Skip to content

A Java binding for KaHyPar (Karlsruhe Hypergraph Partitioning).

License

Notifications You must be signed in to change notification settings

crillab/jkahypar

Repository files navigation

JKaHyPar

Description

This project provides a Java binding for the KaHyPar hypergraph partitioning framework.

This library consists of a Java module named fr.univartois.cril.jkahypar, which provides an object-oriented interface for building and partitioning hypergraphs. Under the hood, Java Native Access is used for invoking the native implementation provided by KaHyPar (currently, only Linux and macOS are supported).

Requirements

To use JKaHyPar, you need to have Boost installed on your computer. If you want to use the bundled shared library of KaHyPar, make sure to install the version 1.79.0 of Boost, which has been used to compile this library.

If you use an unsupported OS or a different version of Boost, you may need to recompile the shared library, and rebuild JKaHyPar.

Adding JKaHyPar as a Dependency to your Project

The JKaHyPar library is available on Maven Central. You may thus declare it as a dependency in your pom.xml, if you use Maven:

<dependency>
    <groupId>fr.cril</groupId>
    <artifactId>jkahypar</artifactId>
    <version>0.2.0</version>
</dependency>

or, if you use Gradle, in your build.gradle file:

implementation "fr.cril:jkahypar:0.2.0"

Building the Library

This library is developed with Gradle. You are not required to install it, since the gradlew script can do the job on its own.

To see how to build the native shared library, go here. Go there to see how to build the Java library.

Using the Library

To use JKaHyPar, you need to declare its module fr.univartois.cril.jkahypar as a requirement in your module-info.java, and put its JAR on the modulepath.

You may also need to put this JAR on the classpath when using it within another application, so as to allow to dynamically load the native library it contains (JNA does not look into the modulepath for such libraries). A possible alternative consists in specifying manually the location of the library (e.g., from this directory), but this approach is much less convenient in practice.

Once your path is properly set, you can use JKaHyPar to create hypergraphs and compute their partitions. The following section describes how to use these various features.

Creating a Hypergraph

The HypergraphBuilder allows to easily create your hypergraphs, which may have weights on their hyperedges or vertices, or both.

The following example shows how to create the hypergraph on page 12 in the hMetis manual.

int nbVertices = 7;
int nbHyperedges = 4;

var hypergraph = createHypergraph(nbVertices, nbHyperedges)
        .withHyperedge(joining(1, 2).withWeight(2))
        .withHyperedge(joining(1, 7, 5, 6).withWeight(3))
        .withHyperedge(joining(5, 6, 4).withWeight(8))
        .withHyperedge(joining(2, 3, 4).withWeight(7))
        .withVertexWeight(1, 5)
        .withVertexWeight(2, 1)
        .withVertexWeight(3, 8)
        .withVertexWeight(4, 7)
        .withVertexWeight(5, 3)
        .withVertexWeight(6, 9)
        .withVertexWeight(7, 3)
        .build();

// Do something with hypergraph.

Note that the methods createHypergraph and joining are statically imported from the classes HypergraphBuilder and UnweightedHyperedge, respectively.

You may also read a graph from a file written using the hMetis format (page 11 of its manual).

try (var parser = new HypergraphParser(inputStream)) {
    parser.parse();
    var hypergraph = parser.getHypergraph();

    // Do something with hypergraph.
}

Manipulating a Hypergraph

The Hypergraph interface defines a wide variety of methods to deal with the hypergraph you have built (or read). You may get its String representation (in the hMetis format mentioned above), its vertices and their weights, or its hyperedges.

You may also get its internal representation. For efficiency reasons, note that the hypergraph does not keep the reference to the hyperedge objects used to build it. Actually, the same representation as hMetis and KaHyPar is used. As such, the hyperedges you can retrieve from the hypergraph are views of its internal representation, and instantiating them has a cost, so think twice before using them!

Partitioning the Hypergraph

The main purpose of this library is to compute a partition of a hypergraph. To do so, you need to write something similar to the following example.

try (var context = new KahyparContext()) {
    // Configure the context using one of KaHyPar's predefined configurations.
    context.configureFrom("/path/to/configuration.ini");

    // Set the imbalance parameter for the partitioning algorithm.
    // This specifies how much the size of the blocks may be different.
    context.setImbalance(0.03);

    // Set the number of blocks you want in your partition.
    context.setNumberOfBlocks(2);

    // Compute the partition of a hypergraph.
    var hypergraph = createAnAwesomeHypergraph();
    var partitioner = context.createPartitionerFor(hypergraph);
    var partition = partitioner.computePartition();

    // Get the identifier of the block in which a vertex is.
    int blockOf1 = partition.blockOf(1);

    // Get all the blocks of the partition.
    var blocks = partition.getBlocks();

    // Get the value of the objective function for the computed partition:
    // the lower the value of the objective, the better the partition.
    int objective = partition.objectiveValue();

    // Improve the previously computed partition, by applying a certain amount
    // of iterations (5 iterations here).
    partition = partitioner.improvePartition(5);
}

Observe that the KahyparContext is instantiated in a try-with-resource. You should always do so, as this ensures that the memory used by the native library is properly freed. If you cannot use a try-with-resource, e.g., because you want to reuse the context, you must make sure that its close() method is called when you no longer need it.

About

A Java binding for KaHyPar (Karlsruhe Hypergraph Partitioning).

Topics

Resources

License

Stars

Watchers

Forks

Packages

No packages published

Languages