Performance analysis of multi-class Internet traffic classifier in a connection oriented router environment

In this work, we analyze the performance of multi-class Internet traffic classifier primarily in a connection-oriented IP router environment. We define the tasks and related concepts of traffic classification in the Internet and then proceed to construct a multi-class traffic classifier using the Learning Vector Quantization algorithm classifier that has been previously used to divide the traffic into two classes. We show how the functionality of the 2-class LVQ classifier can easily be extended to an arbitrary amount of classes, in this work to three: the hard-interactive, the elastic and the best effort service classes. This is done by consecutive 2-class classifications and unanimous decisions on classification. The multi-class LVQ classifier is observed to perform adequately being able to provide network service profiles, sets of classified services, with characteristics relating to respective classes. The multi-class classifier also performs well within the hardware performance limits and is still able to provide the forwarding component of the Internet router a substantial workload reduction. We also observe the individual service class performance and conclude that separately defined service architectures are needed to limit the use of QoS resources in the network. Our belief is that in order to provide end-to-end QoS a connection oriented link layer technology is needed. These technologies are able to offer dedicated forwarding paths for individual packets thus reducing the processor intensive hop count and flattening the network.