Hi RTG WG, In today’s meeting we presented draft-cheng-rtgwg-enhanced-ecmp-00 [1], and slides uploaded as [2]. The draft proposes two enhanced ECMP solutions for AIDC networks and shows clear training-speed gains:
1. Balanced GPU-send → group ECMP by ingress interface. 2. Balanced GPU-recv → group ECMP by Egress-Group. Both have been validated in practice. Authors would appreciate reviews/comments to help refine the draft. [1] https://datatracker.ietf.org/doc/html/draft-cheng-rtgwg-enhanced-ecmp-00 [2] https://datatracker.ietf.org/meeting/123/materials/slides-123-rtgwg-enhanced-ecmp-for-ai-cluster-00.pdf Thanks, Weiqiang Cheng
_______________________________________________ rtgwg mailing list -- [email protected] To unsubscribe send an email to [email protected]
