Enabling Low Latency at Large-Scale Data Center and High-Performance Computing Interconnect Networks Using Fine-Grained All-Optical Switching Technology

In large-scale data center (DC) and high performance computing (HPC) interconnect networks, end-to-end latency becomes a fatal problem due to the processing and queuing delays of electronic packet switching (EPS) at intermediate switching points. Introducing optical switching into DC and HPC networks can provide a potential solution to the latency problem by establishing low-latency optical bypass (end-to-end lightpath). However, the number of connections that can be provided by current coarse-grained optical circuit switching (OCS) technology is far less than the required amount for all-to-all communication in a hundreds-of-thousands-nodes large-scale system, and this will weaken its effect in reducing latency. Optical packet switching has a much finer granularity compared with OCS; however, the lack of adequate technologies for optical buffering makes it difficult to avoid packet collision. In this paper, we investigate the relationship among network scale, granularity and latency, and introduce the use of our proposed fine-grained optical time slice switching (OTSS) in DC and HPC networks over arbitrary topologies. Simulation results under 6x6 2-D Torus topology demonstrate the advantage of OTSS in end-to-end latency compared with conventional EPS and spectrum-flexible wavelength switching (WS).