Virtual Thoughts

Virtualisation, Storage and various other ramblings.

Page 5 of 24

Installing & Using the Nvidia GPU Operator in K3s with Rancher

This post outlines the necessary steps to leverage the Nvidia GPU operator in a K3s cluster. In this example, using a gift from me to my homelab, a cheap Nvidia T400 GPU which is on the supported list for the operator.

Step 1 – Configure Passthrough (If required)

For this environment, vSphere is used and therefore PCI Passthrough is required to present the GPU to the VM. The Nvidia GPU is represented as two devices – one for the video controller, and another for the audio controller – we only need the video controller. Steps after this are still relevant to bare metal deployments.

Step 2 – Create VM

When creating a VM, choose to add a PCI device, and specify the Nvidia GPU:

Step 3 – Install nvidia-container-runtime and K3s

In order for Containerd (within K3s) to pick up the Nvidia plugin when K3s starts, we need to install the corresponding container runtime:

root@ubuntu:~# curl -s -L https://nvidia.github.io/nvidia-container-runtime/gpgkey |   sudo apt-key add -
root@ubuntu:~# distribution=$(. /etc/os-release;echo $ID$VERSION_ID)
root@ubuntu:~# curl -s -L https://nvidia.github.io/nvidia-container-runtime/$distribution/nvidia-container-runtime.list |   sudo tee /etc/apt/sources.list.d/nvidia-container-runtime.list
root@ubuntu:~# apt update && apt install -y nvidia-container-runtime

root@ubuntu:~# curl -sfL https://get.k3s.io | INSTALL_K3S_VERSION="v1.23.7+k3s1" sh

We can validate the Containerd config includes the Nvidia plugin with:

root@ubuntu:~# cat /var/lib/rancher/k3s/agent/etc/containerd/config.toml | grep -i nvidia
[plugins.cri.containerd.runtimes."nvidia"]
[plugins.cri.containerd.runtimes."nvidia".options]
  BinaryName = "/usr/bin/nvidia-container-runtime"

Step 4 – Import Cluster into Rancher and install the nvidia-gpu-operator

Follow this guide to import an existing cluster in Rancher.

After which, Navigate to Rancher -> Cluster -> Apps -> Repositories -> Create

Add the Helm chart for the Nvidia GPU operator:

Select to install the GPU Operator chart by going to Cluster -> Apps -> Charts -> Search for "GPU":

Follow the instructions until you reach the Edit YAML section. At this point add the following configuration into the corresponding section; this is to cater to where K3s stores the Containerd config and socket endpoint:

toolkit:
  env:
    - name: CONTAINERD_CONFIG
      value: /var/lib/rancher/k3s/agent/etc/containerd/config.toml
    - name: CONTAINERD_SOCKET
      value: /run/k3s/containerd/containerd.sock

Proceed with the installation and wait for the corresponding Pods to spin up. This will take some time as it’s compiling the GPU/CUDA drivers on the fly.

Note: You will notice several GPU-Operator Pods initially in a crashloop state. This is expected until the nvidia-driver-daemonset Pod has finished building and installing the Nvidia drivers. You can follow the Pod logs to get more insight as to what’s occurring.

oot@ubuntu:~# kubectl logs nvidia-driver-daemonset-wmrxq
DRIVER_ARCH is x86_64
Creating directory NVIDIA-Linux-x86_64-515.65.01
Verifying archive integrity... OK
root@ubuntu:~# kubectl logs nvidia-driver-daemonset-wmrxq -f
DRIVER_ARCH is x86_64
Creating directory NVIDIA-Linux-x86_64-515.65.01
Verifying archive integrity... OK
Uncompressing NVIDIA Accelerated Graphics Driver for Linux-x86_64 515.65.01............................................................................................................................................
root@ubuntu:~# kubectl get po
NAME                                                              READY   STATUS            RESTARTS      AGE
nvidia-dcgm-exporter-dkcz9                                        0/1     PodInitializing   0             4m42s
gpu-operator-v22-1669053133-node-feature-discovery-master-t4mrp   1/1     Running           0             6m26s
gpu-operator-v22-1669053133-node-feature-discovery-worker-rxxw5   1/1     Running           1 (91s ago)   6m1s
gpu-operator-8488c86579-gf7z8                                     1/1     Running           1 (10m ago)   30m
nvidia-container-toolkit-daemonset-mgn92                          1/1     Running           0             5m59s
nvidia-driver-daemonset-46sdp                                     1/1     Running           0             5m55s
nvidia-cuda-validator-cmt7x                                       0/1     Completed         0             74s
gpu-feature-discovery-4xw2q                                       1/1     Running           0             4m23s
nvidia-device-plugin-daemonset-8czgl                              1/1     Running           0             5m
nvidia-device-plugin-validator-tzpq8                              0/1     Completed         0             37s

