access-experiment-runner

About

The main role of the ACCESS experiment runner is to manage and monitor experiment job runs on the supercomputing environment (e.g., Gadi). It builds on Payu, handling the orchestration of multiple configuration branches, experiment setup, and job lifecycle.

Key features

• Leverages Payu and run multiple experiments from different configuration branches.
• Supports updating parameters even after branches have been created, eliminating the need to delete and recreate entire branches when corrections are required.
• Submits and tracks PBS jobs on Gadi; oversees job lifecycle from submission through completion.
  • When a job completes within expected run times, the tool prints a confirmation and stops further submissions.
  • If a job fails, users may choose to inspect the working directory to diagnose the root cause. The tool will detect the failure and pause further actions, giving the user control over whether to resubmit.
  • Detects already running or queued jobs and avoids redundant submissions—quickly skips duplicates with a user notification.

Installation

User setup

The experiment-runner is installed in the payu-dev conda environment, hence loading payu/dev would directly make experiment-runner available for use.

module use /g/data/vk83/prerelease/modules && module load payu/dev

Alternatively, create and activate a python virtual environment, then install via pip,

python3 -m venv <path/to/venv> --system-site-packages
source <path/to/venv>/bin/activate

pip install experiment-runner

Development setup

For contributors and developers, setup a development environment,

git clone https://github.com/ACCESS-NRI/access-experiment-runner.git
cd access-experiment-runner

# under a virtual environment
pip install -e .

Usage

experiment-runner -i --help

usage: experiment-runner [-h] [-i INPUT_YAML_FILE]

Manage ACCESS experiments using configurable YAML input.
If no YAML file is specified, the tool will look for 'Experiment_runner.yaml' in the current directory.
If that file is missing, you must specify one with -i / --input-yaml-file.

options:
  -h, --help            show this help message and exit
  -i INPUT_YAML_FILE, --input-yaml-file INPUT_YAML_FILE
                        Path to the YAML file specifying parameter values for experiment runs.
                        Defaults to 'Experiment_runner.yaml' if present in the current directory.

One YAML example is provided in example/Experiment_runner_example.yaml

test_path: /g/data/{PROJECT}/{USER}/prototype-0.1.0
repository_directory: 1deg_jra55_ryf
running_branches: [ctrl, perturb_1, perturb_2]
keep_uuid: True
running_branches: # List of experiment branches to run.
  - ctrl
  - perturb_1
  - perturb_2

nruns: # Number of runs for each branch; must match the order of running_branches.
  - 2
  - 0
  - 0

# Starting point for each branch. Options include:
#   cold: start from scratch (cold start).
#   control/restartXXX: start from a specific control run restart index.
#   perturb/restartXXX: start from a specific perturbation run restart index.
startfrom_restart:
  - cold
  - cold
  - cold

where,

test_path: All control and perturbation experiment repositories.

repository_directory: Local directory name for the central repository, where the running_branches are forked from.

running_branches: A list of git branches representing experiments to run.

keep_uuid: Preserve unique identifiers (UUIDs) across runs.

nruns: A list indicating how many runs to perform for each branch listed in running_branches.

startfrom_restart: Starting point for each branch.

Workflow example

1. Trigger the experiment

experiment-runner -i example/Experiment_runner_example.yaml

2. The tool then checks status:

• Completed:

... already completed " {doneruns}, hence no new runs.

• Failed:

Clean up a failed job {work_dir} and prepare it for resubmission.

• Running/Queued:

You have duplicated runs for in the same folder hence not submitting this job!