The advance of internet applications, such as video streaming, big data and cloud computing, is reshaping the telecommunication and internet industries. Bandwidth demands in datacentres have been boosted by these emerging data-hungry internet applications. Regarding inter- and intra-datacentre communications, fine-grained data need to be exchanged across a large shared memory space. Large-scale high-speed optical switches tend to use a rearrangeably non-blocking architecture as this limits the number of switching elements required. However, this comes at the expense of requiring more sophisticated route selection within the switch and also some forms of time-slotted protocols. The looping algorithm is the classical routing algorithm to set up paths in rearrangeably non-blocking switches. It was born in the electronic switch era, where all links in the switches are equal. It is, therefore, not able to accommodate loss difference between optical paths due to the different length of waveguides and distinct numbers of crossings, and bends, leading to sub-optimal performance. We, therefore, propose an advanced path-selection algorithm based on the looping algorithm that minimises the path-dependent loss. It explores all possible set-ups for a given connection assignment and selects the optimal one. It guarantees that no individual path would have a sufficiently substantial loss, therefore, improve the overall performance of the switch. The performance of the proposed algorithm has been assessed by modelling switches using the VPI simulator. An 8×8 Clos-tree switch demonstrates a 2.7dB decrease in loss and 1.9dB improvement in IPDR with 1.5 dB penalty for the worst case. An 8×8 dilated Beneš shows more than 4 dB loss reduction for the lossiest path and 1.4 dB IPDR improvement for 1 dB power penalty. The improved algorithm can be run once for each switch design and store its output in a compact lookup table, enabling rapid switch reconfiguration. Microelectromechanical systems (MEMS) based optical switches have been fabricated with over 1,000 ports which meet the port count requirements in data centre networks. However, the reconfiguration speed of the MEMS switches is limited to the millisecond to microsecond timescale, which is not sufficient for packet switching in datacentres. Opto-electronic devices, such as Mach-Zehnder Interferometers (MZIs) and semiconductor optical amplifiers (SOAs) with nanosecond response time show the potential to fulfil the requirements of packet switching. However, the scalability of MZI switches is inherently limited by insertion loss and accumulated crosstalk, while the scalability of SOA switches is restricted by accumulated noise and distortion. We, therefore, have proposed a dilated Beneš hybrid MZI-SOA design, where MZIs are implemented as 1×2 or 2×1 low-loss switching elements, minimising crosstalk by using a single input, and where short SOAs are included as gain or absorption units, offering either loss compensation or crosstalk suppression though adding only minimal noise and distortion. A 4×4 device has been fabricated and exhibits a mere 1.3dB loss, an extinction ratio of 47dB, and more than 13dB IPDR for a 0.5dB power penalty. When operating with 10 Gb/s per port, 6pJ/bit energy consumption is demonstrated, delivering 20% reduced energy consumption compared with SOA-based switches. The tolerance of the current control accuracy of this switch is very broad. Within a 5 mA bias current range, the power penalty can be maintained below 0.2 dB for 8 dB IPDR and 12 mA for 10 dB IPDR with a penalty less 0.5 dB. The excellent crosstalk and power penalty performance demonstrated by this chip enable the scalability of this hybrid approach. The performance of 16×16 port dilated Beneš hybrid switch is experimentally assessed by cascading 4×4 switch chips, demonstrating an IPDR of 15 dB at a 1 dB penalty with a 0.6 dB power penalty floor. In terms of switches with port count larger than 16×16, the power penalty performance has been analysed with physical layer simulations fitted with state-of-the-art data. We assess the feasibility of three potential topologies, with different architectural optimisations: dilated Beneš, Beneš and Clos-Beneš. Quantitative analysis for switches with up to 2048 ports is presented, achieving a 1.15dB penalty for a BER of 10-3, compatible with soft-decision forward error correction.