Skip to content

GPU pods

❗ In this section you will request GPUs. Make sure you don't waste those and delete your pods when not using the GPUs.

Running GPU pods

Use this definition to create your own pod and deploy it to kubernetes:

apiVersion: v1
kind: Pod
metadata:
  name: gpu-pod-example
spec:
  containers:
  - name: gpu-container
    image: gitlab-registry.nrp-nautilus.io/prp/jupyter-stack/prp:latest
    command: ["sleep", "infinity"]
    resources:
      limits:
        nvidia.com/gpu: 1

This example requests 1 GPU device. You can have up to 8 per node if you're using jobs, and up to 2 for pods. If you request GPU devices in your pod, kubernetes will auto schedule your pod to the appropriate node. There's no need to specify the location manually.

You should always delete your pod when your computation is done to let other users use the GPUs. Consider using Jobs with actual script instead of sleep whenever possible to ensure your pod is not wasting GPU time. If you have never used Kubernetes before, see the tutorial.

Requesting high-demand GPUs

Certain kinds of GPUs have much higher specs than the others, and to avoid wasting those for regular jobs, your pods will only be scheduled on those if you request the type explicitly.

Currently those include:

  • NVIDIA-TITAN-RTX
  • NVIDIA-RTX-A5000
  • Quadro-RTX-6000
  • Tesla-V100-SXM2-32GB
  • NVIDIA-A40
  • NVIDIA-RTX-A6000
  • Quadro-RTX-8000
  • NVIDIA-A100-SXM4-80GB*

*A100 running in MIG mode is not considered high-demand one.

Requesting many GPUs

Since 1 and 2 GPU jobs are blocking nodes from getting 4 and 8-GPU jobs, there are some nodes reserved for those. Once you submit a job with 4 or 8 GPUs request, a controller will automatically add toleration which will allow you to use the node reserved for more GPUs. You don't need to do anything manually for that.

Choosing GPU type

We have a variety of GPU flavors attached to Nautilus. This table describes the types of GPUs available for use, but is not up to date - it's better to use the actual cluster information (f.e. kubectl get nodes -L nvidia.com/gpu.product).

Credit: GPU types by NRP Nautilus

If you need more graphical memory, use this table or official specs to choose the type:

GPU Type Memory size (GB)
NVIDIA-GeForce-GTX-1070 8G
NVIDIA-GeForce-GTX-1080 8G
Quadro-M4000 8G
NVIDIA-A100-PCIE-40GB-MIG-2g.10gb 10G
NVIDIA-GeForce-GTX-1080-Ti 12G
NVIDIA-GeForce-RTX-2080-Ti 12G
NVIDIA-TITAN-Xp 12G
Tesla-T4 16G
NVIDIA-A10 24G
NVIDIA-GeForce-RTX-3090 24G
NVIDIA-GeForce-RTX-3090 24G
NVIDIA-TITAN-RTX 24G
NVIDIA-RTX-A5000 24G
Quadro-RTX-6000 24G
Tesla-V100-SXM2-32GB 32G
NVIDIA-A40 48G
NVIDIA-RTX-A6000 48G
Quadro-RTX-8000 48G

NOTE: Not all nodes are available to all users. You can consult about your available resources in Matrix and on resources page. Labs connecting their hardware to our cluster have preferential access to all our resources.

To use a specific type of GPU, add the affinity definition to you pod yaml file. The example below specifies 1080Ti GPU: 

spec:
 affinity:
   nodeAffinity:
     requiredDuringSchedulingIgnoredDuringExecution:
       nodeSelectorTerms:
       - matchExpressions:
         - key: nvidia.com/gpu.product
           operator: In
           values:
           - NVIDIA-GeForce-GTX-1080-Ti

To make sure you did everything correctly after you've submited the job, look at the corresponding pod yaml (kubectl get pod ... -o yaml) and check that resulting nodeAffinity is as expected.

Selecting CUDA version

In general the higher CUDA versions support the lower and same driver version. The nodes are labelled with the major and minor CUDA and driver versions. You can check those at the resources page or list with this command (it will also choose only GPU nodes):

kubectl get nodes -L nvidia.com/cuda.driver.major,nvidia.com/cuda.driver.minor,nvidia.com/cuda.runtime.major,nvidia.com/cuda.runtime.minor -l nvidia.com/gpu.product

If you're using the container image with higher CUDA version, you have to pick the nodes supporting it. Example:

spec:
  affinity:
    nodeAffinity:
      requiredDuringSchedulingIgnoredDuringExecution:
        nodeSelectorTerms:
        - matchExpressions:
          - key: nvidia.com/cuda.runtime.major
            operator: In
            values:
            - "12"
          - key: nvidia.com/cuda.runtime.minor
            operator: In
            values:
            - "2"

Also you can choose the driver above something if you know which one you need (this will pick drivers above 535):

spec:
  affinity:
    nodeAffinity:
      requiredDuringSchedulingIgnoredDuringExecution:
        nodeSelectorTerms:
        - matchExpressions:
          - key: nvidia.com/cuda.driver.major
            operator: Gt
            values:
            - "535"

MIG mode

A100 GPUs allow slicing those into several logical GPUs ( MIG mode ). This mode is enabled in our cluster. Things can change, but currently we're thinking about slicing those in halves. The current MIG mode can be obtained from nodes via the nvidia.com/gpu.product label: NVIDIA-A100-PCIE-40GB-MIG-2g.10gb means 2 compute instances (out of 7 total) and 10GB memory per virtual GPU.