/hypergraph_partitioner

Collection of (hyper)graph partitioners with a simple python interface for experiments.

Primary LanguageC

Hypergraph Partitioner Collection

This repository contains a collection of commonly used (hyper)graph partitioners and provides scripts to simplify setup and execution.

Installation

Simply execute:

git submodule init
git submodule update
./install.sh

Worflow for Running Experiments

This assumes you are using bash or a compatible shell.

  • Create a folder that contains the graphs for the experiment (and preferably nothing else)
  • Adjust the paths in env.sh as required and execute source <path-to-repo>/env.sh
  • Create a folder for the experiments and an experiment.json file within it
  • Define the experimental setup within experiment.json (you can use experiments/example_experiment.json as orientation). Specifically, the path to the graphs folder is set here
  • Execute <path-to-repo>/experiments/setup_experiments.py experiment.json within the same folder. This will create a new subfolder with a file workload.txt that contains one line for each run of the experiment
  • Run the workload, either directly or by using <path-to-repo>/experiments/execute_experiments.py experiment.json
  • After the experiment is completed: Use <path-to-repo>/experiments/grep_experiment_results.sh <generated-folder> to collect the results into csv files

Notes:

  • The partitioner calls are implemented by the python scripts in scripts/. The scripts can be extended, e.g. to collect additonal stats
  • To add a new partitioner, an according script must be added and the mapping in experiments/setup_experiments.py (as well as experiments/execute_experiments.py) must be updated accordingly

Executing a (Hyper)graph Partitioner as Standalone

Before you can use the python interface scripts for each partitioner (see folder scripts), please make sure that env.sh contains the correct paths to each subprogram required for execution. If so, execute source env.sh to make the changes visible for the python interfaces. Afterwards, you can start a partitioner by executing the following command:

Serial Partitioner:

<partitioner>.py <graph/hypergraph-file> <number-of-blocks> <epsilon> <seed> <objective> <timelimit>

Parallel Partitioner

<partitioner>.py <graph/hypergraph-file> <number-of-threads> <number-of-blocks> <epsilon> <seed> <objective> <timelimit>

Hypergraph Formats

hMetis, KaHyPar, MT-KaHyPar, and Hype require that the input hypergraph file is in hMetis format (see hMetis User Guide). Parkway uses a binary representation of an already distributed hypergraph. However, our python interface expects the hypergraph in hMetis format and the conversion to the internal Parkway format is done inside our python script. Please make sure that the helper tool HgrToParkway is built inside the MT-KaHyPar partitioner (make HgrToParkway).

Mondriaan uses a matrix representation as input hypergraph file. You can convert a hypergraph in hMetis format to Mondriaan form via tool HgrToMondriaanMtx inside the KaHyPar partitioner (make HgrToMondriaanMtx). Please make sure that the corresponding hMetis hypergraph file is inside the same folder than the Mondriaan hypergraph file if you execute the mondriaan python interface script.

PaToH and Zoltan also use different hypergraph file representations. You can convert a hMetis hypergraph file via HgrToPatoh (inside the KaHyPar partitioner) resp. HgrToZoltan (inside the MT-KaHyPar partitioner) to PaToH resp. Zoltan hypergraph file format. Please make sure that Zoltan hypergraph file names ends with .zoltan.hg.