Algorithms for IP network design with end-to-end QoS constraints

Abstract

The new generation of packet-switching networks is expected to support a wide range of communication-intensive real-time multimedia applications. A key issue in the area is how to devise reasonable packet-switching network design methodologies that allow the choice of the most adequate set of network resources for the delivery of a given mix of services with the desired level of end-to-end Quality of Service (e2e QoS) and, at the same time, consider the traffic dynamics of today’s packet-switching networks. In this paper, we focus on problems that arise when dealing with this subject, namely Buffer Assignment (BA), Capacity Assignment (CA), Flow and Capacity Assignment (FCA), Topology, Flow and Capacity Assignment (TCFA) problems. Our proposed approach maps the end-user’s performance constraints into transport-layer performance constraints first, and then into network-layer performance constraints. This mapping is then considered together with a refined TCP/IP traffic modeling technique, that is both simple and capable of producing accurate performance estimates, for general-topology packet-switching design networks subject to realistic traffic patterns. Subproblems are derived from a general design problem and a collection of heuristic algorithms are introduced for compute approximate solutions. We illustrate examples of network planning/dimensioning considering Virtual Private Networks (VPNs).

Introduction

The new generation of packet-switching networks is expected to support a wide range of communication-intensive real-time multimedia applications. These applications will have their own different quality-of-service (QoS) requirements in terms of throughput, reliability, and bounds on end-to-end (e2e) delay, jitter, and packet-loss ratio. It is technically a challenging and complicated problem to deliver multimedia information in a timely, synchronized manner over a decentralized, shared network environment, especially one that was originally designed for best-effort traffic such as the Internet.

Accordingly, a key issue in this area is how to devise reasonable packet-switching network design methodologies that allow the choice of the most adequate set of network resources for the delivery of a given mix of services with the desired level of e2e QoS and, at the same time, consider the traffic dynamics of today’s packet-switching networks.

The traditional approaches to optimal design and planning of packet networks, extensively investigated in the early days of packet-switching networks [1], [2], focus on the network-layer infrastructure thus neglecting e2e QoS issues, and Service Level Agreement (SLA) guarantees. From the end-user’s point of view, QoS is driven by e2e performance parameters, such as data throughput, Web page latency, transaction reliability, etc. Matching the user-layer QoS requirements to the network-layer performance parameters is not a straightforward task. The QoS perceived by end-users in their access to Internet services is mainly driven by the Transmission Control Protocol (TCP), the reliable transport protocol of the Internet, whose congestion control algorithms dictate the latency of information transfer. Indeed, it is well known that TCP accounts for a great amount of the total traffic volume in the Internet [3], [4], and among all the TCP flows, a vast majority is represented by short-lived flows (also called mice), while the rest is represented by long-lived flows (also called elephants).

The description of traffic patterns inside the Internet is a particularly delicate issue, since it is well known that IP packets do not arrive at router buffers following a Poisson process [5]. Traditionally, either M/M/1 or M/M/1/B queueing models were considered as good representations of packet queueing elements in the network. However, the traffic flowing in IP networks is known to exhibit Long Range Dependent (LRD) behavior, which cause queue dynamics to severely deviate from the above model predictions. For these reasons, the usual approach of modeling packet-switching networks as networks of M/M/1 queues [6], [7], [8] appears now inadequate for the design of such networks. Recently, in [9], the authors for the first time abandon the Markovian assumption in favor of a fractional Brownian motion model, i.e., an LRD traffic model. They solve the discrete Capacity Assignment problem under network e2e delay constraints only, using a simulated annealing meta heuristic. Unfortunately, it is difficult to extend this approach to consider more general network problems, because the relation among traffic, capacity and queueing delay is not expressed as a closed-form expression.

Additionally, with the enormous success of the Internet, all enterprises have become dependent upon networks or networked computation applications. In this context the loss of network services is a serious outage, often resulting in unacceptable delays, loss of revenue, or temporary disruption. To avoid loss of network services, communications networks should be designed so that they remain operational and maintain as high a performance level as feasible, even in the presence of network component failures.

In this paper, we focus on several types of problems that arise when dealing with packet-switching network design. We consider the traffic dynamics of packet networks, as well as the effect of protocols at the different layers of the Internet architecture on the e2e QoS experienced by end-users. Of course, in any realistic network problem an “optimal design” is an extremely difficult task. In [10], [11] an IP network design methodology is proposed which is based on a “Divide and Conquer” approach, in the sense that it corresponds to several tasks. Fig. 1 shows the flow diagram of the design methodology. Shaded, rounded boxes represent function blocks, while white parallelograms represent input/output of functions. There are three main blocks, which correspond to the classic blocks in constrained optimization problems: constraints (on the left), inputs (on the bottom right) and optimization procedure (on the top right). Considered as constraints, for every source/destination pair, are the specification of user-layer QoS parameters. Thanks to the definition of the QoS translators, all the user-layer constraints are then mapped into lower-layer constraints, down to the IP layer. The optimization procedure takes as inputs (in accordance to the problem to be solved) the description of the physical topology, the routing algorithm specification, the traffic matrix, and the expression of the cost as function of the design variables. The objective of the optimization is to find the minimum-cost solution that satisfies the user-layer QoS constraints. A second important point of the proposed methodology is the adoption of a refined TCP/IP traffic modeling technique that is both simple and capable of producing accurate performance estimates for packet-switching networks subject to realistic traffic patterns. The main idea behind this approach corresponds to reproducing the effects of traffic correlations on network queueing elements by means of Markovian queueing models with batch arrivals. Hence, using M_[X]/M/1 like queues.

The rest of the paper is organized as follows. Section 2 briefly describes the QoS translation problem as well as the traffic and queueing models. Section 3 outlines the general design problem and provides the formulation of the related optimization subproblems. It also introduces heuristic algorithms to compute approximate solutions, and discusses numerical and simulation results. Finally, Section 4 summarizes the main results obtained in this research.