Configuring workload resource profiles

Each project editor session and deployment consumes compute resources within the Workbench cluster. If you need to run applications that require more memory or compute power than Anaconda provides by default, you can create customized resource profiles and configure the number of cores and amount of memory/RAM available for them. You can then allow users to access your customized resource profiles either globally or based on their role, assigned group, or as individual users. For example, if your installation includes nodes with GPUs, you can add a GPU resource profile so users can access the GPUs to accelerate computation within their projects, which is essential for AI/ML model training. Resource profiles that you create are listed for users when they create or deploy a project. You can create as many resource profiles as needed.

Connect to your instance of Workbench.
View a list of your configmaps by running the following command:

kubectl get cm

Edit the anaconda-enterprise-anaconda-platform.yml file.

kubectl edit cm anaconda-enterprise-anaconda-platform

Anaconda recommends making a backup copy of this file before you edit it. Any changes you make will impact how Workbench functions.

Find the resource-profiles: section of the file.
Add any additional resources using the following examples as a template for your resource profiles, then customize them for your environment:

Resource profile examples

resource-profiles:
  resource-profile:
    description: Custom resource profile (global)
    resources:
      limits:
        cpu: "2"
        memory: 4096Mi
    user_visible: true

  roles_profile:
    description: Custom resource profile (roles)
    resources:
      limits:
        cpu: "3"
        memory: 4096Mi
    user_visible: true
    acl:
      roles:
        - ae-creator

  groups_profile:
    description: Custom resource profile (group)
    resources:
      limits:
        cpu: "4"
        memory: 4096Mi
    user_visible: true
    acl:
      groups:
        - managers

  users_profile:
    description: Custom resource profile (users)
    resources:
      limits:
        cpu: "1"
        memory: 4096Mi
    user_visible: true
    acl:
      users:
        - user2

  not_visible_to_anyone:
    description: Custom resource profile (nv)
    resources:
      limits:
        cpu: "3"
        memory: 4096Mi
    user_visible: false

  gpu-profile:
    description: GPU resource profile
    resources:
      limits:
        cpu: "4"
        memory: 8Gi
        nvidia.com/gpu: 1
      requests:
        cpu: "1"
        memory: 2048Mi
        nvidia.com/gpu: 1
    user_visible: true

Resource profiles display their description: as their name. Profiles are listed in alphabetical order, after the default profile.

(Optional) By default, CPU sessions and deployments are allowed to run on GPU nodes. To reserve your GPU nodes for sessions and deployments that require them, comment out the affinity: specification in the file as shown:

(Optional) If you need to schedule user workloads on a specific node, add a node_selector to your resource profile. Use node selectors when running different CPU types, such as Intel and AMD; or different GPU types, such as Tesla v100 and p100. To enable a node selector, add node_selector to the bottom of your resource profile, with the model: value matching the label you have applied to your worker node.

GPU node selector example

gpu-profile:
  description: GPU resource profile
  resources:
    limits:
      cpu: "4"
      memory: 8Gi
      nvidia.com/gpu: 1
    requests:
      cpu: "1"
      memory: 2048Mi
      nvidia.com/gpu: 1
  user_visible: true
  node_selector:
    model: v100

Save your changes to the file.
Restart the workspace and deploy services by running the following command:

kubectl delete pods -l 'app in (ap-workspace, ap-deploy)'

Open a project and view its settings to verify that the resource profiles you added appear in the Resource Profile dropdown.

Data Science & AI Workbench (5.7.1)