Runs

A run consists of an algorithm stored in a code repository, data stored in a data repository (not yet implemented) and an infrastructure (e.g. HPC) on which the algorithm connected with the data will be run. All runs of a project are stored in the runs repository.

You can view the runs repository by navigating to a public project, or one that you are a member of (see projects), and select “Runs” in the sidebar on the left.

The runs’ details can be viewed and runs can be deleted by selecting the respective action icons - ,.

• Runs

  • Create a new run

  • Run schedules

    • Schedule a run

    • Submitting a run

Create a new run

To create a new run, navigate to the runs -> submissions view in a project and click the “ADD” button.

For creating a new run, you first need an existing connection stored on the Mantik platform. In the form, you then need to provide

• name for your run,

• experiment repository in which the run should be included (choose from the existing experiments from the project),

• code repository with the MLflow project,

• branch within this repository for the algorithm that should be run (choose from the existing code repositories from project),

• relative path to your MLflow project file (relative from the code repository root),

• relative path to the Compute Backend config (relative from the code repository root).

Furthermore, in order to run the algorithm on a remote resource, you need to

• choose the correct connection from the existing connections (see user settings),

• provide a compute budget account (i.e. account/project that is being billed for the compute time).

Note that MANTIK_COMPUTE_BUDGET_ACCOUNT used in CLI or Python API or Compute Backend is the same as the UNICORE compute budget that needs to be specified in the UI.

After clicking RUN!, the run is created, submitted to the remote system, and all details for submitting this run are stored.

Register/unregister Model from a ‘Finished’ run

Click the options menu of a ‘Finished’ run and click the ‘Register Model’ button. Provide a model name that is unique within the project to the pop-up window opened and register the model. Once registered you can also unregister the model using the same menu and click Unregister Model.

Run schedules

Runs are not executed automatically but need to be manually executed. Run Schedules allow to select a prior run for automatic, repeated execution. The interval of execution can be defined as desired.

You can create and view all schedules for runs by selecting Schedules in the sidebar on the left.

Existing schedules can be edited, the schedules’ details can be viewed and schedules can be deleted by selecting the respective action icons - ,,.

Schedule a run

To execute a run, navigate to the schedules view in a project and click the “Add” button.

For creating a new schedule, you need to provide a name for your schedule and a run (choose from the existing runs from the project) Further, you need to choose a connection (choose from the existing connections, see user settings) and UNICORE compute budget account. The timing of the schedule needs to be specified by using Cron expressions (click on “UNIX Cron format” for a reference), a time zone, and an end date.