Energy efficient composable data centres

There is a proliferation of the number of operational data centres (DCs) across the globe to meet present and future demands for on-demand computational offerings. In spite of the many efforts to improve the utilisation and power efficiency of traditional DCs, results achieved remain sub-optimal. This is primarily because of the rigid utilisation boundaries of traditional server architecture. Disaggregation of server resource components and dynamic orchestration of such resources over a suitable network has been proposed to improve efficiency of next generation composable DCs. This thesis conducts a study on the best setup for such composable DC infrastructure to achieve optimal energy efficiency. First, we formulate a mixed integer linear programming (MILP) model to investigate the optimal scale and scope of resource disaggregation for energy efficient composable DCs. Concurrently, we also investigate the most suitable network for optimal energy efficiency. By placing CPU and memory intensive workloads energy efficiently in different composable DCs, we found that implementing logical disaggregation at rack-scale in composable DCs that adopt all-optical network enables optimal energy efficiency. Physical resource disaggregation of traditional DC servers at rack-scale leads to up to 8% and 20% savings in overall power consumption when CPU intensive and memory intensive workloads are provisioned respectively. We also found that adoption of micro-service architecture in conjunction with the logical disaggregation and rack-scale resource disaggregation can further improve efficiency in composable DCs. A combination of disaggregation and micro-services enabled optimal resources utilisation and energy efficiencies. Thus, relative to the traditional DC up to 23% reduction in the total power consumption is achieved by combining both approaches. Secondly, we describe two variants of a practical and scalable network for composable DC that leverages optical technologies and techniques. Additionally, we formulate a MILP model to evaluate the performance of the novel network in rack-scale composable DCs that implement different forms of disaggregation. The electrical-optical variant of the novel topology achieves similar performance as a reference network while utilising fewer transceivers per compute node. The targeted adoption of optical technologies by both variants of the proposed network achieves greater (4 - 5 times greater) utilisation of available network throughput than the reference network which implemented a generic design. Furthermore, we also formulate a MILP model and develop a comparable heuristic to study the benefits of adopting server disaggregation in the fog computing tier. We evaluate the energy efficient placement of interactive apps in a future fog 6G network in our study. Relative to the present practice of deploying traditional servers in the fog computing layer, adoption of disaggregated servers reduces total fog computing power consumption by up to 18% when a network with low delay penalty is considered. Finally, we recommend that logical disaggregation and rack-scale disaggregation should be implemented in composable DCs that desire energy efficiency. This is because of the advantages and flexibility that both approaches jointly offer as reported in this thesis. We also recommend the targeted use of optical network technologies and techniques. Relative to a general-purposed design, this provides a more efficient approach to mitigate network challenges of composable DCs. Furthermore, these recommendations should be extended to the fog computing tier and edge of future networks to enable greater energy efficiency of the cloud-of-things architecture.