Elsevier

Computer Networks

Volume 40, Issue 6, 20 December 2002, Pages 695-709
Computer Networks

MATE: multipath adaptive traffic engineering

https://doi.org/10.1016/S1389-1286(02)00308-0Get rights and content

Abstract

Destination-based forwarding in traditional IP routers has not been able to take full advantage of multiple paths that frequently exist in Internet service provider networks. As a result, the networks may not operate efficiently, especially when the traffic patterns are dynamic. This paper describes a multipath adaptive traffic engineering scheme, called MATE, which is targeted for switched networks such as multiprotocol label switching networks. The main goal of MATE is to avoid network congestion by adaptively balancing the load among multiple paths based on measurement and analysis of path congestion. MATE adopts a minimalist approach in that intermediate nodes are not required to perform traffic engineering or measurements besides forwarding packets. Moreover, MATE does not impose any particular scheduling, buffer management, or a priori traffic characterization on the nodes. This paper presents an analytical model, derives a class of MATE algorithms, and proves their convergence. Several practical design techniques to implement MATE are described. Simulation results are provided to illustrate the efficacy of MATE under various network scenarios.

Introduction

Internet traffic engineering is emerging as an important tool to provide fast, reliable and differentiated services. According to the Internet engineering task force (IETF), Internet traffic engineering is broadly defined as that aspect of network engineering dealing with the issue of performance evaluation and performance optimization of operational IP networks [1]. More specifically, traffic engineering often deals with effective mapping of traffic demands onto the network topology, and adaptively reconfiguring the mapping to changing network conditions. It is more general than QoS routing in the sense that traffic engineering typically aims at maximizing operational network efficiency while meeting certain constraints, whereas the main objective in QoS routing is to meet certain QoS constraints for a given source–destination traffic flow.

The emergence of multiprotocol label switching (MPLS) with its efficient support of explicit routing provides basic mechanisms for facilitating traffic engineering [9]. Explicit routing allows a particular packet stream to follow a pre-determined path rather than a path computed by hop-by-hop destination-based routing such as OSPF or IS–IS. With destination-based routing as in traditional IP network, explicit routing may be provided by attaching to each packet the network-layer address of each node along the explicit path. This approach generally incurs prohibitive overhead. In MPLS, a path (known as a label switched path or LSP) is identified by a concatenation of labels which are stored in the nodes. As in traditional virtual-circuit packet switching, a packet is forwarded along the LSP by swapping labels. Thus, support of explicit routing in MPLS does not entail additional packet header overhead.

In this paper, we propose a state-dependent traffic engineering mechanism called multipath adaptive traffic engineering (MATE). MATE assumes that several explicit LSPs (typically ranges from 2 to 5) between an ingress node and an egress node in an MPLS domain have been established using a standard protocol such as CR-LDP [6] or RSVP-TE [2], or configured manually. This is a typical setting which exists in an operational ISP network that implements MPLS. The goal of the ingress node is to distribute the traffic across the LSPs so that the loads are balanced and congestion is minimized. The traffic to be balanced by the ingress node is the aggregated flow (called traffic trunk in [7]) that shares the same destination (and possibly quality of service). Fig. 1 shows an example of a network environment where there are two ingress nodes, AI and BI, and two egress nodes, AE and BE, in an MPLS domain. MATE would be run on AI and BI to balance traffic destined to AE and BE, respectively, across the LSPs connecting AI to AE and BI to BE. Note that the LSPs connecting the two pairs may share links. In the following, we will derive adaptive MATE algorithms, discuss their implementation, and present simulation results to illustrate their performance. We will prove that it is possible to achieve stability even when ingress–egress (IE) pairs operate asynchronously and in a distributed manner.

