HPC and cluster operation

This section covers production-style runs on MPI clusters. Complete a small single-rank run before moving to scheduler, filesystem, or GPU-specific setup.

Start with the HPC operator guide.

Generic cluster guidance

Topic	Document
CPU or GPU path selection	GPU path decision
Separate CPU, CUDA, and HIP build trees	CPU and GPU build trees
MPI-IO paths and collective layout	MPI-IO layout checklist
Runtime instrumentation	Performance profiling
Profiling output contract	Profiling export schema
First batch submission	Slurm day one

Site-specific guidance

Site	Document
LUMI software and ROCm environment	LUMI installation
LUMI scheduler examples	LUMI Slurm guide
VTT Tohtori modules and paths	Tohtori installation

Site pages may name current modules and filesystem paths. Generic API, CMake, and configuration contracts belong in the main install guide and reference pages rather than being copied into site notes.

Before a production run

Verify that:

• OpenPFC and HeFFTe use the same compiler and MPI stack;

• the launcher visible inside the job matches the linked MPI implementation;

• CPU and device-enabled binaries come from separate build trees;

• output paths are visible and appropriate on compute nodes;

• the selected FFT backend matches the OpenPFC and HeFFTe build;

• a small representative run succeeds before scaling to the target allocation.

Use Troubleshooting for build and runtime failures and Operator playbooks for symptom-oriented checks.