Installing graph-tool from source with Anaconda

System

OS

Arch Linux 4.9.11 x86_64

Python Distribution

Anaconda Python 3 https://www.continuum.io/downloads#linux

These installation instructions are adapted from the official ones at https://git.skewed.de/count0/graph-tool/wikis/installation-instructions, to match the requirements of an Anaconda distribution.

Requirements

• GCC (latest version, 6.3.1 for me)

• Boost libraries

• Python 3.6

• The expat XML library

• Scipy

• Numpy

• CGAL C++ library

• sparsehash (optional)

• cairomm, matplotlib (optional, for plotting)

These packages must be installed via conda and via your package manager (pacman for Arch), in order to satisfy dependencies at build time. For Arch, run:

pacman -S clang boost expat cgal sparsehash cairomm

For other distributions, you should find similar packages from your standard repositories.

Note	Do not use another compiler! You have to use the same compiler as the one used to compile the Boost library on your system, and that is very probably GCC.

Installation steps

• First, if you don’t have Anaconda, install it by running the official installer. By default, it installs to /home/user/anaconda3. We will keep the assumption that the default wasn’t changed.

• Create a conda environment for graph-tool and activate it:

conda create -n graph
source activate graph

• Install the required libraries:

conda install -c conda-forge python boost scipy numpy expat cgal
conda install -c conda-forge sparsehash
conda install -c conda-forge cairomm matplotlib

Note	We use the community conda-forge channel, because packages tend to be more up-to-date compared to the official channel.

• Download graph-tool from the official website (https://graph-tool.skewed.de/), version 2.21 at the time of writing. Decompress it.

cd ~/builds
tar xjvf graph-tool-2.21.tar.bz2
cd graph-tool-2.21

• Compile

./configure --prefix=/home/dimitri/anaconda3/envs/graph/ --with-python-module-path=/home/dimitri/anaconda3/envs/graph/lib/python3.6/site-packages
make -j2
make install

If you have changed the name of the environment, adapt the paths provided to ./configure.

Note

Adapt the -j2 parameter to your system’s number of cores, but more importantly, to your available memory. Compilation uses a lot of memory, and you will run quickly into an internal compiler error if you parallelize too much. To remain conservative, use -j1 or add some swap. Compilation may be very long (approx. 1 hour on my system).

• Final touches: install jupyter to have all the usual tools from the Anaconda distribution.

conda install -c conda-forge jupyter

Testing

• Deactivate and reactivate your conda environment (so that your PATH is updated):

source deactivate
source activate graph

• Open a Jupyter console session

jupyter console

• Inside the console, run:

from graph_tool.all import *

g = Graph()
v1 = g.add_vertex()
v2 = g.add_vertex()
e = g.add_edge(v1, v2)
v1.out_degree()

Issues

• All packages come with OpenBLAS instead of MKL. This is not optimal if you have an Intel CPU, as MKL is usually way faster. I think that this is due to conda-forge, as Continuum probably have an agreement with Intel regarding the MKL license.

• You can not use graph-tool builtin plotting capabilities, as pycairo was not installed.

If you manage to solve any of this, please let me know!