We now comment on related work. Several researchers have proposed to add traffic engineering capabilities in traditional datagram networks using shortest path algorithms (e.g., see [5], [10]). Although such schemes have been shown to improve the efficiency of the network, they suffer from several limitations including:

  • load sharing cannot be accomplished among paths of different costs,

  • traffic of different QoS classes follow the same route,

  • traffic/policy constraints (for example, avoiding certain links for particular source–destination traffic) are not taken into account,

  • modifications of link metrics to re-adjust traffic mapping tend to have network-wide effects and may cause undesirable and unanticipated traffic shifts, and

  • traffic demands must be predictable and known a priori.


The combination of MPLS technology and its traffic engineering capabilities are expected to overcome these limitations. Explicit LSPs and flexible traffic assignment address the first two limitations. Constraint-based routing has been proposed to address the third limitation. Furthermore, network-wide effects can be prevented since LSPs can be pinned down. A change in LSP route limits the disturbance of the traffic for the corresponding source–destination pair. The objective of this paper is to address the final limitation.

In MPLS, traffic engineering mechanisms may be time dependent or state dependent. In a time-dependent mechanism, historical information based on seasonal variations in traffic is used to pre-program LSP layout and traffic assignment. Additionally, customer subscription or traffic projection may be used. Pre-programmed LSP layout typically changes on a relatively long time scale (e.g., diurnal). Time-dependent mechanisms do not adapt to unpredictable traffic variations or changing network conditions. An example of a time-dependent mechanism is a global centralized optimizer where the input to the system is a traffic matrix and multiclass QoS requirements as described in [8].

When there are appreciable variations in actual traffic that could not be predicted using historical information, a time-dependent mechanism may not be able to prevent significant imbalance in loading and congestion. In such a situation, a state-dependent mechanism can be used to deal with adaptive traffic assignment to the established LSPs according to the current state of the network which may be based on utilization, packet delay, packet loss, etc. In this paper, we assume that LSP layout has been determined. The focus is on load balancing traffic among multiple LSPs between an ingress node and an egress node.

The rest of the paper is organized as follows. Section 2 details the overall MATE scheme and discusses several implementation techniques, such as traffic filtering and distribution, traffic measurement, bootstrapping, etc. Section 3 presents an analytical model of MATE and proves its stability and optimality. Section 4 describes an experimental setup to verify the effectiveness of the proposed scheme. Section 5 presents the simulation results that illustrate the behavior of the algorithm in different environments. Conclusions are given in Section 6. Analytical proofs are collected in the Appendix A.

Section snippets

Overview

The basic idea of MATE is as follows. The ingress node of each LSP periodically sends probe packets to estimate a congestion measure on the forward LSP from ingree to egress. The congestion measure can be delay, loss rate, or other performance metrics; see below for measurement details. Each ingress node then routes incoming traffic onto multiple paths to its egress node in a way that equalizes the marginal congestion measure (their derivatives). That is, traffic will be shifted from LSPs with

MATE stability

In this section we present an analytical model of MATE and prove their stability and optimality.

Experimental methodology

In this section, we use simulations to evaluate the effectiveness of MATE. We concentrate on two network topologies: one with a single IE pair connected by multiple LSPs, and the other with multiple IE pairs where some links are shared among the LSPs from different pairs, as shown in Fig. 3, Fig. 4. All links are identical so that the LSPs have the same bottleneck link bandwidth. Note that in the latter case, there is a considerable interaction between the pairs.

We wrote a packet level

Simulation results

In this section, we present simulation results that illustrate the convergence properties of MATE.

First we present two sets of results for the single IE pair. Fig. 5, Fig. 6 show the results of an experiment with Poisson traffic on the network in Fig. 3. Initially, all of the engineered traffic streams are routed on one of the LSPs, and cross-traffic enter the network at the intermediate nodes connecting the ingress and egress nodes. We have an unbalanced situation with one heavily congested

Conclusion

Our focus on this paper is to apply adaptive traffic engineering to utilize network resource more efficiently and minimize congestion. We have proposed a class of algorithms called MATE, which tries to achieve these objectives using minimal assumptions through a combination of techniques such as bootstrap probe packets, which control the amount of extra traffic, and marginal delays that are easily measurable and do not require clock synchronization. Further, we prove the stability and

