Modules

nengo_mpi is composed of several fairly separable modules.

python

The python code consists primarily of alternate implementations of both nengo.Simulator and nengo.Model. nengo_mpi.Simulator is a wrapper around the C++/MPI code which provides an interface nearly identical to nengo.Simulator (see Getting Started). The nengo_mpi.Model class is primarily responsible for adapting the output of the reference implementation’s build step (converting from a high-level model specificationto a concrete computation graph; see How It Works) to work with nengo_mpi.Simulator.

The final major chunk of python code handles the complex task of partitioning a nengo Network into a desired number of components that can be simulated independently.

C++

The directory mpi_sim contains the C++ code. This code implements a back-end for nengo which can use MPI to run simulations in parallel. The C++ code only implements simulation capabilities; the build step is still done in python, and nengo_mpi largely uses the builder provided by the reference implementation. The C++ code can be used in at least three different ways.

mpi_sim.so

A shared library that allows the python layer to access the core C++ simulator. The python code creates an HDF5 file encoding the built network (the operators and signals that need to be simulated), and then makes a call out to mpi_sim.so with the name of the file. mpi_sim.so then opens the file (in parallel if there are multiple processors active) and runs the simulation.

nengo_mpi executable

This is an executable that allows the C++ simulator to be used directly, instead of having to go through python. The executable accepts as arguments the name of a file specifying the operators and signals in a network, as well as the length of time to run the simulation for, in seconds. Removing the requirement that the C++ code be accessed through python has a number of advantages. In particular, it can make attaching a debugger much easier. Also, some high-performance clusters (e.g. BlueGene) provide only minimal support for python. The nengo_mpi exectuable has no python dependencies, and so it can be used on these machines. A typical workflow is to use the python code to create the HDF5 file on a machine with full python support, and then transfer that file over to the high-performance cluster where the network encoded by that file can be simulated using the nengo_mpi executable. See Workflows for further details.

nengo_cpp executable

This is just a version of the nengo_mpi executable which does not have MPI dependencies [1] (which, of course, means that there is no parallelization). It is possible that some users may find this useful in some situations where the MPI dependency cannot be met, as the C++ simulator is often significantly faster then the reference implementation simulator even without parallelization (see our Benchmarks).

[1]	This is currently a lie, but it will be true soon.