MPIPartition

A python module for MPI volume decomposition and particle distribution

• Free software: MIT license
• Documentation: https://argonnecpac.github.io/MPIPartition
• Repository: https://github.com/ArgonneCPAC/MPIPartition

Features

• Cartesian partitioning of a cubic volume (arbitrary dimensions) among MPI ranks
• Equal area decomposition of the spherical shell (S2) among MPI ranks
• distributing particle-data among ranks to the corresponding subvolume / surface segment
• overloading particle-data at rank boundaries ("ghost particles")

Installation

Installing from the PyPI repository:

pip install mpipartition

Installing the development version from the GIT repository

git clone https://github.com/ArgonneCPAC/mpipartition.git
cd mpipartition
python setup.py develop

Requirements

These packages will be automatically installed if they are not already present:

• Python >= 3.8
• mpi4py: MPI for Python
• numpy: Python array library
• numba: Python JIT compiler

Basic Usage

Check the documentation for an in-depth explanation / documentation.

# this code goes into mpipartition_example.py

from mpipartition import Partition, distribute, overload
import numpy as np

# create a partition of the unit cube with available MPI ranks
box_size = 1.
partition = Partition()

if partition.rank == 0:
    print(f"Number of ranks: {partition.nranks}")
    print(f"Volume decomposition: {partition.decomposition}")

# create random data
nparticles_local = 1000
data = {
    "x": np.random.uniform(0, 1, nparticles_local),
    "y": np.random.uniform(0, 1, nparticles_local),
    "z": np.random.uniform(0, 1, nparticles_local)
}

# distribute data to ranks assigned to corresponding subvolume
data = distribute(partition, box_size, data, ('x', 'y', 'z'))

# overload "edge" of each subvolume by 0.05
data = overload(partition, box_size, data, 0.05, ('x', 'y', 'z'))

This code can then be executed with mpi:

mpirun -n 10 python mpipartition_example.py

---

A more applied example, using halo catalogs from a HACC cosmological simulation (in the GenericIO data format):

from mpipartition import Partition, distribute, overload
import numpy as np
import pygio

# create a partition with available MPI ranks
box_size = 64.  # box size in Mpc/h
partition = Partition(3)  # by default, the dimension is 3

# read GenericIO data in parallel
data = pygio.read_genericio("m000p-499.haloproperties")

# distribute
data = distribute(partition, box_size, data, [f"fof_halo_center_{x}" for x in "xyz"])

# mark "owned" data with rank (allows differentiating owned and overloaded data)
data["status"] = partition.rank * np.ones(len(data["fof_halo_center_x"]), dtype=np.uint16)

# overload by 4Mpc/h
data = overload(partition, box_size, data, 4., [f"fof_halo_center_{x}" for x in "xyz"])

# now we can do analysis such as 2pt correlation functions (up to 4Mpc/h)
# or neighbor finding, etc.