# lammps-parallel-hpc

> This skill should be used when users ask about parallel and hpc in lammps; it prioritizes documentation references and then source inspection only for unresolved details.

- Author: Tao E. Li
- Repository: TEL-Research-Group/lammps
- Version: 20260207223556
- Stars: 0
- Forks: 0
- Last Updated: 2026-02-08
- Source: https://github.com/TEL-Research-Group/lammps
- Web: https://mule.run/skillshub/@@TEL-Research-Group/lammps~lammps-parallel-hpc:20260207223556

---

---
name: lammps-parallel-hpc
description: This skill should be used when users ask about parallel and hpc in lammps; it prioritizes documentation references and then source inspection only for unresolved details.
---

# lammps: Parallel and HPC

## High-Signal Playbook
### Route the request
- Use `lammps-build-and-install` for enabling accelerator packages at compile time.
- Use `lammps-simulation-workflows` once launch performance is acceptable.
- Use `lammps-troubleshooting` for MPI/GPU crashes, hangs, or nondeterministic failures.

### Triage questions
- Hardware target: CPU-only, NVIDIA/AMD GPU, or mixed cluster?
- Current MPI tasks per node and OpenMP threads per task?
- Which accelerator path is intended (`OPENMP`, `GPU`, `KOKKOS`)?
- Is MPI GPU-aware (for multi-rank CUDA KOKKOS runs)?
- Is PPPM/KSpace dominating wall time?

### Canonical workflow
- Start from a correct MPI-only baseline and capture timing breakdown (`doc/src/Run_output.rst`, `doc/src/Speed_measure.rst`).
- Test OPENMP via `-sf omp` and thread/task sweeps (`doc/src/Speed_omp.rst`).
- Test GPU package via `-sf gpu -pk gpu Ng` and tasks/GPU sweeps (`doc/src/Speed_gpu.rst`).
- Test KOKKOS via `-k on -sf kk`, then tune `-pk kokkos` options (`doc/src/Speed_kokkos.rst`).
- Pin MPI tasks and threads to hardware locality for stable performance.
- Validate physics consistency after switching precision/backend.

### Minimal working examples
- OPENMP launch (`doc/src/Speed_omp.rst`):
```bash
mpirun -np 4 lmp_omp -sf omp -pk omp 4 -in in.script
```
- KOKKOS CPU launch (`doc/src/Speed_kokkos.rst`):
```bash
mpirun -np 32 -ppn 4 lmp_kokkos_omp -k on t 4 -sf kk -in in.lj
```
- GPU package launch (`doc/src/Speed_gpu.rst`):
```bash
mpirun -np 12 lmp -sf gpu -pk gpu 2 -in in.script
```

### Pitfalls
- Oversubscribing cores (`MPI tasks * threads` too high) (`doc/src/Speed_omp.rst`, `doc/src/Speed_kokkos.rst`).
- Forgetting `-k on` when using KOKKOS-accelerated styles (`doc/src/Speed_kokkos.rst`).
- Multi-rank CUDA KOKKOS runs without GPU-aware MPI (use `gpu/aware off`) (`doc/src/Speed_kokkos.rst`).
- Assuming one MPI task per GPU is always optimal; often 2-10 tasks/GPU is better (`doc/src/Speed_gpu.rst`).
- Interpreting Pair/KSpace timing without accounting for async GPU overlap (`doc/src/Speed_gpu.rst`).

### Validation checklist
- Track `Loop time`, `katom-step/s`, and category percentages across tuning runs.
- Verify CPU utilization is near expected (`MPI x OMP`) for chosen backend.
- Check load-balance histograms (`Nlocal`, neighbors) for imbalance bottlenecks.
- Reconfirm key thermodynamic averages after backend/precision changes.

## Scope
- Handle questions about MPI/OpenMP/GPU execution, scaling, and batch systems.
- Keep responses abstract and architectural for large codebases; avoid exhaustive per-function documentation unless requested.

## Primary documentation references
- `doc/src/Run_output.rst`
- `doc/src/Speed_measure.rst`
- `doc/src/Speed_packages.rst`
- `doc/src/Speed_omp.rst`
- `doc/src/Speed_gpu.rst`
- `doc/src/Speed_kokkos.rst`
- `doc/src/Speed_intel.rst`
- `doc/src/Speed_tips.rst`
- `doc/src/package.rst`

## Workflow
- Start with the primary references above.
- If details are missing, inspect `references/doc_map.md` for additional topic documents (generated inventory plus curated anchors).
- Use tutorials/examples as executable usage patterns when available.
- Use tests as behavior or regression references when available.
- If ambiguity remains after docs, inspect `references/source_map.md` and start with the ranked source entry points.
- Cite exact documentation file paths in responses.

## Tutorials and examples
- `examples`

## Test references
- None discovered.

## Optional deeper inspection
- `fortran`
- `lib`
- `python`
- `src`

## Source entry points for unresolved issues
- `src/NETCDF/dump_netcdf_mpiio.h`
- `src/NETCDF/dump_netcdf_mpiio.cpp`
- `lib/kokkos/core/src/OpenMP/Kokkos_OpenMP_Parallel_Scan.hpp`
- `lib/kokkos/core/src/OpenMP/Kokkos_OpenMP_Parallel_Reduce.hpp`
- `lib/kokkos/core/src/OpenMP/Kokkos_OpenMP_Parallel_For.hpp`
- `src/OPENMP/pair_zbl_omp.h`
- `src/OPENMP/pair_zbl_omp.cpp`
- `src/OPENMP/pair_yukawa_omp.h`
- Prefer targeted source search (for example: `rg -n "<symbol_or_keyword>" fortran lib python src`).