Performance Experiences in a Wide Area ATM Network

Experiments in Asynchronous Transfer Mode (ATM) wide area network (WAN) environments are discussed in this work. The lessons learned through practical experiences with ATM systems and how these can be applied to improving performance are presented. Factors ranging from the transport protocol to the equipment used and the network conditions are considered. With the recent deployment of the rst wide area ATM networks (ATM WANs), experiments in this environment have become possible. This paper discusses lessons learned through practical experiences with ATM and how these can be applied to improving performance. Throughput that is one or two orders of magnitude below link speed has sometimes been observed on newly deployed ATM systems. The problem is not that ATM networks perform below expectations, but that many network parameters must be appropriately set in order to obtain the best possible performance. Some of the most important and well-known 13] factors, which will be further discussed in the following sections, relate to protocols, hosts and networks. Most ATM connections currently utilize TCP/IP as the transport protocol. When using TCP/IP over high speed WANs, throughput may be limited by factors such as excessive retransmissions due to cell losses at the edge of the network, ineeciency due to Maximum Transmission Unit (MTU) size, and TCP window management 2]. In most cases, these limiting factors can be avoided with careful tuning and occasional modii-cations to application code. Hosts are often the limiting factor when transmitting or receiving over high-speed networks. Characteristics to be considered include CPU and memory speed, operating system, system parameters and load. Finally, such factors as propagation delay, switch buuer size, link speeds, bandwidth mismatches, Usage Parameter Control (UPC) and congestion avoidance methods will determine the maximum throughput that can be expected. Many of these factors have already been discussed in detail in the authors' previously published work 5, 6, 7, 14]; this paper describes recent experiments that address some of the remaining factors (a tutorial including some of these results is to be published in 4]). Experiments described in this paper were conducted over the ACTS ATM Internetwork (AAI). This network provides wide area ATM connectivity between several DoD High Performance Computing centers and the MAGIC and ATDnet gigabit testbeds, with ATM service provided by Sprint 1. In Sections 2 through 4 of this paper, we describe and discuss our results in each of the three main areas listed above; section 5 …