Arbor’s Python API will be the most convenient interface for most users.
Arbor requires Python version 3.7 and later. It is advised that you update
pip as well.
We strongly encourage using
pip to install Arbor.
To get help in case of problems installing with pip, run pip with the
--verbose flag, and attach the output
(along with the pip command itself) to a ticket on Arbor’s issues page.
Every point release of Arbor is pushed to the Python Package Index. For x86-64 Linux and MacOS platforms, we provide binary wheels. The easiest way to get Arbor is with pip:
# Recommended but optional: install Arbor in a virtual environment
python -m venv arbor_env
# Download and install Arbor
pip install arbor
To test that Arbor is available, try the following in a Python interpreter to see information about the version and enabled features:
>>> import arbor
For builds from Arbor’s source, you will need to have some development packages installed. Installing Arbor
for any other platforms than listed above,
pip will attempt a build from source and thus require these
packages as well.
If you wish to get the latest Arbor straight from the master branch in our git repository, you can run:
pip install git+https://github.com/arbor-sim/arbor.git
If you want to work on Arbor’s code, you can get a copy of our repo and point pip at the local directory:
# get your copy of the Arbor source
git clone https://github.com/arbor-sim/arbor.git --recursive
# make your changes and then instruct pip to build and install the local source
pip install ./arbor/
Every time you make changes to the code, you’ll have to repeat the second step.
Arbor comes with a few compilation options, some of them related to advanced forms of parallelism and other features.
The options and flags are the same as documented for the CMAKE build, but they are passed differently.
To enable more, they must be placed in the
CMAKE_ARGS environment variable.
The simplest way to do this is by prepending the
pip command with
where you place the arguments separated by space inside the quotes.
The following flags can be used to configure the installation:
ARB_WITH_MPI=<ON|OFF>: Enable MPI support, requires MPI library. Default
OFF. If you intend to use
mpi4py, you need to install the package before building Arbor, as binding it requires access to its headers.
ARB_GPU=<none|cuda|cuda-clang|hip>: Enable GPU support for NVIDIA GPUs with nvcc using
cuda, or with clang using
cuda-clang(both require cudaruntime). Enable GPU support for AMD GPUs with hipcc using
hip. By default set to
none, which disables GPU support.
ARB_VECTORIZE=<ON|OFF>: Enable vectorization. The architecture argument, documented below, may also have to be set appropriately to generated vectorized code. See Architecture for details.
ARB_ARCH=<native|*>: CPU micro-architecture to target. The advised default is
native. See the GNU GCC documentation for a full list of options.
There are more, advanced flags that can be set. We are using
CMake under the hood, so all flags and options valid in
be used in this fashion.
Detailed instructions on how to install using CMake are in the Python configuration section of the installation guide. CMake is recommended if you need more control over compilation and installation, plan to use Arbor with C++, or if you are integrating with package managers such as Spack and EasyBuild.
In the examples below we assume you are installing from a local copy.
Vanilla install with no additional features enabled:
pip install ./arbor
With MPI support. This might require loading an MPI module or setting the
CMAKE_ARGS="-DARB_WITH_MPI=ON" pip install ./arbor
CMAKE_ARGS="-DARB_VECTORIZE=ON -DARB_ARCH=skylake" pip install ./arbor
Enable NVIDIA GPUs (compiled with nvcc). This requires the CUDA toolkit:
CMAKE_ARGS="-DARB_GPU=cuda" pip install ./arbor
Enable NVIDIA GPUs (compiled with clang). This also requires the CUDA toolkit:
CMAKE_ARGS="-DARB_GPU=cuda-clang" pip install ./arbor
Enable AMD GPUs (compiled with hipcc). This requires setting the
CC=clang CXX=hipcc CMAKE_ARGS="-DARB_GPU=hip" pip install ./arbor
Note on performance#
The Python interface can incur significant memory and runtime overheads relative to C++ during the model building phase, however simulation performance is the same for both interfaces.