Network design of cable modem systems for WWW applications

The cable industry has begun testing and deploying high speed cable modems for data delivery. However, practical network designs require the estimation and prediction of performance under realistic traffic models and must consider the dominant factors that affect packet loss probabilities and offered load. In this paper we evaluate the use of a stochastic model of WWW client-server transactions as a tool for network design and analysis. Specifically, a sensitivity analysis of the model is performed to identify the impact various model parameters have on predicted packet loss probabilities in an Internet access network utilizing CATV transport. The traffic model used is a self-similar traffic model in which the inter-arrival times of document requests generated by each source is based on a two-state ON-OFF source model described and the length of each document is given by a Pareto distribution. We investigate the results of using such a model in predicting a number of network performance parameters including buffer requirements and achieved channel utilization. We also evaluate the number of WWW users that a head-end can support for a specific bandwidth. The effects of multiplexing, decreasing the inter-arrival times, increasing the length of the ON period, changing the tail of the document size distribution, and increasing the buffer size are all considered. We also identify which parameters most affect the packet loss probability and the offered load on the system. Such parameters will be the dominant factors that need to be used in network and capacity planning for CATV data delivery systems.