(Baklizi, et al., 2014). Subsequently, it is changed gradually

when the rate changes overtime. However, the same Equations above, Equation 4.2

and Equation 4.3 can be used for normal averaging process by given the weights

values equal to 0.5 and can be used as high-pas filter by given values above 0.5.

Subsequently, the value of DEsti in the optimal case, when the rate of arrival equal

to the rate of departure and only a single packet queued is equal to one. The value of

40

one is the edge between positive/high delay and negative/no-delay. Positive/high

delay has a value above one, which occurs when the arrival rate is more than the

departure rate and the queue is not empty. Negative/no-delay delay has a value below

one, which occurs when the departure rate is more than the arrival rate and the queue

is not full of packets. DEsti is proportional with the term (ArrEsti /DepEsti) as given in

Figure 4.1.

Figure 4.1: DEsti vs. (ArrEsti /DepEsti)

4.1.2 Hybrid-ERED Mechanism:

In the proposed Hybrid-ERED, avg and q indicators are tight up with thresholds.

These indicators according to their values with respect to the threshold values

determine the status of congestion. Based on the value of avg and q Hybrid-ERED

either drops no packets, drop packets based on Dp that is calculated using Equation

4.4, drop packets based on Dp calculated using Equation 4.5, which generates values

below the values generated in 4.4, subsequently drop less packets or drop all packets.

( h)

h h ,

( )

(4.4)

41

(4.5)

Equation 4.5 is bored from ERED as it is. This Equation drops fewer packets than

Equation 4.4 and it is used in non-severe congestion as will be discussed in the

following. Thus, adjusting this Equation is not valuable. The first part Equation 4.4 is

derived from ERED and is connected with the second part that has been developed in

hybrid-ERED to consider the delay. The parameter wD is a weight value determines

the contribution of the delay status in the network to the overall calculated Dp value.

The delay itself affects the calculated Dp value through the calculated delay, DEsti. As

mentioned previously in the optimal network status the value of the DEsti is 1.

4.1.3 Algorithm:

The Hybrid-ERED algorithm that integrates with delay measure is given below in

Algorithm 4.1.

Algorithm 4.1: Hybrid-ERED:

1

2

4

4

5

6

7

8

INITIALIZATION:

avg:= 0

count:= 0

FOR EACH arrival packet

CALCULATE new avg as follows:

IF q = 0 THEN avg:= (1 - wavg) f(idel Time)

* avg // When q = 0 , Idel time =

time

IF q != 0 THEN avg:= (1 - wavg)* avg + wavg *q

42

9

10

11

12

14

15

16

17

18

19

20

21

22

23

24

25

26

27

28

29

30

CALCULATE Dest and its related parameters as follows:

IF q = FULL THEN arrp = arrp-1 and depp:= (1-wdep)* depp-1 + wdep * #departedp

IF q != FULL THEN

arrp:= (1-warr)* arrp-1 + warr * #arrivedp and depp:= (1-wdep)* depp-1 + wdep *

#departedp

Dest:= (1 - depp) * arrp *q

CALCULATE Dp and its related parameters, and implements packet dropping, as:

if (minth2 ≤ avg < max th3) && (q≥min th2)

Dp'= maxp * (avg-min th)/(max th-min th)

Dp = Dp'/ (1-count* Dp') + wd(D est)

with probability Dp

drop packet

count := 0

else if (avg < min th2) && (q>max th2)

Calculate Dp = maxp / (1-count* maxp)

with probability Dp

drop packet

count := 0

else if (avg ≥ max th3)

Drop packet

Count = 0

else

43

Count = -1

Lines 1-4 provide the necessary initialization, in which the parameter avg is set to

0 and count to 0. Lines 4-30 are repeated with each arrival packet, mainly four major

processes is implemented with each arrival packet, there are:

Calculate the value of avg, in lines 5-7 as follows: Case 1 in line 6, if the q is

empty, q is the size of the buffer, and then avg is calculated based on the

function powered by the idle time. The more idle time (the longer the buffer

remains empty), as given in Equation 4.6. Case 2 in line 7, if q is not empty,

avg is calculated as low-pass filter of the average queue size, as given in

Equation 4.7.

( ) ( )

(4.6)

( ) ( )

(4.7)

Calculate the value of DEsti, lines 8-12, as follows: Case 1 in line 9, if the q is

full, then packets will be lost and no packets will arrive, subsequently, the

estimated arrival ArrEsti, cannot be updated and it remains as calculated before

buffer over flooding while estimated departure DepEsti, is calculated as given

in Equation 4.3 and discussed earlier. Case 2 in lines 10-11, if q is not full,

both the estimated arrival ArrEsti, and estimated departure DepEsti, are

calculated as given in Equation 4.2, Equation 4.3 and discussed earlier.

