Concepts overview

To learn how to use Arbor, it is helpful to understand some of its concepts. Arbor’s design aims to enable scalability through abstraction. To achieve this, Arbor makes a distinction between the description of a model, and the execution of a model: a recipe describes a model, and a simulation is an executable instantiation of a model.

To be able to simulate a model, three basic steps need to be considered:

  1. Describe the model by defining a recipe;

  2. Define the computational resources available to execute the model;

  3. Initiate and execute a simulation of the recipe on the chosen hardware resources.

The Python front-end further abstracts away some of these steps for single cell models, where users only need to describe the cell and simulation; and the details of the recipe and computational resources construction are handled under the hood. Generally speaking though, these 3 steps are the building blocks of an Arbor application.

Recipes represent a set of neuron constructions and connections with mechanisms specifying ion channel and synapse dynamics in a cell-oriented manner. This has the advantage that cell data can be initiated in parallel.

A cell represents the smallest unit of computation and forms the smallest unit of work distributed across processes. Arbor has built-in support for different cell types, which can be extended by adding new cell types to the C++ cell group interface.

Simulations manage the instantiation of the model and the scheduling of spike exchange as well as the integration for each cell group. A cell group represents a collection of cells of the same type computed together on the GPU or CPU. The partitioning into cell groups is provided by Domain decomposition which describes the distribution of the model over the locally available computational resources.

In order to visualize the result of detected spikes a spike recorder can be used, and to analyse Arbor’s performance a meter manager is available.