Anwar Elwalid is with Bell Labs, Lucent Technologies, where he is a DMTS (Distinguished Member of Technical Staff). He received the BS degree in Electrical Engineering from Polytechnic Institute of New York, Brooklyn, and the Ph.D. degree in Electrical Engineering from Columbia University, New York. Since 1991 he has been with the Mathematics of Networks and Systems Research Department at Bell Labs, Murray Hill, New Jersey, where he developed theory and algorithms for network resource

References (14)

  • D. Awduche, A. Chui, A. Elwalid, I. Widjaja, X. Xiao, Overview and Principles of Internet Traffic Engineering....
  • D.O. Awduche et al., RSVP-TE: Extensions to RSVP for LSP Tunnels, IETF RFC 3209, December...
  • D. Bertsekas

    Nonlinear Programming

    (1995)
  • D. Bertsekas et al.

    Data Networks

    (1992)
  • B. Fortz, M. Thorup, Internet traffic engineering by optimizing OSPF weights, in: Proceedings of INFOCOM’2000,...
  • B. Jamoussi et al., Constraint-based LSP setup using LDP, IETF RFC 3212, January...
  • T. Li, Y. Rekhter, Provider architecture for differentiated services and traffic engineering (PASTE), RFC 2430, October...
There are more references available in the full text version of this article.

Cited by (0)

Anwar Elwalid is with Bell Labs, Lucent Technologies, where he is a DMTS (Distinguished Member of Technical Staff). He received the BS degree in Electrical Engineering from Polytechnic Institute of New York, Brooklyn, and the Ph.D. degree in Electrical Engineering from Columbia University, New York. Since 1991 he has been with the Mathematics of Networks and Systems Research Department at Bell Labs, Murray Hill, New Jersey, where he developed theory and algorithms for network resource management and QoS support, and for the analysis and engineering of multimedia traffic. He holds several patents. His current research interests include IP and optical network architectures, traffic engineering and stochastic systems. He has been active in the Traffic Engineering and MPLS Working Groups of the IETF, and co-authored Internet Drafts and RFCs. He received best paper award from the ACM and IFIP. He has been a guest editor of the IEEE Journal on Selected Area in Communications, and served on the executive and technical program committees of several conferences. Dr. Elwalid is senior member of IEEE, and member of Tau Beta Pi (National Engineering Honor Society) and Sigma Xi.

Cheng Jin is a Ph.D. candidate in the Department of EECS at the University of Michigan. He received his B.Sc. in Electrical Engineering from Case Western Reserve University in 1996. His current area of research includes the placement of servers of various services inside a network and the modeling of Internet topology. He is a co-developer of the Inet topology generator.

Steven H. Low received his B.S. degree from Cornell University and Ph.D. from the University of California, Berkeley, both in electrical engineering. He was with AT&T Bell Laboratories, Murray Hill, from 1992 to 1996, and was with the University of Melbourne, Australia, from 1996 to 2000, and is now an Associate Professor at the California Institute of Technology, Pasadena. He has held visiting academic positions in the US and Hong Kong, and has consulted with companies and government in the US and Australia. He was a co-recipient of the IEEE William R. Bennett Prize Paper Award in 1997 and the 1996 R&D 100 Award. He is on the editorial board of IEEE/ACM Transactions on Networking. He has been a guest editor of the IEEE Journal on Selected Area in Communications, on the program committee of several conferences. His research interests are in the control and optimization of communications networks and protocols. His home is netlab.caltech.edu.

Indra Widjaja received the Ph.D. degree in Electrical Engineering from the University of Toronto, Toronto, Canada. From 1994 to 1997, he was an Assistant Professor of the Electrical and Computer Engineering Department of the University of Arizona. From 1997 to May 2001, he was with Fujitsu Network Communications where he was manager of systems engineering. He joined Bell Labs Research, Lucent Technologies, in May 2001. His research interests include traffic engineering, high-speed switching, and optical networking.

View full text