A study on QOS routing for providing guaranteed services under multi class traffic loads

MPLS Multi-Protocol Label Switching MAQR Multi-class Adaptive QoS Routing MPCPO Multi-Path-Constrained Path-Optimization routing MBGP Multi-protocol BGP NGN Next Generation Network OSPF

A STUDY ON QOS ROUTING FOR PROVIDING GUARANTEED SERVICES UNDER MULTI-CLASS

TRAFFIC LOADS

JIA LEI

NATIONAL UNIVERSITY OF SINGAPORE

2005

A STUDY ON QOS ROUTING FOR PROVIDING GUARANTEED SERVICES UNDER MULTI-CLASS

TRAFFIC LOADS

JIA LEI

(B.Eng., NANJING UNIVERSITY OF SCIENCE & TECHNOLOGY)

A THESIS SUBMITTED FOR THE DEGREE OF MASTER OF ENGINEERING DEPARTMENT OF ELECTRICAL & COMPUTER

ENGINEERING NATIONAL UNIVERSITY OF SINGAPORE

2005

Acknowledgements

Many thanks are given to my supervisor, Dr Yin Qinghe, who sparks this research and leads me into the networking research world His valuable guidance and kindly help in all aspects of my work and life here was really appreciated

Special thanks are given to my wife, who supports me with her love and many encouragements, which accompany me on this long journey In addition, many of my lab mates and friends had contribution to this work Especially I would like to thank Huang Qijie and Li Dan for their valuable suggestions and assistance Also, I want to thank Li Jianfeng, Wu Zheng, Xiao Haiming and Er Inn Inn for those helpful discussions

Thanks also go to National University of Singapore and Institute of Infocomm Research for providing me this great opportunity and all the facilities to carry out my research

This thesis is dedicated to my parents It is their love and dedication that made everything I have possible

Contents

Acknowledgement i

Contents ii

List of Figures v

List of Tables vii

List of Abbreviations viii

Summary ix

Chapter 1 Introduction 1

1.1 A Brief Introduction to Next Generation Network 1

1.2 Research Motivation 3

1.3 Assumptions 5

1.4 Thesis Contribution………….………6

1.5 Thesis Organization 8

Chapter 2 Preliminary and Related Work 9

2.1 Basic Concepts 9

2.1.1 Weighted Graph Model 9

2.1.2 QoS Metrics and Constraints 10

2.2 QoS Routing Problems 13

2.2.1 Routing Techniques 14

2.2.2 QoS Unicast Routing 16

2.2.3 Multi-class QoS Routing 25

2.2.4 QoS Routing and Other QoS-provision Components 27

2.3 Token Bucket Traffic Model 29

2.4 Weighted Fair Queuing 30

Chapter 3 A Multi-class Adaptive QoS Routing Algorithm 34

3.1 Network Model 35

3.2 ULARAC Path Algorithm for DBCLC Path Problem 37

3.2.1 DBCLC Path Problem 37

3.2.2 ULARAC Cost Function 40

3.2.3 ULARAC Path Algorithm 42

3.2.4 Loop-free Property and Complexity Analysis 45

3.3 A Multi-class Adaptive QoS Routing Algorithm 46

3.3.1 Routing Principle 47

3.3.2 VRB Calculating Method 48

3.3.3 A Multi-class Adaptive QoS Routing (MAQR) algorithm 52

Chapter 4 Multi-class Adaptive QoS Routing Algorithm: Simulation Study 55 4.1 Network Topology 56

4.2 Traffic Load 58

4.3 Performance Metrics 59

4.4 Performance Evaluation 60

4.4.1 Impact of Best-effort Traffic Distribution 61

4.4.1.1 Uneven Distribution of Best-effort Traffic 61

4.4.1.2 Even Distribution of Best-effort Traffic 67

4.4.2 Impact of Delay Constraint 68

4.4.3 Impact of the Characteristics of the Links 73

4.5 Summary 74

Chapter 5 Conclusions 76

Bibliography 79

List of Figures

Figure 2.1: Path selection from s to d with (cost, delay) values indicated

on each link and routing objective is cost minimization and

path delay no more than 40 ms…… .17

Figure 2.2: Token bucket traffic model 30

Figure 2.3: GPS server model 31

Figure 3.1: ULARAC path algorithm .44

Figure 3.2: A constructed loop path 45

Figure 3.3: Representation of calculating credit function 52

Figure 4.1: Mesh (3*3) topology .57

Figure 4.2: vBNS topology 57

Figure 4.3: Average loss rate of best-effort traffic as a function of QoS traffic load: flow 1, Delay bound=0.025s, vBNS topology .64

Figure 4.4: Average loss rate of best-effort traffic as a function of QoS traffic load: flow 2, Delay bound=0.025s, vBNS topology .65

Figure 4.5: Average loss rate of best-effort traffic as a function of QoS traffic load: flow 2, Delay bound=0.048s, cluster topology .65

