Performance analysis of high-speed optical packet switching in high performance computing and datacentre networks

The massive growth in datacentre traffic, due to a huge increase in the deployment of data-intensive applications, is forcing datacentre infrastructure to migrate away from conventional electronic packet-switched networks, where capacity scaling imposes significant financial and technical constraints, and to evolve towards more advanced architectures. Motivated by this, new optical switching technologies and networking architectures, capable of providing very large bandwidth capacity, high scalability, high switching speed and high energy efficiency, are being targeted for building next-generation high-performance datacentre and High Performance Computing (HPC) networks. Optical packet switching technology is considered to be a long-term solution to meet these design challenges, as it can exploit fully the enormous potential capacity enabled by optics and support high switching flexibility at packet level. In this thesis, new wavelength-routed optical packet switching architectures, which exploit the functionalities of key optical switching components such as Tunable Wavelength Converters (TWCs), Arrayed Waveguide Gratings (AWGs) and Wavelength Selective Switches (WSSs), to realise all-optical switching, are proposed for use as next-generation datacentre and HPC networks. To maximise the efficiency of the proposed optical switching architecture, a dynamic bandwidth provisioning algorithm, which allocates switch resources to traffic demands based on application requirements, is developed to further enhance network flexibility, resource utilisation and network performance. Moreover, based on the proposed switch architectures, a large-scale high-performance datacentre network with flexible central control is modelled, with a view of determining the optimal network topology and traffic scheduling methods. This flexible network architecture employs a modular design and combines transparent optical packet switches, based on Arrayed Waveguide Grating (AWG) routers, and a hybrid congestion control scheme using recirculating Fibre Delay Lines (FDLs) along with novel packet retransmission schemes. The work carried out in this thesis indicates that the proposed network structure not only provides high scalability, capable of hosting hundreds of thousands of severs, but also delivers high bandwidth utilisation and network provisioning flexibility. The network offers a promising and viable networking solution to address current and future application needs in datacentre and HPC environments.