Skip to content

Capacity Aware Adaptive Routing#

OCP Global Summit 2025 digest Program Core pick 13–16 October 2025

Abstract Session recording Materials

Speakers: Yikai Lin (Meta), Midhun Somasundaran (Meta)

  • Why: This is exactly the failure-mode analysis a fabric designer needs: what breaks in local-load-aware adaptive routing at Meta's topology scale, and the fix. Real production architecture, open-source linkage (FBOSS).
  • What: Meta's next-generation backend routing design. Their existing adaptive routing scheme (ARS), which relies on local load awareness, degrades on larger topologies — parallel links plus shallow-buffer switches plus silent capacity loss from remote link failures. The new approach combines remote capacity information with local-congestion-aware spray. (from abstract only)
  • How: Dynamically adjusts load balancing based on end-to-end link capacity so partial failures reduce performance gracefully; adds algorithms for merging ARS groups and uses non-load-aware ECMP spray groups to address scaling limits.
  • Where applicable: Meta-specific topology and switch-buffer assumptions (shallow-buffer Ethernet, parallel-link spines), but the technique — capacity-aware rather than purely queue-aware spray — is a general idea for any large Clos fabric. FBOSS reference code is linked from the session materials.