Window-based congestion management in broadband ATM networks: the performance of three access-control policies

Abstract
This report compares the network-level and user-level performance of three proposed access control mechanisms for congestion management in public broadband packet-switched networks: rate-based windows ("leaky buckets"), end-to-end acknowledgement windows, and congestion feedback acknowledgement windows. Their performance in terms of throughput, delay, and patterns of cell loss was studied using a cell-level simulation of an ATM network. When assessed at the network level, leaky buckets perform well while neither acknowledgement-based mechanism performs adequately. However, when assessed from the user perspective, leaky buckets also fail to provide satisfactory service. The simulation results reported here suggest that further study of intranetwork queueing and routing mechanisms is required to control congestion within the network and supply assured quality of service to the end users. This report evaluates the performance of three proposed access-control policies — rate-based ("leaky bucket"), traditional end-to-end acknowledgement, and acknowledgement with congestion feedback — for congestion management in broadband public packet networks. Our emphasis throughout is on the quality of service seen by end users, a perspective that yields significantly different conclusions than previous work that has concentrated on intranetwork performance. When both external and internal performance measures are considered, none of the proposed access control mechanisms appears to be adequate. The distinctive features of the networks considered here are their high transmission bandwidth (150 Mbits/second or greater), the use of a small fixed-size data unit (the "cell" or "packet", of less than 100 bytes), cell-based switching, and the alignment of cells in fixed time slots for transmission. For convenience, we shall refer to the general class of networks satisfying these criteria as "broadband ATM networks", although the results presented here do not depend on the precise details of the ATM specification (1). Congestion management mechanisms are required by the statistical nature of traffic on any switched packet network, including ATM networks. Unacceptable delay and cell loss arise not simply from the unpredictability of cell-based traffic, but also from contention, multiplexed time- correlated traffic, and cell clustering effects at switches. The window-based mechanisms for congestion control addressed here will be discussed in terms of broadband ATM networks, although they are also applicable in more general settings. Ideally, a satisfactory mechanism should guarantee performance to the end user applications: once a connection request is accepted, for example, that connection should continue to observe the quality of service it requested.

This publication has 8 references indexed in Scilit: