Master thesis : Dynamic Network Flow Classification with Hardware offloading

Abstract

Nowadays, network applications and connected devices are calling for a better management of the network infrastructures. With the growth of network load, efficient network classification is playing a key role for providing quality of service.

In this document we present a network classification scheme that allow to contextualize a packet in its flow. The designed classifier is composed of four components: a static classifier, a dynamic classifier, a tag system and queues. The static classifier is responsible for matching the packets to the provided classification rules. The tag mechanism is elaborated to give to the user a mean of retrieving information about the flow of the processed packet. The dynamic classifier associate the packets with their tags and a queue system is used to forward the results to the user.

The implementation is provided under the form of a C library that ease the use of it in network low level applications. The design of it allow a flexible modularity and portability. The results show that our solution give scalability when a large amount of rules is used and allow fast dynamic classification. The use of lock-free algorithms to handle multithreading and concurrency allow to perform the classification of several packets simultaneously with efficiency. The designed tag mechanism allow an easy consultation of the flow of a processed packet. The hardware capabilities offered by advanced network hardware is studied and matched with the need of the solution, enabling possibilities of offloading it for efficiency.

Network classification is a problem hard to solve efficiently, many solutions are available depending on the situation and the requirements. In today's network, portable and scalable solutions are important and need a particular attention. The implementation of our solution give the opportunity for efficient dynamic network classification that ease the processing of packets in their contexts.