pnumpy 1.4.1

A very lightweight implementation of distributed arrays

pnumpy

Parallel computing in N dimensions made easy in Python

Overview

pnumpy is a very lightweight implementation of distributed arrays, which runs on architectures ranging from multi-core laptops to large MPI clusters. pnumpy is based on numpy and mpi4py and supports arrays in any number of dimensions. Processes can access remote data using a "getData" method. This can be used to access neighbor ghost data but is more flexible as it allows to access data from any process--not necessarily a neighboring one. pnumpy is designed to work seamlessly with numpy's slicing operators ufunc, etc., making it easy to transition your code to a parallel computing environment.

Speedup of 512^3 Laplacian on a 4 core desktop

How to get pnumpy

git clone https://github.com/pletzer/pnumpy.git

How to build pnumpy

pnumpy requires:

• python 2.7 or 3.5 or later
• numpy, e.g. 1.10
• MPI library (e.g. MPICH2 1.4.1p1 that comes with Anaconda)
• mpi4py, e.g. 3.0.3s

python setup.py install

or, if you need root access,

sudo python setup.py install

Alternatively you can use

pip install pnumpy

or

pip install pnumpy --user

How to test pnumpy

Run any file under tests/, e.g.

cd tests
mpiexec -n 4 python testDistArray.py

How to use pnumpy

To run script myScript.py in parallel use

mpiexec -n numProcs python <myScript.py>

where numProcs is the number of processes.

A lightweight extension to numpy arrays

Think of pnumpy arrays as standard numpy arrays with additional data members and methods to access neighboring data.

To create a ghosted distributed array (gda) use:

from pnumpy import gdaZeros
da = gdaZeros((4, 5), numpy.float32, numGhosts=1)

The above creates a 4 x 5 float32 array filled with zeros -- the syntax should be familiar to anyone using numpy arrays.

All numpy operations apply to pnumpy distributed arrays with no change and this includes slicing. Note that slicing operations are with respect to local array indices.

In the above, numGhosts describes the thickness of the halo region, i.e. the slice of data inside the array that can be accessed by other processes. A value of numGhosts = 1 means the halo has depth of one. The thicker the halo the more costly communication will be because more data will need to be copied from one process to another.

For a 2D array, the halo can be broken into four regions:

• da[:numGhosts, :] => west
• da[-numGhosts:, :] => east
• da[:, :numGhosts] => south
• da[:, -numGhosts:] => north

(In n-dimensions there are 2n regions.) pnumpy identifies each halo region with a tuple:

• (-1, 0) => west
• (1, 0) => east
• (0, -1) => south
• (0, 1) => north

To access data on the south region of remote process otherRk, use

southData = da.getData(otherRk, winID=(0, -1))

Using a regular domain decomposition

The above will work for any domain decomposition, not necessarily a regular one. In the case where a global array is split into uniform chunks of data, otherRk can be inferred from the local rank and an offset vector:

from pnumpy import CubeDecomp
decomp = CubeDecomp(numProcs, dims)
...
otherRk = decomp.getNeighborProc(self, da.getMPIRank(), offset=(0, 1), periodic=(True, False))

where numProcs is the number of processes, dims are the global array dimensions and periodic is a tuple of True/False values indicating whether the domain is periodic or not. In the case where there is no neighbour rank (because the local da.getMPIRank() rank lies at the boundary of the domain), then getNeighborProc may return None. In this case getData will also return None.

A very high level

For the Laplacian stencil, one may consider using

from pnumpy import Laplacian
lapl = Laplacian(decomp, periodic=(True, False))

Applying the Laplacian stencil to any numpy-like array inp then simply involves:

out = lapl.apply(inp)