Figure 4.6: Average loss rate of best-effort traffic as a function of QoS traffic load: flow 1, Delay bound=0.048s, cluster topology .66

Figure 4.7: Average loss rate of best-effort traffic as a function of QoS traffic load: even distribution, cluster topology 66

Figure 4.8: Average loss rate of best-effort traffic as a function of QoS traffic load: flow 1, Delay bound=0.04s, vBNS topology .71

Figure 4.9: Average loss rate of best-effort traffic as a function of QoS

traffic load: flow 2, Delay bound=0.04s, vBNS topology .71

Figure 4.10:Average loss rate of best-effort traffic as a function of QoS

traffic load: flow 1, Delay bound=0.02s, vBNS topology .72 Figure 4.11:Average loss rate of best-effort traffic as a function of QoS

traffic load: flow 2, Delay bound=0.02s, vBNS topology .72

List of Tables

Table 4.1: Blocking rate for QoS connections: Delay bound=0.025s,

vBNS topology 64 Table 4.2: Performance gains for best-effort traffic under different

delay requirementin vBNS topology .70

DBCLC Delay and Bandwidth Constrained Least Cost

FRED Flow Random Early Detection

FEC Forwarding Equivalence Class

GPS Generalized Processor Sharing

IS-IS Intermediate System-Intermediate System

LARAC Lagrange Relaxation based Aggregated Cost

LBAP Linear Bounded Arrival Processes

MCI Microwave Communications Inc

MPLS Multi-Protocol Label Switching

MAQR Multi-class Adaptive QoS Routing

MPCPO Multi-Path-Constrained Path-Optimization routing

MBGP Multi-protocol BGP

NGN Next Generation Network

OSPF Open Shortest Path First

PGPS Packet-by-packet Generalized Processor Sharing

PSTN Public Switched Telephone Network

QoS Quality of Services

RED Random Early Detection

RIP Routing Information Protocol

SLS Service Level Specification

SFQ Stochastic Fair Queuing

ULARAC Utilization based and Lagrange Relaxation based Aggregated Cost vBNS very high-speed Backbone Network Services

VRB Virtual Residual Bandwidth

WFQ Weighted Fair Queuing

Summary

Much is certain about the Next Generation Network (NGN) nowadays NGN will be a service-driven network to support proliferate services While a large number of Quality of Services (QoS) routing algorithms have been proposed, many of them do not take multi-class traffic loads into account Although several studies addressed this issue, most of them did not include delay-sensitive traffic Therefore, QoS routing under multi-class traffic loads is studied in this thesis More precisely, we study the routing problems based on bandwidth-delay-guaranteed QoS traffic and traditional best-effort traffic The objective of our algorithms is to provide the end-to-end QoS guarantee for QoS traffic and achieve high resource utilization for whole network

First, we present an applicable QoS routing algorithm to solve the Delay and Bandwidth Constrained Least Cost (DBCLC) path problem We define a link’s cost to represent the link utilization ratio and construct an aggregated cost in order for the adoption of Lagrange Relaxation technique Then, we propose the Utilization and Lagrange Relaxation based Aggregated Cost path selection algorithm Next, we extend the algorithm to a multi-class adaptive routing algorithm addressing dynamic resource sharing between two classes of traffic loads The proposed algorithm is simulated using NS2 The simulation results demonstrate that the routing algorithm can lead to certain performance improvements for best-effort traffic in some

scenarios However, the efficiency of the algorithm is undermined by more stringent delay constraints and unique properties of the links in the network

Chapter 1

Introduction

In this chapter, a brief introduction to Next Generation Network is first presented The research motivation of this thesis is then described An overview of the thesis is given

in the last section

1.1 A Brief Introduction to Next Generation Network

The future vision of “information and communication anytime, anywhere and in any form” is starting to come into focus as many key players are positioning themselves for radical transformation of their network and service infrastructures to Next Generation Network (NGN) It has become increasingly clear that a prerequisite for realization of such a vision is the convergence of the current multiple networks into a unified, multi-services, data centric network [25] Therefore, NGN will be such a network that draws on the existing and emerging technologies to deliver multiple

services at different qualities and costs It will be such an open network that is accessible to the third-party developers and various service providers In [25], four classes of services that NGN will most possibly offer are given

