MetaRoCE: Meta’s RDMA Transport
summary
MetaRoCE is a clean‑sheet RDMA transport that sprays packets across many paths, keeps per‑path RTT/congestion state and per‑path retransmission timers, and replaces many QPs with a single multipath connection to avoid hotspots. It targets 90% network efficiency with per-packet multipath spraying, runs on commodity Ethernet with only ECN/ECMP required, and pairs a conformance suite with Keysight Anvil that validates behavior from reference code through RTL emulation to production silicon, catching tag‑pausing and shared‑timer bugs.
AI-written summary · about
chapters (15)
- 0:00MetaRocky protocol introduction
- 1:07AI networking scale challenges
- 2:54Fabric intelligence to load balancing
- 4:18MetaRocky design and mechanisms
- 5:20Paths in MetaRocky foundation
- 6:44Resilience and failure handling
- 7:26Congestion control and reliability
- 8:06Scaling apps with MetaRocky
- 9:10MetaRocky openness and adoption
- 9:52Compliance with Kingshuk
- 11:15MetaRocky conformance test plan
- 13:02RNR timeout case study
- 15:08Retransmission timer bugs case study
- 17:35Why conformance testing matters
- 18:38Future of MetaRocky and AI networking
related talks
21mScaling RoCE Networks for AI Training
18mRoCE BALBOA: Service-Enhanced RDMA Offload Engine for Data Center SmartNICs
12mPreview: Networking
19mBring RoCE (RDMA over Converged Ethernet) Capability to Virtio-net Device
30mgRPC Meets RDMA
20mRDMA at Cloud Scale: The OCI Experience
23mMulti-Path Transport for RDMA in Datacenters
25mRDMA @Scale