发表机构
Shanghai Jiao Tong University（上海交通大学）
机构由 AI 辅助整理，请以论文原文为准。AI 中文总结
THz-SynC是结合上下文博弈辅助协调的新框架，通过为分布式AI训练数据中心的集体通信合成定制拓扑、动态分配资源，在动态网络下实现了更优的延迟-能耗权衡，性能优于多种基线方案。
AI 中文摘要
太赫兹（THz）无线互连具备高容量、低延迟和高能效，可实现按需连接重构，是通信密集型分布式AI训练数据中心中光网络结构的理想补充。然而，由于光拥塞动态变化、THz信道波动以及异构计算掉队节点的存在，协同优化光与THz资源以最小化集体通信完成时间和传输能耗仍具挑战性。现有可重构数据中心设计主要优化网络拓扑和流量路由，较少考虑混合网络结构上分布式AI工作负载的集体通信语义。为应对这些挑战，本文提出THz-SynC，这是一种将集体合成与上下文博弈辅助的混合网络协调相结合的新框架，用于优化集体通信完成时间与传输能耗之间的权衡。THz-SynC利用集体特有的语义，为全对全（All-to-All）和全归约（AllReduce）模式合成定制化通信拓扑，同时动态分配THz资源。此外，上下文博弈协调器会利用网络状态和集体语义的实时观测，自适应地在光和THz链路上路由通信块，并选择机架功率预算。基于轨迹的评估表明，THz-SynC在动态网络条件下优于仅有线、仅无线及混合基线，实现了更优的延迟-能耗帕累托前沿。
英文摘要
Terahertz (THz) wireless interconnects offer high-capacity, low-latency, and energy-efficient rack-to-rack links capable of on-demand connectivity reconfiguration, serving as a promising complement to optical fabrics for communication-intensive distributed AI training datacenters. However, co-optimizing optical and THz resources to minimize collective completion time and transmission energy remains challenging due to dynamic optical congestion, THz channel fluctuations, and heterogeneous compute stragglers. Existing reconfigurable data-center designs predominantly optimize network topology and traffic routing, with limited consideration of collective communication semantics in distributed AI workloads over hybrid fabrics. To address these challenges, we propose THz-SynC, a novel framework that integrates collective synthesis with contextual-bandit-assisted hybrid-fabric coordination to optimize the tradeoff between collective completion time and transmission energy. By exploiting collective-specific semantics, THz-SynC synthesizes tailored communication topologies for All-to-All and AllReduce patterns while dynamically allocating THz resources. Furthermore, a contextual-bandit coordinator adaptively routes communication chunks across optical and THz links and selects rack power budgets leveraging real-time observations of network states and collective semantics. Trace-driven evaluations show that THz-SynC outperforms wired-only, wireless-only, and hybrid baselines, achieving a superior delay-energy Pareto frontier under dynamic network conditions.
