draft-wh-rtgwg-adaptive-routing-arn
Adaptive Routing Notification
Large-scale supercomputing and AI data centers utilize multipath to
implement load balancing and/or improve transport reliability.
Adaptive routing (AR), widely used in direct topologies such as
dragonfly, is growing popular in commodity data centers to
dynamically adjust routing policies based on path congestion and
failures. When congestion or failure occurs, the sensing node can
not only apply AR locally but also send the congestion/failure
information to other nodes in a timely and accurate manner to enforce
AR on other nodes, thus avoiding exacerbating congestion on the
reported path. This document specifies Adaptive Routing Notification
(ARN), a general mechanism to proactively disseminate congestion
detection and congestion elimination information for remote nodes to
perform re-routing policies. Particularly for AI workloads like
DeepSeek's MoE models that exhibit dynamic all-to-all communication
patterns with bursty traffic characteristics, such mechanisms become
crucial to enable immediate network response to transient congestion
conflicts.