A few large RDMA elephant flows from the housing-authority trainers appear stuck on one spine link despite multiple ECMP paths. What load-balancing concern should engineers address?

A small number of large RDMA elephant flows from housing-authority trainers can pin to one spine link under ECMP when the hash sees insufficient flow entropy. The concern to address is ensuring enough entropy—or suitable fabric load-balancing features—so those few large sessions do not all map to a single member while other equal-cost paths stay cold. That is a classic AI-fabric ECMP polarization problem, not a reason to abandon multipath. Relying on standard 5-tuple ECMP hashing or larger ECMP group tables does not solve the pinning problem—with only a few large RDMA flows there is too little 5-tuple variation to spread evenly, so those elephants can still hash to the same spine link while other equal-cost paths stay cold. Enabling unordered per-packet ECMP spraying or aggressive hash rotation timers to force a small number of large RDMA flows onto multiple spine links is unsafe for RDMA—spraying packets out of order or repeatedly re-hashing active elephants breaks lossless in-order delivery assumptions, triggering retransmissions and congestion collapse rather than clean load sharing. Pinning RDMA queue pairs to spine links via static ECMP entries or sticky resilient hashing does not create entropy—it hard-codes placement, so collisions still occur when pinned flows land together and imbalance persists as flows start, stop, or scale without adaptive spreading. Exam caveat: "more ECMP members" alone may not fix elephants if all large flows still hash identically—inspect the five-tuple diversity and platform load-balance options. Operational check: identify elephants stuck on one spine link via telemetry, increase flow entropy or enable suitable AI load-balancing features, retest housing-authority trainers, and confirm traffic spreads across ECMP members without harming RDMA correctness.