/opm-linear-solver-lab

Simple benchmarking tool for the linear solvers of OPM.

Primary LanguageC++GNU General Public License v3.0GPL-3.0

opm-linear-solver-lab

Experimental playground for testing out linear solvers in OPM Flow.

Compiling

We assume you have opm and dune in your prefix path. Compiling should just be

mkdir build
cd build
cmake .. -DCMAKE_BUILD_TYPE=Release
make

Note that you probably want to specify CMAKE_PREFIX_PATH to point to the locations of the OPM and Dune build folder, that is

# from build. 
# IMPORTANT: Notice the quotation marks around the prefix path. 
# This is IMPORTANT when you have multiple paths
cmake .. -DCMAKE_BUILD_TYPE=Release \
    -DCMAKE_PREFIX_PATH="/path/to/opm-common/build;/path/to/opm-grid/build;MOREPATHSHERE"

To simplify this process, we've added the script build_helpers/prefix_path.sh to generate this prefix path. If you directory structure looks like this:

#DUNE
/some/path/for/dune/dune-common/build
/some/path/for/dune/dune-grid/build
/some/path/for/dune/dune-geometry/build
/some/path/for/dune/dune-istl/build

#OPM
/some/path/other/path/for/opm/opm-common/build
/some/path/other/path/for/opm/opm-grid/build
/some/path/other/path/for/opm/opm-models/build
/some/path/other/path/for/opm/opm-simulators/build

you can run cmake as

# from build. 
cmake .. -DCMAKE_BUILD_TYPE=Release \
    -DCMAKE_PREFIX_PATH="$(bash ../build_helpers/prefix_path.sh /some/path/other/path/for/opm/ /some/path/for/dune/dune-common/)"

Note on libfmt

You probably need to have libfmt installed. On Ubuntu this can be accomplished by installing

sudo apt install libfmt-dev

Alternatively you can manually donwload and install libfmt. Make sure to extend your CMAKE_PREFIX_PATH to contain the install directory of libfmt.

Running

From the build folder, you can run eg.

./linsolverlab \
    -x ../examples/matrices/spe1/rhs.mm \
    -y  ../examples/matrices/spe1/rhs.mm  \
    -m  ../examples/matrices/spe1/matrix.mm \
    -c ../examples/configurations/cpu/ilu0.json \
    -g ../examples/configurations/gpu/cuilu0.json

Example output

{
    "CPU": {
        "runtime": "21541",
        "failed_by_exception": "false",
        "iterations": "25",
        "reduction": "9.463892568041427e-13",
        "converged": "true",
        "conv_rate": "0.33040209473507992",
        "elapsed": "0.021522673999999999",
        "condition_estimate": "-1"
    },
    "GPU": {
        "runtime": "21831",
        "failed_by_exception": "false",
        "iterations": "25",
        "reduction": "3.532348061635463e-13",
        "converged": "true",
        "conv_rate": "0.31763074967064442",
        "elapsed": "0.02177946",
        "condition_estimate": "-1"
    }
}

Running only preconditioner

./linsolverlab \
    -x ../examples/matrices/spe1/rhs.mm \
    -y  ../examples/matrices/spe1/rhs.mm  \
    -m  ../examples/matrices/spe1/matrix.mm \
    -c ../examples/configurations/cpu/ilu0_onlypreconditioner.json \
    -g ../examples/configurations/gpu/cuilu0_onlypreconditioner.json

Example output

{
    "CPU": {
        "runtime": "6",
        "failed_by_exception": "false",
        "iterations": "1",
        "reduction": "0",
        "converged": "true",
        "conv_rate": "1",
        "elapsed": "0",
        "condition_estimate": "-1"
    },
    "GPU": {
        "runtime": "2396",
        "failed_by_exception": "false",
        "iterations": "1",
        "reduction": "0",
        "converged": "true",
        "conv_rate": "1",
        "elapsed": "0",
        "condition_estimate": "-1"
    }
}

Running benchmarks

OPM-flow

Generating benchmark data

# usage: run_opm_benchmark.py <subfolder1> [<subfolder2> ...]
python benchmarking_scripts/run_opm_benchmark.py sleipner

Adjusting the experiment in the file

    json_files = [
        'examples/configurations/cpu/ilu0.json --block-size 2',
        'examples/configurations/cpu/dilu.json --block-size 2',
        'examples/configurations/gpu/cuilu0.json --block-size 2',
        'examples/configurations/gpu/cudilu.json --block-size 2'
    ]
    matrix_root = "./examples/matrices"  # folder where the matrix files can be found
    limit = 10  # Adjust this to change the number of matrices processed per directory

Plotting the results

python benchmarking_scripts/process_and_plot_opm.py