Step 5 – Validate and Test

First, check to see the runtimeClass is present:

root@ubuntu:~# kubectl get runtimeclass
NAME     HANDLER   AGE
nvidia   nvidia    30m

kubectl describe node should also list a GPU under the Allocatable resources:

Allocatable:
  cpu:                8
  ephemeral-storage:  49893476109
  hugepages-1Gi:      0
  hugepages-2Mi:      0
  memory:             16384596Ki
  nvidia.com/gpu:     1

We can use the following workload to test. Note the runtimeClassName reference in the Pod spec:

 cat << EOF | kubectl create -f -
apiVersion: v1
kind: Pod
metadata:
  name: cuda-vectoradd
spec:
  restartPolicy: OnFailure
  runtimeClassName: nvidia
  containers:
  - name: cuda-vectoradd
    image: "nvidia/samples:vectoradd-cuda11.2.1"
    resources:
      limits:
         nvidia.com/gpu: 1
EOF

Logs from the Pod will indicate if it was successful:

root@ubuntu:~# kubectl logs cuda-vectoradd
[Vector addition of 50000 elements]
Copy input data from the host memory to the CUDA device
CUDA kernel launch with 196 blocks of 256 threads
Copy output data from the CUDA device to the host memory
Test PASSED

Without providing the runtimeClassName in the spec the Pod will error:

root@ubuntu:~# kubectl logs cuda-vectoradd
[Vector addition of 50000 elements]
Failed to allocate device vector A (error code CUDA driver version is insufficient for CUDA runtime version)!

Using the NSX-T CNI with RKE2

This post outlines the necessary steps to leverage the VMware NSX-T CNI with RKE2

1. Planning

The following illustrates my Lab environment with a single node cluster:

General Considerations

If you have a new NSX-T environment, ensure you have (as a minimum) the following:

  • T0 Router
  • T1 Router
  • Edge Cluster
  • VLAN Transport Zone
  • Overlay Transport Zone
  • Route advertisement (BGP/OSPF) to the physical network

NSX Specific Considerations

  • A network segment (or vds port) for management traffic (NS-K8S-MGMT in this example)
  • A network segment for overlay traffic (NS-K8S-OVERLAY)

Management traffic should be put on a routed network
Overlay traffic does not have to be on a routed network

You will need to acquire and upload the ncp container image to a private repo:

This will contain the NCP image

2. Prepare NSX Objects

  • Create and retrieve the object ID’s for:
  • An IP Block for the Pods (this /16 will be divided into /24’s in our cluster)
  • An IP Pool for loadBalancer service types

3. Create VM

  • Create a VM with one nic attached to the Management network, and one attached to the Overlay network. Note, for ease you can configure NSX-T to provide DHCP services to both
  • Ensure Python is Installed (aka Python2)

4. Install RKE2

  • Create the following configuration file to instruct RKE2 not to auto-apply a CNI:
packerbuilt@k8s-test-node:~$ cat /etc/rancher/rke2/config.yaml 
cni:
  - none
  • Install RKE2
curl -sfL https://get.rke2.io | sh -
systemctl enable rke2-server.service
systemctl start rke2-server.service
# Wait a bit
export KUBECONFIG=/etc/rancher/rke2/rke2.yaml PATH=$PATH:/var/lib/rancher/rke2/bin
kubectl get nodes
  • You will notice some pods are in pending state – this is normal as these reside outside of the host networking namespace and we have yet to install a CNI

