How to set up Charmed Kubeflow on NVIDIA DGX
NVIDIA DGX systems are purpose-built hardware for enterprise AI use cases. These platforms feature NVIDIA Tensor Core GPUs, which vastly outperform traditional CPUs for machine learning workloads, alongside advanced networking and storage capabilities.
This guide contains setup instructions for running Charmed Kubeflow on single-node Nvidia’s DGX-enabled hardware. It supports both single-node and multi-node environments, having a couple of examples of how to use two components: Jupyter Notebooks and Kubeflow Pipelines.
Prerequisites:
- Nvidia DGX-enabled hardware setup with correctly configured/updated BIOS settings, bootloader, OS, drivers, and packages (sample setup instructions provided below).
- Familiarity with Python, Docker, Jupyter notebooks.
- Tools:
juju,kubectl
Sample Ubuntu and Grub setup
NOTE: The following setup instructions are given only as an example. There is no guarantee that they will be sufficient for all environments. Contact hardware distributor for more details on specific system setup.
Ubuntu Setup
This document was tested on Ubuntu 20.04 vanilla
Hint: please make sure you don’t have any Nvidia drivers preinstalled. You can do that with following steps:
Check for apt packages (if empty you OK) :
$ sudo apt list --installed | grep nvidia
If there are some packages presented you can try to remove them with
$ sudo apt remove <package-name>
$ sudo apt autoremove
Check for kernel modules (if empty you OK):
$ lsmod | grep nvidia
If there are some presented you can try to remove them with
sudo modprobe -r <module-name>
Reboot
$ sudo reboot
Grub Setup
Edit /etc/default/grub and add the following options to
GRUB_CMDLINE_LINUX_DEFAULT: modprobe.blacklist=nouveau nouveau.modeset=
$ sudo reboot
Contents:
- Install Kubernetes (MicroK8s)
- Enable GPU add-on and configure MIG
- Deploy Charmed Kubeflow
- Try Kubeflow examples
Install Kubernetes (MicroK8s)
Install microk8s and enable required addons
$ sudo snap install microk8s --classic --channel 1.22
$ sudo microk8s enable dns:10.229.32.21 storage ingress registry rbac helm3 metallb:10.64.140.43-10.64.140.49,192.168.0.105-192.168.0.111
$ sudo usermod -a -G microk8s ubuntu
$ sudo chown -f -R ubuntu ~/.kube
$ newgrp microk8s
Edit /var/snap/microk8s/current/args/containerd-template.toml. Add:
[plugins."io.containerd.grpc.v1.cri".registry.configs]
[plugins."io.containerd.grpc.v1.cri".registry.configs."registry-1.docker.io".auth]
username = "afrikha"
password = "<>"
$ microk8s.stop; microk8s.start
Enable GPU add-on and configure MIG
Install GPU operator
$ sudo microk8s.enable gpu
$ mkdir .kube
$ microk8s config > ~/.kube/config
Check gpu count for k8s
$ kubectl get nodes --show-labels | grep gpu.count
Configure MIG devices
$ kubectl label nodes blanka nvidia.com/mig.config=all-1g.5gb --overwrite
Recheck gpu count (should be increased)
kubectl get nodes --show-labels | grep gpu.count
Troubleshooting: If none of the nodes appear in get nodes command please makes ure to uninstall all GPU drivers form kubernetes nodes and reinstall the microk8s.
Deploy Charmed Kubeflow
Follow the instructions from How to install Charmed Kubeflow to install Charmed Kubeflow.
Once installed, you can Login to Charmed Kubeflow with this guide.
When using Microk8s, sometimes applications may go into an error state because of `“error”: “too many open files” or similar. If you see this, see these suggested fixes Wait before all components are available.
