Resources#

A resource spec defines the number and types of resources required to run the task. The following fields are used to describe a resource spec:

Field	Description
`cpu`	Specify the amount of cores to request
`memory`	Specify the amount of memory (RAM) to use.
`storage`	Specify the amount of disk space to use.
`gpu`	Specify the amount of GPUs to request
`platform`	Specify the platform to target. If no platform is specified, the default platform for the pool is used if the admins have specified a default platform. Learn more at Pool List.
`nodesExcluded`	Specify the nodes to exclude from the resource spec.
`topology`	Specify topology co-location requirements for tasks using this resource spec. Only available on pools with topology keys configured. See Topology-Aware Scheduling for details.

Note

The default resource spec can be configured but requires service-level configuration. If you have administrative access, you can enable this directly. Otherwise, contact someone with pool administration privileges.

Multiple resource specs can be defined in the same workflow and assigned individually to tasks. To define a resource spec in the workflow, use the resources field under workflow. To assign the resource to each task, use the resource field under tasks:

workflow:
  name: my_workflow
  resources:
    default:                # (1)
      cpu: 1
      memory: 16Gi
      storage: 1Gi
      platform: ovx-a40
    x86_gpu:                # (2)
      cpu: 4
      gpu: 1
      memory: 16Gi
      storage: 1Gi
      platform: dgx-a100
  tasks:
  - name: task1
    resource: default       # (3)
    ...
  - name: task2
    resource: x86_gpu       # (4)
    ...
  - name: task3             # (5)
    ...

Defines the default resource spec which targets an A40 node.
Defines the x86_gpu resource spec uses a single GPU which targets an A100 node.
Assigns the default resource to task1.
Assigns the x86_gpu resource to task2.
Since task3 does not define a resource, the default resource spec will be used.

If the resource field is left blank, the default resources are used.

Follow the Resource List CLI for available resources before building the resource spec.

If there are some node with poor performance or network in the pool, you can exclude them using the nodesExcluded field in the resource spec:

resources:
  default:
    cpu: 1
    memory: 16Gi
    storage: 1Gi
    nodesExcluded:
    - worker1
    - worker2

Warning

Excluding too many nodes can lead to the tasks stuck in PENDING forever!

Topology Requirements#

The topology field specifies a list of co-location requirements for tasks that use this resource spec. Each entry has the following fields:

Field	Description
`key`	The topology key to apply the constraint to (e.g., `rack`, `gpu-clique`). Must match a key configured in the pool’s topology keys.
`group`	A user-defined name grouping tasks that must share the same topology value. Tasks with the same `key` and `group` will be co-located together. Defaults to `default`.
`requirementType`	Either `required` (default) or `preferred`. Use `required` to block scheduling unless the constraint is satisfied. Use `preferred` to allow scheduling even if the constraint cannot be met.

For example, to require all tasks using this resource spec to be scheduled on the same gpu-clique:

resources:
  default:
    gpu: 8
    topology:
    - key: gpu-clique