ATM Networks Flow Control by Fuzzy Logic

(1)

中華大學碩士論文

題目：模糊邏輯於非同步傳輸模式網路之應用

ATM Networks Flow Control by Fuzzy Logic

系所別：電機工程學系碩士班計算機組

學號姓名： 8701545 謝宗融

指導教授：周智勳博士劉懷仁博士

中華民國八十九年七月

(2)

中文摘要

隨著電腦網路科技在這幾年快速的進步，流量控制在高速網路上扮演著越來越重要的角色，然而在設計高速網路上流量控制的同時，許多人所提出來的方法大都是架構在流量的辨證上，因此難免會有一些先天上的限制，又加上流量具有突發(burst)的特性，導致在設計上存在著許多的不確定性。近幾年來有越來越多的人應用 Soft Computing 來解決控制方面的問題，但是把它運用在網路上的流量控制卻是相當少見，因此在本篇論文中，我們嘗試著應用模糊理論來處理非同步傳輸中流量控制的問題。首先，我們提出一個以模糊理論為理論基礎的呼叫允諾控制(Fuzzy logic based CAC)，接著我們又設計一個具有適應能力的流量控制 (FATFC)，它可以自動的調整頻寬給來源端使用，並且也能符合服務品質的要求。由實驗的結果可以看出來，我們所提出來的呼叫允諾控制 (Fuzzy logic based CAC)和具有適應能力的流量控制 (FATFC) 能夠保證服務品質的要求，並且也能夠有效使用系統的資源。

(3)

ENGLISH ABSTRACT

As the computer and networking technologies advance at a fast pace in recent years, the traffic control plays a more and more important role in high-speed networks. Most traffic controller design on high speed networks are based on the source modeling, which can suffer from some limitations. In addition to the burst characteristic of the traffic, there are many uncertainties in the design process. In recent years, there are more and more applications of Soft Computing in the control problems, but less for the communication network control. Therefore, in this thesis we apply fuzzy logic to the ATM traffic control. Firstly, we propose a fuzzy logic based call admission control (CAC). Secondly, we present a fuzzy adaptive traffic flow controller (FATFC), which adaptively tunes the bandwidth of the sources to achieve the global QoS. Simulation results exhibit that the proposed CAC and FATFC guarantee the usual QoS requirement and achieve high system utilization.

(4)

誌謝

本論文能夠順利的完成，有許多要感謝的人：

首先要感謝的是我的指導教授周智勳老師和劉懷仁老師以及實驗室林道通老師，謝謝他們這兩年來的耐心教導，不但帶領我進入人工智慧與高速網路的領域，更使我體會作學問的真諦，其治學之道，是我學習的對象。

另外還要感謝中華大學師長及同學們的幫助與鼓勵，尤其是實驗室同伴們（同學美玉、群立、宏德、及學弟耀暄、嘉慶、昌甫、

朝暉等人）有了他們的陪伴和協助，讓研究生活增添了幾分歡笑。

最後感謝家人，有了他們的支持才讓我在研究的過程中沒有後顧之憂，僅將此論文獻給我最親愛的家人及所有幫助我的老師及朋友們。

2000/7 于新竹

(5)

In recent years, quality of service (QoS) was well discussed because more and more customers requested QoS guarantees for their applications in high-speed networks. But what is QoS? From user’s point of view, for instance, when broadcasting a video conferenc ing, we will not endure large variation of frame refresh rate. From application point of view, not only the data transmission delay but also the delay variation is declared for the purpose of good transmission quality. So, more and more people pay attention to the QoS guarantee in high-speed networks [16].

Nevertheless, due to the demands for communication and the kinds of multimedia services are increasing (e.g., data, voice, and video), it is difficult to achieve it.

Multimedia applications are known to be the major traffic sources in future high-speed networks. These multimedia services have characteristics such as various QoS, burst traffic, and application-based bandwidth requirement. Hence, high-speed networks are requested to be capable of handling burst traffic and satisfying various QoS (e.g., delay, cell loss ratio) and bandwidth requirements to support the multimedia sources. As is known asynchronous transfer mode (ATM) is one of the transmission technologies that can integrate multimedia services into high-speed

(11)

networks.

To obtain high system performance, ATM must have a good call admission control (CAC) scheme to determine whether to accept a new call or not. A poor CAC may make a wrong decision and failure of the guaranteed QoS, therefore a simple and robust CAC scheme is important especially in high-speed networks. However, because of the burst traffic behavior, it is hard for a standard ATM to achieve high system performance. Nevertheless, in ATM network, Available Bit Rate (ABR) services have the characteristic of dynamic bandwidth assignment, that is, they can transmit data by dynamic rates according to the traffic load. Hence, to enhance the system performance, designing an adaptive bandwidth assignment as well as a good CAC is necessary.

1.2 Methodology and Contributions

Traditional CAC schemes in ATM are based on the source modeling, which can suffer from some fundamental limitations. Generally, it is difficult to acquire complete statistics of the input traffics because new traffic sources are continuously emerging.

Therefore, these schemes are forced to make a decision with incomplete information and the decision process is full of uncertainties.

Several algorithms have been proposed for the flow control [3,9,13,28]. Most of

(12)

them utilize “binary” thresholds of queue length as congestion indication and make a simple approximation of the fair share converge to the actual fair share [13]. Although these algorithms offer good performance and are easy to implement, they have the following disadvantages:

1. “Binary” thresholds are excessively restrictive and have not enough robustness to deal with the high traffic load variation.

2. When high burst VBR traffic sources exist in the network, the performance of these schemes may degrade and the bandwidth allocation will be unfair.

In 1965 Zadeh devised the fuzzy set concept for the vague problems, after that, fuzzy logic has been successfully applied to many problems, and has become more and more popular than before [18]. The fuzzy logic systems provide simple and effective solutions for controlling nonlinear, time- varying, and ill-defined systems.

Complex networks are usually dynamic, there are great uncertainties associated with the input traffics. Fuzzy logic appears to be a promising approach to address key aspects of networks. It has the ability of modeling networks in the continuous mathematics of fuzzy logic rather than with traditiona l crisp value. In this thesis we propose a fuzzy logic based CAC and a fuzzy logic based bandwidth assignment.

