HiveNet: a high-performance, hybrid, and cost-effective architecture for data center networks

Abstract

Network interconnection critically impacts the performance of data centers (DCs) and high-performance computing (HPC) systems, with scalability becoming vital as computing demands grow. This necessitates interconnection architectures that meet stringent latency, bandwidth, cost, and power consumption requirements. Optical interconnections provide cost-efficiency, reduced power consumption, and scalability to fulfill bandwidth needs. However, optical switches lack optical buffers, complicating the operation of all optical networks. To this end, we propose HiveNet, a novel hybrid interconnect architecture based on dual-port nodes and arrayed waveguide grating routers (AWGRs). HiveNet integrates low-radix electrical switches at lower layers to reduce cable complexity and construction costs, while AWGR-based optical connections at upper layers ensure fast switching and high bandwidth. The dual-port capability enables robust fault tolerance and supports all communication types (node-switch and node-node). Furthermore, a customized routing algorithm significantly enhances performance. Simulations conducted under various traffic patterns demonstrate that HiveNet achieves controlled delay and superior aggregate throughput. For large-scale networks (with 104,976 nodes at 10 Gb/s), HiveNet reduces construction costs by 49.3%, 26.4%, 32.7%, 54.1%, and 59.3% compared to Fat-Tree, H-LION, Leaf-Spine, BCube, and Lotus, respectively. Additionally, HiveNet decreases power consumption by 34.8%, 48.2%, 29.8%, and 23.1% compared to Fat-Tree, BCube, Leaf-Spine, and Lotus, respectively.