Parallel Wavelength Routing Optical Interconnects for High-Performance Computing Systems

This dissertation explores the potential optical devices in the form of an Arrayed Waveguide Grating Router (AWGR) to address low-latency scalable communication need inside large-scale multi-processor computing systems. This dissertation proposes the Low-latency Interconnect Optical Network Switch (LIONS) to realize scalable high-throughput data plane interconnections. The LIONS exploits wavelength parallelism supported by a switching fabric based on the AWGR to accelerate contention resolution. The LIONS adopts electrical loopback buffers so that delayed packets can access the switching resources whenever they are available. The LIONS utilizes a two-stage link layer flow control to effectively eliminate packet drop due to contention in the switch. Simulation results indicate that the LIONS exhibits lower latency and higher throughput even at high input loads compared with electronic switches or previously proposed optical switch architectures. In addition to the studies in the data plane network, this dissertation proposes the Generic AWGR-based optical Global Communication Network (G-AGCNet) to support efficient global communications that are required in many parallel computing primitives, such as barrier synchronization and reduction. Wavelength parallelism and wavelength routing allows the G-AGCNet to realize connectivity as a multi-ary multi-fly flattened butterfly network while the wiring overhead is similar to a 2-ary multi-fly flattened butterfly network. Analyses show that the G-AGCNet is more power efficient than the electrical counterpart and completing one barrier synchronization for one million processors connected by the G-AGCNet takes only hundreds of nanoseconds .