These schemes have the abilities for robust to traffic variation, easily to implement, improving the system utilization, and guaranteeing the QoS requirement.

(13)

1.3 Organization

In this thesis we present a CAC and an adaptive rate controller, both are based on fuzzy logic. This thesis is organized as follows: The ATM traffic engineering and the related literatures survey are introduced in Chapter 2. Chapter 3 and Chapter 4 describe the fuzzy logic based CAC and fuzzy logic for ABR flow control, which include many simulation results. A conclusion and some future works are discussed in Chapter 5.

(14)

CHAPTER 2 ATM TRAFFIC ENGINEERING

Before present ing the fuzzy traffic control, we give a review about ATM, ABR flow control and Traffic models that be used in later chapters. A lot of ATM traffic controls have been proposed, some of them are based on Soft Computing, which will also be examined in this chapter.

2.1 Asynchronous Transfer Mode (ATM)

ATM, a packet switching based technique, is different from circuit switching based synchronous transfer mode (STM) [4]. In STM, bandwidth is organized in a periodic frame consisting of time slot (Figure 2.1(a)). Each slot in an STM frame is assigned to a particular call, which is identified by the position of the slot. Therefore, if a station has nothing to transmit when its assigned time slot comes up, that time slot is wasted. So, STM is suitable for fixed-rate services because of its circuit- like nature.

In ATM, however, the cell transmission time is equal to a slot length, and slots are allocated to a call on demand (Figure 2.1(b)). Because slots are allocated to services on demand, ATM can easily accommodate various bit rate services. The fundamental difference between ATM and STM is that the slot assignments are not fixed. In other words, the time slots are assigned in an asynchronous manner. Hence ATM is considered to be more promising because of its efficiency and flexibility.

(15)

(a) STM Multiplexing

(b) ATM Multiplexing

Figure 2.1 STM and ATM principles

In ATM, information flow is organized into fixed-size blocks called “cells” for transmission, each cell consists of 5 bytes for header and 48 bytes for payload data, as shown in Figure 2.2. Cells in ATM network are transmitted by passing through ATM switches, which analyze information in the header to switch the cell into the output interface. The interface connects the switch to the next appropriate switch to its destination. In the following, we will discuss the ATM service architecture and the call admission control (CAC) and available bit rate (ABR) flow control policies.

Periodic Frame

…

Framing 1 n Framing 1 … n Framing 1 … n

Time Slot

Channel 1 Channel 3 Channel 4 Channel 1

Cell Header

Cell

… … …

(16)

Figure 2.2 The ATM Cell

2.1.1 ATM Service Architecture

ATM network, supporting many kinds of services, is capable of handling burst traffic and satisfying various QoS (e.g., delay, cell loss ratio) and bandwidth requirements. According to the ATM Forum specifications [2], ATM network currently offers five service categories: constant bit rate (CBR), real-time variable bit rate (rt-VBR), non-real- time variable bit rate (nrt-VBR), available bit rate (ARB), and unspecified bit rate (UBR). Of those, ABR and UBR are designed for data traffic exhibiting burst unpredictable behavior.

The CBR service category is used by connections that request a static amount of bandwidth that is continuously available during the connection lifetime. CBR service is intended to support real-time applications requiring strictly constrained delay variation (e.g., voice, video, circuit emulation). The rt-VBR service category is also intended for real-time applications that require sternly constrained delay and delay variation, as would be appropriate for voice and video transmissions. The nrt-VBR service category is intended for non-real-time applications that have burst traffic

Header Payload

53 bytes

5 48

(17)

characteristics. In ABR, only lower and upper bounds on the bandwidth of a Virtual Connection (VC) are specified at call set up. ABR service guarantees a cell loss ratio only to those VC’s whose sources dynamically adapt their traffics in accordance with the feedback received from the network. The detailed information about ABR service will be described later. The UBR service category is intended for non-real-time applications. Examples of such applications are traditional computer communication applications such as file transfer and email. The UBR service is simple and does not give traffic source any guarantee.

2.1.2 Call Admission Control (CAC)

ATM network uses connection admission control, traffic shaping, policing, selective discard, packet discard, and explicit feedback to manage the traffic. In this thesis we focus on connection admission control and explicit feedback.

ATM is connection-oriented and provides QoS guarantee, so before transmitting data, building connections with networks is necessary. When an ATM endstation is connected to the ATM network, it is essentially making a contract with the network based on quality of service (QoS) parameters. This contract specifies an envelope that describes the intended traffic flow including peak bandwidth, average sustained bandwidth, and burst size.

CAC is defined as the set of actions taken by the network during the call set up

(18)

phase in order to determine whether the connection request can be accepted or not. A connection request is accepted when sufficient resources are available to the required QoS of new call and to maintain the contracted QoS of existing calls.

To meet the above requirements, it is necessary to evaluate the degree of availability of the current network loading and the impact of adding a new connection.

CAC is not only to guarantee QoS for existing calls but also to achieve high system utilization. So, CAC is an important function in ATM, an appropriate CAC can achieve good system utilization and QoS guarantee.

2.2 ABR Flow Control

ABR service provides better service for data traffic by periodically advising sources about the rates at which they should be transmitted. The ATM Forum has adopted a rate-based feedback control algorithm as the traffic control mechanism for ABR service and also has standardized it [2]. Rate-based schemes use feedback information from the network to control the rate so that the source can emit cells into the network. The ATM Forum has defined special control cells called Resource Management (RM) cell, to convey information to the sources about the status of congestion of the network. The RM cell is periodically generated by the source and turned around by the destination (Figure 2.3).

(19)

Figure 2.3 The RM cell path

At connection set up phase, Peak Cell Rate (PCR), Minimum Cell Rate (MCR), and Nrm for the ABR connection are negotiated. The source regularly sends a forward RM cell every Nrm data cells to the destination. The RM cell contains the Current Cell Rate (CCR), which is set by the source according to its current Allowed Cell Rate (ACR). The RM cell also contains three bits called Direction (DIR) bit, Congestion Indication (CI) bit, No Increase (NI) bit, and a field called the Explicit Cell Rate field (ER). The DIR bit indicates which direction of data flow is associated with the RM cell (forward or backward RM cell). The CI bit is used to indicate whether or not to increase or decrease the transmission rate of the source by referring the negotiation at call set up. The NI bit provides an indication for a source to keep its rate unchanged.

