Link Flap

Overview

A link flap is formally an unexpected state transition on an optical link, defined by the IEEE 802.3 standard as three uncorrected codewords triggering loss of lock — a threshold originally designed for copper links that fires very quickly at 400G+ optical speeds. Link flaps are the primary reliability threat to large-scale RDMA AI fabrics: each flap triggers a cascade from L1 PCS reset through L2/MAC re-initialization, IGP reconvergence, IB-12 RDMA timeout failures, DLB flow recalculation, and DCQCN congestion disruption, all scaling with workload size. At 2025 AI cluster scale (4.5 GW / 3M GPUs / 10M+ transceivers), the cluster-wide mean time to link flap (MTTL) drops to 48 seconds — compared to 4-day MTBF / 3-hour MTTL in 2023 at 30 MW scale.

Root causes include bad transceivers (lasers, photo detectors, DSPs), connector coupling failures, fiber quality issues, deployment hygiene problems, thermal cycling, ESD damage during installation, assembly defects, firmware bugs, and switch/NIC interoperability issues. Latent damage that does not fail at installation but elevates long-term failure probability is the most dangerous mode: 100,000 such links aggregate into a perpetually flapping population.