[KaHIPApplication] Introduction of KaHIPApplication for high-quality graph partitioning

[loumalouomega]

---

name: ✨ Feature
about: Introduction of KaHIPApplication for high-quality graph partitioning

---

📝 Description

This Pull Request introduces the KaHIPApplication, a new partitioning tool for Kratos Multiphysics based on the Karlsruhe High Quality Partitioning (KaHIP) framework. It is designed as a high-performance, drop-in alternative to the existing MetisApplication, offering superior edge-cut reduction for parallel computations.

Comparison

The resulñts, at least in these examples, are similar to the one obtained with Metis. Feel free to test with more complex meshes.

Cube

Quad

Why KaHIP?

While METIS is a standard in the industry, KaHIP provides several key advantages for our users:

• Efficiency: The eco configuration is comparable in speed to METIS but typically produces 5–20% fewer edge cuts on FEM meshes.
• Flexibility: The strong configuration allows for even lower edge cuts for users who prioritize communication efficiency over partitioning time.
• Licensing: KaHIP is MIT licensed, providing a friendly legal framework for various project types.
• In-Memory Workflow: Supports MPI Scatterv to distribute data directly in-memory, bypassing the need for intermediate .mdpa files on disk.

Basic usage example

Using the new high-level Modeler is straightforward:

import KratosMultiphysics as KM
import KratosMultiphysics.KaHIPApplication as KaHIP

model = KM.Model()
model_part = model.CreateModelPart("Main")
model_part.AddNodalSolutionStepVariable(KM.PARTITION_INDEX)

settings = KM.Parameters("""{
    "model_part_name": "Main",
    "input_filename":  "my_mesh",
    "partition_in_memory": true,
    "kahip_settings": {
        "preconfiguration": "eco",
        "imbalance": 0.03
    }
}""")

modeler = KaHIP.KaHIPPartitioningModeler(model, settings)
modeler.SetupModelPart()

Technical Details & Prerequisites

• Dependencies: Requires KaHIP ≥ v3.25 (included in-tree by default) and OpenMP.

• Build Configuration: Can be enabled via add_app in your configure script. Supports optional 64-bit edge indices for massive meshes:

  -DKAHIP_64BIT=ON

Key changes

1. Drop-in Compatibility: Includes processes like KaHIPDivideHeterogeneousInputProcess that mirror the interface of their Metis counterparts.
2. Advanced Modeler: Introduces KaHIPPartitioningModeler, a high-level API that handles the entire "partition-and-read" workflow in a single step.
3. CSR Integration: Automatically converts mesh nodal connectivity to CSR format for processing.
4. Boundary Synchronization: Optional synchronization to ensure conditions and constraints are co-located with their parent elements.
5. Customizable: Full access to KaHIP's preconfigurations (fast, eco, strong, etc.) and imbalance controls via Kratos Parameters.

Validation

I have included comprehensive tests covering both serial and MPI workflows. You can run them using the following commands:

# Python tests
python -m pytest applications/KaHIPApplication/tests/

# MPI tests
mpirun -n 4 python applications/KaHIPApplication/tests/test_KaHIPApplication_mpi.py

# C++ Unit tests
ctest -R KaHIPApplication

🆕 Changelog

• Core Application: Introduced KaHIPApplication, providing high-quality graph partitioning based on the Karlsruhe High Quality Partitioning (KaHIP) framework.
• Partitioning Processes:
  • KaHIPDivideHeterogeneousInputProcess: A serial partitioner designed as a drop-in replacement for MetisDivideHeterogeneousInputProcess.
  • KaHIPDivideHeterogeneousInputInMemoryProcess: An MPI-aware process for in-memory partitioning and distribution via MPI_Scatterv, eliminating the need for intermediate files.
• High-Level API: Added KaHIPPartitioningModeler, a streamlined modeler to automate the "partition-and-read" workflow in a single Python call.
• KaHIP Integration:
  • Integrated KaHIP v3.25 (in-tree build support under external_libraries/KaHIP).
  • Support for all KaHIP preconfigurations: fast, eco, strong, fastsocial, ecosocial, and strongsocial.
  • Support for configurable node weight imbalance and random seeds.
• Features:
  • Automatic conversion of Kratos mesh connectivity to CSR (Compressed Sparse Row) format.
  • Boundary Synchronization: Logic to ensure conditions, geometries, and master-slave constraints are co-located with their parent elements.
  • 64-bit Support: Added CMake option -DKAHIP_64BIT=ON for partitioning meshes with more than (2^{31}) edges.
• Testing Suite:
  • Comprehensive Python tests for both Serial and MPI execution.
  • C++ unit tests for core partitioning logic.
  • Integration with kratos_run_tests.py.

[loumalouomega]

Okay, this is ready, I tested and the fast mode is not significantly better than Metis, but is MPI ready...

[loumalouomega]

Okay, this is ready for review

[roigcarlo]

Looks good also from my side, not very into MPI nowadays, but if you need something added to the images I can add it

[loumalouomega]

Looks good also from my side, not very into MPI nowadays, but if you need something added to the images I can add it

I added automatic download in case not defined the lib, That's why I could add it to CI, and looks like it works. The library is small so the penalty of compiling is not that high, but if you prefer we can add it to the images

[pooyan-dadvand]

@KratosMultiphysics/technical-committee agrees with the addition of the library and assigns @roigcarlo to review it.