1 The first and perhaps most promising class of services will be interactive

communication services, which include real-time services such as multiple parties

interacting in real time and using multi-media, and non-real-time services such as

messaging of multi-media content between multiple receivers Real-time

communication services have strict Quality of Services (QoS) requirements for

the underlying transport network

2 The second major class of services will be information services These services

may be thought of as the evolution of today’s Internet services such as browsing, information retrieval, online directories, e-commerce, and advertising Although these services will continue to be data-oriented, the evolution of them will involve major improvements in enhancing reliability, providing billing, QoS,

policy enforcement and ease of use

3 The third class of services that NGN will need to offer and/or enable is delivery

of content Typically, such content delivery is for purposes of entertainment and education For example, these services can be video on demand, music on demand, distance learning and multi-player network games Obviously, the various flavors of these services will pose technical challenges from the point of

view of scalability

4 Finally, last class of services will be management of other services Although they may not have revenue-generating potential, they are undoubtedly necessary

to other services These services include subscription service, customer network

management and customer service management

In one word, NGN will be a converged broadband network capable of supporting proliferate services that a service provider can offer or facilitate to its customers Thus, the efficient and QoS-assured service provision is absolutely critical The requirement of supporting broadband multimedia services means that NGN has to be optimized for QoS-enabled packet services, which is in sharp contrast to the best-effort, non-QoS-enabled data services of the current Internet and the narrowband voice-centric services of Public Switched Telephone Network (PSTN)

1.2 Research Motivation

As NGN is a service-driven network and the underlying transport network in NGN is expected to support various new QoS-enabled services such as video conferencing and multi-player network games, the network resources must be managed effectively

to ensure end-to-end QoS while at the same time sustaining high network throughput New resource management mechanisms being proposed include scheduling disciplines, resource reservation protocols, admission control and QoS routing QoS routing has drawn huge interests in the communication community and many routing algorithms hereby have been proposed to take the traffic and QoS characteristics into consideration while selecting a path However, most of existing QoS routing algorithms focus on improving the performance of individual service class while

assuming the service class being studied is sole one in the network For example, routing strategies for the traffic with certain bandwidth guarantee take minimizing call blocking rate as the unique objective while not considering the performance of other classes of traffic loads, especially the traditional best-effort traffic in the network As far as we know, not much effort was devoted to path selection decision policy based

on multi-class traffic loads in the network Note that routing in such an integrated service network is not simply an addition of different routing algorithm for different service class of traffic [23] Service classes that deliver Quality-of-Service (QoS) to applications have priority over others that do not Therefore, high-priority QoS traffic can preempt the network resources by buffer management and scheduling methods For instance, in a network that supports both QoS traffic requiring bandwidth guarantee and best-effort traffic, two classes of traffic compete for the network resources while best-effort traffic can only share the bandwidth left unused by QoS traffic Thus, an “optimal” path selected for a QoS flow may worsen the congestion condition or even create starvation of best-effort flows In order to address this issue, several studies were carried out in the literature For example, Q Ma and P Steenkiste presented a multi-class routing algorithm addressing dynamic resource sharing between different classes of traffic [23] They initially put forward the idea of

‘virtual residual bandwidth’, which is created to adjust each link’s actual residual bandwidth to reflect the best-effort traffic load on the link Similarly, Y Chen and R Hwang presented QoS routing algorithms [3] for multiple classes of traffic by applying both ‘virtual residual bandwidth’ and ‘virtual cost’ In both studies, best effort traffic shares network resources according to the max-min fairness, which thereby increases the computational complexity of the algorithms Finally, Kochkar,

Ikenaga and Oie proposed a simpler method of calculating virtual residual bandwidth

in [17] However, they all studied QoS routing problems based on merely guaranteed traffic and best-effort traffic, which might not be adequate To our understanding, most real-time services in NGN would have strict requirement on the end-to-end network delay On account of that, a study on the multi-class routing is definitely needed based on the coexistence of best-effort traffic and such kind of QoS traffic that not only asks for bandwidth guarantee, but also asks for delay guarantee Therefore, we perform such a study in this thesis and propose an approach aiming to provide satisfied end-to-end delay bound and sustain low blocking rate for QoS traffic while, on the other hand, to improve the performance of low-priority best-effort traffic

bandwidth-1.3 Assumptions

The process of effective network resource management is very complicated and requires a variety of software and hardware support from both network side and end systems in order to successfully support multiple classes of services It involves numerous system functions, including scheduling disciplines, resource reservation protocols, admission control, real-time transport protocol, multimedia devices, real-time operating system and QoS routing The routing function hereby must work closely with other network functions in order to effectively manage network resources

In this thesis, the focus is on the routing mechanism to support end-to-end QoS provision while at the same time sustaining high network throughput Although various difficult problems exist for each of the other relevant system functions, they are out of the scope of this thesis Hence, we simply make the following assumptions:

1 There exists a negotiation mechanism that Service Level Specification (SLS) can

be effectively exchanged between the end users and the network components so that QoS requirements for each flow can be obtained

2 There exists a resource reservation and allocation protocol which allows sufficient resources to be allocated according to QoS requirements at the network routers

3 There exists an automatic traffic engineering manager that combines admission control, resource reservation and scheduling protocols in order to measure, record and keep track of the network state, e.g residual resources and link availability of every router and every link in the network In addition, the traffic engineering manager can spawn off signaling for route setup from the source to the destination

