Requesting Resources#

In this tutorial, you will learn how to request and configure compute resources for your workflows. Understanding resource management is essential for running efficient workflows at scale.

Every task in OSMO runs on compute nodes with specific resources. You can specify:

CPU - Number of CPU cores
Memory - Amount of RAM
Storage - Disk space
GPU - Number of GPUs
Platform - Target specific compute platforms

Understanding Resources#

Resources in OSMO are specified using standard Kubernetes resource notation:

CPU: Cores (e.g., 1, 2, 0.5)
Memory: Bytes with units (e.g., 1Gi, 512Mi, 2Gi)
Storage: Bytes with units (e.g., 10Gi, 100Mi)
GPU: Count (e.g., 1, 2, 4)

Note

OSMO only accepts resource units specified binary units like Gi (gibibytes) or Mi (mebibytes), not in decimal units like GB (gigabytes) or MB.

1 Gi = 1,073,741,824 bytes
1 GB = 1,000,000,000 bytes

Basic Resource Specification#

Resources are defined under the workflow in the resources field and assigned using the resource field in the tasks section. If a task doesn’t specify a resource, it automatically uses the default resource spec:

workflow:
  name: resource-demo
  resources:
    default:
      cpu: 2
      memory: 4Gi
      storage: 10Gi
  tasks:
  - name: my-task
    image: ubuntu:24.04
    command: ["bash", "-c", "echo 'Using default resources'"]

This task will automatically use 2 CPU cores, 4Gi of memory, and 10Gi of storage because no explicit resource field is specified.

Multiple Resource Specifications#

You can define multiple named resource specifications and assign them to different tasks:

workflow:
  name: mixed-resources
  resources:
    default:
      cpu: 1
      memory: 2Gi
      storage: 5Gi
    gpu_resource:
      cpu: 8
      gpu: 1
      memory: 32Gi
      storage: 100Gi
  tasks:
  - name: light-task
    image: ubuntu:24.04
    command: ["bash", "-c", "echo 'Using default resources'"]
    # No resource field specified, so 'default' is used

  - name: gpu-task
    image: ubuntu:24.04
    command: ["bash", "-c", "nvidia-smi"]
    resource: gpu_resource  # Explicitly uses gpu_resource

In this example, light-task uses the default resource spec automatically, while gpu-task explicitly requests the gpu_resource spec.

Targeting Specific Platforms#

Use the platform field to target specific compute platforms. If no platform is specified, the default platform for the pool is used (if configured by administrators):

workflow:
  name: platform-demo
  resources:
    default:
      cpu: 1
      memory: 16Gi
      storage: 1Gi
      platform: ovx-l40
    x86_gpu:
      cpu: 4
      gpu: 1
      memory: 16Gi
      storage: 1Gi
      platform: dgx-h100
  tasks:
  - name: cpu-task
    image: ubuntu:24.04
    command: ["bash", "-c", "echo 'Running on L40'"]
    # Uses default resource with ovx-l40 platform

  - name: gpu-task
    image: nvcr.io/nvidia/pytorch:24.01-py3
    command: ["bash", "-c", "echo 'Running on DGX H100 with GPU'"]
    resource: x86_gpu  # Uses x86_gpu resource with dgx-h100 platform

To see available platforms and pools, use:

$ osmo pool list

This shows all pools, along with their available platforms and resources.

Checking Available Resources#

Before building your resource specifications, check what’s available in your pools:

$ osmo resource list -p <pool_name>

Node             Pool      Platform       Storage [Gi]   CPU [#]   Memory [Gi]   GPU [#]
========================================================================================
hil-ovx-01       default   ovx-a40        0/2028         1/127     1/1006        0/8
hil-ovx-02       default   ovx-a40        14/2028        25/127    15/1006       0/8
hil-ovx-03       default   ovx-a40        2/2028         4/127     8/1006        0/8
orin-01          default   agx-orin-jp6   0/1646         1/11      1/28          0
orin-02          default   agx-orin-jp6   0/1646         1/11      1/28          0
x86-rtx4090-01   default   x86-4090       0/822          1/31      1/123         0/1
========================================================================================
                                          16/10198       33/434    27/3197       0/25

When building your resource specs, ensure:

Your requested resources fit within what’s available in the pool
The resources can fit on a single node, you can check per-node resources via:
```
$ osmo resource info <node_name>
```

Tip

If your workflow is not getting scheduled, check pool quotas with osmo pool list to see if you have reached quota limits. See Why is my workflow not getting scheduled quickly? for detailed troubleshooting.

Note

Please contact your administrator to configure the default resource spec for your pool.

Next Steps#

Now that you understand resource management, learn about working with data in the next tutorial: Working with Data.