Finally, the ER field is used to limit the ACR of a source to a specific value.

The ATM Forum requires that each switch control congestion by implementing at least one of the following three methods:

Forward

RM Cell Data Cell

Switch Switch

Source Destination

Backward RM Cell

(20)

a) EFCI Marking: The switch may set the EFCI (Explicit Forward Congestion Indication) bit in the header of data cell. In EFCI marking, the switch marks a control bit in the data cells to indicate congestion, and the destinations convey information back to the sources by properly marking RM cell.

b) Relative Rate Marking (RR): The switch may set the CI bit or the NI bit in forward and/or backward RM cell according to the network condition. In RR marking, the switch uses RM cell to provide a binary feedback to source that state the condition of congestion.

c) Explicit Rate Marking: The switch uses RM cell to explicitly provide the allowed rate to the source.

2.3 Traffic Models

ATM networks must support various communications sources such as data, voice, and video each has different traffic characteristics. The purpose of this section is to examine several traffic models proposed for these sources. These traffic models include Bernoulli processes, ON-OFF model, and Markov-modulated Bernoulli processes (MMBP), which are frequently used to model real- world traffic data [1,14,19,23,29]. In the following, we describe that traffic models in details.

(21)

Figure 2.4 The Bernoulli process

a) Bernoulli process: In the Bernoulli process, the probability that next cell slot contains data cell isλ (Figure 2.4).

b) ON-OFF model: In an ON-OFF model, each source is characterized by ON and OFF states, which appear in turn. The transitions from ON to OFF and OFF to ON occur with probabilities β and α, respectively. In a discrete time case, ON and OFF periods are geometrically distributed with means 1/

β and 1/α, corresponding. The cell generation rate during the on period is P and no cell is generated during the off period (Figure 2.5). The ON-OFF source model is assumed as a constant peak rate when cell bursts arrive.

This assumption simplifies the analytical computation, and it is enough for general applications.

Figure 2.5 The ON-OFF model Cell generation

rate P

cell arrival rateλ

(22)

c) Markov-modulated Bernoulli processes (MMBP): The state-transition diagram is shown in Figure 2.6. The MMBP is an (M+1)-state birth-death Markov process. The state-transition diagram for MMBP uses the label mrA to indicate the cell generation rate of a state and uses the labels (M-mr)α and mrβ to denote the transition probabilities from state mrA to state (mr+1)A and from state mrA to state (mr-1)A, respectively. The process in Figure 2.6 can be decomposed into a superposition of simpler processes. It can be thought as a superposition of M independent identical ON-OFF minisources, each has been modeled as in Figure 2.7. Each minisource alternates between ON and OFF states. The transition from ON to OFF state is with rateβ and from OFF to ON state is with rate α. The data rate of a minisource at ON state is A and is zero at OFF state.

Figure 2.6 The Markov-Modulated Bernoulli Processes

(23)

Figure 2.7 The Minisource model

In this thesis, three kinds of input traffic sources include data, voice, and video are applied. In the following, we describe the cell arrival processes of the input traffic sources and their corresponding traffic models.

For data source, we assume it has two states: one is the high bit rate state with rate θh and the other is the low bit rate state with rateθl, both are characterized by Bernoulli process. The data source will alternate between high bit rate and low bit rate states; there is a transition rate θh-l in the high bit rate state and a transition rate θl-h

in the low bit rate state (Figure 2.8). The cell generation process for a voice source connection is modeled by ON-OFF model (Figure 2.9). For video traffic source, the cell generation is assumed have two motion states: one is the low motion state called interframe coding, and the other is the high motion state called intraframe coding. The rate of the high motion state is stated by two parts: the first part is the rate of low motion state and the second part, called difference coding, is the difference between the rates of intraframe coding and interframe coding. The interframe coding and difference coding are both modeled as MMBP. The video sources will alternate between interframe and intraframe; there is a transition rate c in the interframe and a

(24)

transition rate d in the intraframe (see figure 2.10).

Figure 2.8 Traffic model for data sources

Figure 2.9 Traffic model for voice sources

Figure 2.10 Traffic model for video sources

(25)

2.4 Literature Survey

Network flow control is an important study in past ten years. There are a lot of researches study in this topic, among them, Soft Computing approaches, however, are still rare. In this section, we examine some of them.

T. Murase et al. [25] proposes a call admission control based on the estimated cell loss quality for individual burst traffic source. The estimated cell loss quality is expressed in terms of virtual cell loss probability. In [22], the authors propose a fast method to estimate the cell loss ratio. Their computation algorithm requires extra memory and needs only two multiplications but one division to judge whether a call request can be accepted or not. The above two papers [22,25] estimate the cell loss probability from the probability theory view of point, and deal with bufferless case.

Many designs on admission control compute the worst case of theoretical queueing delay to guarantee an absolute delay bound for all packets. Nevertheless, [21]

proposes a measurement-based admission control algorithm (ACA) for predictive service, which allows occasional delay violations. Other approaches can be found in [8,24,26,36].

There are many applications of fuzzy system to control problems [18], and applications to communication network control are recently reported [5-7,10-11,31,34]. Ramalho [31] discusses the use of fuzzy logic approach to CAC

(26)

and designs the fuzzy system by genetic algorithms (GA). GA is used to automatically design and tune the fuzzy system basing online measurements of the cell loss values.

In [11], Cheng and Chang achieve a global fuzzy traffic controller via separately designing of admission control and congestion control. The fuzzy congestion controller outputs a signal to indicate the occurrence of congestion. The fuzzy admission controller judges whether a connection request can be accepted or not according to cell loss ratio, bandwidth requirement for the new connection and congestion situation. B. Bensaou et al. propose a fuzzy-based algorithm to predict the CLR and use it with real-time traffic measurement to design an effective measurement-based CAC [5]. Other fuzzy logic approaches can be found in [6-7,10,35].

As for neural network approaches, in [19], Hiramatsu use a back-propagation neural network to learn the relation between the offered traffic and service quality. In some literatures, combination of fuzzy logic (FL) and artificial neural networks (ANNs) has been investigated to design a CAC [12,15,17,27,32]. An extensive review of FL-based and ANNs-based traffic controls is given in [15]. In [12,17], the CAC schemes combine the linguistic control of fuzzy system and the learning abilities of artificial neural network. The ANN is used in the learning phase to tune the rules and the membership functions of the fuzzy system automatically.

