nengo_mpi¶

nengo_mpi is a C++/MPI backend for nengo, a python library for building and simulating biologically realistic neural networks. nengo_mpi makes it possible to run nengo simulations in parallel on thousands of processors, and existing nengo scripts can be adapted to make use of nengo_mpi with minimal effort.

With an MPI implementation installed on the system, nengo_mpi can be used to run neural simulations in parallel using just a few lines of code:

import nengo
import nengo_mpi
import numpy as np
import matplotlib.pyplot as plt

with nengo.Network() as net:
    sin_input = nengo.Node(output=np.sin)

    # A population of 100 neurons representing a sine wave
    sin_ens = nengo.Ensemble(n_neurons=100, dimensions=1)
    nengo.Connection(sin_input, sin_ens)

    # A population of 100 neurons representing the square of the sine wave
    sin_squared = nengo.Ensemble(n_neurons=100, dimensions=1)
    nengo.Connection(sin_ens, sin_squared, function=np.square)

    # View the decoded output of sin_squared
    squared_probe = nengo.Probe(sin_squared, synapse=0.01)

partitioner = nengo_mpi.Partitioner(2)
sim = nengo_mpi.Simulator(net, partitioner=partitioner)
sim.run(5.0)

plt.plot(sim.trange(), sim.data[squared_probe])
plt.show()

There are 3 differences between this script and a script using the reference implementation of nengo. First, we need to import nengo_mpi. Then we need to create a Partitioner object, which specifies how many components the nengo network should be split up into (this corresponds to the maximum number of distinct processors that can be used to run the simulation). Finally, when creating the simulator we need to use nengo_mpi’s Simulator class, and we need to pass in the Partitioner instance.

The script can then be run in parallel using:

mpirun -np 2 python -m nengo_mpi <script_name>

sin_ens and sin_squared will be simulated on separate processors, with the output of sin_ens being passed to the input of sin_squared every time-step using MPI. After the simulation has completed, the probed results from all processors are passed back to the main processor, so all probed data can be accessed in the usual way (e.g. sim.data[squared_probe]).

nengo_mpi is fully featured, supporting all aspects of Nengo Release 2.0.2.