Year: 2026
Type: article
Abstract: In-network aggregation (INA) reduces parameter-server traffic in distributed DNN training, but best-effort INA makes congestion control harder than ordinary packet forwarding. When a packet collides with an occupied switch aggregator, it falls back to end-host aggregation and becomes additional traffic on the parameter-server-bound link. Existing INA transports either react to this mixed load with coarse link-level control or manage aggregator collisions separately, missing the coupling between congestion window, collision probability, and fallback traffic. This paper presents RICA, a resource-integrated congestion control framework for best-effort INA. RICA models the bottleneck load as the sum of aggregated traffic and collision-induced fallback traffic, and derives an adaptive backoff factor from this model. RICA also adds lightweight in-band straggler telemetry to existing ACKs, allowing fast workers to yield aggregator occupancy when a persistent straggler is detected. Packet-level simulations show that RICA reduces average training iteration time by up to 8.6%--12.0% compared with the state-of-art INA system. Our code is available at https://github.com/wzy240/rica.
DOI: 10.1145/3856973
Authors: Ziyu Wang, Jianglong Nie, Yuan Yao, Wenfei Wu
Institution: Peking University
Language: English
Subfield: Computer Networks and Communications
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7236672777
