python-hpc-project

Wall Heating mini-project for 02613.

Dataset: /dtu/projects/02613_2025/data/modified_swiss_dwellings

Structure

• wall_heating/ — library code (data loading, solvers, metrics, parallel scheduling)
• scripts/ — runnable scripts for each task
• jobs/ — LSF job scripts for the HPC cluster
• results/ and figures/ — default output folders

Quick start

Run from the repo root.

source /dtu/projects/02613_2025/conda/conda_init.sh
conda activate 02613_2026
make run-ref N=20
make inspect N=4
make visualize N=4 MAX_ITER=2000

Run a solver

python scripts/run_solver.py 40 --solver reference --workers 1
python scripts/run_solver.py 40 --solver reference --workers 8 --schedule static
python scripts/run_solver.py 40 --solver reference --workers 8 --schedule dynamic
python scripts/run_solver.py 40 --solver numba-cpu --workers 1
python scripts/run_solver.py 40 --solver numba-cuda --workers 1 --max-iter 20000
python scripts/run_solver.py 40 --solver cupy --workers 1

Solver options:

• reference — baseline from the assignment
• numpy — optimized NumPy with ping-pong buffers
• numba-cpu — CPU JIT (requires numba)
• numba-cuda — custom CUDA kernel, fixed iterations (requires numba + CUDA)
• cupy — GPU vectorized (requires cupy)

Task scripts

• Profile Jacobi: kernprof -l -v scripts/profile_jacobi.py
• Speedup + Amdahl: python scripts/benchmark_speedup.py --schedule static / --schedule dynamic
• CuPy nsys profile: make submit-cupy-nsys N=20
• Final analysis: python scripts/analyze_results.py results/all_buildings_stats.csv

LSF jobs

Queues used: hpc (CPU) and c02613 (GPU).

make submit-ref N=20
make submit-static N=80
make submit-dynamic N=80
make submit-numba-cpu N=40
make submit-numba-cuda N=40 MAX_ITER=20000
make submit-cupy N=40
make submit-cupy-nsys N=20
make submit-full

Job files in jobs/:

• lsf_reference_hpc.lsf
• lsf_speedup_static_hpc.lsf
• lsf_speedup_dynamic_hpc.lsf
• lsf_numba_cpu_hpc.lsf
• lsf_numba_cuda.lsf
• lsf_cupy_gpu.lsf
• lsf_cupy_nsys.lsf
• lsf_full_run_hpc.lsf

Notes

• The course conda environment 02613_2026 is shared and pre-provisioned; this repo does not install packages into it.
• GPU tasks require CUDA-enabled packages already present in 02613_2026.