Search

Emerging applications create tremendous volume of data traffic every day, increasingly stressing the capabilities of the networks hosting these applications. Understanding the communication patterns of these applications can significantly benefit the corresponding network management tasks. Motivated by today’s wide use of data analysis techniques, this thesis studies how to perform effective network management through using network data analysis. We investigate problems in managing the QoS targets and resources in enterprize networks and datacenter networks, presenting two applications of traffic analysis. The first application of traffic analysis addresses the utilisation of traffic classification techniques in enterprize networks to differentiate servicing flows of different classes or applications. The traffic classification techniques developed based on statistical analysis produce advantages over the existing techniques based on deep packet inspections or port matching in the scenarios that new applications emerge and packets are encrypted. This dissertation presents enhanced traffic classification techniques based on traffic analysis. The second application addresses the requirement of effective resource allocation in datacenters, where predictable performance is essential to the applications hosted in datacenters. This thesis proposes resource allocation techniques that can allocate datacenter resources in a way that the QoS requirements of Virtual Machines (VM) are met while minimizing the amounts of required resources. This is achieved through the integrated use of bin packing algorithms, effective bandwidth estimation techniques and software-defined networking technologies, where network data are analyzed to identify the minimum amounts of bandwidth required to meet the QoS targets for provisioning VMs on demand.

The knowledge hidden in evolving data may change with time, this issue is known as concept drift. It often causes a learning system to decrease its prediction accuracy. Most existing techniques apply ensemble methods to improve learning performance on concept drift. In this paper, we propose a novel meta learning approach for this issue and develop a method: Multi-Step Learning (MSL). In our method, a MSL learner is structured in a recursive manner, which contains all the base learners maintained in a hierarchy, ensuring the learned concepts are traceable. We evaluated MSL and two ensemble techniques on three synthetic datasets, which contain a number of drastic concept drifts. The experimental results show that the proposed method generally performs better than the ensemble techniques in terms of prediction accuracy.