Skip to content

Latest commit

 

History

History
109 lines (69 loc) · 5.4 KB

README.md

File metadata and controls

109 lines (69 loc) · 5.4 KB
Raw

SuiteSparse:GraphBLAS driver for LDBC Graphalytics

Build Status

Platform driver for the LDBC Graphalytics benchmark using SuiteSparse:GraphBLAS and LAGraph.

To execute the Graphalytics benchmark on GraphBLAS, follow the steps in the Graphalytics tutorial on Running Benchmark with the GraphBLAS-specific instructions listed below.

Project structure

This project implements the GraphBLAS platform driver for the LDBC Graphalytics benchmark. It consists of the following components:

  • The platform driver is written in Java and implements the classes required by the ldbc_graphalytics framework.
  • The algorithms (BFS, PR, etc.) are implemented in C in the LAGraph library (currently on the dev branch).
  • The C++ wrapper for LAGraph is defined in the src/main/c directory.
  • The Java driver uses shell scripts to run the benchmark, see e.g. the execute-job.sh script that invokes the binary program for a given algorithm.
  • The graphs (stored in .v and .e files) are converted to a vertex relabelling file (.vtx) and a matrix stored in Matrix Market format (.mtx). The vertex relabelling is a bijective mapping that maps between the sparse UINT64 IDs in the original data to a dense contiguous set of IDs between 1 and |V|. The mapping is implemented in the relabel.py Python script that internally uses DuckDB.

Dependencies

On Debian/Fedora-based Linux distributions, you may install the prerequisite packages and dependencies listed below using a singe comman:

scripts/install-dependencies.sh

Prerequisite packages

Make sure you have the following software packages installed:

  • Apache Maven 3+
  • CMake 3.10+
  • C++ compiler: GCC, Clang, or ICC
  • Python 3.8+
  • DuckDB Python package (duckdb)

On Linux, you may use the following script to install these dependencies:

scripts/install-prerequisites.sh

SuiteSparse:GraphBLAS and LAGraph libraries

The implementation depends on two C libraries, SuiteSparse:GraphBLAS and LAGraph. We require very recent versions of these libraries, so it is best to compile them from their source code.

  1. To install SuiteSparse:GraphBLAS v8.0.0+, run:

    scripts/install-graphblas.sh

  2. To install LAGraph (dev branch), run:

    scripts/install-lagraph.sh

Building only the C++ wrapper

To only build the C++ wrapper (for quick test builds), run the following script:

bin/sh/build-wrapper-only.sh

Building the project and running the benchmark

  1. To initialize the benchmark package, run:

    scripts/init.sh ${GRAPHS_DIR} ${MATRICES_DIR}

    where

    • GRAPHS_DIR is the directory of the graphs and the validation data. The argument is optional and its default value is ~/graphs.
    • MATRICES_DIR is the directory of the pre-generated matrix files (in Matrix Market format). The argument is optional and its default value is ~/matrices.

    This script creates a Maven package (graphalytics-${GRAPHALYTICS_VERSION}-graphblas-${PROJECT_VERSION}.tar.gz). Then, it decompresses the package, initializes a configuration directory config (based on the content of the config-template directory) and sets default values of the directories (see above) and the number of threads.

    Note that the project uses the Build Number Maven plug-in to ensure reproducibility. Hence, builds fail if the local Git repository contains uncommitted changes. To build it regardless (for testing), run it as follows:

    scripts/init-for-testing.sh ${GRAPHS_DIR} ${MATRICES_DIR}

  2. Navigate to the directory created by the init.sh script:

    cd graphalytics-*-graphblas-*/

  3. Edit the configuration files (e.g. graphs to be included in the benchmark) in the config directory.

    • To conduct benchmark runs, edit the config/benchmark.properties file and replace the include = benchmarks/custom.properties to select the dataset size you wish to use, e.g. include = benchmarks/xl.properties
    • Inspect config/platform.properties and check whether the value of platform.graphblas.num-threads was set correctly.

  4. Run the benchmark with the following command:

    bin/sh/run-benchmark.sh

Workflow

The workflow of the GraphBLAS implementation is illustrated in the following figure. Note that the "raw graph files" and the "configuration" are provided by the user, while the rest of the data artifacts (intermediate data sets, outputs, etc.) are created automatically by the framework.

Workflow of the GraphBLAS implementation