To ensure Charmed Kubeflow is ready to be used, run this command
nice -n 16 watch -n 1 -c juju status --relations --color
Expected state:
App Version Status Scale Charm Channel Rev Address Exposed Message
admission-webhook res:oci-image@84a4d7d active 1 admission-webhook 1.6/stable 50 10.152.183.98 no
argo-controller res:oci-image@669ebd5 active 1 argo-controller 3.3/stable 99 no
argo-server res:oci-image@576d038 active 1 argo-server 3.3/stable 45 no
dex-auth active 1 dex-auth 2.31/stable 129 10.152.183.147 no
istio-ingressgateway active 1 istio-gateway 1.11/stable 114 10.152.183.29 no
istio-pilot active 1 istio-pilot 1.11/stable 131 10.152.183.104 no
jupyter-controller res:oci-image@8f4ec33 active 1 jupyter-controller 1.6/stable 138 no
jupyter-ui res:oci-image@cde6632 active 1 jupyter-ui 1.6/stable 99 10.152.183.222 no
katib-controller res:oci-image@03d47fb active 1 katib-controller 0.14/stable 92 10.152.183.220 no
katib-db mariadb/server:10.3 active 1 charmed-osm-mariadb-k8s latest/stable 35 10.152.183.23 no ready
katib-db-manager res:oci-image@16b33a5 active 1 katib-db-manager 0.14/stable 66 10.152.183.25 no
katib-ui res:oci-image@c7dc04a active 1 katib-ui 0.14/stable 90 10.152.183.197 no
kfp-api res:oci-image@1b44753 active 1 kfp-api 2.0/stable 81 10.152.183.8 no
kfp-db mariadb/server:10.3 active 1 charmed-osm-mariadb-k8s latest/stable 35 10.152.183.148 no ready
kfp-persistence res:oci-image@31f08ad active 1 kfp-persistence 2.0/stable 76 no
kfp-profile-controller res:oci-image@d86ecff active 1 kfp-profile-controller 2.0/stable 61 10.152.183.173 no
kfp-schedwf res:oci-image@51ffc60 active 1 kfp-schedwf 2.0/stable 80 no
kfp-ui res:oci-image@55148fd active 1 kfp-ui 2.0/stable 80 10.152.183.146 no
kfp-viewer res:oci-image@7190aa3 active 1 kfp-viewer 2.0/stable 79 no
kfp-viz res:oci-image@67e8b09 active 1 kfp-viz 2.0/stable 74 10.152.183.16 no
kubeflow-dashboard res:oci-image@6fe6eec active 1 kubeflow-dashboard 1.6/stable 154 10.152.183.61 no
kubeflow-profiles res:profile-image@0a46ffc active 1 kubeflow-profiles 1.6/stable 82 10.152.183.204 no
kubeflow-roles active 1 kubeflow-roles 1.6/stable 31 10.152.183.141 no
kubeflow-volumes res:oci-image@cc5177a active 1 kubeflow-volumes 1.6/stable 64 10.152.183.100 no
metacontroller-operator active 1 metacontroller-operator 2.0/stable 48 10.152.183.169 no
minio res:oci-image@1755999 active 1 minio ckf-1.6/stable 99 10.152.183.193 no
oidc-gatekeeper res:oci-image@32de216 active 1 oidc-gatekeeper ckf-1.6/stable 76 10.152.183.194 no
seldon-controller-manager res:oci-image@eb811b6 active 1 seldon-core 1.14/stable 92 10.152.183.221 no
tensorboard-controller res:oci-image@667e455 active 1 tensorboard-controller 1.6/stable 56 10.152.183.232 no
tensorboards-web-app res:oci-image@914a8ab active 1 tensorboards-web-app 1.6/stable 57 10.152.183.174 no
training-operator active 1 training-operator 1.5/stable 65 10.152.183.14 no
Unit Workload Agent Address Ports Message
admission-webhook/0* active idle 10.1.19.32 4443/TCP
argo-controller/0* active idle 10.1.19.82
argo-server/0* active idle 10.1.19.53 2746/TCP
dex-auth/0* active idle 10.1.19.56
istio-ingressgateway/0* active idle 10.1.19.28
istio-pilot/0* active idle 10.1.19.62
jupyter-controller/0* active idle 10.1.19.60
