MG-αGCD

MG-αGCD is an open source library that optimizing Louvain graph community detection method for efficiency and scalability on Multi-GPU platforms.

Paper information

Shuai Yang and Changyou Zhang. 2025. MG-𝛼GCD: Accelerating Graph Community Detection on Multi-GPU Platforms. In 2025 International Conference on Supercomputing (ICS ’25), June 8–11, 2025, Salt Lake City, UT, USA. ACM, New York, NY, USA, 14 pages. https://doi.org/10.1145/3721145.3725753

Dependencies

1. Install CUDA 11.6 or newer

2. Install NVSHMEM 2.7.0 or newer (Spack install sample)

   • Install spack

     Please check this branch , which provides nvshmem 2.9.0 version.

   • Requirements for nvshmem installation

     • Load spack

       source ~/spack/share/spack/setup-env.sh

     • Add gcc and gfortran compilers to spack

       spack compiler find
       sudo apt-get install gfortran

       Modify /root/.spack/linux/compilers.yaml, Specifies the gfortran path

       f77: /usr/bin/gfortran
       fc: /usr/bin/gfortran

       spack compilers

     • install libibverbs-dev

       sudo apt install libibverbs-dev
       

   • Spack install nvshmem

     spack install nvshmem +gpu_initiated_support +cuda +gdrcopy +ucx +mpi build_system=cmake ^openmpi +cuda fabrics=ucx ^ucx +cuda +gdrcopy +dm +thread_multiple ^pmix@4.2.2

   • NVSHMEM environment load

     spack load nvshmem ^pmix@4.2.2

3. Install OpenMP 5.0 or newer

4. Install CMake 3.15.5 or newer

Getting Start

Our program only supports input graph files in matrix market format (.mtx). All the graph datasets used in our evaluation can be accessed from the SuiteSparse Matrix Collection.

1. Create the build file folder in the root path of the program, and generate makefile with:

   cmake ..

2. Generate the executable file MG_GCD:

   make

3. Run the program with mpirun.

   mpirun -np 4 --allow-run-as-root ./MG_GCD -path_of_graph ../data/wiki-topcats.mtx 

   Options:

   • -np: The number of GPUs
   • -path_of_graph: The path of the input graph data (only .mtx format is acceptable)
   • -version: The different versions of MG-αGCD discussed in the paper.
     • 0: The irregular sparse remote atomic implementation.
     • 1: The global bi-probing and pipelined version.
     • 2: The local bi-probing and pipelined version (default).