Traffic Engineering Technique for Data Center Networks 
Traffic engineering consists in improving the performance of the telecomunication
networks which is evaluated by a large number of criteria. The ultimate objective is
to avoid congestion in the network by keeping its links from being overloaded. In
large Ethernet networks with thousand of servers, such as data centers, improving
the performance of the traditional switching protocols is a crucial but very
challenging task due to an exploration in the size of solution space and the
complexity. Thus, exact methods are inappropriate even for reasonable size
networks.
Local Search (LS) is a powerful method for solving computational optimization
problems such as the Vertex Cover, Traveling Salesman, or Boolean Satisfiability.
The advantage of LS for these problems is its ability to find an intelligent path from a
low quality solution to a high quality one in a huge search space. In this thesis, we
propose different approximate methods based on Local Search for solving the class
of traffic engineering problems in data center networks that implement Spanning
Tree Protocol and Multiple Spanning Tree Protocol.
First, we tackle the minimization of the maximal link utilization in the Ethernet
networks with one spanning tree. Next, we cope with data center networks
containing many spanning trees. We then deal with the minimization of service
disruption and the worst-case maximal link utilization in data center networks with
many spanning trees. Last, we develop a novel design of multi-objective algorithms
for solving the traffic engineering problems in large data centers by taking into
account three objectives to be minimized: maximal link utilization, network total load
and number of used links.
Our schemes reduce significantly the size of the search space by releasing the
dependence of the solutions from the link cost computation in order to obtain an
intended spanning tree. We propose efficient incremental techniques to speed up
the computation of objective values. Furthermore, our approaches show good
results on the credible data sets and are evaluated by the strong assessment
methods.
Membres du jury :
Prof. Yves DEVILLE (UCL), Promoteur
Prof. Olivier BONAVENTURE (UCL), Promoteur
Prof. Pierre DUPONT (UCL), Président
Prof. Bernard FORTZ (ULB), Secrétaire
Prof. Guy LEDUC (ULg)
Prof. Pierre FRANÇOIS (IMDEA, Espagne)
