What Is Link Flap?_
In GPU data-center networking, a link flap occurs when a physical or logical link oscillates between operational and non-operational states, often due to faulty transceivers, dirty connectors, or marginal signal integrity. This behavior triggers continuous renegotiation and route recalculation, degrading fabric performance and potentially causing packet loss. Persistent flapping can lead to network partitions or routing loops if not promptly addressed.
Technical Details
Link flaps are typically detected by switch or NIC firmware through loss of signal or bit-error-rate threshold crossings, prompting link-down events followed by automatic retraining. In NVLink domains, copper spine cables inside the rack are less prone to flapping than fiber links, but MPO connectors on InfiniBand/Ethernet scale-out networks can flap due to dust or insertion damage. Troubleshooting involves inspecting optical power levels per the OEM spec, cleaning connectors, and verifying cable seating in push-pull QSFP/OSFP cages. Persistent flaps may require replacing the transceiver or cable assembly to restore stable link state.
How Leviathan Systems Works with Link Flap
During rack commissioning, we monitor link flaps in the GPU-to-GPU NVLink copper spines and the scale-out fabric to ensure all links stabilize before declaring the node ready. A flapping link on a high-density GPU node’s InfiniBand uplink can delay the entire cluster’s integration into the production network.
Related service
Network Testing & Commissioning →