hendriceh / ogl

The OpenFOAM Ginkgo Layer

Requirements | Compilation | Usage | Known Limitations | Citing | Example | Performance

---

OGL is a wrapper for ginkgo solvers and preconditioners to provide GPGPU capabilities to OpenFOAM.

Requirements

OGL has the following requirements

• cmake 3.9+
• OpenFOAM 6+ or v2106
• Ginkgo 1.5.0+ (recommended to install via OGL)
• C++14 compliant compiler (gcc or clang)

See also ginkgo's documentation for additional requirements.

Compilation

OGL can be build using cmake following the standard cmake procedure.

mkdir build && cd build && ccmake ..

By default OGL will fetch and build ginkgo, to specify which backend should be build you can use the following cmake flags -DGINKGO_BUILD_CUDA, -DGINKGO_BUILD_OMP, or -DGINKGO_BUILD_HIP. For example to build OGL with CUDA and OMP support use

cmake -DGINKGO_BUILD_CUDA=ON -DGINKGO_BUILD_OMP=ON ..

Then, compile and install by

make -j && make install

Ninja builds

If you have Ninja installed on your system we recommend to use ninja over gnu make for better compilation times

cmake -G Ninja -DCMAKE_BUILD_TYPE=Release -DGINKGO_BUILD_HIP=ON  ../..
cmake --build . --config Release
cmake --install .

And make sure that the system/controlDict includes the libOGL.so or libOGL.dyLib file:

libs ("libOGL.so");

Usage

OGL solver support the same syntax as the default OpenFOAM solver. Thus, to use Ginkgo's CG solver you can simply replace PCG by GKOCG. In order to run either with CUDA, HIP, or OMP support set the executor keyword to cuda, hip, or omp in the system/fvSolution dictionary.

Argument	Default	Description
ranksPerGPU	1	gather from n ranks to GPU
updateRHS	true	whether to copy the system matrix to device on every solver call
updateInitGuess	false	whether to copy the initial guess to device on every solver call
export	false	write the complete system to disk
verbose	0	print out extra info
executor	reference	the executor where to solve the system matrix, other options are omp, cuda
adaptMinIter	true	based on the previous solution set minIter to be relaxationFactor*previousIters
relaxationFactor	0.8	use relaxationFactor*previousIters as new minIters
scaling	1.0	Scale the complete system by the scaling factor
forceHostBuffer	false	whether to copy to host before MPI calls

Supported Solver

Currently, the following solver are supported

• CG
• BiCGStab
• GMRES
• IR (experimental)
• Multigrid (experimental)

additionally, the following preconditioners are available

Supported Matrix Formats (Experimental)

Currently, the following matrix formats can be set by matrixFormat

• Coo
• Csr
• Ell (not supported for ranksPerGPU != 1)
• Hybrid (not supported for ranksPerGPU != 1)

Supported Preconditioner

• BJ, block Jacobi
• ISAI, Incomplete Sparse Approximate Inverses,
• ILU, incomplete LU (experimental)
• IC, incomplete Cholesky (experimental)
• Multigrid, algebraic multigrid (experimental)

The following optional arguments are supported to modify the preconditioner. Note some preconditioners like IC or (SPD) ISAI require positive values on the system matrix diagonal, thus in case of the pressure equation the complete system needs to be scaled by a factor of -1.0.

Argument	Default	Preconditioner
SkipSorting	True	all
Caching	1	all
MaxBlockSize	1	block Jacobi
SparsityPower	1	ISAI
MaxLevels	9	Multigrid
MinCoarseRows	10	Multigrid
ZeroGuess	True	Multigrid

Known Limitations

Currently, only basic cyclic boundary conditions are supported. Block-coupled matrices are not supported.

Citing

When using OGL please cite the main Ginkgo paper describing Ginkgo's purpose, design and interface, which is available through the following reference:

@misc{anzt2020ginkgo,
    title={Ginkgo: A Modern Linear Operator Algebra Framework for High Performance Computing},
    author={Hartwig Anzt and Terry Cojean and Goran Flegar and Fritz Göbel and Thomas Grützmacher and Pratik Nayak and Tobias Ribizel and Yuhsiang Mike Tsai and Enrique S. Quintana-Ortí},
    year={2020},
    eprint={2006.16852},
    archivePrefix={arXiv},
    primaryClass={cs.MS}
}

Example

Below an animation of a coarse 2D simulation of a karman vortex street performed on a MI100 can be seen. Here both the momentum and Poisson equation are offloaded to the gpu.