PyMPDATA 1.1.3

Numba-accelerated Pythonic implementation of MPDATA with examples in Python, Julia and Matlab

PyMPDATA

PyMPDATA is a high-performance Numba-accelerated Pythonic implementation of the MPDATA algorithm of Smolarkiewicz et al. used in geophysical fluid dynamics and beyond. MPDATA numerically solves generalised transport equations - partial differential equations used to model conservation/balance laws, scalar-transport problems, convection-diffusion phenomena. As of the current version, PyMPDATA supports homogeneous transport in 1D, 2D and 3D using structured meshes, optionally generalised by employment of a Jacobian of coordinate transformation. PyMPDATA includes implementation of a set of MPDATA variants including the non-oscillatory option, infinite-gauge, divergent-flow, double-pass donor cell (DPDC) and third-order-terms options. It also features support for integration of Fickian-terms in advection-diffusion problems using the pseudo-transport velocity approach. In 2D and 3D simulations, domain-decomposition is used for multi-threaded parallelism.

PyMPDATA is engineered purely in Python targeting both performance and usability, the latter encompassing research users', developers' and maintainers' perspectives. From researcher's perspective, PyMPDATA offers hassle-free installation on multitude of platforms including Linux, OSX and Windows, and eliminates compilation stage from the perspective of the user. From developers' and maintainers' perspective, PyMPDATA offers a suite of unit tests, multi-platform continuous integration setup, seamless integration with Python development aids including debuggers and profilers.

PyMPDATA design features a custom-built multi-dimensional Arakawa-C grid layer allowing to concisely represent multi-dimensional stencil operations on both scalar and vector fields. The grid layer is built on top of NumPy's ndarrays (using "C" ordering) using the Numba's @njit functionality for high-performance array traversals. It enables one to code once for multiple dimensions, and automatically handles (and hides from the user) any halo-filling logic related with boundary conditions. Numba prange() functionality is used for implementing multi-threading (it offers analogous functionality to OpenMP parallel loop execution directives). The Numba's deviation from Python semantics rendering closure variables as compile-time constants is extensively exploited within PyMPDATA code base enabling the just-in-time compilation to benefit from information on domain extents, algorithm variant used and problem characteristics (e.g., coordinate transformation used, or lack thereof).

A separate project called PyMPDATA-MPI depicts how numba-mpi can be used to enable distributed memory parallelism in PyMPDATA.

The PyMPDATA-examples package covers a set of examples presented in the form of Jupyer notebooks offering single-click deployment in the cloud using mybinder.org or using colab.research.google.com. The examples reproduce results from several published works on MPDATA and its applications, and provide a validation of the implementation and its performance.

Dependencies and installation

To install PyMPDATA, one may use: pip install PyMPDATA (or pip install git+https://github.com/open-atmos/PyMPDATA.git to get updates beyond the latest release). PyMPDATA depends on NumPy and Numba.

Running the tests shipped with the package requires additional packages listed in the test-time-requirements.txt file (which include PyMPDATA-examples, see below).

Examples (Jupyter notebooks reproducing results from literature):

PyMPDATA examples are bundled with PyMPDATA and located in the examples subfolder. They constitute a separate PyMPDATA_examples Python package which is also available at PyPI. The examples have additional dependencies listed in PyMPDATA_examples package setup.py file. Running the examples requires the PyMPDATA_examples package to be installed. Since the examples package includes Jupyter notebooks (and their execution requires write access), the suggested install and launch steps are:

git clone https://github.com/open-atmos/PyMPDATA-examples.git
cd PyMPDATA-examples
pip install -e .
jupyter-notebook

Alternatively, one can also install the examples package from pypi.org by using pip install PyMPDATA-examples.

Package structure and API:

In short, PyMPDATA numerically solves the following equation:

$$ \partial_t (G \psi) + \nabla \cdot (Gu \psi) + \mu \Delta (G \psi) = 0 $$

where scalar field $\psi$ is referred to as the advectee, vector field u is referred to as advector, and the G factor corresponds to optional coordinate transformation. The inclusion of the Fickian diffusion term is optional and is realised through modification of the advective velocity field with MPDATA handling both the advection and diffusion (for discussion see, e.g. Smolarkiewicz and Margolin 1998, sec. 3.5, par. 4).

