Adaptive Flow Classification in IP Switching – The Measurement Based Approach

In this work, we first briefly introduce the concept of IP flow classification on a general conceptual level. The intention is to rise above the technological details and create a conceptual point of view on flow classification and closely related issues. Then we move on to study and compare earlier flow classification methods such as the all and selected flow classifier and the packet count flow classifier. The comparison of these methods is done with actual network traffic and various performance metrics are presented. It is found that while the traditional methods of flow classification are found to reduce the resource usage of the network elements, they provide the user with an ambiguous traffic profile at the best. A measurement based learning approach to flow classification is then presented. We first introduce the list based flow classification algorithm to act as the reference point to the novel approach of using learning vector quantization in flow classification. It is found that both the list classifier and the learning vector quantization algorithm, when used in flow classification, require only moderate performance from the network elements while producing an intuitive and user-comprehensible traffic profile being able to adapt to traffic profile changes. The learning vector quantization flow classifier is more sensitive to changing network traffic profiles and functions somewhat more accurately than the list classifier. While all measurement-based approaches suffer the delay of analyzing the measurement data our results indicate that measurement-based approach to flow classification is able to provide users more accurate service profiles in changing traffic environment while stating reasonable performance demands to the network equipment.