Skip to content

Part 4: Workload

What Will You Do

In this part of the self-paced exercise, you will deploy a "GPU workload" to your Amazon EKS cluster that has a GPU node group.

Step 1: Namespace

In a typical production environment, administrators will have already created a "Kubernetes Namespace" for your workload. In this exercise, let us go ahead and create a namespace.

  • Open Terminal (on macOS/Linux) or Command Prompt (Windows) and navigate to the folder where you forked the Git repository
  • Navigate to the folder "/getstarted/gpueks/workload"
  • Type the command
rctl create ns -f namespace.yaml 

This step creates a namespace in your project. The controller can create a namespace on "multiple clusters" based on its placement policy.

rctl publish ns gputest


To verify that then namespace was successfully created on your EKS cluster

rctl get ns gputest 

You should see results like the following. Note that the namespace was successfully created on your EKS cluster.

| NAME    | TYPE        | CREATEDAT                    | DEPLOYEDCLUSTERS | ID      | LABELS | ANNOTATIONS |
| gputest | RafayWizard | Mon Nov 22 01:58:20 UTC 2021 | demo-gpueks      | kre63zm | []     | []          |

Step 2: Deploy Workload

The "gputest.yaml" file contains the declarative specification for our GPU Workload. Let us review it.

name: gputest
namespace: gputest
project: default
type: NativeYaml
clusters: demo-gpueks
payload: ./gpu-job.yaml

Note that the workload's name is "gputest" and it is of type "k8s YAML". The actual k8s YAML file is in the payload "gpu-job.yaml".

rctl create workload gputest.yaml

If there were no errors, you should see a message like below

Workload created successfully

Now, let us publish the newly created workload to the downstream clusters. The workload can be deployed to multiple clusters as per the configured "placement policy". In this case, you are deploying to a single EKS cluster with the name "demo-gpueks".

rctl publish workload gputest

Step 4: Verify

In the web console, click on Applications -> Workloads. You should see something like the following.

Published Workload

Step 5: GPU Dashboard

The GPU workload you deployed is a "cron job" that is scheduled to automatically wake up "every 5 minutes" and consume the GPU attached to the EKS cluster. Administrators that wish to view GPU metrics have access to an "integrated GPU dashboard"

  • Navigate to Infrastructure -> Clusters
  • Click on "GPU" in the cluster card to open the GPU selector

GPU Selector

  • Select the GPU by clicking on it to view its dashboard. You should see something like the following

GPU Dashboard


Congratulations! At this point, you have successfuly

  • Configured and provisioned an Amazon EKS cluster with a GPU node group
  • Deployed a "GPU Workload" to the EKS Cluster and reviewed the integrated "GPU Dashboards"