MetaRoCE: From Spec to NIC to Open Source
summary
MetaRoCE ships a full software reference implementation and compliance suites and runs on an AMD P4 programmable AI NIC to prototype features before silicon. It sprays packets across paths with per-path congestion windows (example: 72-packet baseline, dynamically bootstraps 30–50 paths for long-RTT links), replaces PFC with receiver-driven scheduling, sustains line rate on an 8×100Gb multiplane FPF lab, tolerates ~1/1,000 drops with minimal slowdown, and is being released as open-source spec and tests.
AI-written summary · about
chapters (9)
- 0:00MetaRoCE introduction
- 1:29Spec to hardware loop
- 4:41AMD P4 NIC architecture
- 10:25Congestion-aware multi-pathing
- 12:34Scale out and connections
- 14:23Receiver-based congestion control
- 15:47Selective ACK and retransmission
- 16:53Multi-plane performance
- 18:38Open source plans
related talks
21mScaling RoCE Networks for AI Training
18mRoCE BALBOA: Service-Enhanced RDMA Offload Engine for Data Center SmartNICs
19mBring RoCE (RDMA over Converged Ethernet) Capability to Virtio-net Device
24mMetron: NFV Service Chains at the True Speed of the Underlying Hardware
12mPreview: Networking
28mSafe sharing of the network with eBPF
17mTraffic Engineering for AI Training Networks
18mRevisiting Congestion Control for Lossless Ethernet