Calculate the value of Dp and implements packet dropping or queuing, lines

14-30, as follows: Case 1 in lines 15-20, if the avg between minth2 and

44

maxth3, and q above the minth2 the dropping probability Dp is calculated as

given in Equation 4.4 and discussed earlier. If the arrived packet will be

dropped according to the calculated Dp value as given in lines 16-17, the value

of the variable count is set to 0, as given in line 20. Case 2 in lines 21-25, if

the avg is less than minth2 and q above the maxth2 the dropping probability

Dp is calculated as given in Equation 4.5 and discussed earlier. If the arrived

packet will be dropped according to the calculated Dp value the value of the

variable count is set to 0. Case 3 in lines 26-28, if the avg is more than

maxth3, regardless of the q vale, the dropping probability Dp is set to 1, the

arrived packet is dropped and the value of the variable count is set to 0. Case 4

in lines 29-30, if the avg is between minth2 and maxth3 and q vale is less than

min th2 or if the avg is less than minth2 and q vale is between minth2 and

maxth3 the dropping probability Dp is set to 0 and the value of the variable

count is set to -1.

Keep track of the starting of idle time, the comment in line 6, the idle time

starts when the buffer gets empty.

The proposed method control congestion by calculating a Dp value based on

some input measures that estimated from the current buffer and traffic statues, there

parameters are: Queue size q, Average Queue Length avg, Estimated Arrival Rate

ArrEsti, Estimated Departure Rate DepEsti, Estimated Delay DEsti, A fixed value maxp

and count since the last dropped packet count. Mainly, the dropping is controlled

when q, avg values are between some thresholds that are: minth, minth2, maxth,

maxth2, maxth3. The utilized, measures, parameters and thresholds are identified as

follows:

1. Queue (q): The number of packets queued in the buffer.

45

2. Average Queue Length (avg): The average number of packets over a time

window calculated using low-pass filtering.

3. Estimated Arrival Rate (ArrEsti): The average number of packets received in a

time window. This value is updated with every arrival packet, when the buffer

overflowed, ArrEsti value freeze.

4. Estimated Departure Rate (DepEsti): The average number of packets departure

in a time window. This value is updated with every arrival packet.

5. Estimated Delay (DEsti): the amount of time units the packets stayed in the

buffer. This value is updated with every arrival packet

6. The count since the last dropped packet (count): number of packets queued

after the last packets dropped.

7. Full: packet number stored in router buffer equal buffer capacity.

8. Empty: no packet in router buffer.

The Thresholds are:

1. minth: A fixed value that refers to a position in the router buffer, if not

exceeded by the queued packets, according to avg and q, no dropping is

necessary as the queue will be considered of fair size.

2. minth2: A value that refers to a position in the router buffer, if not exceeded

by the queued packets, according to avg and q, no dropping is necessary as the

queue will be considered of fair size. minth2: is not a fixed value and it is

calculated as it is calculated as given in Equation 4.8.

3. maxth: A fixed value that refers to a position in the router buffer, if exceeded

by the queued packets, according to avg and q, the arrived packets should be

dropped as the queue will be considered to be in risk stage.

46

4. maxth2: A value that refers to a position in the router buffer, if exceeded by

the queued packets, according to avg and q, all the arrived packets should be

dropped. maxth2 is not a fixed value and it is calculated as it is calculated as

given in Equation 4.9.

5. maxth3: A value that refers to a position in the router buffer, if exceeded by

the queued packets, according to avg and q, the arrived packets should be

dropped. This stage is more risky than previous one. Maxth3 is not a fixed

value and it is calculated as it is calculated as given in Equation 4.10. These

parameters and equations have been borrowed from ERED (Abbasov and

Korukoglu, 2009).

(4.8)

(4.9)

(4.10)

Dp is calculated based on the status of the buffer and with reference to the

parameters identified previously based on the following scenarios:

1. When avg between minth2 and maxth3, and q above the minth2, packets are

dropped randomly with probability Dp.

2. When avg is less than minth2 and q above the maxth2, packets are dropped

randomly with probability Dp, however, with probability value less than the

one calculated in the previous case.

3. When avg above maxth3, all arrived packets are dropped (e.g.: Dp = 1.0).

47

4. When avg is between minth2 and maxth3 and q vale is less than min th2 or if

the avg is less than minth2 and q vale is between minth2 and maxth2, no

packets are dropped (e.g.: Dp = 0.0).

4.2 Fuzzy-based hybrid-ERED:

The proposed hybrid-ERED is wrapped with fuzzy inference process, to solve the

parameter initialization problem, in this section. Specifically, the thresholds and the

maxp parameters' values should be eliminated. The proposed method relies on two

input linguist variables that are avg and q. The output is initial Dp value. As a FIP-

based, the proposed method will be implemented in sequential process that are,

initialization, fuzzification, rules evaluation, aggregation and defuzzification.

4.2.1 Initialization:

The fuzzy sets is determined as: q: {empty, low, moderate, full}, avg:{low,

moderate, high}, Dp: {zero, low, medium, high}. The membership function can be

formulated for the input and output linguistic variables, using well-established

approaches of trapezoidal (Klir, 1995) or triangular (Negnevitsky, 2009). For the

simplicity triangular membership function was selected. Figure 4.2, Figure 4.3 and 4.4

illustrates the membership function of the linguistic variables according to the

selected triangular function.

The membership function should be trained to determine the boundary and the

slop of each linguistic term.

48

Figure 4.2: Initial Membership Functions for avg

Figure 4.3: Initial Membership Functions for q

Figure 4.4: Initial Membership Functions for Dp

4.2.2 Fuzzification:

low Moderate High

Empty Moderate Full Low

Zero Medium High Low

49

The Fuzzification step converts the input crisp values of the linguist variables avg

and q into equivalent linguistic terms in the fuzzy set.

