智算中心光电混合组网下的流量差异化调度技术研究

摘要：光电混合组网凭借“电交换的灵活可控”与“光交换的高带宽低时延”双重特性，已成为智算中心网络的核心发展趋势。聚焦智算中心光电混合组网场景，深入分析AI训练流量的差异化特征，并提出一种基于流量特征的差异化调度机制。该机制通过构建“感知—决策—执行—反馈”的闭环调度框架，实现了对数据并行（DP）流量、流水线并行（PP）流量等多种类型流量的精准识别与链路优势匹配。研究成果为智算中心光电混合组网下的流量差异化调度提供了理论依据与实践参考。

关键词：智算中心；光电混合组网；差异化调度；AI训练

Abstract: Optical-electrical hybrid networking, leveraging the dual advantages of flexible controllability of electrical switching and high-bandwidth low-latency of optical switching, is a key development trend in data center networks for artificial intelligence (AI). Focusing on the optical-electrical hybrid networking scenario in AI data centers, the differentiated characteristics of AI training traffic are analyzed in depth, and a traffic-characteristic-based differentiated scheduling mechanism is proposed. By constructing a closed-loop scheduling framework comprising perception, decision, execution, and feedback, accurate identification of different traffic types—including data parallelism (DP) and pipeline parallelism (PP) traffic—as well as matching with the advantages of optical or electrical links, is achieved. The findings provide both theoretical foundations and practical references for differentiated traffic scheduling in optical-electrical hybrid networks within AI data centers.

Keywords: intelligent computing center; optical-electronic hybrid networking; differentiated scheduling; AI training