4 There exists an efficient QoS-based routing protocol that is able to disseminate and keep up to date the network state information by flooding the information to all the nodes in the network The network state information sent to the node is assumed to be always accurate and up-to-date

1.4 Thesis Contribution

The major contribution of this thesis appears in [42], which studied the QoS routing problem based on multiple traffic loads We summarize our contributions as follows:

1 An industry applicable path selection algorithm is proposed, which not only aims

to provide end-to-end delay guarantee for the traffic, but also to achieve efficient resource utilization for whole network Besides, an improved ULARAC path/link cost function is proposed, which can incorporate the advantages of both shortest-distance path cost function [22] and Lagrange Relaxation based Aggregated Cost function [15] In addition, Weighted Fair Queuing (WFQ) is introduced into routing in the NGN as the scheduling discipline and also contributes to the cost function WFQ is described in detail in Section 2.4

2 A mathematical computation method is proposed to calculate the ‘virtual residual bandwidth’ (VRB) as VRB is vital to implement multi-class QoS routing This method is quite simpler than that of [23], in which max-min fair-share rate of best-effort connections is used thus increasing the complexity of the algorithm

3 The performance of a multi-class routing algorithm is studied thoroughly through simulations based on bandwidth-delay-constrained QoS traffic and best-effort traffic To our knowledge, previous research effort did not consider delay-constrained traffic coexisting with best-effort traffic Our experimental results demonstrate, although the efficiency of the algorithm is undermined by more stringent delay constraints and unique properties of the links in the network, that the routing algorithm can lead to certain or even obvious performance

improvement for best-effort traffic for most of the scenarios while not deteriorating the performance of QoS traffic

1.5 Thesis Organization

In Chapter 2, we first introduce the basic concepts and terminology used in this thesis Then, we describe in detail Weighted Fair Queuing (WFQ) scheduling algorithm and the concept of token bucket as both are used as fundamental disciplines for end-to-end QoS provisioning We also review various QoS routing algorithms in the literature in this chapter Next in chapter 3, we discuss Delay and Bandwidth Constrained Least Cost path problem and present Utilization and Lagrange Relaxation based Aggregated Cost (ULARAC) path algorithm to solve this problem We then introduce our multi-class adaptive QoS routing algorithm We evaluate the proposed routing algorithm through simulations and discuss the simulation results in Chapter 4 We conclude this thesis and point out possible future work in Chapter 5

Chapter 2

Preliminary and Related Work

In this chapter, some basic concepts are first introduced together with terminology used in the rest of the thesis Next, routing techniques commonly used in the network are discussed In addition, different QoS unicast routing algorithms in the literature are reviewed Especially, the Delay Constraint Least Cost path problem and its solutions are highlighted Then, other components that can help ensure end-to-end QoS are briefly introduced In the last section, Weighted Fair Queuing (WFQ) scheduling algorithm and the concept of token bucket are described in detail

2.1 Basic Concepts

2.1.1 Weighted Graph Model

Generally, a computer communication network can be modeled as a directed and

connected graph, N = G(V, E), where E denotes the set of directed communication

links and V represents the set of network devices, typically those with routing

capabilities, e.g routers or layer 3 switches Letn V=| | andm=| |E be the number of network nodes and the number of the edges respectively Also, lete u v be the link ( , )from router u to router v Usually, we have duplex links: the existence of a

link ( , )e u v implies the existence of a link ( , ) e v u for any v u V, ∈ Such a pair of links may or may not be symmetric They are symmetric only if they have same properties such as capacity, propagation delay and etc Thus in this case, the edges are undirected It should be noted that, although some examples in this thesis use the undirected graphs for fewer edges, the routing algorithms under discussion were designed for both symmetric and asymmetric networks

Each link e E∈ is characterized by the following values: delayd , which denotes the e propagation delay of a link e; Cap , which denotes the link speed or link capacity; e

andcost , which denotes the link cost and can be represented by e.g a simple hop e

count, a policy cost or some measure of link’s residual capacity

2.1.2 QoS Metrics and Constraints

Various real-time applications, like webcasting and telemedicine, are being deployed over the internet, which requires the network to provide the guarantee of the quality of the service to the receivers The quality of the service can be estimated and specified

in terms of some parameters (also known as metrics) which are used to deliver the

application’s requirements to the underlying network These parameters or metrics could be delay, bandwidth, delay jitter, packet loss that can be tolerated by the receiver and/or number of hops, etc The first four parameters are better known as constituting the Quality of Service (QoS) metrics However, the QoS requirement can

be either specified in terms of user response time, audio/video quality, etc which can

in turn, be mapped into the QoS metrics

In this thesis, we are particularly interested in the guarantee of certain end-to-end delay One of the most important performance measures of a data network is the average delay required to deliver a packet from origin to destination [1] Thinking of a voice application, there are numerous sources of delay: queuing delay, propagation delay, serialization delay, codec delay and others e.g., shaping or reshaping delay In this thesis, we focus on end-to-end delay that a packet will experience within the communication network (i.e., in terms of the network layer) This kind of delay is the sum of delays on each network link traversed by the packet Each link delay in turn consists of the following four components:

