Skip to main content

Utils for interfacing to MPI libraries using mpi4py and dask

Project description

Utils for interfacing to MPI libraries using mpi4py and dask

python pypi license build & test codecov pylint black

This package provides tools for interfacing to MPI libraries based on mpi4py and dask:

  • mpi4py is a complete Python API of the MPI standard.

  • dask is a flexible library for parallel computing in Python. Particurally, we use the following sub-modules of the latter:


NOTE: lyncs_mpi requires a working MPI installation. This can be installed via apt-get:

sudo apt-get install libopenmpi-dev openmpi-bin

OR using conda:

conda install -c anaconda mpi4py

The package can be installed via pip:

pip install [--user] lyncs_mpi


In this package we implement several low-level tools for supporting classes distributed over MPI. These are described in this guide for developers. In the following we describe the high-level tools provided in this package.


The Client is a wrapper of dask.distributed.Client made MPI compatible following the instructions of dask-mpi documentation.

from lyncs_mpi import Client

client = Client(num_workers=4)

If the above script is run in a interactive shell, the Client will start an MPI server in the background running over num_workers+1 processes. The workers are the effective processes involved in the calculation. The extra process (+1) is the scheduler that will manage the task scheduling.

The client, the interactive shell in this example, will proceed processing the script: submitting tasks to the scheduler that will run them on the workers.

The same script can be run directly via mpirun. In this case one needs to execute

mpirun -n $((num_workers + 2)) python

that will run on num_workers+2 processes (as above +1 for the scheduler and +1 for the client that processes the script).


Another feature that make lyncs_mpi.Client MPI compatible is the support of MPI communicators.

comm = client.comm
comm1 = client.create_comm(num_workers=2)
comm2 = client.create_comm(exclude=comm1.workers)

In the example, comm = client.comm is the main MPI communicator involving all the workers. The second comm1 and third comm2 communicators, instead, are communicators over 2 workers each. The first two workers have been optimally chosen by the client, the other two instead are the remaining one excluding the workers of comm1.

Another kind of communicators are Cartesian MPI communicators. They can be initialized doing

cart = comm.create_cart([2,2])

where [2,2] are the dimensions of the multi-dimensional grid where the processes are distributed.

Cartesian communicators directly support Dask arrays and e.g. cart.zeros([4,4,3,2,1]) instantiates a distributed Dask array assigned to the workers of the communicator with local shape (chunks) (2,2,3,2,1).

Project details

Download files

Download the file for your platform. If you're not sure which to choose, learn more about installing packages.

Files for lyncs-mpi, version 0.1.4
Filename, size File type Python version Upload date Hashes
Filename, size lyncs_mpi-0.1.4.tar.gz (18.0 kB) File type Source Python version None Upload date Hashes View

Supported by

Pingdom Pingdom Monitoring Google Google Object Storage and Download Analytics Sentry Sentry Error logging AWS AWS Cloud computing DataDog DataDog Monitoring Fastly Fastly CDN DigiCert DigiCert EV certificate StatusPage StatusPage Status page