Year

2000

Degree Name

Doctor of Philosophy

Department

School of Electrical, Computer and Telecommunications Engineering

Abstract

The label switching forwarding mechanism, which was derived from the connection oriented ATM protocol, is currently being applied to IP networks. Use of this mechanism brings significant advantages in terms of simplifying forwarding decisions as well as enabling implementation of comprehensive traffic engineering mechanisms. Label switching simplifies forwarding by replacing the longest-match prefix search algorithms with a simple lookup table.

This thesis concentrates on examining the use of label switching techniques for scalable best effort unicast forwarding. Most label switching protocols are capable of operating over cell based switched networks such as ATM , and we constrain this thesis to examining this case. In essence w e examine the implementation of label switching techniques in an Internet wide environment. The aim of this work is to examine ways of maximizing the percentage of packets following label switched paths and therefore reducing the number of packets that require traditional IP forwarding. We concentrate on core Internet routers.

There is a significant body of literature proposing and examining implementations several label switching approaches such as: Multi-Protocol Label Switching (MPLS), IP Switching, IP Navigator, Aggregate Route Based IP Switching, and Tag Switching. W e present an extensive literature survey. In addition we present a comprehensive classification of label switching techniques to provide scalable best effort IP forwarding in the Internet. From this work we isolate several areas of work for further consideration. Specifically, these areas concern the use of label switching forwarding in future IP networks including: an IP version 6 environment with highly aggregated routing tables; use of label switching in conjunction with congestion sensitive routing; and an optically switched network.

The examination of label switching in hierarchical networks examines the impact a high level of IP routing table aggregation that will be associated with IP version 6. We show that the popular MPLS best effort forwarding mechanism will perform poorly in this environment and will require a high level of forwarding in core gateway routers. We propose and examine the performance of a hybrid label switching protocol, called Destination Site Label Switching (DSLS), which performs well in this environment. W e show using traffic traces that the network layer forwarding requirement of DSLS can be reduced to below 0.2% of total packets regardless of routing table aggregation.

Some label switching protocols, such as IP switching, abstract forwarding from routing information. We examine the use of these protocols in conjunction with congestion sensitive routing protocols, where routing information will change frequently. Experiments, using traffic traces, were performed to examine the response of these protocols to changes in underlying routing information. We investigate several mechanisms to improve the sensitivity. It was found that altering flow detection parameters did not significantly improve the performance. However, the introduction of a maxim u m flow length in the order of 200 seconds improved the performance significantly even for highly aggregated DSLS flows with minimal (0.13% for D S L S ) additional packet layer forwarding required.

The examination of label switching in optical cell switched networks highlights problem with a mechanism called " VC Merge" which is used by many label switching protocols. VC Merge requires additional buffers for packet reassembly, and increases switch complexity. Previous studies have shown that VC Merge does not add significantly to total buffer requirements. However, these studies examined electronic switches which are capable of implementing large switch output buffers and complex forwarding mechanisms. We use analytical and simulation techniques to examine the buffer requirements of VC merge when traffic smoothing techniques are used to reduce buffer requirements. In this scenario w e find the VC merge mechanism buffer size becomes a significant part of buffer requirements and actually exceeds the average output buffer size.

Share

COinS