Deploying vGPU workloads on CREODIAS Kubernetes

Utilizing GPU (Graphical Processing Units) presents a highly efficient alternative for fast, highly parallel processing of demanding computational tasks such as EODATA processing, machine learning and many others.

In cloud environment, virtual GPU units (vGPU) are available with certain Virtual Machine flavors. This guide provides instructions how to attach such VMs with GPU as Kubernetes cluster nodes and utilize GPU from Kubernetes pods.

We will present three alternative ways for adding vGPU capability to your Kubernetes cluster, based on your required scenario. For each, you should be able to verify the vGPU installation and test it by running vGPU workload.

What Are We Going To Cover

  • Scenario No. 1 - Add vGPU nodes as a nodegroup on a non-GPU Kubernetes clusters created after June 21st 2023

  • Scenario No. 2 - Add vGPU nodes as nodegroups on non-GPU Kubernetes clusters created before June 21st 2023

  • Scenario No. 3 - Create a new GPU-first Kubernetes cluster with vGPU-enabled default nodegroup

  • Verify the vGPU installation

  • Test vGPU workload

  • Add non-GPU nodegroup to a GPU-first cluster

Prerequisites

No. 1 Hosting

You need a CREODIAS hosting account with Horizon interface https://horizon.cloudferro.com.

No. 2 Knowledge of RC files and CLI commands for Magnum

You should be familiar with utilizing OpenStack CLI and Magnum CLI. Your RC file should be sourced and pointing to your project in OpenStack. See article

How To Install OpenStack and Magnum Clients for Command Line Interface to CREODIAS Horizon.

No. 3 Cluster and kubectl should be operational

To connect to the cluster via kubectl tool, see this article How To Access Kubernetes Cluster Post Deployment Using Kubectl On CREODIAS OpenStack Magnum.

No. 4 Familiarity with the notion of nodegroups

Creating Additional Nodegroups in Kubernetes Cluster on CREODIAS OpenStack Magnum.

No. 5 Cloud environment with both GPU flavors and Kubernetes Magnum support

Currently, WAW3-1 is the only CREODIAS cloud that has

  • GPU flavors available and is

  • supporting Kubernetes, through OpenStack Magnum.

There are four gpu flavors:

Name

RAM

Disk

VCPUs

vm.a6000.1

14336

40

2

vm.a6000.2

28672

80

4

vm.a6000.3

57344

160

8

vm.a6000.4

114688

320

16

Scenario 1 - Add vGPU nodes as a nodegroup on a non-GPU Kubernetes clusters created after June 21st 2023

In order to create a new nodegroup, called gpu, with one node vGPU flavor, say, vm.a6000.2, we can use the following Magnum CLI command. Adjust the node-count and flavor to your preference, adjust the $CLUSTER_ID to the one of your clusters (this can be taken from Clusters view in Horizon UI), and ensure the role is set as worker.

The key setting is adding a label worker-type=gpu:

The command is:

openstack coe nodegroup create $CLUSTER_ID gpu --labels "worker_type=gpu" --merge-labels --role worker --flavor vm.a6000.2 --node-count 1

Your request will be accepted:

../_images/request_for_nodegroup.png

Now list the available nodegroups:

openstack coe nodegroup list $CLUSTER_ID_RECENT --max-width 120

We get:

../_images/result_of_creating_nodegroup.png

The result is that a new nodegroup called gpu is created in the cluster and that it is using the GPU flavor.

Scenario 2 - Add vGPU nodes as nodegroups on non-GPU Kubernetes clusters created before June 21st 2023

The instructions are the same as in the previous scenario, with the exception of adding an additional label:

existing_helm_handler_master_id=$MASTER_0_SERVER_ID

where $MASTER_0_SERVER_ID is the ID of the master0 VM from your cluster. The uuid value can be obtained

  • in Horizon, through the Instances view

  • or using a CLI command to isolate the uuid for the master node:

openstack coe nodegroup list $CLUSTER_ID_OLDER -c uuid -c name -c status -c role
../_images/older_uuid.png

In this example, uuid is 413c7486-caa9-4e12-be3b-3d9410f2d32f. Set up the value for master handler label:

export MASTER_0_SERVER_ID="413c7486-caa9-4e12-be3b-3d9410f2d32f"

and execute the following command to create an additional nodegroup in this scenario:

openstack coe nodegroup create $CLUSTER_ID_OLDER gpu --labels "worker_type=gpu,existing_helm_handler_master_id=$MASTER_0_SERVER_ID" --merge-labels --role worker --flavor vm.a6000.2 --node-count 1

There may not be any space between the labels.

The request will be accepted and after a while, a new nodegroup will be available and based on GPU flavor. List the nodegroups with the command:

openstack coe nodegroup list $CLUSTER_ID_OLDER --max-width 120

Scenario 3 - Create a new GPU-first Kubernetes cluster with vGPU-enabled default nodegroup

To create a new vGPU-enabled cluster, you can use the usual Horizon commands, selecting one of the existing templates with vgu in their names:

../_images/gpu_templates.png

In the below example, we use the CLI to create a cluster called k8s-gpu-with_template with k8s-1.23.16-vgpu-v1.0.0 template. The sample cluster has

  • one master node with flavor eo1.medium and

  • one worker node with vm.a6000.2 flavor with vGPU enabled.

