Skip to content

suzumura/graph500

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

35 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Please cite the following paper if you use our code.

Koji Ueno, Toyotaro Suzumura, Naoya Maruyama, Katsuki Fujisawa, Satoshi Matsuoka: Efficient Breadth-First Search on Massively Parallel and Distributed-Memory Machines. Data Science and Engineering 2(1): 22-35 (2017)

Koji Ueno, Toyotaro Suzumura, Naoya Maruyama, Katsuki Fujisawa, Satoshi Matsuoka (2016-12-01). "Extreme scale breadth-first search on supercomputers". 2016 IEEE International Conference on Big Data (Big Data): 1040–1047. doi:10.1109/BigData.2016.7840705.

The license of the code is Apache License, Version 2.0.

Here is how you build and run our code. All the detailed technical information is described in the above our paper.

Build

Prerequisites

Libraries/Frameworks shown below are needed to be installed.

  • libnuma
  • C/C++ compiler supports OpenMP (e.g. gcc/g++)
  • GNU make
  • MPI library & runtime
    • OpenMPI
    • MPICH
    • MVAPICH

Build instruction

  1. Change directory to root of this repository.
  2. Run this commands.
cd mpi
make cpu
  1. Now you can run benchmarking with command like this.
# OpenMPI >= 1.8
mpirun -np 1 -bind-to none -output-filename ./log/lP1T8S16VR0BNONE -x OMP_NUM_THREADS=8 ./runnable 16
# OpenMPI <= 1.6.5
mpirun -np 1 -bind-to-none -output-filename ./log/lP1T8S16VR0BNONE -x OMP_NUM_THREADS=8 ./runnable 16
# MPICH / MVAPICH
mpirun -n 1 -outfile-pattern ./log/lP1T8S16VR0BNONE -genv OMP_NUM_THREADS 8 ./mpi/runnable 16

Available options

Specifying number of threads and number of scales

# OpenMPI >= 1.8
mpirun -np 1 -bind-to none -output-filename ./log/lP1T8S16VR0BNONE -x OMP_NUM_THREADS=<nthreads> ./runnable <nscale>
# OpenMPI <= 1.6.5
mpirun -np 1 -bind-to-none -output-filename ./log/lP1T8S16VR0BNONE -x OMP_NUM_THREADS=<nthreads> ./runnable <nscale>
# MPICH / MVAPICH
mpirun -n 1 -outfile-pattern ./log/lP1T8S16VR0BNONE -genv OMP_NUM_THREADS <nthreads> ./mpi/runnable <nscale>
  • nthreads : number of threads
  • nscale : number of scale

make options

make [VERBOSE=<bool>] [VERTEX_REORDERING=<0|1|2>] [REAL_BENCHMARK=<bool>] cpu
  • VERBOSE : toggle verbose output. true = enable, false = disenable.
  • VERTEX_REORDERING : specify vertex reordering mode. 0 = do nothing (default), 1 = only reduce isolated vertices, 2 = sort by degree and reduce isolated vertices.
  • REAL_BENCHMARK : change BFS iteration times. true = 64 times, false = 16 times (for testing).

Benchmarking support script

In this repository, we provides support script written in python. This script includes these features:

  • spawn rebuilding if nesessary
  • generate logging file name automatically
  • generate and spawn mpirun command (supports OpenMPI(>= 1.8, <= 1.6.5)/MPICH/MVAPICH)
  • increasing-scale mode (iterate benchmarking with increasing scale unless runnable returns non-zero exit status)

For usage, please run this command:

# change directory to root of this repository and ...
./run-benchmark.py -h
# or ...
python run-benchmark.py -h

Examples

# scale 20, thread x 8, real benchmarking, MPI runtime : OpenMPI (>= 1.8)
./run-benchmark.py -s 20 -t 8 --logfile-dest ./log201606XX -m OpenMPI
# Notes : rebuilding `runnable` binary is executed in this script, so manual rebuilding is not needed.

# scale 20, thread x 8, test mode (BFS x 16), vertex reordering mode : 1, MPI runtime : MPICH
./run-benchmark.py -s 20 -t 8 --logfile-dest ./log201606XXtest --test --vertex-reordering 0 -m MPICH

Tips

  • If you need to rebuild forcibly, remove file prev_build_options.

Appendix : Tested environment

  • OS : Ubuntu 14.04 LTS (64bit)
  • Linux kernel version : 3.13.0-88-generic
  • libnuma, mpich are installed via apt
  • CPU : Intel Core i7-2640
  • gcc version : 4.8.4 (Ubuntu 4.8.4-2ubuntu1~14.04.3)
  • tested MPI runtime : MPICH 3.0.4