NetworKit is an open-source tool suite for high-performance network analysis. Its aim is to provide tools for the analysis of large networks in the size range from thousands to billions of edges. For this purpose, it implements efficient graph algorithms, many of them parallel to utilize multicore architectures. These are meant to compute standard measures of network analysis. NetworKit is focused on scalability and comprehensiveness. NetworKit is also a testbed for algorithm engineering and contains novel algorithms from recently published research (see list of publications below).

NetworKit is a Python module. High-performance algorithms are written in C++ and exposed to Python via the Cython toolchain. Python in turn gives us the ability to work interactively and a rich environment of tools for data analysis and scientific computing. Furthermore, NetworKit's core can be built and used as a native library if needed.

In addition to this Readme, the NetworKit_UserGuide provides an introduction to the NetworKit tools, in the form of an interactive IPython Notebook. The DevGuide is meant for developers who would like to contribute. When using NetworKit as a Python module, refer to the docstrings of classes, methods and functions.

C++ sources are also documented in Doxygen format, while the documentation for the Python sources can be generated with Sphinx. If you have both utilities installed, the documentation can be easily generated by calling the script make_docs.sh in Doc/docs.

To convert the documentation markdown files to PDF install the pandoc utility and call the script docs2pdf.sh.

For questions regarding NetworKit, subscribe to our e-mail list (networkit@ira.uka.de) and feel free to ask.

A C++ compiler supporting C++11 (we use GCC 4.8). The compiler and linker flags -fopenmp -std=c++11 are required.

• OpenMP for parallelism

To avoid possible binary incompatibilities, try to build these libraries with the same compiler that will be used to build NetworKit.

To build the module, the following is required:

• Python 3 (>= 3.4 recommended, 3.3 supported)
• [for developers: Cython (>= 0.21)]

Additionally, the module uses the following external Python packages:

• networkx
• tabulate
• scipy
• matplotlib

NetworKit is still usable without these packages, but with reduced functionality.The dependencies are best installed via easy_install or pip. (If you have multiple Python installations, be sure you use the commands matching your Python 3 version. In the following examples, python3 and pip3 will be used.)

pip3 install package_name

Then, switch to the top folder and run the script setup.py with the following options:

python3 setup.py build_ext --inplace [--optimize=V] [-jX]

The script will call scons to compile NetworKit as a library and then build the extensions in the top folder. By default, NetworKit will be built with the amount of available cores in optimized mode. It is possible the add the options --optimize=V and -jN the same way it can be done to a manual scons call, to specify the optimization level and the number of threads used for compilation. The setup script provides more functionality and can be used with pip aswell:

pip3 install -e .

will compile NetworKit, build the extensions and on top of that temporarily install NetworKit so that it is available on the whole system. pip3 uninstall networkit will remove networkit.

python3 setup.py clean [--optimize=V]

will remove the extensions and its build folder as well as call scons to remove the NetworKit library and its build folder specified by --optimize=V.

Note: All of the above installation command may require root privileges depending on your system, so try this accordingly. If you do not have root privileges, add --user to your command.

Assuming the extension module _NetworKit exists, the front end can be imported in python:

python3
>>> import networkit

With NetworKit as a Python extension module, you get access to native high-performance code and can at the same time work interactively in the Python ecosystem. Although the standard Python interpreter works fine, we recommend IPython as a great environment for scientific computing. IPython can also be installed via pip or easy_install. For tab completion in ipython you may also need to install readline. The following should work: pip3 install readline ipython.

After the requirements are satisfied, start IPython and import NetworKit.

ipython3
>>> from networkit import *

Now you should be able to use NetworKit interactively. For usage examples, refer to the UserGuide.

It is now possible to install Networkit via [sudo] pip[3] install networkit. While this is supposed to be convenient, there are some restrictions. Before you use pip to install NetworKit, Cython, SCons and g++ need to be installed. During installation, the setup will check if the external packages networkit uses are available and print warnings at the end of the installation process. If you don't see any warnings, your system should be ready to use NetworKit.

You can remove NetworKit completely by using the command [sudo] pip[3] uninstall networkit.

