Live Group Q&A Session
summary
Meta builds a new transport called metari because RoCEv2 cannot tolerate out‑of‑order multipath delivery; a working reference implementation with AMD and Keysight uses lab fiber spools and validates four concrete promises: entropy for path splitting, ACKs follow multipath, multipath bookkeeping, and path failover and scaling. Tests show up to eight live paths (72 sprayed in lab), end‑to‑end RDMA across ~1.5 ms RTT, and support for ~4K QPs (design headroom toward 8K); congestion control runs per tiny path and it works on existing switches using ECN.
AI-written summary · about
chapters (9)
- 0:00Motivation for Meta-Rocky
- 1:05Multipath validation
- 4:13Scaling and long-distance
- 5:46Evolving protocol challenges
- 9:18Designing tiny network pipes
- 11:44Hard design trade-offs
- 13:31AI workload requirements
- 15:22Infrastructure compatibility
- 16:43Future OCP release