The key classes constituting the PyMPDATA interface are summarised below with code snippets exemplifying usage of PyMPDATA from Python, Julia and Matlab.

A pdoc-generated documentation of PyMPDATA public API is maintained at: https://open-atmos.github.io/PyMPDATA

Options class

The Options class groups both algorithm variant options as well as some implementation-related flags that need to be set at the first place. All are set at the time of instantiation using the following keyword arguments of the constructor (all having default values indicated below):

• n_iters: int = 2: number of iterations (2 means upwind + one corrective iteration)
• infinite_gauge: bool = False: flag enabling the infinite-gauge option (does not maintain sign of the advected field, thus in practice implies switching flux corrected transport on)
• divergent_flow: bool = False: flag enabling divergent-flow terms when calculating antidiffusive velocity
• nonoscillatory: bool = False: flag enabling the non-oscillatory or monotone variant (a.k.a flux-corrected transport option, FCT)
• third_order_terms: bool = False: flag enabling third-order terms
• epsilon: float = 1e-15: value added to potentially zero-valued denominators
• non_zero_mu_coeff: bool = False: flag indicating if code for handling the Fickian term is to be optimised out
• DPDC: bool = False: flag enabling double-pass donor cell option (recursive pseudovelocities)
• dimensionally_split: bool = False: flag disabling cross-dimensional terms in antidiffusive velocity
• dtype: np.floating = np.float64: floating point precision

For a discussion of the above options, see e.g., Smolarkiewicz & Margolin 1998, Jaruga, Arabas et al. 2015 and Olesik, Arabas et al. 2020 (the last with examples using PyMPDATA).

In most use cases of PyMPDATA, the first thing to do is to instantiate the Options class with arguments suiting the problem at hand, e.g.:

Julia code (click to expand)

using Pkg
Pkg.add("PyCall")
using PyCall
Options = pyimport("PyMPDATA").Options
options = Options(n_iters=2)

Matlab code (click to expand)

Options = py.importlib.import_module('PyMPDATA').Options;
options = Options(pyargs('n_iters', 2));

Python code (click to expand)

from PyMPDATA import Options
options = Options(n_iters=2)

Arakawa-C grid layer and boundary conditions

In PyMPDATA, the solution domain is assumed to extend from the first cell's boundary to the last cell's boundary (thus the first scalar field value is at $[\Delta x/2, \Delta y/2]$. The ScalarField and VectorField classes implement the Arakawa-C staggered grid logic in which:

• scalar fields are discretised onto cell centres (one value per cell),
• vector field components are discretised onto cell walls.

The schematic of the employed grid/domain layout in two dimensions is given below (with the Python code snippet generating the figure as a part of CI workflow):

Python code (click to expand)

import numpy as np
from matplotlib import pyplot

dx, dy = .2, .3
grid = (10, 5)

pyplot.scatter(*np.mgrid[
        dx / 2 : grid[0] * dx : dx, 
        dy / 2 : grid[1] * dy : dy
    ], color='red', 
    label='scalar-field values at cell centres'
)
pyplot.quiver(*np.mgrid[
        0 : (grid[0]+1) * dx : dx, 
        dy / 2 : grid[1] * dy : dy
    ], 1, 0, pivot='mid', color='green', width=.005,
    label='vector-field x-component values at cell walls'
)
pyplot.quiver(*np.mgrid[
        dx / 2 : grid[0] * dx : dx,
        0: (grid[1] + 1) * dy : dy
    ], 0, 1, pivot='mid', color='blue', width=.005,
    label='vector-field y-component values at cell walls'
)
pyplot.xticks(np.linspace(0, grid[0]*dx, grid[0]+1))
pyplot.yticks(np.linspace(0, grid[1]*dy, grid[1]+1))
pyplot.title(f'staggered grid layout (grid={grid}, dx={dx}, dy={dy})')
pyplot.xlabel('x')
pyplot.ylabel('y')
pyplot.legend(bbox_to_anchor=(.1, -.1), loc='upper left', ncol=1)
pyplot.grid()
pyplot.savefig('readme_grid.png')