Consider that the avg membership has been trained and determined as given in

Equation 4.2, subsequently; the following Equation 4.11, Equation 4.12 and Equation

4.14 are representative for low, moderate and high, respectively.

* +

(4.11)

* +

(4.12)

* +

(4.14)

Where, each function is represented by 4-values. The first value is the lower-

bound of the first line of the triangle and the second value is the upper-bound of that

line. The third and fourth values are the upper-bound and lower-bound of the second

line, respectively. Each function is represented by two lines, the low-bound is referred

to "a" and the upper-bound is "b". Subsequently, each of the two lines in each

function is represented, the low-bound is referred to "a1" and the upper-bound is "b1"

for the first line and low-bound is referred to a2 and the upper-bound is b2 for the

second line. For an input crisp value x, if (x<a1 or x>b2) then the output probability

of x to belong to that function is 0.0, while if (a1=b1=0.0) or (x>b1 and x<a2) then

the output probability of x to belong to that function is 1.0, finally if a1<x<b2, then

the output probability of x to belong to that function is x-a/b-a, depends on the line at

which x is occurred.

50

Subsequently, if an input crisp value for avg is given as 0.3, then the output fuzzy

term will be as follows: µavg = (low, 0.3) = 0.3-0.2/0.4-0.2 = 0.5. µavg = (moderate,

0.3) = 0.3-0.2/0.5-0.2 = 0.33. µavg = (high, 0.3) =0.0. This example is represented in

Figure 4.5. In summary, in this step, the fuzzy sets and the fuzzy membership

functions are determined.

Figure 4.5: Mapping Crisp into Linguistic Term

4.2.3 Rules Evaluation:

Rules Evaluation determines the rules utilized to obtain the linguistic terms of the

output variable from the linguistic terms of the input variables. As mentioned, the

linguistic terms of the input are obtained in the previous step. Example of the fuzzy

rules is given in Table 4.1 .More rules are added to be tested as given in Table 4.2.

However, all these rules will be tested and verified, the final set of rules will be

obtained when the experiments are implemented.

low Moderate High

0.3

51

Table 4.1: Examples of the Set of Fuzzy Logic Rules

Rules

IF q is empty AND avg is low THEN Dp is zero

IF q is empty AND avg is moderate THEN Dp is zero

IF q is empty AND avg is high THEN Dp is zero

IF q is low AND avg is low THEN Dp is zero

IF q is low AND avg is high THEN Dp is zero

IF q is low AND avg is moderate THEN Dp is zero

IF q is moderate AND avg is low THEN Dp is low

IF q is moderate AND avg is moderate THEN Dp is low

IF q is moderate AND avg is high THEN Dp is medium

IF q is full AND avg is low THEN Dp is medium

IF q is full AND avg is moderate THEN Dp is high

IF q is full AND avg is high THEN Dp is high

Table 4.2: Complete Set of Fuzzy Logic Rules

Avg Q

52

Empty Low Moderate Full

Low Zero Zero low/ moderate/ low/ moderate/

high

Moderate zero/ low/

moderate

zero/ low/

moderate

low/ moderate/

high

low/ moderate/

high

High zero/ low/

moderate/ high

zero/ low/

moderate/ high

moderate/ high low/ moderate/

high

4.2.4 Aggregation:

The next step is to aggregate the output of all rules, multiple outputs will be

obtained from previous step. Each of these will also have a probability value which is

obtained in this step. This step is start with obtain a probability value for each output

in which the input can be anticipated. The probability value of the output is obtained

as the maximum value of the probability of the input linguistic term, as given in

Equation 4.14.

( ( ) ( ))

(4.14)

4.2.5 Defuzzification:

Defuzzification, the final step generates a crisp value for the output linguistic

term based on its fuzzy set. One of the popular defuzzification techniques is the center

of gravity (COG) method, which aims to find out the point located on the center of the

aggregate fuzzy set for each output linguistic. Given that multiple output linguistic

terms is obtained as multiple rules are anticipated, each with a probability value, the

53

final output crisp result is obtained mostly by the center of gravity, which is

calculated as given in Equation 4.15. Figure 4.11 shows how the multiple linguistic

terms can be assigned for a single output variable.

∑ ( )

∑ ( )

(4.15)

Where the upper term represents the value of the probability of the output

linguistic term x multiplied by all the values covered by this term, e.g.: 0.0, 0.1 and

0.2. The lower term represents the number of values multiplied by the probability.

Figure 4.6: Aggregating Output

4.2.6 Calculate Final Dropping Probability

As a final fuzzy inference process, the output of the fuzzy inference process is

then merged with the delay measure to produce the final dropping probability. The

final dropping probability is calculated as given in Equation 4.16.

( )

(4.16)

zero medium high low

54

Where Dp' is the output of the fuzzy inference process and Dp is the final

dropping probability.

4.3 Conclusion:

Generally, original ERED and its ancestor method, RED, harm delay and it does

not take into consideration the rates of traffic over time. Generally, the queue-size

measure is helpful to avoid congestion and thus, packet loss. However, delay is useful

to avoid long packet queuing time and dropping packets unnecessary especially, when

the load decreases significantly. Hybrid-ERED method is proposed by extended

ERED method to consider the delay measure. Fuzzy hybrid-ERED wraps the

proposed method with fuzzy inference process, to solve the parameter initialization

problem.

