As large-scale applications are demanding more computation power while Moore’s Law is slowing down, rack-scale computing systems are being developed to meet the increasing computation and energy requirements. With the mainstream paradigm in high-performance computer architecture being shifted considerably from single-core systems toward multi-core computing systems, system performance and energy efficiency are extensively challenged by the communication capacity among processing units and memories/storage. The emerging silicon photonics, on the other hand, promises exceedingly high bandwidth, low latency, and low energy consumption by enabling optical interconnects between entities. However, the inherent distinction between optical interconnects and electrical ones requires fundamental...[ Read more ]

As large-scale applications are demanding more computation power while Moore’s Law is slowing down, rack-scale computing systems are being developed to meet the increasing computation and energy requirements. With the mainstream paradigm in high-performance computer architecture being shifted considerably from single-core systems toward multi-core computing systems, system performance and energy efficiency are extensively challenged by the communication capacity among processing units and memories/storage. The emerging silicon photonics, on the other hand, promises exceedingly high bandwidth, low latency, and low energy consumption by enabling optical interconnects between entities. However, the inherent distinction between optical interconnects and electrical ones requires fundamentally different methods to compose a system-level communication network. We systematically analyze the rack-scale optical network (RSON) architecture with different path reservation schemes and optical inter-chip networks and the most commonly used architecture for high-performance computing systems. We explore the RSON architecture, floorplan optimized delta optical network (FODON) switch architecture and the preemptive chain feedback (PCF) scheme to optimize multi-domain path reservation. We also propose a systematic approach to optimizing the overall energy-delay product of the photonic backplane network for rack-scale computing systems. Experimental results show that the RSON with FODON switch and PCF scheme can improve system performance per energy consumption by up to 5x, and around 4x on average, while still maintaining better scalability than state-of-the-art systems. Also, the system with proposed 2-layer photonic backplane network with the floorplan optimization improves system performance by 3x and improve system performance per energy consumption by around 46 %, which can also maintain good scalability.