Internet Traffic Mid-term Forecasting: A Pragmatic Approach Using Statistical Analysis Tools

Abstract

Network planning is usually based on long-term trends and forecasts of Internet traffic. However, between two large updates, telecommunication operators deal with resource allocation in contracts depending on the mid-term evolution of their own traffic. In this paper, we develop a methodology to forecast the fluctuations of Internet traffic in an international IP transit network. We do not work on traffic demands which can not be easily measured in a large network. Instead, we use link counts which are much simpler to obtain. If needed, the origin-destination demands are estimated a posteriori through traffic matrix inference techniques. We analyze link counts stemming from France Telecom IP international transit network at the two hours time scale over nineteen weeks and produce forecasts for five weeks (mid-term). Our methodology relies on Principal Component Analysis and time series modeling taking into account the strain of cycles. We show that five components represent 64% of the traffic total variance and that these components are quite stable over time. This stability allows us to develop a method that produce forecasts automatically without any model to fit.