To adjust these parameters to your requirements, you will need to replace the $KEYPAIR to your own. Also, to verify that the nvidia labels are correctly installed, first create a namespace called nvidia-device-plugin. You can then list the namespaces to be sure that it was created properly. So, the preparation commands look like this:

export KEYPAIR="sshkey"
kubectl create namespace nvidia-device-plugin
kubectl get namespaces

The final command to create the required cluster is:

openstack coe cluster create k8s-gpu-with_template --cluster-template "k8s-1.23.16-vgpu-v1.0.0" --keypair=$KEYPAIR --master-count 1 --node-count 1

Verify the vGPU installation

You can verify that vGPU-enabled nodes were properly added to your cluster, by checking the nvidia-device-plugin deployed in the cluster, to the nvidia-device-plugin namespace. The command to list the contents of the nvidia namespace is:

kubectl get daemonset nvidia-device-plugin -n nvidia-device-plugin
../_images/daemonset_01.png

See which nodes are now present:

kubectl get node

Each GPU node, should have several nvidia labels added. To verify, you can run one of the below commands, the second of which will show the labels formatted:

kubectl get node k8s-gpu-cluster-XXXX --show-labels
kubectl get node k8s-gpu-cluster-XXXX -o go-template='{{range $key, $value := .metadata.labels}}{{$key}}: {{$value}}{{"\n"}}{{end}}'

Concretely, in our case, the second command is:

kubectl get node k8s-gpu-with-template-lfs5335ymxcn-node-0 -o go-template='{{range $key, $value := .metadata.labels}}{{$key}}: {{$value}}{{"\n"}}{{end}}'

and the result will look like this:

../_images/go_template_image.png

Also, GPU workers are tainted by default with the taint:

node.cloudferro.com/type=gpu:NoSchedule

This can be verified by running the following command, in which we are using the name of the existing node:

kubectl describe node k8s-gpu-with-template-lfs5335ymxcn-node-0 | grep 'Taints'

Run test vGPU workload

We can run a sample workload on vGPU. To do so, create a YAML manifest file vgpu-pod.yaml, with the following contents:

vgpu-pod.yaml

apiVersion: v1
kind: Pod
metadata:
  name: gpu-pod
spec:
  restartPolicy: Never
  containers:
    - name: cuda-container
      image: nvcr.io/nvidia/k8s/cuda-sample:vectoradd-cuda10.2
      resources:
        limits:
          nvidia.com/gpu: 1 # requesting 1 vGPU
  tolerations:
  - key: nvidia.com/gpu
    operator: Exists
    effect: NoSchedule
  - effect: NoSchedule
    key: node.cloudferro.com/type
    operator: Equal
    value: gpu

Apply with:

kubectl apply -f vgpu-pod.yaml
../_images/apply_yaml.png

This pod will request one vGPU, so effectively it will utilize the vGPU allocated to a single node. For example, if you had a cluster with 2 vGPU-enabled nodes, you could run 2 pods requesting 1 vGPU each.

Also, for scheduling the pods on GPU, you will need to apply the two tolerations as per the example above, That, effectively, means that the pod will only be scheduled on GPU nodes.

Looking at the logs, we see that the workload was indeed performed:

kubectl logs gpu-pod


[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
Done

Add non-GPU nodegroup to a GPU-first cluster

We refer to GPU-first clusters as the ones created with worker_type=gpu flag. For example, in cluster created with Scenario No. 3, the default nodegroup consists of vGPU nodes.

In such clusters, to add an additional, non-GPU nodegroup, you will need to:

  • specify the image ID of the system that manages this nodegroup

  • add the label worker_type=default

  • ensure that the flavor for this nodegroup is non-GPU.

In order to retrieve the image ID, you need to know with which template you want to use to create the new nodegroup. Out of the existing non-GPU templates, we select k8s-1.23.16-v1.0.2 for this example. Run the following command to extract the template ID, as that will be needed for nodegroup creation:

openstack coe cluster template show k8s-1.23.16-v1.0.2 | grep image_id

In our case, this yields the following result:

../_images/template_show_non_gpu.png

We can then add the non-GPU nodegroup with the following command, in which you can adjust the parameters. In our example, we use cluster name from Scenario 3 (the one freshly created with GPU) above and set up worker node flavor to eo1.medium:

export CLUSTER_ID="k8s-gpu-with_template"
export IMAGE_ID="42696e90-57af-4124-8e20-d017a44d6e24"
openstack coe nodegroup create $CLUSTER_ID default --labels "worker_type=default" --merge-labels --role worker --flavor "eo1.medium" --image $IMAGE_ID --node-count 1

Then list the nodegroup contents to see whether the creation succeeded:

openstack coe nodegroup list $CLUSTER_ID --max-width 120
../_images/nodegroups_in_cluster_id.png