高性能智算网络架构探讨与实践

摘要：智算网络对超高吞吐、超低时延、超高可用性的极致性能需求，是传统网络架构的核心挑战和瓶颈。论述了一种新型高性能智算网络架构，其关键特征包括端网精细化协作、精准流控、无状态组播、轻量级在网计算等。基于智算网络的关键性能堵点，提出了一种新智算网络架构，其主要特征是端网双域协同、网侧定量反馈以及无状态组播，概述并分析了原型测试及主要结论数据。

关键词：智算网络；端网协同；无状态组播；精准流控；轻量级在网计算

Abstract: The extreme performance requirements of artificial intelligence network lie at super-throughput, super low-latency and super high availability, which turn out to be the key challenges and bottlenecks of the traditional network. This document illustrates a new AI network architecture with such key features as fine-grained host and network coordination, high-accuracy flow control, stateless multicast and lightweight in-network computing. Against the key network performance choke points, a new architecture perspective has been demonstrated with corresponding enhanced solutions and prototype testing as well as resulting data.

Keywords: artificial intelligence network, host and network coordination; stateless multicast; high accuracy flow control; lightweight in-network computing