Traffic Class Prediction and Prioritization on a Diversified IP Network Using Machine Learning

As more homes are becoming broadband enabled, a huge strain is placed on the underlying IP infrastructure. In being able to predict the types of traffic that nodes might generate, preference can be given to high priority, latency sensitive traffic. Traditionally, a home had a single computer utilizing a link to the Internet. In recent years this single computer was replaced by a multitude of smart devices with the same ultimate goal: embedding the user into the social fabric of the Internet. With the home still connected via a single link, these nodes are all contending for bandwidth and a novel QoS scheme is needed. In this paper, various models characterizing packet flow on a network will be developed. These are subsequently used to predict as accurately as possible, the immediate future traffic classes to be expected from nodes. This will eventually be used as a dynamic QoS criteria for pre-emptively assigning transmission slots to specific hosts. A number of techniques to accomplish this important task of node traffic pattern prediction, are presented. Lastly, evaluation of our approach, based on an 802.11b/g SOHo implementation, is also covered.