55

Chapter 5

Results

This Chapter presents, compares and discusses the results of the Fuzzy Hybrid

ERED method that has been proposed in previous Chapter. The proposed and

compared methods are developed, executed and compared in this chapter. The pre-

evaluation execution of the developed methods that aims to set up some parameters'

values is presented in Section 5.3. The actual results, the comparison and the

discussion are given in Section 5.4. Finally the summary and conclusion are given in

Section 5.5.

5.1 Introduction:

In order to prove the efficiency of the proposed method, an experiment is

conducted. The proposed and compared methods are developed in the same

environment and then executed, finally the results are collected which forms the

outcome of the conducted experiments. In order to highlight the value added by the

proposed method, the results are compared with the results obtained by other

methods, these are: ERED which is the base of the proposed method and the one with

the best results so far as given in (Abbasov and Korukoglu, 2009). RED is the root of

most of the AQM methods.

5.2 Environment:

The simulation of the network process is implemented using one of the well-

known approaches called Discrete Time Queue (Alfa, 2010). The Discrete Time

Queue tracks, measures and evaluates the status of the network and network resources

56

at equal time intervals called slots. At each slot, either a packet arrive event or

departed event or both events may occur. Two subsequent packets arrival without

departure makes two time slots and so on. Several methods have been modeled and

tested using discrete-time queues. The other approach, called continuous model,

measures and evaluates the network performance periodically with equally length

periods. However, this approach does not efficiently address the events of packet

arrival and departures accurately. As the AQM is based on calculating Dp with each

arrival packets (event based), Discrete time queue was chosen to verify the proposed

work.

Experiments are simulated under the discrete time queue in Java. The

programming code is developed in NetBeans 7.5 IDE. The results are obtained in a

machine with Windows 7 operating system, Core i7 2500GHz and 6 GB RAM.

For the discrete time queue, packet arrival and packet departure rates are

established as probability values. Probability for arrival, and similarly for departure,

with 0 values, means no packets will be arrived at any time slot, with probability of

1.0 a packet surely will arrive at each time slot. For a value of 0.5, there is equal

chance for the packet to arrive or not at each time slot. The probability of the packet

departure in the conducted experiments was set to 0.5. And the probability of packet

arrival was set 0.33, 0.5, 0.66 and 0.93. When the arrival probability is below the

departure, no congestion is expected, while congestion or pre-congestion is expected

when the arrival rate is higher than the departure rate. Two Million time slots were

used in the experiments. This value allows for sufficient results, part of these slots are

used as a warm-up before the steady state, with total number of 800000 slots, and the

rest is for experimental measures. The buffer size of 20 packets was used (Baklizi, et

57

al., 2014). minth, maxth, maxp, and weight values are set to 3, 9, 0.1, and 0.002,

respectively, as recommended in RED(Floyd and Jacobson, 1993).

5.3 Parameter settings:

Besides the common parameters introduced above, the proposed method involves

other parameters that are: The delay weight, which is the value given to the delay in

the overall Dp calculation. The weight used in the calculation of the arrival rate, this

weight is used to balance between the previous and current arrival values to come out

with weighted average value for the estimated arrival rate. Note that, in real

implementation, the arrival rate is not fixed; the fixed value for the arrival probability

cannot be used with the proposed method. Thus, the estimated arrival is used as

discussed in Chapter four and the weight used by this estimation is set in this section.

The weight used in the calculation of the departure rate. These parameters are set up

empirically side by side with the fuzzy sets and memberships values used in the Fuzzy

Inference Process (FIP). The weights values attempted and the final weights set ups

are given in Table 5.2. While the final sets and memberships are given in Table 5.3.

The FIP values are set up in trial and error approach.

Table 5.2: Empirical weight settings

Parameter Attempted Final Value

Delay Weight

1.0, 0.9, 0.8, .., 0.1, 0.09,

0.08, ..0.01

0.05

Arrival Weight

1.0, 0.9, 0.8, .., 0.1, 0.09,

0.08, ..0.01

0.2

Departure Weight 1.0, 0.9, 0.8, .., 0.1, 0.09, 0.2

58

0.08, ..0.01

Table 5.3: Empirical FIP settings

Parameter Attempted Final Value

Queue Set

2, 3, 4, 5, 6 Linguistic Variables 4 Linguistic Variable: {zero, low,

moderate, full}

Queue Membership {0,0,0.6,0.7},{0.6,0.7,0.7,0.8},{0.7,0.8,0.8,0.9},{0.8,0.9,0.9,1.0}

Average Queue Length Set

2, 3, 4, 5, 6 Linguistic Variables 3 Linguistic Variable: { low,

moderate, high}

Average Queue Length

Membership

{0,0,0.6,0.7},{0.6,0.8,0.8,0.9},{0.8,0.9,0.9,1.0}

Dropping Probability

2, 3, 4, 5, 6 Linguistic Variables 4 Linguistic Variable: {zero, low,

moderate, high}

Dropping Probability

Membership

{0,0,0.3,0.5},{0.3,0.5,0.5,0.6},{0.5,0.6,0.8,0.9},{0.8,0.9,1.0,1.0}

5.4 Results:

The overall results of the compared and proposed methods are given in Table 5.4.

