Running an experiment

Once CRAB is installed, your benchmarks are set up, and you have a preset for your cluster, you can run experiments two ways: the CLI (for automation and scripting) or the TUI (for interactive configuration).

Both submit a Slurm job — make sure you're on a system where sbatch is available and you have an allocation to draw from.

CLI

crab run <config.json> -p <preset>

Argument	Meaning
<config.json>	Path to the experiment config.
-p, --preset	Preset name from config/presets.json. Optional — see preset resolution.
--log-level	DEBUG / INFO / WARNING / ERROR / CRITICAL (default INFO).

Example:

crab run examples/leonardo/mega_test.json -p leonardo --log-level DEBUG

What happens: the orchestrator prepares a timestamped output directory, writes the resolved config.json and environment.json into it, generates a crab_job.sh batch script, and submits it with sbatch. The job itself runs the worker, which executes your experiments on the allocated nodes. See Architecture for the full flow.

Monitoring a run: because the work happens inside a Slurm job, track it with the usual tools — squeue for queue status, and the job's slurm_output.log / slurm_error.log plus the CRAB logs inside the run's output directory (see Reading results).

TUI

crab tui

The Textual interface organizes a run into four tabs:

Tab	Purpose
Applications	Add and configure the applications to run (wrapper path, args, collect, start/end).
Options	Global options (nodes, ppn, allocation, convergence, …).
Environment	Select the preset / environment settings.
Log	Live output once a run is launched.

Key bindings:

Key	Action
s	Save the current configuration to a JSON file
l	Load a configuration from a JSON file
space	Launch the benchmark
q	Quit

The TUI requires the optional textual dependencies; if they're missing, crab tui offers to install them on first launch (see Installation).

Configurations you build in the TUI are saved as ordinary config files — interchangeable with hand-written ones and runnable with crab run.