5. Install additional CNI binaries

  • NSX-T also requires access to the portmap CNI binary. this can be acquired by:
wget https://github.com/containernetworking/plugins/releases/download/v1.1.1/cni-plugins-linux-amd64-v1.1.1.tgz
  • Extract the contents to /opt/cni/bin/

6. Tag the overlay network port on the VM

The NSX-T container plugin needs to identify the port used for container traffic. In the example above, this is the interface connection to our Overlay switch

img_1.png
  • In NSX-T navigate to Inventory -> Virtual Machines -> Select the VM
  • Select the port that’s connected to the overlay switch
  • Add the tags as appropriate

7. Download the NCP operator files

  • git clone https://github.com/vmware/nsx-container-plugin-operator
  • Change directory – cd /deploy/kubernetes/

8. Change the Operator yaml

  • Operator.yaml – replace where the image resides in your environment. Example:
            - name: NCP_IMAGE
              value: "core.harbor.virtualthoughts.co.uk/library/nsx-ncp-ubuntu:latest"

9. Change the Configmap yaml file

Which values to change will depend on your deployment topology, but as an example:

@@ -11,7 +11,7 @@ data:
 
     # If set to true, the logging level will be set to DEBUG instead of the
     # default INFO level.
-    #debug = False
+    debug = True
 
 
 
@@ -52,10 +52,10 @@ data:
     [coe]
 
     # Container orchestrator adaptor to plug in.
-    #adaptor = kubernetes
+    adaptor = kubernetes
 
     # Specify cluster for adaptor.
-    #cluster = k8scluster
+    cluster = k8scluster-lspfd2
 
     # Log level for NCP modules (controllers, services, etc.). Ignored if debug
     # is True
@@ -111,10 +111,10 @@ data:
     [k8s]
 
     # Kubernetes API server IP address.
-    #apiserver_host_ip = <None>
+    apiserver_host_ip = 172.16.100.13
 
     # Kubernetes API server port.
-    #apiserver_host_port = <None>
+    apiserver_host_port = 6443
 
     # Full path of the Token file to use for authenticating with the k8s API
     # server.
@@ -129,7 +129,7 @@ data:
     # Specify whether ingress controllers are expected to be deployed in
     # hostnework mode or as regular pods externally accessed via NAT
     # Choices: hostnetwork nat
-    #ingress_mode = hostnetwork
+    ingress_mode = nat
 
     # Log level for the kubernetes adaptor. Ignored if debug is True
     # Choices: NOTSET DEBUG INFO WARNING ERROR CRITICAL
@@ -254,7 +254,7 @@ data:
 
 
     # The OVS uplink OpenFlow port where to apply the NAT rules to.
-    #ovs_uplink_port = <None>
+    ovs_uplink_port = ens224
 
     # Set this to True if you want to install and use the NSX-OVS kernel
     # module. If the host OS is supported, it will be installed by nsx-ncp-
@@ -318,8 +318,11 @@ data:
     # [<scheme>://]<ip_adress>[:<port>]
     # If scheme is not provided https is used. If port is not provided port 80
     # is used for http and port 443 for https.
-    #nsx_api_managers = []
-
+    nsx_api_managers = 172.16.10.43
+    nsx_api_user = admin
+    nsx_api_password = SuperSecretPassword123!
+    insecure = true
+    
     # If True, skip fatal errors when no endpoint in the NSX management cluster
     # is available to serve a request, and retry the request instead
     #cluster_unavailable_retry = False
@@ -438,7 +441,7 @@ data:
     # support automatically creating the IP blocks. The definition is a comma
     # separated list: CIDR,CIDR,... Mixing different formats (e.g. UUID,CIDR)
     # is not supported.
-    #container_ip_blocks = []
+    container_ip_blocks = IB-K8S-PODS 
 
     # Resource ID of the container ip blocks that will be used for creating
     # subnets for no-SNAT projects. If specified, no-SNAT projects will use
@@ -451,7 +454,7 @@ data:
     # creating the ip pools. The definition is a comma separated list:
     # CIDR,IP_1-IP_2,... Mixing different formats (e.g. UUID, CIDR&IP_Range) is
     # not supported.
-    #external_ip_pools = []
+    external_ip_pools = IP-K8S-LB
 
