TCP performance estimation using neural networks modelling
Faculty of Computing, Health and Science
School of Computer and Information Science / Centre for Security Research
In recent years, the Internet has evolved from providing connectivity to hosts towards providing performance to applications, leading to significant interest in TCP performance modelling. Hitherto, mathematical approaches have been used to evaluate data transfers performance. Although such models were built upon a sound theoretical basis, they were validated largely using synthetic or endpoint controlled traffic and they are not suitable for short-lived transfers or clients with unknown behaviour, characteristics that are typical for the current Internet. The research presented in this paper aims to overcome these problems by using a supervised adaptive learning approach to build the relationship between TCP performance and the influencing parameters. The proposed model, using knowledge of past connections, was trained and tested using both synthetic and real network traffic. Comparison against the mathematical models showed that the proposed model provides more accurate estimates in the majority of the cases. The tests results indicated that the average error of the TCP throughput, estimated using the proposed model, was reduced to more than half the value obtained using the mathematical approach. The robustness and accuracy of the model allows its use in a variety of areas, such as network planning and evaluation of new TCP implementations.