Writing
Problems before solutions, trade-offs before recommendations. New here? Start here.
EssaysDecisionsExplainersLabsResearch notes
Essays
Principles and trade-offs that outlast any product cycle.
- One link fails, every job slows: what ECMP hashing really does
Plain modulo hashing reshuffles almost every flow when one link fails. Why that hurts AI fabrics, and what resilient, symmetric and weighted hashing change.
October 17, 2026 Load balancingECMPFabric design - The six places an AI fabric congests, and which fix works where
Every congestion point is an egress queue. What decides the fix is whether the traffic had a choice of path: spread it better, or slow the senders down.
October 17, 2026 Fabric designRoCEv2Load balancing - Where the GPU's own network ends: scale-up versus scale-out
Rack-scale systems turned NVLink into a network of its own. Why the size of that domain now shapes the design of the back-end fabric.
October 17, 2026 Scale-upGPUFabric design - Why averages lie: oversubscription in AI fabrics
The airline-overbooking bet behind every oversubscribed network, why it fails for GPU training traffic, and a five-step method for sizing the rest.
October 3, 2026 Fabric designSizing
Decisions
One design question: context, options, trade-offs, what I would do and what would change my mind.
- Decision: one spine ASN or many?
A short design decision for eBGP fabrics. Why a shared spine ASN turns BGP loop prevention into protection against valley paths after a failure.
October 17, 2026 BGPFabric designCCDE
Explainers
A technology from the ground up, picture first.
- Ultra Ethernet in plain English
What the Ultra Ethernet transport changes compared with RoCEv2: delivery modes, sender-driven spraying, selective retransmission and receiver credits, and what it asks of your switches.
October 17, 2026 Ultra EthernetRoCEv2Transport - BGP unnumbered, explained without the hand-waving
How leaf-spine fabrics run BGP over links with no IPv4 addresses, why AI back-ends rely on it, and what actually happens in the routing table.
October 3, 2026 BGPFabric design - Lossless RoCE in one picture: ECN, DCQCN and PFC
The three mechanisms that keep RDMA from stalling on Ethernet, the order they should act in, and the counters that tell you whether they do.
October 3, 2026 RoCEv2QoS
Labs
A reproducible lab with real outputs: build it, break it, and see the protocol behave.
- Two exits, one idle link: a hands-on look at BGP ADD-PATH
A ten-minute FRR lab: behind a route reflector, ADD-PATH and multipath-relax do nothing alone and fix an idle link together. Packet captures and a failover myth, tested.
October 22, 2026 BGPFabric designLab
Research notes
An important paper in plain language, and what it means for your design.
- Research note: DCQCN, the paper behind RoCE congestion control
A plain-language reading of 'Congestion Control for Large-Scale RDMA Deployments' (SIGCOMM 2015): the problem it solved, how DCQCN works, and what it means for your fabric today.
October 17, 2026 RoCEv2Congestion controlResearch