 
 
@@ -461,7 +464,7 @@ data:
     # Name or ID of the top-tier router for the container cluster network,
     # which could be either tier0 or tier1. If policy_nsxapi is enabled, should
     # be ID of a tier0/tier1 gateway.
-    #top_tier_router = <None>
+    top_tier_router = T0
 
     # Option to use single-tier router for the container cluster network
     #single_tier_topology = False
@@ -472,13 +475,13 @@ data:
     # policy_nsxapi is enabled, it also supports automatically creating the ip
     # pools. The definition is a comma separated list: CIDR,IP_1-IP_2,...
     # Mixing different formats (e.g. UUID, CIDR&IP_Range) is not supported.
-    #external_ip_pools_lb = []
+    #external_ip_pools_lb = IP-K8S-LB
 
     # Name or ID of the NSX overlay transport zone that will be used for
     # creating logical switches for container networking. It must refer to an
     # already existing resource on NSX and every transport node where VMs
     # hosting containers are deployed must be enabled on this transport zone
-    #overlay_tz = <None>
+    overlay_tz = nsx-overlay-transportzone
 
 
     # Resource ID of the lb service that can be attached by virtual servers
@@ -500,11 +503,11 @@ data:
 
     # Resource ID of the firewall section that will be used to create firewall
     # sections below this mark section
-    #top_firewall_section_marker = <None>
+    top_firewall_section_marker = 0eee3920-1584-4c54-9724-4dd8e1245378
 
     # Resource ID of the firewall section that will be used to create firewall
     # sections above this mark section
-    #bottom_firewall_section_marker = <None>
+    bottom_firewall_section_marker = 3d67b13c-294e-4470-95db-7376cc0ee079
 
 
 
@@ -523,7 +526,7 @@ data:
 
     # Edge cluster ID needed when creating Tier1 router for loadbalancer
     # service. Information could be retrieved from Tier0 router
-    #edge_cluster = <None>
+    edge_cluster = 726530a3-a488-44d5-aea6-7ee21d178fbc

10. Apply the manifest files

kubectl apply -f /nsx-container-plugin-operator/deploy/kubernetes/*

You should see both the operator and NCP workloads manifest

root@k8s-test-node:/home/packerbuilt/nsx-container-plugin-operator/deploy/kubernetes# kubectl get po -n nsx-system
NAME                       READY   STATUS    RESTARTS   AGE
nsx-ncp-5666788456-r4nzb   1/1     Running   0          4h31m
nsx-ncp-bootstrap-6rncw    1/1     Running   0          4h31m
nsx-node-agent-6rstw       3/3     Running   0          4h31m
root@k8s-test-node:/home/packerbuilt/nsx-container-plugin-operator/deploy/kubernetes# kubectl get po -n nsx-system-operator
NAME                               READY   STATUS    RESTARTS   AGE
nsx-ncp-operator-cbcd844d4-tn4pm   1/1     Running   0          4h31m

Pods should be transitioning to running state, and loadbalancer services will be facilitated by NSX

root@k8s-test-node:/home/packerbuilt/nsx-container-plugin-operator/deploy/kubernetes# kubectl get svc
NAME            TYPE           CLUSTER-IP     EXTERNAL-IP     PORT(S)        AGE
kubernetes      ClusterIP      10.43.0.1      <none>          443/TCP        4h34m
nginx-service   LoadBalancer   10.43.234.41   172.16.102.24   80:31848/TCP   107m
root@k8s-test-node:/home/packerbuilt/nsx-container-plugin-operator/deploy/kubernetes# curl 172.16.102.24
<!DOCTYPE html>
<html>
<head>
<title>Welcome to nginx!</title>
<style>
    body {
      ......
    }

The end result is a topology where every namespace has its own T1 router, advertised to T0:

Evaluating Harvester in vSphere

Disclaimer – The use of nested virtualisation is not a supported topology

Harvester is an open-source HCI solution aimed at managing Virtual Machines, similar to vSphere and Nutanix, with key differences including (but not limited to):

  • Fully Open Source
  • Leveraging Kubernetes-native technologies
  • Integration with Rancher

Testing/evaluating any hyperconverged solution can be difficult – It usually requires having dedicated hardware as these solutions are designed to work directly on bare metal. However, we can circumvent this by leveraging nested virtualisation – something which may be familiar with a lot of homelabbers (myself included) – which involves using an existing virtualisation solution provision workloads that also leverage virtualisation technology.