The __init__ methods of ScalarField and VectorField have the following signatures:

• ScalarField(data: np.ndarray, halo: int, boundary_conditions)
• VectorField(data: Tuple[np.ndarray, ...], halo: int, boundary_conditions) The data parameters are expected to be Numpy arrays or tuples of Numpy arrays, respectively. The halo parameter is the extent of ghost-cell region that will surround the data and will be used to implement boundary conditions. Its value (in practice 1 or 2) is dependent on maximal stencil extent for the MPDATA variant used and can be easily obtained using the Options.n_halo property.

As an example, the code below shows how to instantiate a scalar and a vector field given a 2D constant-velocity problem, using a grid of 24x24 points, Courant numbers of -0.5 and -0.25 in "x" and "y" directions, respectively, with periodic boundary conditions and with an initial Gaussian signal in the scalar field (settings as in Fig. 5 in Arabas et al. 2014):

Julia code (click to expand)

ScalarField = pyimport("PyMPDATA").ScalarField
VectorField = pyimport("PyMPDATA").VectorField
Periodic = pyimport("PyMPDATA.boundary_conditions").Periodic

nx, ny = 24, 24
Cx, Cy = -.5, -.25
idx = CartesianIndices((nx, ny))
halo = options.n_halo
advectee = ScalarField(
    data=exp.(
        -(getindex.(idx, 1) .- .5 .- nx/2).^2 / (2*(nx/10)^2) 
        -(getindex.(idx, 2) .- .5 .- ny/2).^2 / (2*(ny/10)^2)
    ),  
    halo=halo, 
    boundary_conditions=(Periodic(), Periodic())
)
advector = VectorField(
    data=(fill(Cx, (nx+1, ny)), fill(Cy, (nx, ny+1))),
    halo=halo,
    boundary_conditions=(Periodic(), Periodic())    
)

Matlab code (click to expand)

ScalarField = py.importlib.import_module('PyMPDATA').ScalarField;
VectorField = py.importlib.import_module('PyMPDATA').VectorField;
Periodic = py.importlib.import_module('PyMPDATA.boundary_conditions').Periodic;

nx = int32(24);
ny = int32(24);
  
Cx = -.5;
Cy = -.25;

[xi, yi] = meshgrid(double(0:1:nx-1), double(0:1:ny-1));

halo = options.n_halo;
advectee = ScalarField(pyargs(...
    'data', py.numpy.array(exp( ...
        -(xi+.5-double(nx)/2).^2 / (2*(double(nx)/10)^2) ...
        -(yi+.5-double(ny)/2).^2 / (2*(double(ny)/10)^2) ...
    )), ... 
    'halo', halo, ...
    'boundary_conditions', py.tuple({Periodic(), Periodic()}) ...
));
advector = VectorField(pyargs(...
    'data', py.tuple({ ...
        Cx * py.numpy.ones(int32([nx+1 ny])), ... 
        Cy * py.numpy.ones(int32([nx ny+1])) ...
     }), ...
    'halo', halo, ...
    'boundary_conditions', py.tuple({Periodic(), Periodic()}) ...
));

Python code (click to expand)

from PyMPDATA import ScalarField
from PyMPDATA import VectorField
from PyMPDATA.boundary_conditions import Periodic
import numpy as np

nx, ny = 24, 24
Cx, Cy = -.5, -.25
halo = options.n_halo

xi, yi = np.indices((nx, ny), dtype=float)
advectee = ScalarField(
  data=np.exp(
    -(xi+.5-nx/2)**2 / (2*(nx/10)**2)
    -(yi+.5-ny/2)**2 / (2*(ny/10)**2)
  ),
  halo=halo,
  boundary_conditions=(Periodic(), Periodic())
)
advector = VectorField(
  data=(np.full((nx + 1, ny), Cx), np.full((nx, ny + 1), Cy)),
  halo=halo,
  boundary_conditions=(Periodic(), Periodic())
)

Note that the shapes of arrays representing components of the velocity field are different than the shape of the scalar field array due to employment of the staggered grid.

