CONTENTS OF THIS FILE

• Introduction
• Requirements
• Installation
• Maintainers

To learn more about VGL framework, please visit vgl.parallel.ru website.

INTRODUCTION

VGL is a high-performance graph processing framework, designed for modern NEC SX-Aurora TSUBASA vector architecture. NEC SX-Aurora TSUBASA is equipped with world highest memory bandwidth of 1.2-1.5 TB/s, which allows to significantly accelerate graph-processing.

VGL significantly outperforms many state-of the art graph-processing frameworks for modern multicore CPUs and NVIDIA GPUs, such as Gunrock, CuSHA, Ligra, Galois, GAPBS

Requirements

GCC compiler, version >= 8.3.0 NCC compiler, version >= 3.1.0 ASL (Advanced Scientific Library), version >= 2.1.0 Python, version >= 3.6

(additional for GPU API) In order to use VGL GPU API you must also have CUDA toolkit, version >= 10.0 (additional for automatic benchmarking) GNU bash, version >= 4.0

Installation

1. Download VGL:

   • source files from GitHub:

     git clone https://github.com/afanasyev-ilya/VectorGraphLibrary

   • or .tar arcieve from this website:

     wget https://afanasyev-ilya.github.io/vgl_site/VectorGraphLibrary.zip

     wget https://afanasyev-ilya.github.io/vgl_site/VectorGraphLibrary.tar.gz

2. Check sample applications in apps folder:

   cd VectorGraphLibrary/apps

   vim bfs/bfs.cpp

3. Modify Makefile.nec if neccesarry:

   vim Makefile.nec

   you may change compiler paths:

   VE_CXX = nc++

   VH_CXX = g++

   or the path to ASL:

   -I /opt/nec/ve/nlc/2.0.0/include/

4. Build VGL samples:

   • all sample applications:

     make -f Makefile.nec all

   • or a specific application (for example shortest paths):

     make -f Makefile.nec sssp

5. Run automatic verification and performance testing

   • prepare testing data and compiler sources:

     python3 ./run_tests.py --arch=sx --compile --prepare

   • verification:

     python3 ./run_tests.py --arch=sx --verify

   • performance evaluation:

     python3 ./run_tests.py --arch=sx --benchmark file tests_results_sx.xlsx is created with performance and verification results for all algorithms

6. Using VGL on Lomonosov-2 supercomputer:

   • Lomonosov-2 supercomputer is now equipped with NEC SX-Aurora TSUBASA vector engines. To start using VGL load the following modules:

     (1) nec/2.13.0 (2) anaconda3/3.7.0
     (3) slurm/15.08.1 (4) gcc/9.1

   • Download VGL to your home directory /home/user_name/ (Warning! you can'd download it to _scratch or its subfolders)

     git clone https://github.com/afanasyev-ilya/VectorGraphLibrary

   • Compile VGL on head node using either:

     make -f Makefile.nec all

   • Download input graphs from head node:

     python3 ./run_tests.py --arch=sx --download-only

   • Prepare VGL input graphs on vector host of any partition (nec/pascal/test/etc):

     sbatch -p nec ./graph_library.sh python3 ./run_tests.py --arch=sx --prepare

   • Run benchmarks and tests on vector engines:

     sbatch -p nec ./graph_library.sh python3 ./run_tests.py --arch=sx --benchmark --verify

Maintainers

VGL is maintained by Ilya Afanasyev, afanasiev_ilya@icloud.com.