For bandwidth management, Chiussi et al. propose DMRCA_VQ scheme for ABR service. They introduce the concept of virtual queue that is robust to the

(27)

instantaneous behavior of VBR traffic [13]. Aweya et al. present a direct adaptive neural network control strategy for flow control in computer network. The control signals are directly obtained by minimizing a cost function, which is the difference between a reference and the output of the neural model [3]. In [9] the bandwidth allocation scheme is based on the application- layer performance measure called utility.

The goal of their allocation scheme is to provide good applications- layer service to a wide diversity of applications sharing available bandwidth. In [28], they propose an efficient and stable algorithm employing FL for ABR service in ATM network. The fuzzy system estimates the queue length of the ABR source one step ahead according to the delayed source rates.

(28)

CHAPTER 3 Fuzzy Logic Based CAC

The Call Admission Control (CAC) schemes in ATM can be categorized into two types: one is an indirect type, it designs the CAC based on the source modeling; the other is a direct one, which tunes the controller’s parameters by monitoring the traffic behavior. In this chapter, we present a fuzzy logic based direct CAC illustrated in the following sections.

3.1 Main idea

Control schemes based on source modeling are apt to suffer from some fundamental limitations. Generally, it is difficult for a network to acquire complete statistics of input traffics. Therefore, a model-based scheme is forced to make a decision with incomplete information and the decision process is full of uncertainties.

Bearing the aforementioned points in mind, we think that CAC has better basing on actual measurements of the traffic behavior rather than parametric models. The fuzzy based CAC makes use of online estimations of bandwidth and maximum buffer utilization to decide whether accept or reject a new call. It monitors both of the bandwidth and buffer utilizations in order to judge whether there is sufficient bandwidth available for a new call under the QoS requirements of the current existing

(29)

calls. For this purpose, the bandwidth and maximum buffer utilization are stated in details in the following.

3.1.1 Bandwidth requirement

In a switch, traffic management can be separated into statistical and non-statistical operation models. If the summation of peak bit rates (PBR_i) of input traffic sources is smaller than link transmission rate C (see equation 3.1), we call this link operation model non-statistical. On the other hand, if this summation is greater than link transmission rate C (see equation 3.2), we call this link operation model statistical. The advantage of using non-statistical operation model is that it achieves small cell delay. The bandwidth utilization, however, is very low. On the contrary, using of statistical operation model can achieve high system utilization, but result in more complex control algorithm. In this thesis we use statistical operation model for the purpose of high system utilization.

C PBR_i

k i

∑

≤

=1

C PBR_i

k i

∑

>

=1

For CAC, it needs to check whether the system has enough bandwidth for existing calls and the new call. But what is “enough” bandwidth? Assume the system

(3.2) (3.1)

K: total connections

(30)

has k existing connections and is in steady state. The mean bandwidth requested by the k connections is C_k^declared as in equation 3.3, where MBR_i is the mean bit rate for connection i, which is negotiated at call set up phase.

∑

=

= ^k

i

i declared

k MBR

C

1

Next, we can obtain mean bandwidth utilization C_k^used for all traffic sources as in (3.4), where m_i is mean bandwidth utilization for connection i, which can be referred from usage parameter control (UPC) function in ATM network.

∑

=

= ^k

i i used

k m

C

1

Now that those k connections request bandwidthC_k^declared and use bandwidthC_k^used , from average point of view, if we defineC_k^next as the required mean bandwidth of existing connections after the new call, then we claim that:

2

next k used declared k

k

C

C =C +

Hence, the proper bandwidth reserved for the k existing connections is as shown in equation 3.6.

used k declared k next

k C C

C =2* −

There is no adequate information to estimate the exact bandwidth for a new call (3.3)

(3.4)

(3.6) (3.5)

(31)

because we do not know it traffic behavior and the influence of it to current existing calls. So the fuzzy logic based CAC algorithm must estimate the bandwidth by using the declared parameters of the call. Generally, the declared traffic parameters are peak bit rate (PBR), mean bit rate (MBR), and peak bit rate duration (PBRD). These parameters are recommended by the ITU to describe VBR services and are assumed to be enforced by the usage parameter control function [32]. Here we use peak bit rate (PBR) and mean bit rate (MBR) to estimate the bandwidth of the new call as in equation 3.7. According to equation 3.7, it assumes that the new call has a bandwidth requirement ofC_k^new₊₁ . C_k^new₊₁ is computed by reserving a portion p_r of PBR and (1-p_r) of MBR, in which the pr is a ratio with value between 0 and 1.

PBR p

MBR p

C_k^new₊1 =(1− _r)* + _r *

Finally, the fuzzy logic based CAC algorithm estimates the whole required bandwidth of the exiting connections and the new call by equation 3.8.

new k next k

a C C

C = + +1

3.1.2 Maximum buffer utilization

Besides bandwidth, we also consider the maximum buffer utilization Bmax. Cell loss ratio (CLR) is one of QoS, it is caused by buffer overflow occurring as traffic source is burst. So if we manage buffer well, we can control CLR and guarantee QoS.

Although using maximum buffer utilization to calculate the remaining available buffer (3.7)

(3.8)

(32)

is conservative, however, it can deal width the following problems:

1. When burst traffic is happen, it always bring a lot of cells and causes a high CLR.

2. If we use average buffer utilization instead of maximum one, there is no adequate information to estimate CLR.

3.2 Quality of Service

In ATM network, cell segmentation delay and propagation delay are fixed for long-distance transmission. Those delays have nothing to do with traffic characteristic s and CAC scheme. Furthermore, the cell loss and cell delay caused by switch fabrics can be ignored. Therefore, there are two important factors effect the quality of service in ATM network, those are the cell loss and cell delay at the instance of transmitting cells from output buffer to transmission link at every switch in ATM network.

The cell delay can be managed by controlling the buffer size. In other words, we can set the appropriate buffer size according to the maximum cell delay variation. For instance, let the link transmission rate and the output buffer capacity in a switch is C (cells per second) and K (cells). Assume the switch adopts a first in first out (FIFO) manner to transmit the cells stored in output buffer whose maximum available transmission delay must be smaller than D seconds, then C, K, and D must satisfy the following equation.

