| #!/bin/bash | |
| set -e | |
| KO_DOCKER_REPO=${KO_DOCKER_REPO:-kserve} | |
| TAG=${TAG:-latest} | |
| # If running from different directory, override with directory of KServe source code | |
| DEFAULT_DIR=${DEFAULT_DIR:-$(pwd)} | |
| echo |
KernelCache supports two mount types: imageVolume and pvc. The imageVolume path
mounts the OCI image directly in Kubernetes (1.33+) as a read-only volume — no PV/PVC
management is involved, and no extraction job is needed. It is the simpler and preferred
path on modern clusters.
The pvc path extracts the kernel cache from the OCI image into a PersistentVolumeClaim
LocalModelCache (LMC) distributes large model files to the local disk of every targeted
node in a Kubernetes cluster so that InferenceService (ISVC) pods can load models from
fast local storage rather than pulling them over the network on each startup. The design
is built around three CRDs — LocalModelCache, LocalModelNodeGroup, and
LocalModelNode — and a DaemonSet agent that runs on every target node.
| $ make run-on-kind | |
| if [ ! -f /home/bmcfall/src/GKM/bin/kind-gpu-sim.sh ]; then \ | |
| echo "Downloading /home/bmcfall/src/GKM/bin/kind-gpu-sim.sh"; \ | |
| wget -P /home/bmcfall/src/GKM/bin https://raw.githubusercontent.com/maryamtahhan/kind-gpu-sim/refs/heads/main/kind-gpu-sim.sh; \ | |
| chmod +x /home/bmcfall/src/GKM/bin/kind-gpu-sim.sh; \ | |
| fi | |
| Deleting Kind GPU cluster: kind-gpu-sim | |
| cat /home/bmcfall/src/GKM/bin/kind-gpu-sim.sh | bash -s delete --cluster-name=kind-gpu-sim | |
| Using Podman as container runtime | |
| using podman due to KIND_EXPERIMENTAL_PROVIDER |
| Requirement: | |
| * Same logical data on all nodes | |
| * Multiple nodes | |
| * Underlying storage only supports RWO | |
| * Pods should not need to know node-specific PVC names | |
| * Operator should abstract this away | |
| * Data pre-populated before workload starts | |
| * Pod startup must be fast (no on-demand copy) | |
| Does GKM Need to support: |
| ------------------------ | |
| I am coding a CSI Driver in Golang that will mount a PVC into a pod volumemount. | |
| Can you write a golang set of functions that will mount a PVC in a pod? | |
| ------------------------ | |
| This is an important clarification: | |
| 👉 **A CSI driver does NOT mount a PVC into a Pod.** |
| $ kubectl explain clusterbpfapplication | |
| GROUP: bpfman.io | |
| KIND: ClusterBpfApplication | |
| VERSION: v1alpha1 | |
| DESCRIPTION: | |
| ClusterBpfApplication is the schema for the cluster scoped BPF Applications | |
| API. Using this API allows applications to load one or more eBPF programs on | |
| a Kubernetes cluster using bpfman to load the programs. | |
Using Operator-SDK v1.27.0 (from Makefile at the time)
Goal is to run ./bin/operator-sdk create api --kind TcxNsProgram ... to add a new kind object.
As is, can't run in our repo without hitting errors.
$ ./bin/operator-sdk create api --group bpfman --version v1alpha1 --kind TcxNsProgram --resource --controller --force --namespaced=true
Error: failed to create API: unable to inject the resource to "base.go.kubebuilder.io/v3": multiple groups are not allowed by default, to enable multi-group visit https://kubebuilder.io/migration/multi-group.html
| BROKEN first, then WORKS after | |
| ------------- | |
| BROKEN | |
| ------------- | |
| # From main:latest | |
| $ sudo /usr/share/bcc/tools/capable | |
| TIME UID PID COMM CAP NAME AUDIT |