Congestion Signaling
Overview
Congestion Signaling (CSIG) is a network traffic management mechanism developed within the OCP and Ultra Ethernet Consortium (UEC) ecosystem. CSIG provides explicit congestion feedback from switches to endpoint NICs, enabling more intelligent load balancing decisions in AI cluster fabrics compared to traditional reactive approaches (e.g., DCQCN) that wait for packet loss before reacting.
The mechanism works by having switches emit congestion signals that NICs use to proactively redistribute traffic across available paths, avoiding hotspots before they cause job slowdowns. This is particularly valuable in all-reduce and collective communication workloads that are highly sensitive to stragglers.
Google introduced CSIG at OCP approximately two years before the October 2025 summit, and by that point had deployed and validated it in Google production at scale. Broadcom announced support for CSIG, and UEC standardization is proceeding with CSIG targeted for inclusion in UEC 1.1.
Sign in to read the full article.
Sign In