(33)

K D C ^* ≥

It means that we can control the maximum cell delay by appropriately setting the buffer size. So, in this thesis we only consider the cell loss as the quality of service (QoS) because the cell delay can be ignored.

3.3 Fuzzy logic

Fuzzy sets are well known for their usefulness in representing linguistic information rigorously and their ability to tackle problems that conventional control theory cannot approach successfully. Fuzzy logic has been successfully employed in many real-world automatic control systems including subway systems, auto- focus cameras, washing machines, automobile transmissions, air-conditioners, industrial robots, aerospace, and autonomous robot navigation [18]. Fundamentally, fuzzy set theory provides a robust mathematical framework for dealing with “real-world”

imprecision and non-statistical uncertainty.

A fuzzy logic controller as shown in Figure 3.1, has three functional blocks: a fuzzifier, a defuzzifier, and an inference engine containing a fuzzy rule base. The fuzzifier performs the function of fuzzification that translates the crisp values of each input linguistic variable X into fuzzy linguistic terms. These fuzzy linguistic terms are defined in a term set T(X) and are specified by a set of membership functions. The defuzzifier describes the output linguistic variable by a term set T(Y) characterized by

(3.9)

(34)

a set of membership functions, and adopts a defuzzification strategy to convert the linguistic terms of T(Y) into a crisp value that represents the control action Y. The term set should be determined at an appropriate level of granularity to describe the values of linguistic variables, and the number of terms in a term set is selected under the compromise between the complexity and the controlled performance.

The fuzzy rule base is a knowledge base characterized by a set of linguistic statements in the form of “if - then” rules. It describes a fuzzy logic relationship between crisp input X and crisp output Y. The inference engine contains the decision- marking logic, it acquires the input linguistic terms from the fuzzifier and uses an inference method to obtain the output linguistic terms.

Figure 3.1 Block diagram of a typical fuzzy controller.

Crisp Input (X)

Crisp Output(Y) Fuzzifier

T(X)

Inference engine

Defuzzifier T(Y)

Rule Base

(35)

3.4 The proposed fuzzy logic based CAC

To demonstrate our fuzzy logic based CAC, the system model that contains customer side and network side is shown in Figure 3.2. In customer side, traffic sources have types such as CBR and VBR. When there is a new call, a “call request”

is sent to the Network side, which contains the QoS such as PBR, MBR and CLR required by the new call. When network side receives the “call request”, fuzzy based CAC determines whether to reject or accept the call based on the available bandwidth and maximum buffer utilization of the network and the requested QoS of the new call.

In the following we describe the fuzzy logic based CAC in details.

Figure 3.2 System Model Accept/reject

Network Side Customer Side

Call request

CBR, VBR Output

Fuzzy CAC Buffer

(36)

3.4.1 Definition of membership functions

The fuzzy logic based CAC uses bandwidth requirement Ca and maximum buffer utilization as input linguistic variables. For the input variable Ca, we use the term set T(C_a) = {Enough(E), Not Enough(NE)} to describe whether the bandwidth is enough or not. The terms used for maximum buffer utilization is T(Bmax) = {High(H), Low(L)}

which demonstrates the degree of maximum buffer utilization. In order to provide a more soft control, not only the control actions of “accept” and “reject” but also “weak accept” and “weak reject” are considered. Thus the terms stated for the accept levels are “accept”, “weak accept”, “weak reject”, and “reject”, hence the term set is defined as T(A_level) = {Accept(A), Weak Accept(WA), Weak Reject(WR), Reject(R)}.

The membership functions for terms in the term set are defined by trapezoidal functions, because it is suitable for real-time operation and is simple to implement.

The normalized membership functions defined for the bandwidth requirement, maximum buffer utilization, and accept level are shown in Figure 3.3 to 3.5, respectively. The bandwidth ratio is defined as the bandwidth requirement divided by the available bandwidth and the buffer ratio is defined as the maximum buffer utilization divided by the buffer size. Among these membership functions, the parameters Ene, Ee, El, Eh, EA, EWA, EWR, and ER are set by experience and can be optimally searched by genetic algorithm [30,34] in the future works.

(37)

Figure 3.3 The membership function for bandwidth requirement

Figure 3.4 The membership function for maximum buffer utilization

Figure 3.5 The membership function for the Accept level 1

0

L H

R WR WA A

EA

EWA

EWR

ER

Buffer ratio

Degree of acceptance Degree

Degree

0

Bandwidth ratio

E NE

1

Ee Ene 1

0

1

El Eh

(38)

3.4.2 Construction of fuzzy rules

By intuition and experience, we establish the following four control rules in our fuzzy logic based CAC:

1. IF Bandwidth requirement is Enough AND Maximum buffer utilization is Low THEN Accept level is Accept.

2. IF Bandwidth requirement is Enough AND Maximum buffer utilization is High THEN Accept level is Weak Accept.

3. IF Bandwidth requirement is Not Enough AND Maximum buffer utilization is Low THEN Accept level is Weak Reject.

4. IF Bandwidth requirement is Not Enough AND Maximum buffer utilization is High THEN Accept level is Reject.

These four fuzzy rules are summarized in Table 3.1, in which E, NE, L, H, A, WA, WR, and R stand for the corresponding fuzzy sets. The proposed fuzzy logic based CAC adopts max- min inference method for inference engine and center of gravity method for defuzzification.

(39)

Table 3.1 The applied fuzzy rules

Rule Ca Bmax Alevel

1 E L A

2 E H WA

3 NE L WR

4 NE H R

3.5 Experimental results

Before examining the performance of proposed CAC scheme, many parameters in the simulation environment need to set.

3.5.1 Simulation Environment

In the simulation we assume that the link capacity is 155 Mbps, the buffer size is 200 cells, and is operated in a first in first out (FIFO) manner. There are four types of traffic source applied in this simulation, these are high bit rate data (H_D), low bit rate data (L_D), voice, and video sources. The declared transmission characteristics of these input traffic sources are summarized in Table 3.2. The corresponding parameters of these traffic models are shown in Table 3.3 to Table 3.5. For example, for high bit rate data source, it is assumed that PBR=2480 kbps, MBR=1446.66 kbps, holding time is 18 seconds, which give θh =1.6* 10^-2, θl= 6*10^-3, θh-l =8*10^-5,θl-h=4*10^-5. These parameters can be obtained by using some equations or program simulations.

