Data-centre traffic optimisation using software-defined networks

Traffic demands in data centres are steadily increasing. Simple, yet commonly used round-robin algorithms are not able to meet these demands since they distribute the demands unevenly. Manually planning and redistribution, also called traffic engineering (TE), alleviates this problem. Traffic engineering is common in wide area networks (WANs). The high amount of manual work and associated cost in special hardware makes traffic engineering unpractical in data centre networks. This thesis looks into the question whether software-defined networking (SDN) can be used to bring traffic engineering to the data centre. To reach this goal, the amount of manual labour and requirements for hardware need to be reduced. I first show that label switching, which enables better traffic distribution, is possible with SDN without special hardware. I present SynRace that infers network status by combining probes with SDN. The GlobalFIB framework allows applications to report directly their demands to a central instance. Both allow traffic redistribution to be automated. With running multiple SDN application on the network and trying to fulfil all their demands and rules, conflicts are inevitable. In this thesis I analyse the nature of these conflicts and possibilities to handle them. The NetIDE core implements this conflict resolution and dynamic reconfiguration of the network.