MP4GNN is an MPI-based library for partitioning large irregular graphs. MP4GNN is based on ParMETIS' multilevel multi-constraint k-way algorithm developed in our lab and contains various efficiency and memory optimizations. Many of these optimizations were developed while working on DistDGL and DistDGLv2.
You can download MP4GNN by simply cloning it using the command:
git clone [email protected]:KarypisLab/PM4GNN.git
To build MP4GNN you can follow the instructions below:
General dependencies for building MP4GNN are: gcc, cmake, build-essential, and an MPI library. In Ubuntu systems these can be obtained from the apt package manager (e.g., apt-get install cmake, mpich, etc)
sudo apt-get install build-essential
sudo apt-get install cmake
In addition, you need to download and install GKlib and METIS by following the instructions there.
MP4GNN is primarily configured by passing options to make config. For example:
make config cc=mpicc prefix=~/local
make install
will configure MP4GNN to be built using mpicc and then install the binaries, header files, and libraries at
~/local/bin
~/local/include
~/local/lib
directories, respectively.
cc=[compiler] - The C compiler to use [default is determined by CMake]
shared=1 - Build a shared library instead of a static one [off by default]
prefix=[PATH] - Set the installation prefix [~/local by default]
gklib_path=[PATH] - Set the prefix path where GKlib has been installed. You can skip
this if GKlib's installation prefix is the same as that of
MP4GNN.
metis_path=[PATH] - Set the prefix path where METIS has been installed. You can skip
this if METIS' installation prefix is the same as that of
MP4GNN.
gdb=1 - Build with support for GDB [off by default]
debug=1 - Enable debugging support [off by default]
assert=1 - Enable asserts [off by default]
assert2=1 - Enable very expensive asserts [off by default]
make uninstall
Removes all files installed by 'make install'.
make clean
Removes all object files but retains the configuration options.
make distclean
Performs clean and completely removes the build directory.
MP4GNN uses the same data types for integers and floating point numbers (32/64 bit integers and single/double precision floating point numbers) as used when configuring and building METIS.
Copyright 1998-2023, Regents of the University of Minnesota and is licensed under the Apache License, Version 2.0.
There are several papers describing ParMetis' underlying algorithms [1] [2] [3] [4].
For citing this particular repository, use the following:
@misc{pm4gnn23,
author = {George Karypis},
title = {PM4GNN -- Parallel Graph Partitioning for Distributed GNNs},
howpublished = "\url{https://github.com/KarypisLab/PM4GNN/}",
year = {2023}
}