(40)

For example, for Bernoulli process, we assume that the link capacity is C (Mbps), then PBR (Mbps), MBR (Mbps),θh, θl, θh-l, and θl-h must satisfy the following equations, where LOW is the cell generation rate in low bit rate state.

C h

PBR = *θ

C l

LOW = *θ

h l l h

l h h

l l h

h

l LOW

PBR MBR

−

+ +

= +

θ θ

θ

θ *

*

Therefore if we know the PBR and MBR of a traffic source then we can obtain the corresponding parameters of the traffic model.

There will be a new call arrived for about every 1.25*10⁴cell time. The cell loss ratio defined for QoS in the simulation is set as CLR=10^-5 [11,12].

Table 3.2 Traffic sources description

Type Traffic model PBR (kbps) MBR (kbps) Holding time

H_D Bernoulli 2480 1446.66 18 s

L_D Bernoulli 496 330.6 18 s

Voice ON-OFF 64 25.7 3 min

Video MMBP 5100 1168.81 15 min

(3.10) (3.11) (3.12)

(41)

Table 3.3 The parameters of data source

Type θh θl θh-l θl-h

H_D 1.6*10^-2 6*10^-3 8*10^-5 4*10^-5

L_D 3.2*10^-3 1.6*10^-3 8*10^-5 4*10^-5

Table 3.4 The parameters of voice source

Type P (kbps) α β

Voice 64 4.869*10^-3 2*10^-6

Table 3.5 The parameters of video source

Coding M A (kbps) α β

Interframe coding 20 207 3.77*10^-6 5.65*10^-6

Different coding 20 48 2.83*10^-5 2.83*10^-5

3.5.2 Simulation results and Discussion

By experience, we set the parameters of the membership functions as in Table 3.6 to Table 3.8. The parameter p_r in equation 3.7 is set as 0.8. The threshold is defined as 0.5, which means that if the defuzzified value is greater than 0.5 then the new call is accepted else it is rejected.

(42)

In the simulation, the CLR is defined as the total loss cells divided by the total arrived cells during the testing time. The simulation results for CLR and bandwidth utilization are shown in Figure 3.6 to Figure 3.8. From Figure 3.6, we can find that the CLR is always less then 10^-5 and guarantees the QoS. Figure 3.8 displays that the system utilization (refer to capacity utilization) reaches almost 89﹪at steady state.

The fuzzy logic based CAC we proposed has the following advantages

1. It only use four rules, thus it is easy to implement and is suitable for high-speed network.

2. Generally, the declared traffic parameters are peak bit rate (PBR), mean bit rate (MBR), and peak bit rate duration (PBRD), but the fuzzy logic based CAC we proposed only need the peak bit rate (PBR) and the mean bit rate (MBR).

3. In the simulation, traffic models include ON-OFF, Bernoulli, and MMBP models are considered, which correspond to many types of traffic sources such as voice, data, and video. Hence the controller is applicable to multimedia traffic data.

4. The simulation results exhibit that the proposed CAC guarantees the usual QoS requirement and achieves high system utilization.

Table 3.6 The parameters of the membership function for Ca

Linguistic variable Ee Ene

Ca 0.8 0.95

(43)

Table 3.7 The parameters of the membership function for Bmax

Linguistic variable El Eh

Bmax 0.15 0.4

Table 3.8 The parameters of the membership function for Alevel

Linguistic Variable EA EWA EWR ER

A_level 1 0.75 0.25 0

4 0 0 6 0 0 8 0 0 1 0 0 0 1 2 0 0 1 4 0 0

0 0 . 2 0 . 4 0 . 6 0 . 8 1

T i m e ( u n i t : 1 0⁶ c e l l s )

Cell loss ratio

↑ Q o S r e q u i r e m e n t 1 x 1 0^{- 5}

↓ F u z z y b a s e d C A C

Figure 3.6 The cell loss ratio

(44)

4 0 0 6 0 0 8 0 0 1 0 0 0 1 2 0 0 1 4 0 0 0

0 . 1 0 . 2 0 . 3 0 . 4 0 . 5 0 . 6 0 . 7 0 . 8 0 . 9 1

H i g h D a t a L o w D a t a V i d e o

V o i c e

T i m e ( u n i t : 1 0⁶ c e l l s )

Capacity Utilization

Figure 3.7 System utilization for each type of traffic sources

4 0 0 6 0 0 8 0 0 1 0 0 0 1 2 0 0 1 4 0 0

0 . 5 0 . 5 5 0 . 6 0 . 6 5 0 . 7 0 . 7 5 0 . 8 0 . 8 5 0 . 9 0 . 9 5 1

T i m e ( u n i t : 1 0⁶ c e l l s )

Capacity Utilization

Figure 3.8 System utilization

(45)

CHAPTER 4 Fuzzy Logic for ABR Flow Control

In the previous chapter, we have brought up a fuzzy based CAC scheme, it only consider CBR and VBR services in ATM network, in this chapter the ABR service is also taken into account.

4.1 Main idea

In this chapter, we consider CBR, VBR, and ABR services in ATM network at the same time. Among these services, the transmission rates of CBR and VBR are not controlled because their transmission rate is according to the negotiation (such as PBR and MBR … etc.) at call set up phase. For ABR sources, however, only lower and upper bounds on the bandwidth are specified at call set up, so that we can dynamically control their traffic rates in accordance with the feedback information from the network. In general, the transmission rate for VBR service is usually not equal to PBR, so our scheme will assign ABR services the bandwidth that CBR and VBR have declared but not used.

As stated in chapter 3, it would be hard to cover all the traffic situations as new traffic sources are continuously emerging. Therefore, we believe that basing on actual measurements of the traffic dynamics is a better way to know the behavior of traffic

(46)

sources. So, the aim of our fuzzy logic controller is how to use the measured data to optimally allocate the bandwidth to all the traffic sources. For instance, assume the system has k existing connections (include CBR and VBR but not ABR services) and the link transmission rate is C. Then, the global peak bit rate (GPBR) and global mean bit rate (GMBR) declared by these connections are shown in equation 4.1 and equation 4.2:

∑

⁼

=

= ⁱ ^k

i

PBRi

GPBR

1

,

∑

⁼

=

= ⁱ ^k

i

MBRi

GMBR

1

,

where k is number of connections, PBRi and MBRi is Peak Bit Rate and Mean Bit Rate for connections i.

