The Jet Partitioner is a parallel graph partitioner that runs on most CPU and GPU systems (via Kokkos, a required dependency). This partitioner was developed in a collaboration between Sandia National Labs and Pennsylvania State University. For details about the algorithm, please see https://arxiv.org/abs/2304.13194
Kokkos (https://github.com/kokkos/kokkos): Enables performance portable parallelism.
KokkosKernels (https://github.com/kokkos/kokkos-kernels): Necessary only for KokkosSparse::CrsMatrix class.
Metis (https://github.com/KarypisLab/METIS): Used for initial partitioning of coarsest graph.
(Circumstantial) GKLib (https://github.com/KarypisLab/GKlib.git): Needed to link against the github distribution of Metis. Not needed for older distributions of Metis.
Standard CMake build process. If your Metis build requires GKlib, add -DLINK_GKLIB=True to your cmake command when building Jet. Example build scripts are provided for macOS with OpenMP and Linux systems with Cuda. These scripts handle all required dependencies.
Each partitioner executable requires 2 parameters. The first is a graph file in metis format, the second is a config file. Multiple sample config files are provided in the "configs" directory. Optionally, a third parameter can be used to specify an output file for the partition, and a fourth parameter for runtime statistics in JSON format.
Although the partitioner itself supports weighted edges and vertices, the import method currently does not support weighted vertices.
jet: The primary partitioner exe. Coarsening algorithm can be set in config file. Runs on the default device.
jet_host: jet but runs on the host device.
jet_serial: jet but runs on the host on a single thread.
pstat: Given a metis graph file, partition file, and k-value, will print out quality information on the partition.
We can not provide an option to compile a library due to the use of templates. However, you can import "jet.hpp" into your code to use the partitioner via the "jet_partitioner::partition" method. Note that this requires you to add our source directory to your include path and also to link our dependencies. This method currently always uses the default coarsening algorithm.
We do not yet support vertex weights within metis graph files.
<Coarsening algorithm> (0 for 2-hop matching)/(1 for HEC)/(2 for pure matching)/(default is 2-hop matching)
<Number of parts>
<Partitioning attempts>
<Imbalance value>