Tag: kubernetes
All the articles with the tag "kubernetes".
NVIDIA is not a GPU company anymore
A GPU-sharing problem I never solved cleanly years ago stayed with me, so I spent a weekend back in the NVIDIA world to see what had changed. What I found is that NVIDIA is no longer a GPU vendor. It is a full stack, from CUDA and the warp scheduler at the bottom, through NCCL, NVLink, NVSwitch, and the Mellanox fabric, through Grace CPUs and DGX systems and DGX Cloud, through TensorRT, vLLM, Triton, and Dynamo, up to build.nvidia.com with NIMs, Nemotron, Blueprints, Skills, and NeMo Guardrails. These are my weekend findings, and they end where my old problem started: how you share a GPU.
The Kubernetes cluster that had to survive the open ocean
A pre-sale asked for Kubernetes in the server room of a cruise ship. The ship is a distraction. The real constraint is that the network becomes a scheduled resource, present in port and gone at sea, and that splits the system into a runtime that has to survive for days with no link home and a lifecycle of updates and image pulls that can only happen while the ship is connected. EKS Anywhere runs the whole cluster on the boat, which is why it fit and why keeping the control plane onshore would not have.
The private cloud with no requirements
A government customer in the Gulf asked for a private cloud with Kubernetes and handed over no requirements. The design work turned out to be inventing the requirements: imagining the tenants who would show up and what they would ask for, then choosing technology against the two questions that stand in for the word best when there is no spec, whether it locks you in and whether it is already paid for.
The OVSDB ORM I wasn't allowed to open-source
The seventh and last post in the CertaScale series, and the one I'm still not over. Driving OVN means talking to several databases over the OVSDB wire protocol by hand. So I built a full ORM in Go: every table as a typed object, references as pointers, and an ACL match compiler built on Reverse Polish Notation. I wanted to open-source it. The company said no, and it died with the company.
From 12 to 60 Gbps with DPDK, when we tried to become a 5G-edge box
The sixth post in the CertaScale series. When private cloud proved a hard sell, the owners pivoted to 5G edge, and edge means line-rate on commodity hardware. Our kernel data path topped out around 12 Gbps on 100 Gbps NICs. Taking it out of the kernel with DPDK, hugepages, poll-mode drivers, and userspace-bound NICs got the same servers to roughly 60 Gbps.
Running VMs inside pods in 2016, before Kata existed
The fifth post in the CertaScale series, and the one the earlier posts kept promising: running full virtual machines inside Kubernetes pods in 2016, before Kata Containers existed and before RuntimeClass was a thing. A pod is just a namespace boundary and a lifecycle. Put KVM inside it and a VM schedules, migrates, and dies like a pod.
A flat network, GCP-style, on early Kubernetes
The fourth post in the CertaScale series, and the idea I'm proudest of: a flat overlay borrowed from how Google Cloud addresses machines. /32 addresses plus routes, an address operator handing out CRDs from a pool, and the reason we built all of it — moving a workload between nodes without it losing its IP.
Giving Kubernetes an enterprise network
The third post in the CertaScale series. Our customer was building a private cloud that had to drop into any enterprise's network, so we spoke the language every enterprise already speaks and extended it to Kubernetes pods: VLANs, static and DHCP addressing, and RFC 4594 QoS.
Finding ovn-kubernetes when it was seven days old
The second post in the CertaScale series. Handed a failed project and a title, I inherited a broken Vagrant file that created a single bridge, panic-googled my way to a week-old repo, learned OVS internals the hard way, and rewrote our side from Python to Go when the team changed.
The private cloud we built before it was cool
From 2016 to 2019 I helped build a private cloud on top of very early Kubernetes, with a drag-and-drop canvas, VMs running inside pods before Kata existed, and an enterprise-grade network layer I led, from GCP-style flat overlays to a 12-to-60 Gbps DPDK data path. The company is gone. The story shouldn't be.