Optical Packet Switched (OPS) interconnection networks are a promising means of routing high-bandwidth optical data messages between shared memory elements and processors. Depending on the OPS network size, applications can range from data communications and storage to High Performance Computing (HPC) systems. Maintaining routed packet traffic in the optical domain enables full exploitation of the bandwidth provided by Wavelength Division Multiplexing (WDM). This thesis first presents a functional OPS interconnection network based on the data vortex architecture, subsequently used in this work as a platform study investigating the data plane interface to the physical layer. Utilizing Semiconductor Optical Amplifier (SOA)-based switching nodes and conventional fiber-optic technology, a fully implemented 12 × 12 data vortex was designed, demonstrating routing capacity of 160 Gb/s with a median port-to-port latencies of 110 ns.

One class of proposed OPS network topology, Multistage Interconnection Networks (MINs), enables scaling the number of ports using simple switching nodes. The number of stages is proportional to log₂(N) for an N × N interconnection network. In such systems, the routing effectiveness of the switching node directly impacts the network scalability. An optimized robust 2 × 2 switching node was designed, resulting in 40% faster switching times compared to current technology. The use of the novel implemented switching node results in a 100-fold increase in the number of interconnected ports of MINS, such as the data vortex.

Large-scale HPC systems rely on ultra-low communication latency with short and bursty message exchanges among thousands of interconnected elements. In this part of the work, bit-parallel message exchange demonstrated reduce processor/memory access time by exploiting the high degree of parallelism afforded by WDM. A sequence of 32 messages are entirely recovered and processed at the destination node using an embedded clock signal with a measured clock-to-data skew tolerance window of 0.375 UI.

The critical measure of physical layer scalability is investigated for SOA-based OPS interconnection networks. Through a recirculating loop testbed environment, experiments show that bit-error rates (BERs) below 10 ⁻⁹ can be maintained through 58 switching nodes for the entire eight-channel 10-Gb/s-per-channel payload distributed over 24.2 nm of the C-band. This is sufficient for a heavily loaded 10k × 10k port interconnection network. The recirculating loop testbed also enables the investigation of the impact of cumulative polarization dependent gain (PDG) in MINs. It is shown that for nodes based on commercial SOA switches with PDG of less than 0.35 dB, the maximum number of cascaded nodes changes by as much as 20 nodes, depending on both the packet wavelength and its state of polarization. Proposed approaches to minimize cumulative PDG, such as an optimized channel placement, enable more robust OPS interconnection networks.