Kubeflow Sizing

Describes the resource allocation for different Kubeflow components.

Kubeflow sets default resource usage for each workload and component. You can customize the values for resource consumption for Katib experiments, model serving, and Kubeflow pipelines using the YAML file before applying the YAML file to a cluster. You can customize the resource consumption values for the Notebook while creating a Notebook in the Kubeflow UI.

Katib Experiments
Katib experiments create a pod for each trial, and allocates the following resources:
  • vCPU: 50m
  • Memory: 10Mi
Model Serving
Model serving creates a serving pod for each model, and allocates the following resources:
  • vCPU: 100m
  • Memory: 128Mi
Kubeflow Pipeline
Kubeflow Pipeline creates a workload pod for each step, and allocates the following resources:
  • vCPU: 1
  • Memory: 1Gi
Notebook
The default notebook is allocated with the following resources:
  • vCPU: 1
  • Memory: 2Gi
When you create a new notebook, the following resources will be allocated by default:
  • vCPU: 0.5
  • Memory: 1Gi
However, you can change the value of these resources during the notebook creation step in the Kubeflow UI.