We recommend that you familiarize yourself with NetworKit through experimenting with the interactive IPython Notebook NetworKit_UserGuide.ipynb located in the folder Doc/Notebooks. To display and work with these notebooks, you have to start a local notebook server from the terminal with:

ipython3 notebook

It can occur, that ipython3 itself works, but the notebook server won't start. In that case, head over to the IPython website and make sure, you have the listed packages installed. If the notebook server starts as it is supposed to, your default browser should open a web interface named "IPython Dashboard". You can either add NetworKit_UserGuide.ipynb from the above mentioned location, or you can point IPython to the location by starting it with

ipython3 notebook --notebook-dir=Doc/Notebooks

The notebook appears in the list and you can start it by clicking on it.

To show plots within the notebooks, place the following two lines at the beginning of your notebook:

%matplotlib
matplotlib.pyplot as plt

In case you do not need NetworKit's Python functionality, this section describes how to build the C++ parts only.

We recommend SCons for building the C++ part of NetworKit. Individual settings for your environment will be read from a configuration file. As an example, the file build.conf.example is provided. Copy this to build.conf and edit your environment settings. Then call scons.

The call to SCons has the following options:

scons --optimize=<level> --target=<target>

where <level> can be

• Dbg debug
• Opt optimized
• Pro profiling

and <target> can be

• Core build NetworKit as a library, required by the Python shell
• Tests build executable for the unit tests
• Lib build NetworKit as a library and create symbolic links

For example, to build NetworKit as an optimized library, run scons --optimize=Opt --target=Core

To speed up the compilation on a multicore machine, you can append -jX where X denotes the number of threads to compile with.

Logging is enabled by default. If you want to disable logging functionality, add the following to your scons call:

--logging=no

You actually don't need to build and run our unit tests. However if you experience any issues with NetworKit, you might want to check, if NetworKit runs properly. Please refer to the Unit Tests and Testing section in our DevGuide.

It is also possible to use NetworKit as a library. Therefore, choose the target Lib when compiling NetworKit. The include directives in your C++-application look like the following

#include <NetworKit/graph/Graph.h>

NetworKit in the directory include is a symlink to the directory networkit/cpp, so the directory structure from the repository is valid. To compile your application, you need to add the paths for the header files and the location of the library. Note, that it is possible to link the different builds (debug, profiling, optimized) of the library. There is a simple source file to demonstrate this. Feel free to compile LibDemo.cpp as follows:

g++ -o LibDemo -std=c++11 -I/path/to/repository/include
	-L/path/to/repository LibDemo.cpp -lNetworKit -fopenmp

• Mac OS X 10.10 "Yosemite": Some users have reported compilation problems on Yosemite with g++ 4.9. The compiler errors mention register problems. While the exact reason remains unclear, the actual issue seems to be that the compiler tries to perform a dual architecture build. Fix: Enforce a 64-bit build by prepending ARCHFLAGS="-arch x86_64" to your setup/pip command, e.g. as in sudo ARCHFLAGS="-arch x86_64" python3 setup.py build_ext --inplace -j4 or sudo ARCHFLAGS="-arch x86_64" pip3 install networkit.

We would like to encourage contributions to the NetworKit source code. See the development guide (DevGuide.mdown) for instructions. For support please contact christian.staudt @ kit.edu.

• Christian L. Staudt - christian.staudt @ kit.edu - Homepage
• Henning Meyerhenke - meyerhenke @ kit.edu - Homepage

• Maximilian Vogel - maximilian.vogel @ student.kit.edu

• Lukas Barth
• Miriam Beddig
• Elisabetta Bergamini
• Stefan Bertsch
• Pratistha Bhattarai
• Andreas Bilke
• Simon Bischof
• Guido Brückner
• Mark Erb
• Kolja Esders
• Patrick Flick
• Michael Hamann
• Lukas Hartmann
• Daniel Hoske
• Gerd Lindner
• Moritz v. Looz
• Yassine Marrakchi
• Mustafa Özdayi
• Marcel Radermacher
• Klara Reichard
• Marvin Ritter
• Aleksejs Sazonovs
• Arie Slobbe
• Florian Weber
• Michael Wegner
• Jörg Weisbarth

The program source includes:

• the The Lean Mean C++ Option Parser by Matthias S. Benkmann
• the TTMath bignum library by Tomasz Sowa

The source code of this program is released under the MIT License. We ask you to cite us if you use this code in your project. Feedback is also welcome.

The [NetworKit publications page lists the publications on NetworKit as a toolkit, on algorithms available in NetworKit, and simply using NetworKit. We ask you to cite the appropriate ones if you found NetworKit useful for your own research.