miniWeather

This package simulates weather-like flows and can be used for training in parallelizing and porting numerical Julia code to accelerated HPC architectures. It is derived from and inspired by this miniWeather mini app developed by Matt Norman, which provides example serial code in C, C++ and Fortran and solutions for how to parallelize it with MPI, OpenACC-Offloading, OpenMP-Threading and OpenMP-Offloading.

Installation

The miniWeather package can be installed from the Julia REPL by:

julia> ] 
pkg> add "https://github.com/wikfeldt/miniWeather.jl"

To download the package for further development it is handier to clone the repository:

git clone https://github.com/wikfeldt/miniWeather.jl

and work on the source code under miniWeather.jl/src. Your developmental version can then be imported by

julia> ]
pkg> add /your/path/to/miniWeather.jl

Usage

An example for how to use the package is found in this example.

First load the package and set simulation parameters:

using miniWeather

# number of cells in x-direction
nx_glob = 100
# number of cells in x-direction
nz_glob = 50
# total simulation time in seconds
sim_time = 1000.0
# output frequency in seconds
output_freq = 10.0
# what type of phenomenon to simulate 
# (options: collision, thermal, mountain_waves, turbulence, density_current, injection)
weather_type = "thermal"

Then initialize the model and calculate an initial reduction (summation) of total mass and kinetic energy (which should be conserved):

model, grid = init(nx_glob, nz_glob, sim_time, weather_type);
mass0, te0 = reductions(model, grid)

Run the simulation with

# simulation results will be saved in netCDF file which can be visualized
run!(model, grid, output_freq, "output.nc")

Finally, compute another reduction and make sure that mass and kinetic energy are relatively well conserved (see below for rigorous integration testing):

mass, te = reductions(model, grid)
println("Δ_mass: ", (mass - mass0) / mass0)
println("Δ_te:   ", (te - te0) / te0)

Testing

An end-to-end test can be run by

julia --project=. test/runtests.jl

This test runs a simulation using

nx = 100
nz = 50
sim_time = 400

and asserts that the change in mass and kinetic energy are smaller than 1.0e-9 and 4.5e-5, respectively.

Parallelization

WRITEME

License and credit

This project is inspired by and derived from this miniWeather mini app developed by Matt Norman in C, C++ and Fortran. miniWeather.jl is licensed under the BSD 2-Clause "Simplified" License.