If we reserve bandwidth GPBR for these connections then the system utilization is very low because of the over allocating of bandwidth to these connections. On the contrary, if we reserve bandwidth GMBR for those connections then the CLR will be high because the traffic source usually has burst characteristics. Hence, neither GPBR nor GMBR is benefit for those connections. Nevertheless there should exist an optimal trade off between the allocations of GPBR and GMBR that guarantees the QoS and achieves high system utilization (see Figure 4.1). The optimal point is dynamic changed according the behaviors of these connections. So the aim of our fuzzy logic is to find out the optimal point by measured data.

(4.1) (4.2)

(47)

Figure 4.1 Optimal point between GPBR and GMBR

A simple flow procedure for ABR services bandwidth allocation is shown in Figure 4.2. First the fuzzy adaptive traffic flow controller (FATFC) calculates the reserved bandwidth for CBR and VBR services. The Max-Min allocation process allocates the bandwidth for ABR services according to the fairness of Max-Min criterion. In the following, we describe the FATFC and Max-Min allocation in details.

FATFC

Is RM cell arriving No

Start

Yes Max-Min allocation

Figure 4.2 Simple bandwidth allocation for ABR services GMBR

Optimal point

Search point (p) GPBR

(48)

4.2 The Fuzzy Adaptive Traffic Flow Controller (FATFC)

In this chapter, we proposed a fuzzy traffic controller, which supervises the maximum buffer utilization (a situation of buffer utilization) and the Variation of buffer utilization so as to tune the search point (p) as closed to the optimal point as possible. Then, we can obtain the bandwidth of CABR for ABR sources from equation 4.3 in the following:

^CABR =^C−

[ (

1−^p

)

*^MBR+ ^p*^PBR

]

Figurer 4.3 is our system block that contains customer side and network side.

Customer side contains traffic sources such as CBR, VBR and ABR. For CBR and VBR sources, the PBR, MBR, and QoS requirements are negotiated with network side. For ABR traffic sources, however, only the lower and upper bounds on bandwidth are declared. The FATFC will reserve appropriate bandwidth for CBR and VBR traffic sources, and allocate the remaining bandwidth to ABR sources by Max-Min allocation algorithm described in the following sections.

(4.3)

(49)

Figure 4.3 System block of FATFC

4.2.1 Definition of membership functions

The FATFC uses maximum buffer utilization and the Variation of buffer utilization as input linguistic variables. As in chapter 3, we use the terms “low” and

“high” to describe the maximum buffer utilization, and thus the term set is defined as T(Bmax) = {Low(L), High(H)}. The term set used to describe the Variation of buffer utilization is T(Varbuffer) = {Positive Small(PS), Positive Large(PL), Negative Small(NS), Negative Large(NL) }. The term set of the output variable Actionadjust is defined as T(Actionadjust) = {Zero(Z), Positive Small(PS), Positive Large(PL), Negative Small(NS), Negative Large(NL) }.

Network Side Customer Side

Output Transmission Rate

Control

CBR, VBR

FATFC

Buffer ABR

(50)

The membership functions corresponding to the fuzzy sets are shown in Figure 4.4 to Figure 4.6. Among the membership functions, all the variables are normalized.

For example, Bmax is normalized as

size buffer

Bmax

, Varbuffer is normalized as

size buffer

buffer

Var , and Adjust level is set between 0 and 1. The tuning of the ten parameters

in the membership functions will be examined in the simulation experiment.

Figure 4.4 Membership function for the input variable Bmax

Figure 4.5 Membership function for the input variable Varbuffer

NL NS PL

PS 1

El Eh

0 1

H L

0 1

-1

1

Enl Epl

Degree

(51)

Figure 4.6 Membership function for the output variable Actionadjust

4.2.2 Construction of fuzzy rules

Recall that our aim is to assign the bandwidth for each ABR traffic source as closed to its PBR as possible under the QoS of all VBR and CBR sources. Hence, we establish the following eight inference rules in our FATFC.

1. IF Maximum buffer utilization is High AND Variation of buffer utilization is Positive Small THEN Adjust the bandwidth of VBR sources with Positive Small degree.

2. IF Maximum buffer utilization is Low AND Variation of buffer utilization is Positive Small THEN Adjust the bandwidth of VBR sources with Negative Small degree.

Z PS PL

NL NS

Ens

Adjust level

Enl Ez Eps Epl 1

-1

(52)

3. IF Maximum buffer utilization is High AND Variation of buffer utilization is Positive Large THEN Adjust the bandwidth of VBR sources with Positive Large degree.

4. IF Maximum buffer utilization is Low AND Variation of buffer utilization is Positive Large THEN Adjust the bandwidth of VBR sources with Negative Small degree.

5. IF Maximum buffer utilization is High AND Variation of buffer utilization is Negative Small THEN Adjust the bandwidth of VBR sources with Positive Small degree.

6. IF Maximum buffer utilization is Low AND Variation of buffer utilization is Negative Small THEN Adjust the bandwidth of VBR sources with Negative Large degree.

7. IF Maximum buffer utilization is High AND Variation of buffer utilization is Negative Large THEN Adjust the bandwidth of VBR sources with Zero degree.

8. IF Maximum buffer utilization is Low AND Variation of buffer utilization is Negative Large THEN Adjust the bandwidth of VBR sources with Negative Large degree.

