Hi all, Packet-granular adaptive routing, also known as congestion-aware packet spray, has been widely recognized as an ideal load-balancing mechanism for AI Ethernet networks. Some cloud providers have implemented their in-house packet spray approaches, which are mainly built on proactive and real-time congestion detection along all possible ECMP paths.
In order to achieve a non-blocking network fabric, it seems more suitable for network switches for perform packet spray since they could obtain the information about network congestion between switches more quickly and easily. Some major network chip vendors are developed their proprietary congestion notification mechanisms built on their proprietary data-plane signaling. However, to meet the aim of the UEC to deliver an Ethernet-based open, interoperable, high-performance full-communications stack for the growing network demands of AI and HPC at scale, it is meaningful for us to pursue an open standard-based approach for packet spray. This draft is a step towards that goal indeed, any comments and suggestions are welcome. Best regards,rt Xiaohu 发件人: [email protected] <[email protected]> 日期: 星期一, 2024年1月29日 16:36 收件人: Hang Wu <[email protected]>, Hongyi Huang <[email protected]>, Junjie Wang <[email protected]>, Qingliang Zhang <[email protected]>, Xiaohu Xu <[email protected]>, Yadong Liu <[email protected]>, Yinben Xia <[email protected]>, Zongying He <[email protected]> 主题: New Version Notification for draft-xu-lsr-fare-01.txt A new version of Internet-Draft draft-xu-lsr-fare-01.txt has been successfully submitted by Xiaohu Xu and posted to the IETF repository. Name: draft-xu-lsr-fare Revision: 01 Title: Fully Adaptive Routing Ethernet Date: 2024-01-29 Group: Individual Submission Pages: 9 URL: https://www.ietf.org/archive/id/draft-xu-lsr-fare-01.txt Status: https://datatracker.ietf.org/doc/draft-xu-lsr-fare/ HTMLized: https://datatracker.ietf.org/doc/html/draft-xu-lsr-fare Diff: https://author-tools.ietf.org/iddiff?url2=draft-xu-lsr-fare-01 Abstract: Large language models (LLMs) like ChatGPT have become increasingly popular in recent years due to their impressive performance in various natural language processing tasks. These models are built by training deep neural networks on massive amounts of text data, often consisting of billions or even trillions of parameters. However, the training process for these models can be extremely resource- intensive, requiring the deployment of thousands or even tens of thousands of GPUs in a single AI training cluster. Therefore, three- stage or even five-stage CLOS networks are commonly adopted for AI networks. The non-blocking nature of the network become increasingly critical for large-scale AI models. Therefore, adaptive routing is necessary to dynamically load balance traffic to the same destination over multiple ECMP paths, based on network capacity and even congestion information along those paths. The IETF Secretariat
_______________________________________________ rtgwg mailing list [email protected] https://www.ietf.org/mailman/listinfo/rtgwg
