AI-based TCP Performance Modelling
This project aims to analyse the efficiency of artificial neural networks when modelling thePiger B, Ghita BV
performance of Transmission Control Protocol (TCP). First of all we need to understand what
TCP performance depends on. We have therefore researched existing mathematical models.
We have chosen to implement the Cardwell (2000) model. By using this model we can
retrieve the data transfer time thanks to the timeout, loss rate, round trip time and initial
congestion window. We have next built a network features dataset thanks to tcpdump and
tcptrace. This set has enabled us to test our mathematical model and to build our artificial
neural network. We have chosen to include the result of the mathematical model, the data
packet number sent per connection and the average data packet size in the input of the
artificial neural network in order to try to improve its efficiency. In order to analyse their
performance we have chosen to use the correlation and the average relative error. By analysing
our model we have shown the importance of the data. Indeed, we have to choose carefully
their type and scope. We have shown also that our implementation of the mathematical model
was inefficient. At the same time, we have reached a better accuracy with our artificial neural
network model: 86.45% of correlation and 37.4% of average relative error.