AI
SRv6 uSID in the AI data center: let the NIC pick the path
How production fabrics balance RoCE today AI training traffic is a handful of huge RDMA flows between GPUs, and ECMP was built for millions of small ones.…
Blog
Field notes and write-ups, filter by technology.
23 posts
All posts
Everything, field notes and write-ups.
AIDC
AI Fabric Congestion Control
How PFC, ECN and DCQCN actually behave once you build the lab and measure it yourself. Every number here came off the local simulator, and every equation is…
AI
Three tiny neural networks that finally made it click
A network engineer’s notebook, from a toy example to a router crash predictor
AI
Before the Jam
Teaching a network to predict bandwidth and steer traffic around congestion, with PyTorch and zero dropped packets.
AI
What I Learned Running LLMs Locally and in the Cloud, and Then Taking One Apart
A hands-on build on an Apple Silicon Mac: local Llama versus Claude versus cloud Llama, where the tokens actually go, a bug that cost me an afternoon, and…
Service Provider
SRv6 Traffic Engineering on Cisco IOS-XR, Part 1: Bending a Path with uSID and SR-PCE
Most SRv6 write-ups stop at “look, a VPN pings across an IPv6 core.” That is the easy part. The interesting question is the one traffic engineering asks:…
AI
Teaching My Lab to See the Future: Bandwidth Prediction with Two Routers, a Time-Series Database…
How a CiscoU post sent me down a rabbit hole, and what came out the other side: a tiny, fully reproducible lab where real packets train a real neural…
AIDC
Scaling to 256 GPUs: The Rail-Optimized AI Pod
From the 32-GPU pilot to the canonical production pod. What changes, what stays, and why the numbers finally resolve cleanly.
Service Provider
Proof of Concept: SRv6 Migration in a Homogeneous Cisco Network
The multivendor PoC (the companion document) proved that an L3VPN can run across Nokia, Cisco and Juniper on an SR-MPLS core, and it tried to move that live…
Service Provider
Proof of Concept: SRv6 Migration in a Homogeneous Nokia Network
The multivendor PoC proved an L3VPN could run across Nokia, Cisco and Juniper on an SR-MPLS core, then tried to move that live service to SRv6. That…
Service Provider
Proof of Concept: Multivendor SR-MPLS L3VPN and the Move to SRv6
We wanted a straight answer to a question that comes up every time someone proposes SRv6: can you take a live L3VPN that runs across Nokia, Cisco and…
Service Provider
SRv6 on Nokia SR OS, End to End: From Why to the Packets on the Wire
The story of why networks are moving to SRv6, then a real SR OS build in Containerlab with configs, control plane, data plane and live packet captures.
AIDC
Designing an AI Data Center: Start Small, Then Scale
The eight design dimensions every AI DC build resolves, applied to the smallest fabric that has every problem a 30,000-GPU cluster has.
Service Provider
I Spent Three Days Getting Nokia and Juniper VPLS to Talk. Here’s What Actually Works.
A field report on cross-vendor VPLS BGP auto-discovery, FEC 128 vs FEC 129 dead-ends, and the one Nokia knob that fixes everything.
AI
I Built a Chat Bot That Talks to My Network. Here’s What Actually Worked.
A weekend experiment that turned into a sandbox I keep coming back to. Three retrievers, one LLM, and one absolute rule even in a lab: nothing changes…
AIDC
Adaptive Routing and Packet Spraying
Four vendor implementations, head to head, with measured numbers
AIDC
Reading AI Training in Switch Counters
Five traffic signatures, what each one looks like in telemetry, and what to do when you spot them.
AIDC
How RoCEv2 Stays Lossless
PFC, ECN, and DC-QCN. Three layers, three timescales, one fabric that doesn’t drop packets
AIDC
RoCEv2, InfiniBand, and Ultra Ethernet
Two clusters. Same 256 H100s. Same 70-billion-parameter model. Same NCCL build at the same git hash. One ran InfiniBand HDR end to end, the other ran RoCEv2…
AIDC
GPU Fabric Topologies Rail-Optimized, Fat-Tree, Dragonfly
Two networks. Same switch BOM. Same rack space. Same power envelope. Same 256 H100s.
AIDC
Why AI Traffic Breaks Classical DC Networks
At 02:14 UTC on a Tuesday in March, an all-reduce step on a 64-GPU training job started taking 4.2 seconds instead of 0.6. The checkpoint was due in nine…