Constructing stochastic model of TCP throughput on basis of stationarity analysis

Abstract

End-to-end TCP throughput on best-effort networks fluctuates widely depending on cross traffic and radio interference, which degrades the quality of public cloud services. However, if throughput variation is predictable on the application layer, service providers can be proactive in solving the problem of quality deterioration. We argue that TCP throughput fluctuation occurs in two states, stationary and non-stationary, and developed a method for characterizing these states. We also constructed a stochastic model of TCP throughput, which is a mixture of a stationary process model and a non-stationary one based on stationarity analysis. An evaluation experiment showed that our proposed model could predict the next 200 seconds of TCP throughput fluctuation through the Internet on three types of access networks (HSDPA/LTE/Wi-Fi) with an average accuracy of about 90%.