Optical Wireless Communication-Enabled Data Centre Networks

Abstract

Data Centres (DCs) are witnessing significant traffic surges due to the increasing popularity of cloud computing-based applications. To handle this a surge, Data Centre Network (DCN) designs are required to provide high throughput, reconfigurability, scalability, energy efficiency, and low latency. To satisfy these requirements, technological and structural advancements and effective placement of workload are required in DCs. Optical Wireless Communication (OWC) offers potential solutions for the challenges faced by DCNs by enabling improved scalability and flexibility, reducing wiring complexities, and lowering the power consumption while providing high bandwidth. Passive Optical Networks (PONs) have also been considered for DCs as they offer high bandwidth, flexibility, cost-effectiveness, and energy efficiency due to the use of passive components. In this thesis, we propose incorporating OWC and PON technologies into next generation spine-and-leaf DCNs. OWC is used to connect the racks of a DC through Wavelength Division Multiplexing (WDM) Infrared (IR) transceivers that are placed at the top of the racks and at distributed Access Points (APs) in the ceiling. Each transceiver at a rack is connected to a leaf switch that connects the servers within the rack. The transceivers use Angle Diversity Transmitters (ADTs) and Angle Diversity Receivers (ADRs) to enable connecting each rack with all APs. We propose two PON designs to replace the spine switches by providing passive connectivity between APs and between APs and the Optical Line Terminal (OLT) that connect the DCs to external networks. The first PON design uses Arrayed Waveguide Grating Routers (AWGRs) to connect APs to each other and connect APs to the OLT while the second PON design uses Network Interface Cards (NIC) to connect APs to each other and to connect subsets of the APs to the OLT. In addition, we present channel modelling for the uplinks and downlinks for a DC with eight racks and ADTs with ADRs showing data rates between 10 and 20 Gbps for uplink and between 7.5 and 16 Gbps for downlink. Additionally, we study the impact of air turbulences resulting from the temperature variations in temperature on the channel performance. Moreover, we optimise the routing and wavelength assignment in the AWGR-based PON using Mixed Integer Linear Programming (MILP). Furthermore, we develop a MILP model with the objective to maximise the data rates in both the uplink and downlink directions and determine the optimal paths between the racks and between racks and the OLT based on the achievable data rates. We benchmark the power consumption of the proposed OWC-PON-based spine-and-leaf DC designs against the traditional spine-and-leaf DC. Considering eight racks, the results show that in comparison to iii traditional spine and leaf architectures, the AWGR-based architecture minimised the power consumption by 50%, and the AP-to-AP architecture minimised it by 46%. Furthermore, we further develop a MILP model to optimise Virtual Machine (VM) placement in the proposed OWC-PON-based spine-and-leaf DCs architectures to minimise the power consumption of networking and processing. Our results show that jointly minimising the networking and processing power consumption, decreases the networking power consumption by 75% compared to minimising the processing power consumption only. In this work, the proposed model serves as a pivotal contribution., providing a valuable framework for evaluating and benchmarking different architectures and settings.