( ) 1 [(1 ( , )) (1 ( , )) (1 ( , ))]

m P = − −m s i × × −m u v × × −m z t

However, loss probability metric can be easily transformed to an equivalent metric, i.e the probability of a successful transmission, which obviously follows multiplicative composition rule

The QoS requirement of a specific application is given as a set of constraints, which can be link constraints, path constraints and/or tree constraints A link constraint specifies the restriction on the use of links For example, a bandwidth constraint of a unicast connection requires that the links constituting the path must have certain amount of residual bandwidth available for the connection A link constraint is often specified on a concave metric while a path or tree constraint is often specified on an additive metric Furthermore, a path constraint needs to be satisfied from the sender to the receiver along the path A tree constraint needs to be satisfied over the entire multicast tree

2.2 QoS Routing Problems

Routing basically has two main functions: 1) the selection of routes for various destination pairs and 2) the delivery of messages to their correct destination once the routes are selected The second function is conceptually straightforward using a variety of protocols and data structures (known as routing tables) [1] In this thesis, the focus is on the first function (selection of the routes) and how it affects network performance

origin-In general, routing involves two entities: routing protocols and routing algorithms [20] The routing protocol collects the network topology and state information, e.g link/node availability/connectivity information and/or available resources, and keeps the information up to date and disseminates the information throughout the network The dual of the routing protocol, the routing algorithm, assumes a temporarily static

or frozen view of the network state and topology information provided by the routing protocol The routing algorithm then computes appropriate paths according to the information While current best-effort routing deployed in the Internet simply performs route selection based on a relatively static and single measure, QoS routing takes into account both the application’s requirements and the availability of network resources, which definitely increases the number and complexity of network state information It is assumed that for all the QoS routing algorithms discussed in this thesis, the network state information is captured, updated and disseminated

throughout the network using some underlying QoS-based routing protocols (e.g., QoS-enhanced Open Shortest Path First, Q-OSPF) The routing algorithm is the focus

of this thesis

The routing problems can be divided into two major classes: unicast routing and multicast routing The unicast routing problem is to find a feasible path from source

node to destination node while the multicast routing is to find a feasible tree covering

a source node and all nodes in a set of destination nodes In this thesis, the focus is on the unicast routing problem

2.2.1 Routing Techniques

Routing in the network can be classified as either intra-autonomous-system (intra-AS) routing (or simply intra-domain routing), or inter-AS (or inter-domain) routing [40] Intra-domain routing includes Routing Information Protocol (RIP), Open Shortest Path First (OSPF), and Intermediate System-Intermediate System (IS-IS) etc Inter-domain routing includes Border Gateway Protocol (BGP) and Multiprotocol BGP (MBGP) etc In this thesis, we focus on the intra-domain routing

Routing strategies can be classified in terms of either the mechanisms for triggering a search for feasible paths (satisfying constraints), or the amount of state maintained [4, 18] at each node In [40], routing strategies are classified as follows according to different mechanism for triggering a search for feasible paths

z Pro-active (Pre-Computation) Routing: This routing approach stores the routes

to all the destinations at all times Usually, a frequent updated routing table is maintained This kind of pre-computation approaches is highly responsive since the overall average path setup time is significantly reduced However, it incurs high processing and storage overhead Current Internet routing protocols such as RIP, OSPF and BGP are using this pre-computation technique

z Reactive (On-Demand) Routing: This routing approach always computes routes

to the destinations when they are needed, which thus reduces overhead at the expense of slower response times Examples of this technique include Ad-hoc On-demand Distance Vector (AODV) routing used in the mobile ad-hoc networks and most of the QoS routing protocols introduced later

Routing strategies can also be classified as follows in terms of the amount of global state maintained at each node [40]:

z Distance Vector Routing: In this approach, each node periodically exchanges distance vector information (distance and next hop from itself to all destinations) with its neighbors Then each node computes the paths according to the distance vector information The main problem of this approach is the lack of global knowledge, leading to slow convergence and routing loops

z Link State Routing: In this approach, the states of all local links are periodically flooded to all network nodes Based on these states, the required feasible paths are locally determined Compared to Distance Vector Routing, this approach has the advantage of simplicity, accuracy and avoidance of loops Nevertheless, it suffers from three major problems:

¾ High storage overhead

¾ High path computation overhead

¾ High link-state update overhead

Among all possible combinations of the above routing techniques, reactive link state routing seems to be the least popular in terms of high cost incurred by the computation overhead and amount of maintained state However, recent proposals attempt to alleviate some of these problems, e.g [31] We believe that due to the advance of more and more powerful processors and high-density but cheaper storage devices, these problems can be further alleviated