The results are compared under a heavy congestion, where the arrival probability is

0.93, while the departure probability is 0.5. As given in the table, the number of

packet arrival and number of packet departure for used with the execution of each

method are listed first; the packet arrival is almost equal. As for the packet departure,

not all packets arrived will be departed as these packets are subject to packet loss and

59

packet dropping, thus differs from one method to another. The first 800,000 time slots

out of the total times slots 2,000,000, used in the experiments are for warm-up. The

measures and all factors in warm-up period are simply discarded. The numbers of

packet loss and packet drops are then given and the total number of aggregating these

values are given as packet missed. Note that, if a similar number of packets missed in

all methods, it does not means that there performance are equal, because dropping a

packet is better than losing it, as simple fact in network performance evaluation.

Subsequently, a method that has ability to predict congestion in early stage and drop a

packet is highly better than the one that will finally queued-up and start losing

packets. The aim of showing the overall packet missed is to show that having more

packet dropping is an advantage.

Table 5.4: Performance comparison with arrival rate 0.93

Method

#Packe

t

arrived

#Packet

departe

d

#Packe

t Loss

#Packet

droppe

d

Overal

l

Packet

Missed

Delay

RED

111555

7 598376 323055 195136

517181 33.6367

5

ERED

111580

2 599997 69893 555915

515807 32.2766

5

HybridERED

111580

2 599997 3385 511787 515172

23.2277

FuzzyHybridERE 111559599275 0 516318 516318 21.5663

60

D 2 8

As noted, the proposed HybridERED improves the packet loss by dropping more

packets. This dropping cannot be considered as disadvantage, because finally the

overall missed packets for ERED and HybridERED are equal. Subsequently, the

Hybrid ERED dropped packets in-order to prevent the loss, which will for sure occur

in this case as the arrival rate is higher than departure rate. Moreover, the delay is

enhanced significantly by the proposed HybridERED.

As for the fuzzy based method, it is notably that using the FIP to enhance the

results in packet loss and delay. Subsequently, the proposed method is its parametric

and FIP-based forms perform highly better than ERED.

The results of the compared and proposed methods where the arrival probability

is 0.66, 0.5 and 0.33, while the departure probability is 0.5 are given in Table 5.5,

Table 5.6 and Table 5.7. As noted the number of packet arrivals in the following

tables are decreased as the arrival probability decreased.

As noted, when the arrival probability decreased from 0.93 to 066 (Table 5.4 and

Table 5.5), the performance of the HybridERED is better than of the fuzzy method in

term of the packet dropping. Packet loss and delay are still better in the fuzzy method.

However, dropping packets unnecessary is disadvantage. Thus the best performance

here is the HybridERED method. If the HybridERED method is compared to ERED

and RED method loss less packets and has less overall packet missed. Overall, the

proposed method is its parametric form performs better than ERED, when arrival

probability is equal to 0.66.

61

Table 5.5: Performance comparison with arrival rate 0.66

Method

#Packe

t

arrived

#Packet

departe

d

#Packe

t Loss

#Packet

droppe

d

Overal

l

Packet

Missed

Delay

RED

791269 595965 77569 117828

195297 28.3759

1

ERED 789265 601956 8765 178555 187320 25.1288

HybridERED 788967 602121 3385 183563 186858 20.237

FuzzyHybridERE

D 791685 566053 0 225632

225632

9.55655

When the arrival probability decreased to be 0.5 which is equal to the departure

probability, this case means no congestion occurs. In Table 5.6 it is noted that the

performance of the HybridERED is better than of the fuzzy method in term of the

packet dropping. Packet loss and delay are still better in the fuzzy method. If

compared to ERED and RED, HybridERED method loss less packets and has less

overall packet missed. Overall, the proposed method is its parametric form performs

better than ERED, when arrival probability is equal to 0.5

62

Table 5.6: Performance comparison with arrival rate 0.5

Method

#Packe

t

arrived

#Packet

departe

d

#Packe

t Loss

#Packet

droppe

d

Overal

l

Packet

Missed

Delay

RED

601386 575795 5070 20535

25605 16.5393

7

ERED

602113 581530 2055 18526

20581 17.8527

5

HybridERED

601959 581365 2055 18526

20581 17.6903

5

FuzzyHybridERE

D 597857 560056 0 29787

29787

8.66399

When the arrival probability decreased to be 0.33 which is less than the departure

probability, this case means no congestion occurs and the congestion control method

should not drop packets. In Table 5.7 it is noted that the performance of all the

compared methods are same.

63

Table 5.7: Performance comparison with arrival rate 0.33

Method

#Packe

t

arrived

#Packet

departe

d

#Packe

t Loss

#Packet

droppe

d

Overal

l

Packet

Missed

Delay

RED

395363 395363 0 0

0 3.9295

6

ERED

395363 395363

0 0 0 3.9295

6

HybridERED

395363 395363

0 0 0 3.9295

6

FuzzyHybridERE

D 395551 395363 0 0

0 3.9295

6

64

To give more depth on the performance of the proposed and compared methods,

these methods are compared with different arrival probabilities, with mql, dropping,

PL, missed, delay and throughput as given in Figure 5.1, Figure 5.2, Figure 5.3,

Figure 5.4, Figure 5.5 and Figure 5.6.

