Energy Efficient PON Data Centre Architectures

Passive Optical Network (PON) technology can provide energy-efficient, high capacity, low cost, scalable, and flexible designs to support connectivity inside modern data centres. The main focus of this thesis is to implement and optimise passive optical network architectures to resolve many issues in current data centre designs such as high cost and high power consumption resulting from the large number of access and aggregation switches used in the connectivity, switch oversubscription and unbalanced traffic in data centres. Two PON data centre designs were considered in this thesis. The first is a server-centric PON-based data centre network that uses Network Interface Cards in the servers so they forward other servers traffic. This architecture provides multiple paths between servers and hence provides high resilience. Experimental evaluations were carried out for this architecture considering the delay in end-to-end communication which was between two instances of the data centre connected by a core network in the middle and access networks. This work introduced for the first time an experimental set up for the architectures in the passive optical-based data centre networks patent [9].It provided a platform to test how latency is affected over long distances and different platforms. The results showed that the proposed architecture is efficient in lowering latency with latencies of as low as 0.39ms compared to conventional data centre architectures with latencies of 1ms to 4ms, which results in an overall latency reduction of 87.8%. The second architecture is an architecture that uses Time Division Multiplexing/Wavelength Division Multiplexing (TDM/WDM). This PON data centre architecture provides higher data rates at lower power consumption. A Mixed Integer Linear Programming (MILP) model was developed to optimise the resource allocation and assignment of time slots and wavelengths within the architecture. To reflect on the energy-efficiency of a WDM/TDM PON based architecture, a MILP model was developed to optimise the allocation of Virtual Machines (VMs) within the architecture with the objective of minimising the networking and processing power consumption, or minimising the processing power consumption only. Various numbers of VMs were considered in addition to the heterogeneity of the servers processing capacity and the VM processing demands . The results show that by minimising the total power consumption, a reduction of up to 70% in networking power consumption was achieved compared to minimising only the CPU power consumption. This resulted in an overall power consumption reduction of 82%, which shows that by effectively provisioning the allocation of resources within the proposed architecture challenges such as underutilisation of resources and high deployment cost can be mitigated.