2.2.2 QoS Unicast Routing

Interest in Constraint-Based Routing (CBR) has been steadily growing in the Internet community particularly spurred by MPLS traffic engineering [40] CBR denotes a class of routing algorithms that select feasible paths based on a set of requirements or constraints, in addition to the destination If the constraints are imposed by policies, the associated routing is referred to as policy routing If the constraints are imposed

by QoS requirements such as bandwidth, delay or cost, the associated routing is referred to as QoS routing Therefore, the basic function of QoS unicast routing is to

find such a feasible path that has sufficient residual (unused) resources to satisfy the

QoS requirements of a connection An example of QoS routing is illustrated in Figure 2.1 The resources indicated on the links correspond to the cost and delay of each link The routing objective is cost minimization while subject to the constraint on path

delay no more than 40 ms Thus the feasible and best selected path from s to d is

(8,22)(5,12)

(6,14)

(4,19)(3,10)

Figure 2.1:Path selection from s to d with (cost, delay) values indicated on each link and routing objective is cost minimization and path delay no more than 40 ms

One key problem with QoS routing is tractability As path selection decision is based

on various constraints on QoS metrics, the computation complexity is primarily determined by the composition rules of the QoS metrics, i.e additive, multiplicative and/or concave Wang and Crowcroft [38] initially proved that the problem of finding

a path subject to two or more independent additive and/or multiplicative constraints in any possible combination is NP-Complete The tractable combinations are only the concave constraint and the other additive/multiplicative constraints [38] However, Mieghem and Kuipers indicated in [24], that QoS routing in realistic networks may not be NP-Complete in nature For instance, in the network where all the link metric vectors are identical [24], QoS routing can be solved in polynomial time

Recent research on QoS routing has proceeded in two directions [40] Initially, the main focus was on solving the routing problem satisfying various QoS constraints, or combinations of constraints (multi-constrained QoS routing), which is often NP-complete Recently, the focus has shifted to practical problems and optimizations addressing efficient resource utilization Resource efficiency is an important issue that should be considered in designing QoS routing algorithm This requires that the algorithm select a path that can lead to better overall resource efficiency if more than one path is available yet feasible in the network The most common way for routing algorithm to achieve resource efficiency is to keep network load balanced and to limit resource consumption The network load can be balanced by selecting the least loaded path while the resource consumption can be optimized by simply restricting the hop count or in some cases selecting the least cost path In the rest of this section, we review briefly progresses in both research directions

z Single-metric based routing:

This group of routing algorithms selects the feasible path based on single QoS metric, such as bandwidth or delay Several solutions have been proposed to the bandwidth-bounded and delay-bounded routing [8, 21]

Bandwidth-bounded routing: An interesting approach proposed in [10, 30] considers the imprecision in the network while selecting the bandwidth-bounded path The imprecision model is based on the probability distribution The heuristic algorithm assigns every link a weight that represents the probability of success in the link having

certain units of bandwidth For example, a given link l is assigned a weight log

w = − p , where p l = p w l( ) is the probability of success in the link having the w

units of bandwidth Then the heuristic algorithm tries to find a path that has the highest probability to accommodate a new connection with a given bandwidth requirement

Some routing algorithms also address efficient resource utilization while providing bandwidth guarantee for connections They include widest-shortest path [9], shortest-widest path [38], and utilization-based shortest-distance path selection algorithm [22]

As we have mentioned before, there are commonly two ways to achieve resource efficiency One is to limit resource consumption by selecting a path with as few hops

as possible or a path with least ‘cost’ The other way is to balance network load by selecting the least loaded path However, these two “optimality” criteria can conflict each other For example, a shortest path in terms of the hop count may be heavily loaded with traffic Therefore, there should be a tradeoff between the two criteria, which requires routing algorithms put proper weight on limiting hop count and on balancing network load

¾ Shortest-widest path: a path with the maximum ‘bottleneck’ bandwidth (minimum of the residual bandwidth of all the links along the path) among all the feasible paths is selected first If more that one such path exists, the one with the fewest hops is then selected

¾ Widest-shortest path: a path with the fewest hops among all the feasible paths is selected first If more than one such path exists, the one with maximum

‘bottleneck’ bandwidth is then selected

¾ Shortest-distance path: a path with the shortest distance among all the feasible paths is selected The distance function is the summation of all the costs of links that constitute the path, where the cost of each link is defined as the reciprocal of the available bandwidth of the link

A systematic evaluation of the above algorithms together with ‘Dynamic-alternative’ path selection algorithm [22] has been carried out by Q Ma and P Steenkiste They pointed out that the widest-shortest path gives the high priority to limiting the hop count, while the shortest-widest path does contrarily They also drew a conclusion that shortest-distance path algorithm, which selects a path with minimum path cost based on link utilization, outperforms other ones as a whole Shortest-distance path algorithm is able to not only find a load-balanced path, but also constrain the length of selected path dynamically In this thesis, we adopt this algorithm into designing our own QoS routing algorithm Some advantages of this algorithm are listed as follows:

1 Under evenly distributed traffic load, shortest-distance path algorithm performs slightly better than all others, especially in U.S MCI network topology [37]

2 Under unevenly distributed traffic load, shortest-distance path algorithm still performs consistently well Only dynamic-alternative algorithm has similar performance

3 When call holding time is either exponential or long-tail distributed, distance path algorithm still has slightly better performance

shortest-4 Shortest-distance path algorithm has a consistent performance when routing information is not accurate

5 Using Shortest-distance path algorithm for QoS routing results in better performance for best-effort connections when the network supports different type

of traffic at the same time

Delay-bounded routing: In [36], Shin and Chen proposed a distributed route selection approach for delay-constrained routing The approach floods routing messages from source towards destination node Each message accumulates the total delay of the path it has traversed so far When reaching an intermediate node, the message continues to be forwarded only if at least one of the two conditions is satisfied:

1 First such message is received by the node

2 The message holds a better accumulated delay than the previously received message

If either condition is true, the routing message will only be forwarded along the outgoing links whose delay plus the message’s accumulated delay does not exceed the delay constraint Once a message reaches the destination, a delay-constrained path is found Although some techniques like selective probing were analyzed, this approach bears a high communication overhead

z Dual-metric based routing:

Bandwidth-bounded, delay-bounded routing: Wang-Crowcroft’s algorithm in [38]

first eliminates all the links not satisfying the bandwidth requirement by setting their

delay to infinity Next, it finds a shortest path with respect to delay using Dijkstra’s algorithm in the modified graph

Bandwidth-bounded, cost-bounded routing: Chen and Nahrstedt proposed a heuristic algorithm for this routing problem The algorithm maps the cost of every link from an unbounded real number to a bounded small integer value, and then solves the problem in polynomial time by an Extended Bellman-Ford (EBF) or Extended Dijkstra Shortest Path (EDSP) algorithm [4]

Bandwidth-optimized, delay-optimized routing: Wang and Crowcroft considered the

order of importance of the two constraints in [38] They gave the higher priority to link residual bandwidth Thus, given any two nodes, the path with the maximum

bottleneck bandwidth is first selected, which is called the widest path If there are several such paths, the one with the smallest delay is finally selected as the shortest- widest path

Delay-bounded, cost-optimized routing: This problem has witnessed significant interest from research community [15, 32] In this thesis, it is of particular interest as

well First of all, we give the formal definition of the equivalent problem, the Delay Constrained Least Cost path problem (referred hereafter simply as DCLC) as follows

[15]:

Given a directed, connected graph G(V,E), a non-negative total cost of a path c(p) and

a total delay of a path d(p), a source node s, a destination node t, and a positive delay constraint delay∆ The constrained minimization problem is:

' ( , )

min coste

where P (s,t) is the set of path from s to t for which the end-to-end delay is bounded '

by delay∆ ThereforeP (s,t) P s t ' ⊆ ( , ) Namely, a p P s t∈ ( , )is in P (s,t) if and only if '

( )

The DCLC problem is NP-hard [7, 15], however it is worth mentioning that an

important special case is solvable in polynomial time if all link costs or all link delays

are equal, or have a few constant different values [16] Although a number of heuristic

algorithms have been proposed to solve the DCLC problem, the appropriate choice of

routing algorithms is still an open issue [15]

In [15], LARAC, i.e Lagrange Relaxation based Aggregated Cost algorithm was

presented to solve DCLC problem in polynomial time This algorithm is based on

Lagrange Relaxation, which is a common technique for calculating lower bounds and

finding satisfactory solutions for this problem (For a more detailed description,

please refer to [12] [13].)

LARAC algorithm firstly defines a mixed cost function of any path p :

c p = c p +λ λ×d p , where λ is the weight assigned to delay with respect to cost

If λ=0and the delay constraint is satisfied, an optimal solution for the original

problem (DCLC problem) can be found If this is not the case, λ is increased to boost

the dominance of delay in the mixed cost function until the best result of cλ suits the

delay constraint Then LARAC algorithm gives a lower bound to original DCLC for any λ≥0 in order to find the value of λ that gives the best result:

( ) : min{ ( ) : ( , )}

L λ = c p p P s tλ ∈ − × ∆λ delay (2.3)

To obtain the best lower bound, LARAC needs to maximize the function ( )L λ , namely looking for the value L* =max ( )L λ for any λ≥0 and the maximizing variable λ* Since L is a concave piecewise linear function and for anyλ≥0, cλ-

minimal path pλ, (d pλ)is a supgradient of L in the point λ , the following claims can

Summing up, the constrained conditions can be neglected and built into an object function (this is the relaxation) Since the solution feasible to the original problem suits the relaxation conditions as well, a lower bound of the original problem can be found If the path found is not feasible for the constrained conditions, the dominance

of it in the modified cost function is increased thus enforcing the solution to approach

to the optimal solution and decreasing the difference between the obtained lower bound and the optimum of the original problem as well This is the basic principle of Lagrange Relaxation

z Multiple-metric based routing:

