Cluster jobs#

examples/das6_generate_small_gpu.sh is a Slurm job for the DAS-6 cluster that generates randomised Sionna radio maps in an empty scene and in central Amsterdam on one A6000 GPU, 500 per scene unless OWRT_RADIO_MAPS_PER_SCENE says otherwise. Prepare the owrt environment once with the rt extra, then submit from the repository root:

./scripts/setup.sh
sbatch examples/das6_generate_small_gpu.sh

The job forces device="cuda", verifies the allocated GPU before running, and enables deterministic resume so an interrupted submission can be repeated unchanged; the repeated run also rebuilds index.parquet, which a killed job leaves stale (Resume and extend). Its defaults use the schema’s highest supported detail: a 1 m receiver grid, reflection depth 8, 100,000,000 samples per transmitter, 0.5 m voxel pitch and 1 m terrain spacing. Override run-scale and output settings at submission time:

OWRT_RESOLUTION_M=5 OWRT_MAX_DEPTH=5 OWRT_SAMPLES_PER_TX=5000000 \
  OWRT_DATASET_NAME=das6-amsterdam-empty-medium \
  sbatch examples/das6_generate_small_gpu.sh

Submit from another directory by exporting OWRT_PROJECT_ROOT as the absolute repository path. The scheduler directives follow the DAS-6 GPU job guidance; change the typed GPU request for another site, where sinfo -o '%P|%N|%f|%G' reports the available feature and GRES names.

The job wraps examples/generate_dataset.py, so the same OWRT_EXAMPLE_* variables apply on any machine without Slurm:

OWRT_EXAMPLE_ENGINE=sionna OWRT_EXAMPLE_RADIO_MAPS_PER_SCENE=3 \
OWRT_EXAMPLE_RESOLUTION_M=5 OWRT_EXAMPLE_PREVIEWS=true \
  python examples/generate_dataset.py