Besides the exemplified Periodic class representing periodic boundary conditions, PyMPDATA supports Extrapolated, Constant and Polar boundary conditions.

Stepper

The logic of the MPDATA iterative solver is represented in PyMPDATA by the Stepper class.

When instantiating the Stepper, the user has a choice of either supplying just the number of dimensions or specialising the stepper for a given grid:

Julia code (click to expand)

Stepper = pyimport("PyMPDATA").Stepper

stepper = Stepper(options=options, n_dims=2)

Matlab code (click to expand)

Stepper = py.importlib.import_module('PyMPDATA').Stepper;

stepper = Stepper(pyargs(...
  'options', options, ...
  'n_dims', int32(2) ...
));

Python code (click to expand)

from PyMPDATA import Stepper

stepper = Stepper(options=options, n_dims=2)

or

Julia code (click to expand)

stepper = Stepper(options=options, grid=(nx, ny))

Matlab code (click to expand)

stepper = Stepper(pyargs(...
  'options', options, ...
  'grid', py.tuple({nx, ny}) ...
));

Python code (click to expand)

stepper = Stepper(options=options, grid=(nx, ny))

In the latter case, noticeably faster execution can be expected, however the resultant stepper is less versatile as bound to the given grid size. If number of dimensions is supplied only, the integration might take longer, yet same instance of the stepper can be used for different grids.

Since creating an instance of the Stepper class involves time-consuming compilation of the algorithm code, the class is equipped with a cache logic - subsequent calls with same arguments return references to previously instantiated objects. Instances of Stepper contain no mutable data and are (thread-)safe to be reused.

The init method of Stepper has an optional non_unit_g_factor argument which is a Boolean flag enabling handling of the G factor term which can be used to represent coordinate transformations and/or variable fluid density.

Optionally, the number of threads to use for domain decomposition in the first (non-contiguous) dimension during 2D and 3D calculations may be specified using the optional n_threads argument with a default value of numba.get_num_threads(). The multi-threaded logic of PyMPDATA depends thus on settings of numba, namely on the selected threading layer (either via NUMBA_THREADING_LAYER env var or via numba.config.THREADING_LAYER) and the selected size of the thread pool (NUMBA_NUM_THREADS env var or numba.config.NUMBA_NUM_THREADS).

Solver

Instances of the Solver class are used to control the integration and access solution data. During instantiation, additional memory required by the solver is allocated according to the options provided.

The only method of the Solver class besides the init is advance(n_steps, mu_coeff, ...) which advances the solution by n_steps timesteps, optionally taking into account a given diffusion coefficient mu_coeff.

Solution state is accessible through the Solver.advectee property. Multiple solver[s] can share a single stepper, e.g., as exemplified in the shallow-water system solution in the examples package.

Continuing with the above code snippets, instantiating a solver and making 75 integration steps looks as follows:

Julia code (click to expand)

Solver = pyimport("PyMPDATA").Solver
solver = Solver(stepper=stepper, advectee=advectee, advector=advector)
solver.advance(n_steps=75)
state = solver.advectee.get()

Matlab code (click to expand)

Solver = py.importlib.import_module('PyMPDATA').Solver;
solver = Solver(pyargs('stepper', stepper, 'advectee', advectee, 'advector', advector));
solver.advance(pyargs('n_steps', 75));
state = solver.advectee.get();

Python code (click to expand)

from PyMPDATA import Solver

solver = Solver(stepper=stepper, advectee=advectee, advector=advector)
state_0 = solver.advectee.get().copy()
solver.advance(n_steps=75)
state = solver.advectee.get()

Now let's plot the results using matplotlib roughly as in Fig. 5 in Arabas et al. 2014:

