A transaction-level tool for predicting TCP performance and for network engineering

Traditional traffic engineering for communication networks assumes that the sources generate traffic with known statistics. Using simulation and/or analysis, one then studies the quality of service that the network offers to these traffic streams. However, most of the traffic in the Internet is TCP-based. TCP, the Transmission Control Protocol, regulates the delivery of packets to correct errors and avoid saturating the destination of the packets or routers inside the network. Accordingly, the traffic that a TCP source generates is not specified a priori. Instead, this traffic depends in an essential way on the QoS that the network provides. We propose a simple model and we use it to predict the QoS that a network provides to TCP connections. The tool can compute the probability that each connection gets at least some given throughput. Conversely, the tool can be used to design the network to ensure that every connection gets a given throughput with a specified probability. The tool incorporates uncertainties about the load on the network and about the model.