muphys-cpp

Description

C++ prototypes of muphys project using heterogeneous libraries and C extensions.

Documentation

Online documentation is generated automatically using doxygen .

Installation

Dependencies

• NetCDF for CXX
  • for Levante: spack load [email protected]

Other dependency like googletest is built in-tree from github archives.

Available compile options

• Implementation - The sequential implementation is selected by default. The user can choose of the following options:
  • MU_IMPL=seq - C++ serial implementation
• Precision (default is double)
• MU_ENABLE_SINGLE - to switch to float
• Unit-test - compile tests together with the main executable (default is true)
  • MU_ENABLE_TESTS

Compile the project (with default flags and Seq frontend)

cmake -DMU_IMPL=seq -B build -S .

cmake --build build

Usage

./build/bin/graupel input_file.nc

Automated tests

• Run tests manually: cd build && ctest

License

muphys-cpp is available under a BSD 3-clause license. See LICENSES/ for license information and AUTHORS.TXT for a list of authors.

Coding Challenge

• Task: Creat an optimized parallel implementation of muphys

  • Requirements:
    • 4 Versions: Correctness/Performance x CPU/GPU
    • Directive Programming: Use OpenMP (Offload to GPU) !No CUDA Kernels
  • Parallelisation Task:
    • Parallelize the sequential implementation
    • Can rewrite the algorithm entirely
  • Optimization Task:
    • Look reordering/collapse/rewrte
    • Minimize critical regions
    • Data structure to optimize cache utilization
    • Optimal CPU-GPU memory transfer

• Layout:

  • code:
    • /core: physics interactions
    • /io: read/write/ I/O files
    • /implementations/sequential
  • external dependencies:
    • /extern
  • validation:
    • /test: automated unit tests
    • /tasks: input files
    • /reference_results: output fiels
    • /scripts: Levante batch files

• Validation:

  • CPU-CPU: cdo diffv bit-identical
  • CPU-GPU: cdo sub within floating point tolerance

• Submission:

  • Path to the implementation
  • Script to build the 4 versions
  • Slurm logs to confirm the results
  • Summary list of optimizations performed
  • Plots to confirm the performance results
  • (Optional) Profiler analysis output & interpretation
  • (Optional) Experience report for using OpenMP

• Grading:

  • Correctness(50%): results are correct on both CPU and GPU
  • Performance(50%): the code is faster than the sequential version
    • How to determine?
    • Will the teams be ranked?
      • No, but the fastest may get a bonus
  • Bonus(20%): readability, portability, extreme performance, optional documentation ...