Skip to content

Submit a Run:AI Job via YAML

The easiest way to submit Jobs to the Run:AI GPU cluster is via the Run:AI Command-line interface (CLI). Still, the CLI is not a must. It is only a wrapper for a more detailed Kubernetes API syntax using YAML.

There are cases where you want to forgo the CLI and use direct YAML calls. A frequent scenario for using the Kubernetes YAML syntax to submit Jobs is integrations. Researchers may already be working with an existing system that submits Jobs, and want to continue working with the same system. Though it is possible to call the Run:AI CLI from the customer's integration, it is sometimes not enough.


We differentiate between two types of Workloads:

  • Train workloads. Training is characterized by a deep learning session that has a start and an end. A Training session can take anywhere from a few minutes to a couple of weeks. It can be interrupted in the middle and later restored. Training workloads typically utilize large percentages of GPU computing power and memory.
  • Build workloads. Build workloads are interactive. They are used by data scientists to write machine learning code and test it against subsets of the data. Build workloads typically do not maximize usage of the GPU.

The internal Kubernetes implementation of Train is a CRD (Customer Resource) named RunaiJob which is similar to a Kubernetes Job. The internal implementation of Buils is a Kubernetes StatesfulSet.

  • A Kubernetes Job is used for Train workloads. A Job has a distinctive "end" at which time the Job is either "Completed" or "Failed"
  • A Kubernetes StatefulSet is used for Build workloads. Build workloads are interactive sessions. StatefulSets do not end on their own. Instead, they must be manually stopped

Run:AI extends the Kubernetes Scheduler. A Kubernetes Scheduler is the software that determines which Workload to start on which node. Run:AI provides a custom scheduler named runai-scheduler.

The Run:AI scheduler schedules computing resources by associating Workloads with Run:AI Projects:

  • A Project is assigned with a GPU quota through the Run:AI Administrator user interface.
  • Each workload must be associated with a Project name and will receive resources according to the defined quota for the Project and the currently running Workloads

Internally, Run:AI Projects are implemented as Kubernetes namespaces. The scripts below assume that the code is being run after the relevant namespace has been set.

Submit Workloads

Regular Jobs

  • <JOB-NAME>. The name of the Job.

  • <IMAGE-NAME>. The name of the docker image to use. Example:

  • <USER-NAME> The name of the user submitting the Job. The name is used for display purposes only when Run:AI is installed in an unauthenticated mode.

  • <REQUESTED-GPUs>. An integer number of GPUs you request to be allocated for the Job. Examples: 1, 2

Copy the following into a file and change the parameters:

kind: RunaiJob (* see note below)
  name: <JOB-NAME>
    priorityClassName: "build" (** see note below)
        user: <USER-NAME>
      - name: <JOB-NAME>
        image: <IMAGE-NAME>
      restartPolicy: Never
      schedulerName: runai-scheduler


kubectl apply -f <FILE-NAME>

to submit the Job.


  • You can use either a regular Job or RunaiJob. The later is a Run:AI object which solves various Kubernetes Bugs and provides a better naming for multiple pods in Hyper-Parameter Optimization scenarios ** Using 'build' in this field is equivilant to run a job with '--interactive' flag. To run Train job, delete this line
  • The runai submit CLI command includes many more flags. These flags can be correlated to Kubernetes API functions and added to the YAML above.

Using Fractional GPUs

Jobs with Fractions requires a change in the above YAML. Specifically the limits section:

limits: <REQUESTED-GPUs>

should be omitted and replaced with:

        gpu-fraction: "0.5"

where "0.5" is the requested GPU fraction.

Delete Workloads

To delete a Run:AI workload you need to delete the Job or StatefulSet according to the workload type

kubectl delete job <JOB-NAME>


kubectl delete sts <STS-NAME>

See Also

Last update: January 18, 2021