Predicting TCP Throughput Using Time Series Forecasting

Forecasting end-to-end throughput accurately can be vital for several applications. Peer-to-peer applications like BitTorrent or Skype will be able to select the peers more efficiently if the throughput in the short-term is known. On the other hand applications can adapt their behaviour accordingly, if they know how the end-to-end throughput is going to vary in the lifetime of a connection. This paper focuses on predicting throughput using time series forecasting techniques: Exponentially Weighted Moving Average and Holt-Winter. In this work, current data of throughput were collected using the Planet-lab simulator. It was found that for both EWMA and Holt-Winters, the accuracy patterns vary a lot from path to path. After comparing the prediction errors of EWMA and Holt-Winter, we demonstrated that if the data shows trend and seasonality, the prediction error of Holt-Winter is much lower than that of EWMA, while in other cases the accuracy of EWMA and HoltWinter are quite similar. Finally, we analysed the correlation between the prediction and RTT. We identified that, for the predictions obtained from both EWMA and Holt-Winter, the correlation follows a power function.