AI Networking · Darren Quhttps://darrenqu.net/ai-networking/InfiniBand, RDMA, RoCE, NVIDIA Spectrum-X, Cumulus Linux — the fabric layer underneath GPU clusters.Hugoen-USDarren QuDarren Qu2026 Darren QuWed, 16 Sep 2026 00:00:00 +0000EVPN Multihoming Across Vendors, Part 2: Verifying the Forwarding Planehttps://darrenqu.net/ai-networking/evpn-multihoming-across-vendors-part-2/Wed, 16 Sep 2026 00:00:00 +0000https://darrenqu.net/ai-networking/evpn-multihoming-across-vendors-part-2/AI NetworkingevpnThe control plane agrees on the Designated Forwarder and the host still receives every broadcast twice. Multi-point captures, a same-vendor control group, and a packet capture that was quietly lying.EVPN Multihoming Across Vendors, Part 1: From Configuration to Control Plane Alignmenthttps://darrenqu.net/ai-networking/evpn-multihoming-across-vendors-part-1/Wed, 16 Sep 2026 00:00:00 +0000https://darrenqu.net/ai-networking/evpn-multihoming-across-vendors-part-1/AI NetworkingevpnFour leaves, two vendors, one Ethernet Segment. Five configuration differences between Cumulus and Arista stand between the RFC and a working all-active bond — four get solved here, and none of them are in the RFC.BGP Unnumbered and Router Advertisements: Disabling RA Is Not Silence, It Is Withdrawalhttps://darrenqu.net/ai-networking/bgp-unnumbered-router-advertisements/Thu, 10 Sep 2026 00:00:00 +0000https://darrenqu.net/ai-networking/bgp-unnumbered-router-advertisements/AI NetworkingbgpRA drives dynamic peer discovery in BGP unnumbered — and turning it off tears down established sessions instantly, with no flap and no timeout. Why, and what the ND cache does and does not tell you.A BGP Unnumbered Underlay Across Two Vendors: The Route Was There, the Ping Was Nothttps://darrenqu.net/ai-networking/bgp-unnumbered-underlay-across-vendors/Fri, 04 Sep 2026 00:00:00 +0000https://darrenqu.net/ai-networking/bgp-unnumbered-underlay-across-vendors/AI NetworkingbgpArista and Cisco NX-OS, an unnumbered eBGP session, an IPv4 route installed with an IPv6 next hop — and no forwarding. The fix is one interface command, and the reason is RFC 8950.InfiniBand Routing Engines: Adaptive Routing and Credit Loopshttps://darrenqu.net/ai-networking/infiniband-routing-engines/Sat, 07 Mar 2026 00:00:00 +0000https://darrenqu.net/ai-networking/infiniband-routing-engines/AI NetworkinginfinibandThe subnet manager picks the routing engine, and the routing engine decides whether your fabric deadlocks. Min-hop, Up/Down, fat-tree and DragonFly+, and what each one is actually protecting you from.Diagnosing an InfiniBand Fabric: Counters, Tools, and a Workflowhttps://darrenqu.net/ai-networking/infiniband-fabric-diagnostics/Sun, 22 Feb 2026 00:00:00 +0000https://darrenqu.net/ai-networking/infiniband-fabric-diagnostics/AI Networkinginfinibandibdiagnet is the first command to run on a sick fabric. What its output files contain, which counters actually matter, and how to get from a slow training job to a specific port.Link-Layer Troubleshooting with MLXLink and MLXCableshttps://darrenqu.net/ai-networking/mlxlink-mlxcables-troubleshooting/Sun, 08 Feb 2026 00:00:00 +0000https://darrenqu.net/ai-networking/mlxlink-mlxcables-troubleshooting/AI NetworkinginfinibandWhen a fabric link flaps or runs at the wrong width, the answer is usually in the optics. How to read PCIe and fabric port state, per-lane signal quality and cable EEPROM with the NVIDIA firmware tools.RoCE on NVIDIA Spectrum-X: Building Lossless Ethernet for AI Fabricshttps://darrenqu.net/ai-networking/roce-on-nvidia-spectrum-x/Sun, 18 Jan 2026 00:00:00 +0000https://darrenqu.net/ai-networking/roce-on-nvidia-spectrum-x/AI NetworkingroceWhat RoCE actually needs from the network, how PFC and ECN divide the work, and how the lossless and lossy modes are configured and verified on Cumulus Linux.