Congestion Control for AI Workloads in Packet-Sprayed Datacenter Networks

PhD Qualifying Examination


Title: "Congestion Control for AI Workloads in Packet-Sprayed Datacenter
Networks"

by

Miss Yuxuan LI


Abstract:

Packet spray has emerged as a promising solution for mitigating ECMP hash 
collisions in AI clusters. By distributing packets from the same flow across 
multiple paths, packet spray shifts the network abstraction from multiple 
individual paths to one unified pathset. Adaptive packet spray can further 
rebalance traffic within the pathset automatically, providing an intrinsic 
self-healing capability that absorbs partial congestion when healthy paths 
retain sufficient spare capacity. However, a fundamental mismatch remains 
between this pathset abstraction and existing congestion control (CC) 
mechanisms. Conventional single-path CCs are oblivious to the pathset 
abstraction and misinterpret localized hotspots as pathset-wide global 
congestion, leading to unnecessary flow-wide rate reductions. Multipath-aware 
CCs improve tolerance to partial congestion, but can still overreact to 
hotspot feedback or require tight coupling with specific load-balancing 
mechanisms.

This survey provides a comprehensive examination of congestion control for AI 
workloads in packet-sprayed datacenter networks. We first review the 
communication characteristics that motivate packet spray deployment, then 
classify packet spray mechanisms and examine how packet spray changes the 
network abstraction from a single path to a pathset. We next examine 
conventional ECN-, delay-, INT-based CCs, receiver-driven transports, and 
multipath-aware designs, focusing on their congestion signals, control 
mechanisms, and interactions with load balancing. Drawing on simulation 
results, we show how existing CCs overreact to partial congestion and leave 
healthy paths underutilized despite packet spray rebalancing. Finally, we 
discuss the implications of these findings for congestion control that 
accounts for pathset-level congestion and coordinates rate adjustments with 
diverse host-based and in-network packet spray implementations.


Date:                   Monday, 26 October 2026

Time:                   11:00am - 1:00pm

Venue:                  Room 5504
                        Lifts 25/26

Committee Members:      Prof. Kai Chen (Supervisor)
                        Dr. Xiaomin Ouyang (Chairperson)
                        Prof. Song Guo