The analytical model, the math in one function, the canvas, and an honest accounting of the corners cut.

Four failure presets, their blast radius, and why a ring collective plus a spine outage is the validator's worst finding.

Every line on the graph maps to a real counter you'd pull from DCGM or a switch. Here's how to diagnose from them.

Part 2 of the congestion-control deep dive — how InfiniBand sidesteps most of RoCE's problems by design.

Part 1 of the congestion-control deep dive — the Ethernet story, and why disabling ECN under load is a trap.

Spine-leaf vs Clos, credits vs pause frames, and why the simulator hands InfiniBand a small head start.

The vocabulary of distributed-training communication — and why AllToAll is the one that scares people.

Past a certain scale, distributed training is a networking problem. So I built a browser toy to feel it.

First post — setting up the blog and what to expect here.