Step 1 – Planning

To mimic what a production-like system may look like, two NICs will be leveraged – one that facilitates management traffic, and the other for Virtual Machine traffic, as depicted below

MGMT network and VM Network will manifest as VDS Port groups.

Also, download and make available the latest ISO for harvester

Step 2 – Create vDS Port Groups

It is highly recommended to create new Distributed Port groups for this exercise, mainly because of the configuration we will be applying in the next step.

Create a new vDS Port Group:

Give the port group a name, such as harvester-mgmt

Adjust any configuration (ie VLAN ID) to match your environment (if required). Or accept the defaults:

Repeat this process to create the harvester-vm Port group. We should now have two port groups:

  • harvester-mgmt
  • harvester-vm

Step 3 – Enable MAC learning on Port groups [Critical]

William Lam has an excellent post on how to accomplish this. This is required for Harvester (or any hypervisor) to function correctly when operating in a nested environment.

Set-MacLearn -DVPortgroupName @("harvester-mgmt") -EnableMacLearn $true -EnablePromiscuous $false -EnableForgedTransmit $true -EnableMacChange $false

Set-MacLearn -DVPortgroupName @("harvester-vm") -EnableMacLearn $true -EnablePromiscuous $false -EnableForgedTransmit $true -EnableMacChange $false

Step 4 – Creating a Harvester VM

Our Harvester VM will operate like any other VM, with some important differences. In vSphere, go through the standard VM creation wizard to specify the Host/Datastore options. When presented with the OS type, select Other Linux (64 bit).

When customising the hardware, select Expose hardware assisted virtualization to the guest OS – This is crucial, as without this selected Harvester will not install.

Add an additional network card so that our VM leverages both previously created port groups:

And finally, mount the Harvester ISO image.

Step 4 – Install Harvester

Power on the VM and providing the ISO is mounted and connected, you should be presented with the install screen. As this is the first node, select create a new Harvester Cluster

Select the Install target and optional MBR partitioning

Configure the hostname, management nic and IP assignment options.

Configure the DNS config:

Configure the Harvester VIP. This is what we will use to access the Web UI. This can also be obtained via DHCP if desired.

Configure the cluster token, this is required if you want to add more nodes later on.

Configure the local Password:

Configure the NTP server Address:

If desired, the subsequent options facilitate importing SSH keys, reading a remote config, etc which are optional. A summary will be presented before the install begins:

Proceed with the install.

Note : After a reboot, it may take a few minutes before harvester reports as being in a ready state – Once it does, navigate to the reported management URL.

At which point you will be prompted to reset the admin password

Step 5 – Configure VM Network

Once logged in to Harvester navigate to Hosts > Edit Config

Configure the secondary NIC to the VLAN network (our VM network)

Navigate to Settings > VLAN > Edit

Click “Enable” and select the default interface to the secondary interface. This will be the default for any new nodes that join the cluster.

To create a network for our VM’s to reside in, select Network > Create:

Give this network a name and a VLAN ID. Note – you can supply VLAN ID 1 if you’re using the native/default VLAN.

Step 6 – Test VM Network

Firstly, create a new image:

For this example, we can use an ISO image. After supplying the URL Harvester will download and store the image:

After downloading, we can create a VM from it:

Specify the VM specs (CPU and Mem)

Under Volumes, add an additional volume to act as the installation target for the OS (Or leave if purely wanting to use a live ISO):

Under Networks, change the selection to the VM network that was previously created and click “Create”:

Once the VM is in running state, we can take a VNC console to it:

At which point we can interact with it as we would expect with any HCI solution:

« Older posts Newer posts »

© 2025 Virtual Thoughts

Theme by Anders NorenUp ↑

Social media & sharing icons powered by UltimatelySocial
RSS
Twitter
Visit Us
Follow Me