jupyter-ui/0* active idle 10.1.19.55 5000/TCP
katib-controller/0* active idle 10.1.19.46 443/TCP,8080/TCP
katib-db-manager/0* active idle 10.1.19.2 6789/TCP
katib-db/0* active idle 10.1.19.57 3306/TCP ready
katib-ui/0* active idle 10.1.19.36 8080/TCP
kfp-api/0* active idle 10.1.19.85 8888/TCP,8887/TCP
kfp-db/0* active idle 10.1.19.23 3306/TCP ready
kfp-persistence/0* active idle 10.1.19.80
kfp-profile-controller/0* active idle 10.1.19.83 80/TCP
kfp-schedwf/0* active idle 10.1.19.22
kfp-ui/0* active idle 10.1.19.87 3000/TCP
kfp-viewer/0* active idle 10.1.19.51
kfp-viz/0* active idle 10.1.19.78 8888/TCP
kubeflow-dashboard/0* active idle 10.1.19.86 8082/TCP
kubeflow-profiles/0* active idle 10.1.19.49 8080/TCP,8081/TCP
kubeflow-roles/0* active idle 10.1.19.31
kubeflow-volumes/0* active idle 10.1.19.63 5000/TCP
metacontroller-operator/0* active idle 10.1.19.1
minio/0* active idle 10.1.19.47 9000/TCP,9001/TCP
oidc-gatekeeper/0* active idle 10.1.19.89 8080/TCP
seldon-controller-manager/0* active idle 10.1.19.24 8080/TCP,4443/TCP
tensorboard-controller/0* active idle 10.1.19.90 9443/TCP
tensorboards-web-app/0* active idle 10.1.19.88 5000/TCP
training-operator/0* active idle 10.1.19.38
Access the dashboard at http://10.64.140.43.nip.io. If you are running it on cloud and want to access it from local machine you can create socks proxy as follows.
Ssh to instance (the from which you ran commands above):
ssh -D9999 <user>@<IP>
On your computer’s browser, go to Settings > Network > Network Proxy, and enable SOCKS proxy pointing to: 127.0.0.1:9999
On a new browser window, access the link given in the previous step, appended by .nip.io, for example: http://10.64.140.43.nip.io
Try Kubeflow examples
Charmed Kubeflow can be run on single-node and multi-node DGX hardware. Depending on the environment, there are different requirements that should be followed. There are multiple examples that can be tried out.
Single-node DGX with Charmed Kubeflow examples
There is a GitHub repository that includes all the details about the Single-node DGX with Charmed Kubeflow.
The following examples can be found and tested:
- Jupyter Notebook example on a single-node DGX in the file
gpu-notebook.ipynbfrom the repository. It also uses multi GPU setup. - Kubeflow Pipeline example on a single-nod DGX, that uses the same classifier as the Notebook. It is available in the file
gpu-pipeline.ipynb.
Multi-node DGX with Charmed Kubeflow examples
There is a GitHub repository that includes all the details about the Multi-node DGX with Charmed Kubeflow.
The following examples can be found and tested:
- Training Tensorflow models with multi-GPUs in a Jupyter Notebook using Charmed Kubeflow in the folder
multi-gpu-in-notebook, where Jupyter Notebook file is available,gpu-notebook.ipynb. - Training Tensorflow models with GPUs in a Kubeflow Pipeline in the folder
multi-gpu-in-pipeline. - A simulated example of multi-node training in Tensorflow, but using just a single node in the folder
multi-node-gpu-simulated. There are going to be multiple files describing the workload distribution and how to run it. - Multi-node training in Tensorflow using the Kubeflow Training Operator’s TFJob in the folder
multi-node-gpu-tfjob.
Last updated 5 days ago.