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.
We support three ways to install NetworKit:
- NetworKit Virtual Machine: Download and try NetworKit preinstalled on a virtual machine. This is recommended for users using Microsoft Windows.
- Pip install: Download the NetworKit Python package with pip. This is the easier way to get NetworKit but you can only use NetworKit via Python this way.
- Build NetworKit from Source: Clone or download the source code of NetworKit and build the C++ and Python modules from source.
More detailed instructions follow after the requirements section. 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 and jupyterhub as great environments for scientific computing.
Once you have installed NetworKit, please make sure to check out our NetworKit UserGuide for an overview of the features provided in NetworKit.
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 ([email protected]
) and feel free to ask.
You will need the following software to install NetworKit as a python package:
- A modern C++ compiler, e.g.: g++ (>= 4.8) or clang++ (>= 3.7)
- OpenMP for parallelism (usually ships with the compiler)
- Python 3 (>= 3.4 is recommended, 3.3 supported)
- Pip
- SCons: Please note that SCons is only available for Python 2. For installation via pip, we have a script that builds the C++ part of NetworKit, so you can try it without SCons. If you are interested in building different configurations and targets (e.g. unittests) from source, SCons is necessary.
- [for developers: Cython (>= 0.21)]
pip3 install [--user] networkit
Depending on your system, you might need to install python3-tk.
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
.
To check that everything works as expected, open a python terminal and run the example from the beginning of this document or the following lines:
>>> import networkit
>>> G = networkit.Graph(5)
>>> G.addEdge(0,1)
>>> G.toString()
Additionally, we recommend that you familiarize yourself with NetworKit through experimenting with the interactive IPython Notebook NetworKit_UserGuide.ipynb
located in the folder Doc/Notebooks
. The user guide also introduces a large portion of NetworKits functionality with usage examples. To display and work with these notebooks, you have to install jupyterhub and start a local notebook server from the terminal with:
jupyterhub --no-ssl
If you run into any problems with jupyterhub, head over to the jupyterhub documentation 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 or you have to open it manually. Then you can add NetworKit_UserGuide.ipynb
from the above mentioned location or browse to the location through the web interface.
To show plots within the notebooks, place the following two lines at the beginning of your notebook:
%matplotlib
matplotlib.pyplot as plt
Note: Instead of running jupyterhub, it may still be possible to run ipython3 notebook
. However, the notebook functionality of the ipython package is deprecated and has been moved to jupyterhub, which we strongly recommend.
Now that you are done installing NetworKit, you might want to try the following example:
>>> from networkit import *
>>> g = generators.HyperbolicGenerator(1e5).generate()
>>> overview(g)
Network Properties for: G#5
nodes, edges 100000, 302148
directed? False
weighted? False
isolated nodes 1859
self-loops 0
density 0.000060
clustering coefficient 0.718261
min/max/avg degree 0, 1045, 6.042960
degree assortativity 0.000725
number of connected components 4237
size of largest component 77131 (77.13 %)
>>> communities = community.detectCommunities(g, inspect=True)
PLM(balanced,pc,turbo) detected communities in 0.3781468868255615 [s]
solution properties:
------------------- -----------
# communities 4468
min community size 1
max community size 1820
avg. community size 22.3814
modularity 0.989285
------------------- -----------
>>>
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
debugOpt
optimizedPro
profiling
and <target>
can be
Core
build NetworKit as a library, required by the Python shellTests
build executable for the unit testsLib
build NetworKit as a library and create symbolic links
For example, to build NetworKit as an optimized library, run
scons --optimize=Opt --target=Lib
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
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
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. The unit tests can only be run from a clone or copy of the repository and not from a pip installation. Please refer to the Unit Tests and Testing
section in our DevGuide
.
- 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 insudo ARCHFLAGS="-arch x86_64" python3 setup.py build_ext --inplace -j4
orsudo ARCHFLAGS="-arch x86_64" pip3 install networkit
. - NetworKit has not yet been successfully built on Windows. This is partially due to the fact that Windows ships without a C++ compiler which is necessary to build the Python extensions. Even with the Visual C++ Redistributable our attempts were not successful. Any help is appreciated. It may be possible to build NetworKit as a library on Windows in environments like MinGW or Cygwin.
We would like to encourage contributions to the NetworKit source code. See the development guide (DevGuide.mdown
) for instructions. For support please contact the mailing list.
NetworKit is maintained by the Research Group Parallel Computing of the Institute of Theoretical Informatics at Karlsruhe Institute of Technology (KIT).
- Christian L. Staudt
- Henning Meyerhenke
- Maximilian Vogel
- 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
- Matteo Riondato
- 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 (c.f. the publications section below and especially the technical report). 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.