Python code (click to expand)

def plot(psi, zlim, norm=None):
    xi, yi = np.indices(psi.shape)
    fig, ax = pyplot.subplots(subplot_kw={"projection": "3d"})
    pyplot.gca().plot_wireframe(
        xi+.5, yi+.5, 
        psi, color='red', linewidth=.5
    )
    ax.set_zlim(zlim)
    for axis in (ax.xaxis, ax.yaxis, ax.zaxis):
        axis.pane.fill = False
        axis.pane.set_edgecolor('black')
        axis.pane.set_alpha(1)
    ax.grid(False)
    ax.set_zticks([])
    ax.set_xlabel('x/dx')
    ax.set_ylabel('y/dy')
    ax.set_proj_type('ortho') 
    cnt = ax.contourf(xi+.5, yi+.5, psi, zdir='z', offset=-1, norm=norm)
    cbar = pyplot.colorbar(cnt, pad=.1, aspect=10, fraction=.04)
    return cbar.norm

zlim = (-1, 1)
norm = plot(state_0, zlim)
pyplot.savefig('readme_gauss_0.png')
plot(state, zlim, norm)
pyplot.savefig('readme_gauss.png')

Debugging

PyMPDATA relies heavily on Numba to provide high-performance number crunching operations. Arguably, one of the key advantage of embracing Numba is that it can be easily switched off. This brings multiple-order-of-magnitude drop in performance, yet it also make the entire code of the library susceptible to interactive debugging, one way of enabling it is by setting the following environment variable before importing PyMPDATA:

Julia code (click to expand)

ENV["NUMBA_DISABLE_JIT"] = "1"

Matlab code (click to expand)

setenv('NUMBA_DISABLE_JIT', '1');

Python code (click to expand)

import os
os.environ["NUMBA_DISABLE_JIT"] = "1"

Contributing, reporting issues, seeking support

Submitting new code to the project, please preferably use GitHub pull requests (or the PyMPDATA-examples PR site if working on examples) - it helps to keep record of code authorship, track and archive the code review workflow and allows to benefit from the continuous integration setup which automates execution of tests with the newly added code.

As of now, the copyright to the entire PyMPDATA codebase is with the Jagiellonian University, and code contributions are assumed to imply transfer of copyright. Should there be a need to make an exception, please indicate it when creating a pull request or contributing code in any other way. In any case, the license of the contributed code must be compatible with GPL v3.

Developing the code, we follow The Way of Python and the KISS principle. The codebase has greatly benefited from PyCharm code inspections and Pylint code analysis (Pylint checks are part of the CI workflows).

Issues regarding any incorrect, unintuitive or undocumented bahaviour of PyMPDATA are best to be reported on the GitHub issue tracker. Feature requests are recorded in the "Ideas..." PyMPDATA wiki page.

We encourage to use the GitHub Discussions feature (rather than the issue tracker) for seeking support in understanding, using and extending PyMPDATA code.

Please use the PyMPDATA issue-tracking and dicsussion infrastructure for PyMPDATA-examples as well. We look forward to your contributions and feedback.

Credits:

Development of PyMPDATA was supported by the EU through a grant of the Foundation for Polish Science (POIR.04.04.00-00-5E1C/18).

copyright: Jagiellonian University
licence: GPL v3

Other open-source MPDATA implementations:

• mpdat_2d in babyEULAG (FORTRAN) https://github.com/igfuw/bE_SDs/blob/master/babyEULAG.SDs.for#L741
• mpdata-oop (C++, Fortran, Python) https://github.com/igfuw/mpdata-oop
• apc-llc/mpdata (C++) https://github.com/apc-llc/mpdata
• libmpdata++ (C++): https://github.com/igfuw/libmpdataxx
• AtmosFOAM: https://github.com/AtmosFOAM/AtmosFOAM/tree/947b192f69d973ea4a7cfab077eb5c6c6fa8b0cf/applications/solvers/advection/MPDATAadvectionFoam

Other Python packages for solving hyperbolic transport equations

• PyPDE: https://pypi.org/project/PyPDE/
• FiPy: https://pypi.org/project/FiPy/
• ader: https://pypi.org/project/ader/
• centpy: https://pypi.org/project/centpy/
• mattflow: https://pypi.org/project/mattflow/
• FastFD: https://pypi.org/project/FastFD/
• Pyclaw: https://www.clawpack.org/pyclaw/