PREDICTING INTERNET TRAFFIC BURSTS USING EXTREME VALUE THEORY

Computer networks play an important role in today’s organization and people life.
These interconnected devices share a common medium and they tend to compete for
it. Quality of Service (QoS) comes into play as to define what level of services users
get. Accurately defining the QoS metrics is thus important.
Bursts and serious deteriorations are omnipresent in Internet and considered as an
important aspects of it. This thesis examines bursts and serious deteriorations in
Internet traffic and applies Extreme Value Theory (EVT) to their prediction and
modelling. EVT itself is a field of statistics that has been in application in fields like
hydrology and finance, with only a recent introduction to the field of
telecommunications. Model fitting is based on real traces from Belcore laboratory
along with some simulated traces based on fractional Gaussian noise and linear
fractional alpha stable motion. QoS traces from University of Napoli are also used in
the prediction stage.
Three methods from EVT are successfully used for the bursts prediction problem.
They are Block Maxima (BM) method, Peaks Over Threshold (POT) method, and RLargest
Order Statistics (RLOS) method. Bursts in internet traffic are predicted using
the above three methods. A clear methodology was developed for the bursts
prediction problem. New metrics for QoS are suggested based on Return Level and
Return Period. Thus, robust QoS metrics can be defined. In turn, a superior QoS will
be obtained that would support mission critical applications.