Checklist: MPI layout, paths, and binary I/O

Use this before debugging “missing file” or hang issues on clusters.

MPI and launcher

• Same MPI at build and run: mpirun / srun comes from the same prefix as mpicc used for OpenPFC and HeFFTe (INSTALL.md, troubleshooting.md).

• Rank count matches what you expect for decomposition tests (mpirun -n vs #SBATCH tasks).

Working directory and config paths

• Job script **cd**s to the directory where relative paths in JSON are valid.

• argv[1] config path is correct from that directory (or use an absolute path).

• fields[].data directories exist or are creatable; rank 0 creates parents in the default wiring (binary_field_io_spec.md).

Binary writers (BinaryWriter)

• saveat > 0 and fields present, or you intentionally rely on custom writers only (io_results.md).

• Every rank in the writer’s communicator participates in each collective write() — skipping ranks causes deadlock (binary_field_io_spec.md).

• Filename template % placeholders match the simulator increment you expect.

GPU / device MPI (if applicable)

• GPU-aware MPI env vars set per site (e.g. LUMI: INSTALL.LUMI.md).

• plan_options.backend matches how the binary was built (build_cpu_gpu.md, tutorials/gpu_app_quickstart.md).

See also

• tutorials/hpc_slurm_day_one.md — minimal Slurm script

• spectral_app_config_reference.md — JSON keys

• learning_paths.md — run track