Research on Key Technologies for Carrying AI Computing Services in Cloudified IP Metropolitan Area Network

Hu Zehua; Zhu Yongqing; Gong Xia; Lai Daoning; Zhang Fan

doi:10.11959/j.issn.1000-0801.DXKX260138

您当前的位置：

首页 >

文章列表页 >

Research on Key Technologies for Carrying AI Computing Services in Cloudified IP Metropolitan Area Network

更新时间：2026-06-23

- Research on Key Technologies for Carrying AI Computing Services in Cloudified IP Metropolitan Area Network
- Telecommunications Science (2026)
- 作者机构：
  
  中国电信研究院，广州 510630
- 作者简介：
- 基金信息：
- DOI：10.11959/j.issn.1000-0801.DXKX260138
  CLC： TP393
- Received：01 March 2026，
  
  Revised：2026-04-24，
  
  Accepted：23 June 2026，
- 稿件说明：
移动端阅览
Hu Zehua, Zhu Yongqing, Gong Xia, et al. Research on Key Technologies for Carrying AI Computing Services in Cloudified IP Metropolitan Area Network[J/OL]. Telecommunications Science, 2026.
DOI：

Hu Zehua, Zhu Yongqing, Gong Xia, et al. Research on Key Technologies for Carrying AI Computing Services in Cloudified IP Metropolitan Area Network[J/OL]. Telecommunications Science, 2026. DOI： 10.11959/j.issn.1000-0801.DXKX260138.

摘要

随着人工智能大模型训练与推理业务向城域网延伸，城域网络在承载高突发、高敏感的智算业务流量上面临严峻挑战。为此，本文提出以城域云网为底座、以大象流子流级负载分担与精细化流控技术为核心技术的协同增强方案。该方案依托城域云网弹性组网、无阻塞转发等架构优势，以及面向RDMA流量的细粒度负载均衡和精细化流控技术创新，有效保障了智算业务的高效、无损承载。本文基于路由器设备开展组网验证，实验表明该方案在云边协同推理等典型场景下可实现百公里拉远算效下降在4%以内。

Abstract

As Artificial Intelligence (AI) Large Language Model (LLM) training and inference services extend into Metropolitan Area Networks (MANs)

MANs face significant challenges in carrying highly bursty and sensitive AI workload traffic. To address these challenges

this paper proposes a Cloudified IP MAN-based solution driven by two key mechanisms: subflow-level elephant flow load balancing and fine-grained flow control. Leverage architectural advantages of Cloudified IP MAN such as elastic networking and non-blocking forwarding

the proposed solution introduces RDMA-oriented fine-grained load balancing and congestion control mechanisms to ensure efficient and lossless transmission of AI training and inference services. This solution is implemented on routers and conducted networking validation. Experimental results demonstrate that the proposed solution limits computational efficiency degradation to within 4% over transmission distances of up to 100 km.

关键词

Keywords

references

国务院 . 关于深入实施人工智能+行动的意见 [EB ] . 2025 .

侯伟胜 . 算力租赁浪潮来袭,引领数字化创新 [J ] . 商业观察 , 2025 , 11 ( 22 ): 9 - 12 .

吴楠 , 雷波 , 唐静 . 运营商智算网络技术研究与实践 [J ] . 邮电设计技术 , 2025 ,( 03 ): 1 - 6 .

中国电信股份有限公司研究院 , 华为技术有限公司等 . 算力城域网白皮书（2025） [R ] . 广州：中国电信股份有限公司研究院 , 2025 .

H. K. Dhaliwal and C . -H. Lung. Load Balancing Using ECMP in Multi-Stage Clos Topology in a Datacenter [J ] . IEEE Conference on Dependable and Secure Computing , 2018 .

X. Zhong , J. Zhang , Y. Zhang , et al . PACC: Proactive and Accurate Congestion Feedback for RDMA Congestion Control [J ] . IEEE INFOCOM , 2022 .

Yibo Zhu , Haggai Eran , et al . Congestion Control for Large-Scale RDMA Deployments [J ] . SIGCOMM , 2015

龚霞 , 陈华南 , 朱永庆 , 袁世章 , 杨冰 . 城域云网中vBRAS池化部署研究 [J ] . 电信科学 , 2023 , 39 ( 1 ): 146 - 152 .

尹远阳 , 王志中 , 姜有强 , 孙嘉琪 , 梁筱斌 . 面向新型城域网的UMR技术方案研究与应用 [J ] . 电信科学 , 2022 , 38 ( 2 ): 149 - 156 .

Yiren Zhao , Ran Shu , et al . SRC: A Scalable Reliable Connection for RDMA with Decoupled QPs and Connections [J ] . In Proceedings of the 9th Asia-Pacific Workshop on Networking , 2025 .

Calvignac J L , Goetzinger W J , Handlogten G H , et al . High speed network processor [J ] . 2002 .

孟丽珠 , 宋盈 , 陈燕 . SRv6 Policy业务可靠性方案探讨 [J ] . 邮电设计技术 , 2024 ( 4 ): 66 - 70 .

RFC . IPv6 Flow Label Specification [R ] . 2011

IEEE Standard for Local and Metropolitan Area Networks--Bridges and Bridged Networks , IEEE Std 802.1Qbb , 2011 .

RFC . Internet Control Message Protocol (ICMPv6) for the Internet Protocol Version 6 (IPv6) Specification . 2006 .

王波 . 5G SA的网络架构和关键技术 [J ] . 移动通信 , 2020 , 44 ( 1 ): 9 .

Views

下载量

CSCD

Alert me when the article has been cited

提交

Tools

Publicity Resources

Multi-service routing algorithm for satellite networks based on deep Q-network with long short-term memory

GhostMamba-SAN: an efficient DDoS attack detection model

FANET routing optimization method based on CPO algorithm

ARDR: an adaptive service selection and robust decision framework for heterogeneous executor

Optimization method of deep time series prediction model based on meta-heuristic RIME algorithm

Related Author

No data

Related Institution

No data

⁰