(53)

Table 4.1 Fuzzy rules of the FATFC Rule Bmax Varbuffer Actionadjust

1 H PS PS

2 L PS NS

3 H PL PL

4 L PL NS

5 H NS PS

6 L NS NL

7 H NL Z

8 L NL NL

For example, rule 1 indicates that when maximum buffer utilization is high and buffer utilization is increasing slowly then we increase small part of bandwidth for VBR services. These eight rules are listed in Table 4.1 for clearness

4.2.3 Max-Min allocation algorithm

We obtain the bandwidth CABR for ABR sources by FATFC and equation 4.3. in the following, we allocate the bandwidth CABR to these ABR services according to the fairness criterion. According to the ATM Forum, the fairness of Max-Min criterion is that the available bandwidth B is equally shared among the n connections, which are bottlenecked at the link. Hence, for fairness, let B(i) be the allocated bandwidth for connection i (equation 4.4), n is the number of active connections bottlenecked on this

(54)

link, and B is the available bandwidth to be shared by these connections, then

n

i B

B⁽ ⁾= ,

B can be obtained by the following equation:

U C

B = _ABR − ,

in which U is the sum of bandwidths of the ABR connections bottlenecked elsewhere.

4.3 Experimental results

In this section, we present two experiments. In the first, the numbers of VBR and ABR services in network are fixed, we control the bandwidth of ABR services. In the second, in addition to the existing VBR and ABR services, new traffic sources are dynamically added, so that it is more practical for real world applications.

4.3.1 The Simulation Environment

Consider the network topology in Figure 4.7. This network topology is an upstream-bottleneck model and is suitable for testing max- min fairness [28]. It consists of three switches and five connection paths (the pair of source and

(4.5) (4.4)

(55)

destination). In the simulations, both VBR and ABR traffics are presented in the network because we want to observe the interaction between them. The switches are assumed to be output-buffered and non-blocking [33].

The transmission rates are equal to 155 Mbps for all links, so that the time for transmitting one cell is about 2.75 μs. The propagation delay between source and switch, switch and switch, and switch and source for each link is 100 cell time. Buffer sizes in the three switches are all limited to 250 cells.

Figure 4.7 Network topology SW

1

SW 2

SW 3 S1

S2 S3

S4 S5

D2

D3

D1

D4

D5 Path 1: S1-SW1-SW2-SW3-D1

Path 2: S2-SW1-SW2-D2 Path 3: S3-SW1-SW2-D3 Path 4: S4-SW1-SW2-SW3-D5 Path 5: S5-SW2-SW3-D5

: First link : Second link

(56)

4.3.2 Simulation 1

In this simulation, there are 6 virtual connections (VC1-VC6) in the network, which include 2 VBR and 4 ABR traffic sources. VC1 (Path 1) and VC2 (Path 1) are the VBR connections and the others are ABR connections (VC3(Path 2), VC4(Path 3), VC5(Path 4), and VC6(Path 5)) are shown in Table 4.2. We model each VBR VC with Bernoulli process (see section 2.2 for details). The declared traffic rates of these VBR traffic sources are summarize in Table 4.3 and the corresponding parameters are shown in Table 4.4. We set the initial cell rates of all the ABR connections as 15 Mbps, the maximum cell rate to be 155 Mbps, the minimum cell rate to be 0 Mbps, and the value of Nrm is equal to 32. The ABR sources are always assumed to transmit at the maximum allowed transmission rate. Therefore we increases the cell rates for ABR services when system is in congestion-free state, and on the contrary, we reduce the cell rate when system is in congestion state.

Table 4.2 The paths of traffic sources

Connection Service type Path First link Second link

VC1 VBR Path1 ● ●

VC2 VBR Path1 ● ●

VC3 ABR Path2 ●

VC4 ABR Path3 ●

VC5 ABR Path4 ● ●

VC6 ABR Path5 ●

(57)

Table 4.3 Traffic sources description

Connection Traffic model PBR (Mbps) MBR (Mbps)

VC1 Bernoulli 30 20

VC2 Bernoulli 30 23.33

Table 4.4 The parameters of the traffic sources

Connection θh θl θh-l θl-h

VC1 1.935*10^-1 9.67*10^-2 6.67*10^-3 3.34*10^-3 VC2 1.935*10^-1 6.45*10^-2 3.34*10^-3 6.67*10^-3

By experience, we set the parameters of the membership functions for FATFC as in Table 4.5 to Table 4.7. The proposed FATFC adopts max- min inference method for inference engine and center of gravity method for defuzzification.

Table 4.5 The parameters of the membership function for Bmax

Linguistic variable El Eh

B_max 0.2 0.4

(58)

Table 4.6 The parameters of membership function for Varbuffer

Table 4.7 The parameters of the membership function for Actionadjust

Linguistic variable Enl Ens Ez Eps Epl

Actionadjust -1/8 -1/16 0 1/8 1/4

The simulation results are shown in Figure 4.8 to Figure 4.10. Figure 4.8 shows the cell rates of VC3, VC4, and VC5. We discover that the three ABR connections are bottlenecked at SW1 and share the available bandwidth. From the network topology, the first link is shared by VC1 to VC5 and the second link is shared by VC1, VC2, VC5, and VC6. However, because VC1 and VC2 are both VBR connections, so the available bandwidths in SW1 and SW2 are almost equal. In the first link, the available bandwidth is shared by VC3, VC4, and VC5, while in the second link, the available bandwidth is shared by VC5 and VC6. Besides, because the throughput of VC5 is limited to 1/3 of the available bandwidth in the first link, VC6 obtains 2/3 of the available bandwidth in the second link. The cell rates for VC5 and VC6 are shown in Figure 4.9. Figure 4.10 shows the system utilizations for SW1 and SW2. We find that both switches achieve the system utilizations up to almost 95% and the CLR is equal to 0.

Linguistic variable Enl Epl

Varbuffer -0.3 0.3

(59)

0 5 0 1 0 0 1 5 0 2 0 0 2 5 0 3 0 0 3 5 0 4 0 0 4 5 0 5 0 0 3 1 0

3 2 0 3 3 0 3 4 0 3 5 0 3 6 0 3 7 0 3 8 0

T i m e ( u n i t : 2 0 0 c e l l s )

Rate (0.1Mbps)

Figure 4.8 Cell rates for VC3, VC4, and VC5

0 1 0 0 2 0 0 3 0 0 4 0 0 5 0 0

1 0 2 0 3 0 4 0 5 0 6 0 7 0 8 0

T i m e ( u n i t : 2 0 0 c e l l s )

Rate (Mbps)

↑ V C 5

↑ V C 6

Figure 4.9 Cell rates for VC5 and VC6

ATM Networks Flow Control by Fuzzy Logic

中 華 大 學 碩 士 論 文

題目：模糊邏輯於非同步傳輸模式網路之應用