Network & Infrastructure / EVPN, VXLAN & Data Center Fabrics
VXLAN underlay ECMP polarization despite many tenant flows
A fabric has several equal-cost uplinks, yet one is congested while others are quiet. Encapsulated traffic may expose insufficient entropy to the underlay's hashing algorithm.
Environment and inputs: BGP EVPN overlays on a routed underlay; supported route types and multihoming behavior depend on the network OS. Diagnostic/design example; not tested on production equipment.
Collect the evidence
Compare outer source/destination addresses, UDP source-port variation, ECMP group membership and per-link counters. Use multiple tenant flow tuples rather than repeatedly testing one TCP session.
Explain the result
Underlay distribution depends on visible hash fields and encapsulation entropy. Equal-cost paths do not guarantee equal byte counts, particularly when a small number of elephant flows dominate.
Plan the fix
Review supported VXLAN entropy and hashing settings along with workload flow diversity. Change the measured bottleneck mechanism instead of assuming an additional physical link automatically redistributes an existing flow.
Check the result and rollback
Validate distribution with representative flow sizes and one uplink failure. Revert if packet reordering, tunnel compatibility or tenant latency worsens under the revised hash behavior.