As noted, fuzzy method trends to drop packets unnecessarily, unlike the

parametric-based Hybrid method. Subsequently, the parametric-based can be used

when all the performance measures are demanded, while the performance of the fuzzy

is still better in case that delay is a critical factor, because it produces better delay

minimization compares to the other methods. Finally, the throughput which depends

on the number of packet departed and the utilization of queue, is almost equal for all

methods excepts it is less for FIP-based form as it drops more packets.

65

As illustrated in Figure 5.1 the mql for RED, ERED and Parametric-based Hybrid

are similar in arrival probability 0.5. Above this value, RED getting more value for

mql compared to ERED and the Parametric-based Hybrid. This is because RED tends

to avoid packet dropping. Thus, accumulates more packets in buffer when arrival

packets above 0.5. ERED and Parametric-based Hybrid has less mql as both tend to

drop more packets to avoid congestion. ERED mql is higher compared to Parametric-

based Hybrid as it also tries to avoid packet dropping. The FIP Hybrid get the lowest

mql when the arrival probability above and below 0.5. This is because the proposed

takes into consideration the delay factor. From the results, it is clear that the delay of

the fuzzy approach will be the best. However, this may indicates that the dropping

rate of the fuzzy is high.

Figure 5.1: (a) Mean Queue Length Comparison

66

In Figure 5.2 the dropping rate for RED, ERED and Parametric-based Hybrid are

similar in arrival probability 0.5. Above this value, RED drops less packets compared

to ERED and the Parametric-based Hybrid, for the same reason as discussed in the

mql. At arrival probability 0.75, Parametric-based Hybrid starts to drop more packets

compared to ERED. FIP Hybrid method trends to drop more packets in all cases.

However, this dropping is necessary at this case, where the arrival packets exceed the

departed packets. This dropping cannot be considered as disadvantage. Subsequently,

the Hybrid ERED dropped packets in-order to prevent the loss, which will sure occur

in this case as the arrival rate is higher than departure rate.

Figure 5.2: (b) Number of Packet Dropped Comparison

67

Figure 5.3 RED starts losing packets at 0.5 arrival probability with high rate.

ERED starts losing packets with higher probability arrival of 0.75. Parametric-based

and FIP HybridERED decreased packet loss significantly with arrival probability

above 0.5. These results depend on the analysis of previous figures of mql and packet

dropping.

Figure 5.3: (c) Number of Packet Lost

Comparison

68

Figure 5.4 all the compared methods almost missed the same amount of packets.

This is normal, as the arrival packets exceeding the departure packets. The packets

that have no room to be accommodated will be dropped or will be lost.. As the

numbers of packets in all the compared methods are similar, the total of packet missed

in each one is almost similar. However, FIP is tends to miss more packets because

high rate of packet dropping.

Figure 5.4: (d) Number of Packet Missed

Comparison

69

Figure 5.5 the throughput which depends on the number of packet departed and

the utilization of queue is almost equal for all methods except it is less for FIP-based

form as it drops more packets.

Figure 5.5: (e) Throughput Comparison

70

Figure 5.6 the FIP is better for achieving the best delay with arrival probability

above 0.5. Parametric-based HybridERED gives also better delay achievements

compared to RED and ERED.

5.5 Summary:

This chapter presented the performance evaluation of the proposed method in its

parametric and FIP-based forms. The performance evolutions show that the

parametric-based gives a better performance results, for all the performance measures.

Chapter 6

Conclusion and Future Works

Congestion control management was addressed in this thesis. This chapter

summarizes the findings and the future work.

Figure 5.6: (f) Delay Comparison

71

6.1 Conclusion:

In this thesis, a method for congestion control was developed and tested. The

proposed method extends a well known method called ERED in two ways:

First, the proposed method extends ERED by considering the delay issues in

the router buffer. The delay has been estimated using an equation with the

inputs of estimated arrival and estimated departure rates. Subsequently, the

extended method has transferred the original ERED method from being queue-

based into queue-based and load-based (Hybrid) method.

Second, the proposed Hybrid method has been implemented with a Fuzzy

Inference Process (FIP) to solve the problem of parameter initialization.

Experimental has been implemented. The result shows that the proposed method

in its two forms; parametric-based and fuzzy-based have improved the performance of

congestion control in terms of packet dropping, packet loss and delay. The

performance was enhanced by decreasing the packet loss, the queuing delay and

solves the parameter initialization problem.

The loss has been decreased by 10-100% depends on the arrival probability; Delay

has been decreased by 30-60%. The performance of the fuzzy is better than the

parametric-based form and ERED in terms of delay and packet loss.

6.2 Future Work:

Using the estimated delay as one of the thresholds that used to determine the

amount of dropping probability.

72

Proposing a new fuzzy logic-based congestion control method that uses the

delay, average queue length and queue length as input linguistic variables to

control the congestion in the networks.

Estimating the performance of the proposed method in wireless network

environment.

73

REFERENCES

Abbasov, B. and Korukoglu, S. (2009), Effective RED: An Algorithm To Improve

RED's Performance By Reducing Packet Loss Rate. Journal of Network

and Computer Applications, Vol. 32, No.3, pp. 703-709.

