Running workloads with GPUs in Yandex Managed Service for Kubernetes

A Managed Service for Kubernetes cluster allows running workloads on graphics processing units (GPUs), which may be of use in tasks with special computing requirements.

To run workloads using GPUs on Managed Service for Kubernetes cluster pods:

1. Create a pod with a GPU.
2. Test the pod.

If you no longer need the resources you created, delete them.

Required paid resourcesRequired paid resources

• Managed Service for Kubernetes master (see Managed Service for Kubernetes pricing).
• Managed Service for Kubernetes cluster nodes: use of computing resources and storage (see Yandex Compute Cloud pricing).
• Public IP addresses for the Managed Service for Kubernetes cluster's master and nodes with public access enabled (see Yandex Virtual Private Cloud pricing).

Getting startedGetting started

1. If you do not have the Yandex Cloud CLI yet, install and initialize it.

   The folder used by default is the one specified when creating the CLI profile. To change the default folder, use the yc config set folder-id <folder_ID> command. You can also specify a different folder for any command using --folder-name or --folder-id.

   If you access a resource by its name, the search will be limited to the default folder. If you access a resource by its ID, the search will be global, i.e., through all folders based on access permissions.

2. Create security groups for the Managed Service for Kubernetes cluster and its node groups.

   Warning

   The configuration of security groups determines performance and availability of the cluster and the services and applications running in it.

3. Create a Managed Service for Kubernetes cluster of any suitable configuration. When creating, specify the security groups prepared earlier.

4. Create a Managed Service for Kubernetes node group with the following settings:

   • Platform: Select With GPU → Intel Broadwell with NVIDIA® Tesla v100.
   • GPU: Specify the required number of GPUs.
   • Security groups: Select the security groups you created earlier.

   Note

   The cgroup v2 resource management mechanism is not supported for nodes on the Gen2 (gpu-standard-v3i) platform.

5. Install kubectl and configure it to work with the new cluster.

Create a pod with a GPUCreate a pod with a GPU

1. Save the GPU pod specification to a YAML file named cuda-vector-add.yaml:

   apiVersion: v1
   kind: Pod
   metadata:
     name: cuda-vector-add
   spec:
     restartPolicy: OnFailure
     containers:
       - name: cuda-vector-add
         # https://github.com/kubernetes/kubernetes/blob/v1.7.11/test/images/nvidia-cuda/Dockerfile
         image: "registry.k8s.io/cuda-vector-add:v0.1"
         resources:
           limits:
             nvidia.com/gpu: 1 # Request for 1 GPU.
   

   Learn more about the pod creation specification in this Kubernetes guide.

2. Create the pod with a GPU:

   kubectl create -f cuda-vector-add.yaml
   

Test the podTest the pod

1. View the information about the new pod:

   kubectl describe pod cuda-vector-add
   

   Result:

   Name:         cuda-vector-add
   Namespace:    default
   Priority:     0
   ...
     Normal  Pulling    16m   kubelet, cl1i7hcbti99********-ebyq  Successfully pulled image "registry.k8s.io/cuda-vector-add:v0.1"
     Normal  Created    16m   kubelet, cl1i7hcbti99********-ebyq  Created container cuda-vector-add
     Normal  Started    16m   kubelet, cl1i7hcbti99********-ebyq  Created container
   

2. View the pod logs:

   kubectl logs -f cuda-vector-add
   

   Result:

   [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
   

Delete the resources you createdDelete the resources you created

Delete the resources you no longer need to avoid paying for them:

1. Delete the Managed Service for Kubernetes cluster.
2. Delete the public static IP address for your Managed Service for Kubernetes cluster if you reserved one.