Workloads Overview¶
Workloads¶
Run:ai schedules Workloads. Run:ai workloads are comprised of:
- The Kubernetes object (Job, Deployment, etc) which is used to launch the container, inside which the data science code runs.
- A set of additional resources that are required to run the Workload. Examples: a service entry point that allows access to the Job, a persistent volume claim to access data on the network, and more.
All of these components are created together and deleted together when the Workload ends.
Run:ai currently supports the following Workloads types:
| Workload Type | Kubernetes Name | Description |
|---|---|---|
| Interactive | InteractiveWorkload | Submit an interactive workload |
| Training | TrainingWorkload | Submit a training workload |
| Distributed Training | DistributedWorkload | Submit a distributed training workload using TensorFlow, PyTorch or MPI |
| Inference | InferenceWorkload | Submit an inference workload |
Values¶
A Workload will typically have a list of values (sometimes called flags), such as name, image, and resources. A full list of values is available in the runai-submit Command-line reference.
How to Submit¶
A Workload can be submitted via various channels:
- The Run:ai user interface.
- The Run:ai command-line interface, via the runai submit command.
- The Run:ai Cluster API.
Workload Policies¶
As an administrator, you can set Policies on Workloads. Policies allow administrators to impose restrictions and set default values for Researcher Workloads. For more information see Workload Policies.