Abdel-Jaber, H., Mahafzah, M., Thabtah, F., Woodward, M. (2008), Fuzzy Logic

Controller Of Random Early Detection Based On Average Queue Length

and Packet Loss Rate. In Performance Evaluation of Computer and

Telecommunication Systems.

Agarwal, M., Gupta, R., and Kargaonkar, V. (2004), Link Utilization Based AQM

And Its Performance. In Global Telecommunications Conference, IEEE.

Alfa, A.S. (2010), Queuing Theory for Telecommunications.

Allen, A.O. (1990), Probability, Statistics, And Queuing Theory. Academic Press.

Baklizi, M., Abdel-Jaber, H., Ahmad Abu-Shareha, Mosleh M. Abualhaj, and

Sureswaran Ramadass (2014), Fuzzy Logic Controller Of Gentle Random

Early Detection Based On Average Queue Length And Delay Rate.

International Journal of Fuzzy Systems. Vol. 16, No. 1, pp. 9-19.

Bazaz, Y., kumar, S., and Anand, S. (2013), Congestion Control Mechanism using

Fuzzy Logic. International Journal of Emerging Trends & Technology in

Computer Science (IJETTCS), Vol. 2, No.3.

Changbiao, X. and Fengfeng, L. (2008), A Congestion Control Algorithm Of Fuzzy

Control In Routers, in International Conference on Wireless

Communications, Networking and Mobile Computing.

74

Chengnian Long , Bin Zhao, Xinping Guan and Jun Yang (2005), The Yellow Active

Queue Management Algorithm. Computer Networks, Vol. 47, No.4, pp.

525-550.

Chitra, K. and Padamavathi, D.G. (2010), Adaptive Choke: An Algorithm To

Increase The Fairness In Internet Routers. Int. J. Advanced Networking

and Applications, Vol.01, No.06, pp. 382-386.

Chonggang Wang, Bo Li, Y. Thomas Hou and Kazem Sohraby (2005), A Stable

Rate-Based Algorithm For Active Queue Management. Computer

Communications. Vol.28, No.15, pp. 1731-1740.

Chrysostomou, C., Pitsillides, A., Rossides, L., Polycarpou, M. and Sekercioglu, A.

(2003), Congestion Control In Differentiated Services Networks Using

Fuzzy-RED. Control Engineering Practice Vol.11 No.10, pp. 1153-1170.

Chrysostomou, C., Pitsillides A. and Sekercioglu Y.A. (2003), Fuzzy Explicit

Marking For Congestion Control In Differentiated Services Networks, in

Proceedings of the Eighth IEEE International Symposium on Computers and

Communications.

Deng, X., Sungwon Yi, Kesidis, G. and Das, C.R (2002), Stabilized Virtual Buffer

(SVB)-An Active Queue Management Scheme For Internet Quality-Of-

Service. In Global Telecommunications Conference, GLOBECOM'02, IEEE.

Fakharian, A. and Abbasi, A. (2015), Design of Congestion Controller for TCP

Networks Based On LMI Formulation. Journal of Optimization in Industrial

Engineering, Vol. 8, No.17, pp. 51-56.

75

Feng, W.-c., Dilip D. Kandlur , Debanjan Saha and Kang G. Shin (1999), BLUE: A

New Class of Active Queue Management Algorithms. University of

Michigan.

Floyd, S. (2000), Recommendations On Using The Gentle Variant Of RED.

http://www.aciri.org/floyd/red/gentle.html Access date: 12/10/2015.

Floyd, S. and Jacobson, V. (1993), Random Early Detection Gateways For

Congestion Avoidance. IEEE/ACM Trans. Netw., Vol. 1 No.4, pp. 397-413.

Floyd, S., Gummadi, R., and Shenker, S. (2001), Adaptive RED: An Algorithm For

Increasing The Robustness Of RED's Active Queue Management. AT&T

Center for Internet Research at ICSI.

Hiok, H. and Qiu, B. (2000), Fuzzy Logic Target Utilization And Prediction For

Traffic Control. in Global Telecommunications Conference.

Hollot, C.V., Vishal Misra, Don Towsley and Wei-BoGong (2001), On Designing

Improved Controllers For AQM Routers Supporting TCP Flows.

Twentieth Annual Joint Conference of the IEEE Computer and

Communications Societies. Proceedings.

Ingoley, S.N. and Nashipudi, M. (2012), A Review: Fuzzy Logic In Congestion

Control Of Computer Network, in International Conference in Recent

Trends in Information Technology and Computer Science.

Jamali, S., Alipasandi, N., and Alipasandi, B. (2014), An Improvement Over

Random Early Detection Algorithm: A Self-Tuning Approach. Journal of

Electrical and Computer Engineering Innovations, Vol. 2, No.2 pp. 57-61.

76

Kadhum, M. and Manickam, S. (2014), Control Update Frequency Of Queue

Changes Velocity For Congestion Management In High-Speed Networks.

in Communication, Networks and Satellite (COMNETSAT), IEEE

International Conference.

Klir, G.J. (1995), Fuzzy Logic. Potentials, IEEE. Vol.14, No.4, pp. 10-15.

Koo, J., Byunghun Song , Kwangsue Chung and Hyukjoon Lee (2001), MRED: A

New Approach To Random Early Detection. In Information Networking,

Proceedings. 15th International Conference, IEEE.

