Engineering A Workload-balanced Push-relabel Algorithms for Extremely Large Graphs on GPUs

1. Getting started Instructions.

• Clone this project
• Hardware:
  • CPU x86_64 (Test on Intel(R) Core(TM) i9-10900K CPU @ 3.70GHz)
  • NVIDIA GPU (arch>=86) with device memory >= 12GB.(Support NVIDIA RTX3080(sm_86)). Note that we mainly evaluate our experience on RTX3090. The execution time could be different with different devices.
• OS & Compler:
  • Ubuntu 22.04
  • CUDA >= 11.6
  • nvcc >= 11.6
• Important Files/Directories
  • maxflow-cuda/: contains all source code of parallel push-relabel algorithm using reversed CSR (RCSR), including thread-centric(TC) and vertex-centric(VC) approach.
  • maxflow-bcsr: contain TC, VC push-relabel algorithms upon bidirectional CSR (BCSR). Backward edges of a vertex are continuously appended to the end of the its forward edges.
  • (TODO) maxflow-bcsr2: the optimized version of BCSR, where the backward edge is placed right after a forward edges, namely, flow[2i] stores the forward edge value, and flow[2i+1] stores the corresponding backward edge.
  • maxflow-serial/: contains all source code of serial version of push-relabel algorithm on CPU.
  • network/: contains the network generator.
    • DIMACS: the first DIMACS implementation challenge source
    • GTgraph: the modified version from here

2. Setup & Experiments

(0) Choose the right branch

Please choose the underlying CSR implementation

git checkout <branch_name>

branch_name:

• csr: the original CSR implementation
• rcsr: the reversed CSR implementation
• bcsr: the bidirectional CSR implementation

(1) Configure the Makefile

Please uncomment the compiling flag of WORKLOAD in both CC_FLAGS and NVCC_FLAGS to record each warp's execution time and print out the result.

# Profiling flags
# CC_FLAG += -DWORKLOAD
CC_FLAG += -UWORKLOAD
# NVCC_FLAG += -DWORKLOAD
NVCC_FLAG += -UWORKLOAD

(2) Compile source code

Run the makefile to compile source code and create necessary directories:

cd maxflow-cuda
make

(3) Prepare the networks

• 1st DIMACS Challenge synethesized networks (types: Washiongton, Genrmf):

  % Washiongton
  cd network/DIMACS/washiongton
  cc washington.c -o gengraph
  
  % Genrmf
  cd network/DIMACS/genrmf
  make
  

• SNAP

(4) Execute the binary

Usage: ./maxflow [-h] [-v file type] [-f filename] [-s source] [-t sink]
Options:
	-h		Show this help message and exit
	-v		Specify which kind of file to load
				0: SNAP txt file (default)
				1: SNAP binary file
				2: DIMACS file
	-f filename	Specify the file path (binary or txt)
	-s source	Source node
	-t sink		Sink node
	-a algorithm	Specify the algorithm to use
				0: Thread-centric push-relabel (default)
				1: Vertex-centric push-relabel

(5) Use the scripts

• maxflow-cuda/bash_run.py: Execute all files in the given directory and capture the total execution time and each warp's accumulated execution time

  usage: bash_run.py [-h] [--log LOG] --dir DIR --times TIMES [--stats STATS]
  
  optional arguments:
  -h, --help     show this help message and exit
  --log LOG      Path to the log file or 'stdout' for console output.
  --dir DIR      Path to the directory where the input files to execute.
  --times TIMES  Path to the file where execution times of each warp will be logged. (Please add the WORKLOAD flag.)
  --stats STATS  Path to the file where statistics of warp execution time will be logged. (Please add the WORKLOAD flag.)
  

• plots/boxplot: Plot the boxplot of workload

  usage: boxplot.py [-h] --file FILE --output OUTPUT
  
  optional arguments:
  -h, --help       show this help message and exit
  --file FILE      Path to the statistics file generated from bash_run.py.
  --output OUTPUT  Path to the output image file.
  

Citation

If this project helps you well, please give us a star or cite our paper.

@article{hsieh2024engineering,
  title={Engineering A Workload-balanced Push-Relabel Algorithm for Massive Graphs on GPUs},
  author={Hsieh, Chou-Ying and Lin, Po-Chieh and Kuo, Sy-Yen},
  journal={arXiv preprint arXiv:2404.00270},
  year={2024}
}