Multi-constrained routing: The objective of multi-constrained routing is to find a

multi-constrained path (MCP) simultaneously satisfying a set of constraints The MCP problem has some variants such as multi-constrained optimal path (MCOP) problem and restricted shortest path (RSP) problem Unluckily, all of them are known

to be NP-complete In the literature, many heuristics have been proposed for these problems, e.g [16, 19, 26]

2.2.3 Multi-class QoS Routing

Next Generation Network is supposed to support various services and applications in one core network where resources are shared by versatile classes of traffic In such kind of multi-class service network, routing mechanism for high-priority QoS traffic might have an impact on what resources are available for traditional best-effort traffic

In the literature, some effort has been devoted to the study of the consequences of QoS routing on best-effort traffic First, Ma and Steenkiste proved in [22][23] that to apply short-widest path algorithm to QoS traffic might lower the throughput of

coexisting best-effort traffic They further proposed the concept of ‘virtual residual bandwidth’ for the first time, which takes account of the congestion conditions of the best-effort traffic when calculating actual residual bandwidth of a link seen by QoS traffic Their virtual residual bandwidth is calculated by means of the minimum and the mean max-min fair share rates, as best-effort traffic in their network shares the resources according to the max-min fairness Another approach, Chen and Hwang proposed in [3], is that COL-based cost function should be adopted for routing high-priority bandwidth-guaranteed traffic They proposed two alternatives: soft routing and hard routing The basic idea of soft routing algorithm is to adopt a “virtual cost” that consists of the original COL cost for reserving the required bandwidth and an additional cost that takes the existing best-effort traffic into account The hard routing simply adopts the “virtual residual bandwidth” concept The above two approaches both calculate virtual residual bandwidth or virtual cost based on the max-min fair-share rate of best-effort traffic in the network, which definitely increases the computational complexity Finally, Kochkar, Ikenaga and Oie proposed a different yet simpler method of calculating virtual residual bandwidth, which is based on best-effort traffic loads and total link utilization of both classes of traffic [17] Although the method is simple and efficient, it can be further improved by defining a stricter

function of credit for each link and specifying the relationship of some parameters

We will describe this improvement in Section 3.3.2, page 49 and page 51 respectively Summing up, all of above research effort focused on merely one QoS requirement, i.e bandwidth guarantee for high-priority QoS traffic, but not considered delay requirement

2.2.4 QoS Routing and Other QoS-provision Components

In order to fulfill end-to-end QoS provision, in addition to QoS routing, other network components such as admission control, resource reservation and management, buffering and scheduling of the packets from all the connections are necessarily required Before we introduce these modules respectively, we depict the end-to-end

QoS-provision scheme in brief When a nodes s wants to send data with certain to-end QoS requirements to another node t, it initiates a QoS connection request The

end-network first decides whether to accept this request according to current end-network residual resources (This is called admission control.) If the request is accepted, the

network identifies a path from s to t which can satisfy the QoS requirements, (this is

called QoS routing) and then reserves resources along the path to establish the

connection (this is called resource reservation) After the connection is established, s sends its data along that path to t The quality of service is guaranteed by the

resources reserved through packet buffering and scheduling

QoS Negotiation and Admission Control: As a connection is initiated, it will send its setup request to the network with application’s QoS requirements Thus QoS negotiation takes place The network then performs admission control and routing process to decide whether the request should be accepted or rejected The admission control module can be either a measurement-based mechanism [14] or based on some analytic model, e.g., equivalent capacity [11] Both QoS negotiation and admission control are often regarded as byproducts of QoS routing and resource reservation If

the resource reservation is successfully performed along the path selected by the routing algorithm, the connection request is accepted; otherwise, the request is

rejected

Resource Reservation: Once a connection request has been accepted, the feasible path with best chance of providing the required end-to-end QoS for data transmission should be selected by the QoS routing algorithm The required resources (CPU time, buffer, bandwidth, etc.) must be reserved along the path Hence, the data transmission

of the connection will not be affected by the traffic dynamics of other connections sharing the common links Resource reservation can be part of the path set-up process

or it can be a separate process by itself

Packet buffering and Scheduling: Depending on the network load, as the packets of data arrive at the incoming interface, the network devices, e.g router will transfer it to the outgoing interface, buffer it or drop it Appropriate buffering (or queue management algorithms) and packet scheduling algorithms are necessary for minimal packet drop and bandwidth savings due to reduced retransmissions as well as providing guaranteed resources reserved for the connection Some examples of queue management algorithms are Random Early Detection (RED) and Flow Random Early Detection (FRED) while the proposed algorithms for packet scheduling include Stochastic Fair Queuing (SFQ) and Weighted Fair Queuing (WFQ), etc