Kunniyur, S. and Srikant, R. (2003), End-To-End Congestion Control Schemes:

Utility Functions, Random Losses and ECN Marks. Networking,

IEEE/ACM Transactions, Vol.11, No.05, pp. 689-702.

Kwon, M. and Fahmy, S. (2002), Comparison Of Load-Based And Queue-Based

Active Queue Management Algorithms. The Convergence of Information

Technologies and Communications.

Lim, L.B., L.Guan, A. Grigg, I.W. Phillips, X.G.Wang and I.U. Awan. (2011),

Controlling Mean Queuing Delay Under Multi-Class Bursty and

Correlated Traffic. Journal of Computer and System Sciences, Vol. 77,

No.05, pp. 898-916.

Lin, D. and Morris, R. (1997), Dynamics of Random Early Detection. Vol.27, No.4,

pp. 127-137.

Long, C.-N., Zhao, B., and Guan, X.-P. (2005), SAVQ: Stabilized Adaptive Virtual

Queue Management Algorithm. Communications Letters, IEEE, Vol.9,

No.1, pp.78-80.

77

Loukas, R., Kohler, S., Andreas, P. and Tran-Gia Phuoc (2000), Fuzzy RED:

Congestion Control For TCP/IP Diff-Serv. in Electrotechnical Conference,

MELECON, 10th Mediterranean.

Masoumzadeh, S.S., Taghizadeh, G., Meshgi, K. and Shiry, S. (2009), Deep Blue: A

Fuzzy Q-Learning Enhanced Active Queue Management Scheme, in

International Conference on Adaptive and Intelligent Systems (ICAIS'09),

IEEE. pp. 43-48.

Moghaddam, M.H.Y. (2010), A Fuzzy Active Queue Management Mechanism For

Internet Congestion Control, in Third International Workshop on Advanced

Computational Intelligence (IWACI), IEEE. pp. 203-208.

Morris, R. (2000), Scalable Tcp Congestion Control. in Proceedings of the IEEE

INFOCOM.

Naixue, X., Athanasios V. Vasilakos, Laurence T. Yang, Cheng-Xiang Wang

,Rajgopal Kannan, Chin-Chen Chang and Yi Pan (2010), A Novel Self-

Tuning Feedback Controller For Active Queue Management Supporting

TCP Flows. Inf. Sci., Vol.180, No.11, pp. 2249-2263.

Negnevitsky, M. (2009), Artificial Intelligence: A Guide To Intelligent Systems.

Computer Networks Vol.53, pp. 650–667

Ott, T.J., Lakshman, T.V., and Wong, L. (1999), SRED: Stabilized RED. in

INFOCOM '99.Eighteenth Annual Joint Conference of the IEEE Computer

and Communications Societies.

Pedrycz, W. and Vasilakos, A. (2000), Computational Intelligence In

Telecommunications Networks.

78

Seifaddini, O., A. Abdullah, and a.H. Vosough (2013), RED, GRED, AGRED

Congestion Control Algorithms in Heterogeneous Traffic Types. In

International Conference on Computing and Informatics.

Sun, J. and Zukerman, M. (2007), RAQ: A Robust Active Queue Management

Scheme Based On Rate and Queue Length. Computer Communications,

Vol.30, No.8, pp. 1731-1741.

Sun, J., Zukerman, M., and Palaniswami, M. (2007), Stabilizing RED Using a Fuzzy

Controller, in International Conference on Communications (ICC'07), IEEE,

pp. 266-271.

Wang C., Li B., Sohraby K., and Peng Y. (2003), AFRED: An Adaptive Fuzzy-

Based Control Algorithm For Active Queue Management, in International

Conference on Local Computer Networks (LCN'03), IEEE. pp. 12-20.

Woodward, M.E. (1993), Communication and Computer Networks: Modeling

With Discrete-Time Queues. Wiley-IEEE Computer Society Press.

Wu-chang Feng, Shin, K.G., Kandlur, D.D. and Saha, D. (2002), The BLUE Active

Queue Management Algorithms. Networking, IEEE/ACM Transactions,

Vol.10, No.4, pp. 513-528.

Yaghmaee, M.H. and AminToosi, H. (2003), A Fuzzy Based Active Queue

Management Algorithm. Computer Department, Ferdowsi University of

Mashhad, Faculty of Engineering, pp. 458-462.

Yanping, Q., Hohai Univ, Lin Xiangze and Ji Wei (2007), A Stable Enhanced

Adaptive Virtual Queue Management Algorithm For TCP Networks. In

Control and Automation, International Conference, IEEE.

79

Yongju, X., Lianzhi, W., and Yalin, W. (2009), An Adaptive Target Queue Length

FREM Algorithm, in International Conference on Information Science and

Engineering (ICISE), IEEE. pp. 1634-1638.

Zadeh, L.A., (1965), Fuzzy Sets. Information and Control, Vol.8, No.3, pp. 338-

353.

Zadeh, L.A., (1975), The Concept Of A Linguistic Variable And Its Application

To Approximate Reasoning I, II, III. Information Sciences, Vol. 8, No.3, pp.

199-249.

Zheng, B. and Atiquzzaman, M. (2000), DSRED: An Active Queue Management

Scheme For Next Generation Networks. in Local Computer Networks.

25th Annual IEEE Conference.