Internet Traffic Volumes Are Not Gaussian -- They Are Log-Normal: An 18-Year Longitudinal Study With Implications for Modelling and Prediction (Complete Version).

摘要: Getting good statistical models of traffic on network links is a well-known, often-studied problem. A lot attention has been given to correlation patterns and flow duration. The distribution the amount per unit time an equally important but less studied We study large number traces from many different networks including academic, commercial residential using state-of-the-art techniques. show that obeys log-normal which better fit than Gaussian commonly claimed in literature. also investigate alternative heavy-tailed (the Weibull) its performance worse log-normal. examine anomalous exhibit poor for all distributions tried this often due outages or hit maximum capacity. demonstrate data we look at stationary if consider samples 15- minute long even 1-hour long. This gives confidence can use estimation modelling purposes. utility our findings two contexts: predicting proportion will exceed level (for service agreement link capacity estimation) 95th percentile pricing. predictor Weibull both contexts.