计算机科学
光通信
电子工程
计算机网络
集成光学
实时计算
三维光学数据存储
光学滤波器
波分复用
钥匙(锁)
数据传输
光开关
光纤
光子学
信号处理
光交叉连接
光传送网
光突发交换
网络拓扑
光学性能监测
电信网络
作者
Qian Kong,Jianning Su,Xiaowei Lu,Guiyuan Zhang,Liming Liu
摘要
As artificial intelligence (AI) and high-performance computing (HPC) clusters scale, interconnect performance increasingly becomes a primary bottleneck. Small-world networks offer logarithmic average shortest-path length (ASPL), yet realizing random long-range shortcuts in physical data centers is constrained by structured cabling and hard per-node limits. This paper presents the hierarchical hybrid optical network (HHON), a hierarchical hybrid optical network that overlays a regular fiber skeleton with non-planar free-space optical (FSO) shortcuts. A dual-track architecture is introduced to decouple logical-topology stability from physical shortcut placement, together with a global-aware greedy placement rule that combines hop-based gain with a stress-centrality signal approximated via k-Pivots. On hexagonal grids up to 8827 nodes, HHON exhibits small-world-like behavior: ASPL maintains logarithmic scaling with network size and stays within 11.7%–18.7% above an Erdős–Rényi baseline, while the Watts–Strogatz small-worldness metric σ increases from 6.23 to 171.36. Discrete-event simulations under stylized all-to-all and ring all-reduce workloads further suggest stable end-to-end delay within the evaluated scales and settings.
科研通智能强力驱动
Strongly Powered by AbleSci AI