High-Performance Computing for Weather and Climate (HPC4WC)

High Performance Computing for Weather and Climate is a hands-on course on performance-aware scientific programming. The material uses progressively more complex versions of stencil computations to introduce hardware concepts, parallel programming models, and performance-portable abstractions used in weather and climate applications.

The course material is organized by day:

• Day 1: single-core performance, stencil programs, performance metrics, memory hierarchy, memory bandwidth, peak floating-point performance, arithmetic intensity, roofline model, array storage in memory, and data-locality optimizations such as blocking, fusion, and inlining.
• Day 2: shared-memory parallelism with OpenMP, speedup, Amdahl's law, parallelization, synchronization, and variable scoping.
• Day 3: distributed-memory parallelism with MPI, message passing, point-to-point communication, deadlock, non-blocking communication, gather/scatter operations, domain decomposition, halo points, and halo updates.
• Day 4: GPU programming, hybrid node architecture, high-level GPU programming with CuPy, data management, synchronization, vectorization, platform-agnostic code, and lower-level GPU programming.
• Day 5: high-level programming models, domain-specific languages, GT4Py, performance portability, and abstractions for stencil computations.

Running Checks

Some course exercises require specific HPC hardware, compilers, MPI, GPUs, or a CSCS/Alps environment. The repository also contains lightweight checks that can run locally and in GitHub Actions without those systems.

Run the material check for all course days with:

python tools/check_material.py

This validates notebooks and course files, checks Python and shell syntax, compiles selected Fortran/C++ material when suitable compilers are available, verifies local notebook image references, and runs small smoke tests for selected scripts. The same check runs in GitHub Actions. It is intentionally lightweight; full performance runs and GPU/cluster-specific exercises should still be validated in the appropriate course environment.

Pre-commit Checks

The repository provides conservative pre-commit checks for course preparation:

python -m pip install pre-commit
pre-commit install

The hooks block generated solution bundles from being committed, verify that generated student material is current for days that use generated student material, and check formatting for Python tooling in tools/. They intentionally do not auto-format notebooks or Fortran course material.