DPBench - Benchmarks to evaluate Data-Parallel Extensions for Python

• <benchmark>_numba_<mode>.py : This file contains Numba implementations of the benchmarks. There are three modes: nopython-mode, nopython-mode-parallel and nopython-mode-parallel-range.
• <benchmark>_numba_dpex_<mode>.py : This file contains Numba-Dpex implementations of the benchmarks. There are three modes: kernel-mode, numpy-mode and prange-mode.
• <benchmark>_dpnp_<mode>.py : This file contains dpnp implementations of the benchmarks.
• <benchmark>_native_ext/<benchmark>_sycl/_<benchmark>_kernel.hpp : This file contains native dpcpp implementations of the benchmarks.
• <benchmark>_numpy.py : This file contains numpy implementations of the benchmarks. It should take benefits of numpy arrays and should avoid loops over arrays.
• <benchmark>_python.py : This file contains naive python implementations of the benchmarks. Should be run only for small presets, otherwise it will take long execution time.
• <benchmark>_numba_mlir_<mode>.py : This file contains Numba-MLIR implementations of the benchmarks. There are three modes: kernel-mode, numpy-mode and prange-mode. Experimental.

Examples of setting up and running the benchmarks

Using prebuilt version

1. Create conda environment

   conda create -n dpbench dpbench -c dppy/label/dev -c conda-forge -c https://software.repos.intel.com/python/conda -c nodefaults --override-channels
   conda activate dpbench

2. Run specific benchmark, e.g. black_scholes

   dpbench -b black_scholes run

Build from source (for development)

1. Clone the repository

   git clone https://github.com/IntelPython/dpbench
   cd dpbench

2. Setting up conda environment and installing dependencies:

   conda env create -n dpbench -f ./environments/conda.yml

   If you want to build sycl benchmarks as well:

   conda env create -n dpbench -f ./environments/conda-linux-sycl.yml

3. Build DPBench

   pip install --no-index --no-deps --no-build-isolation -e . -v

   Alternatively you can build it with setup.py, but pip version is preferable:

   python setup.py develop

   For sycl build use:

   CC=icx CXX=icpx DPBENCH_SYCL=1 pip install --no-index --no-deps --no-build-isolation -e . -v

   or

   CC=icx CXX=icpx DPBENCH_SYCL=1 python setup.py develop

4. Run specific benchmark, e.g. black_scholes

   dpbench -b black_scholes run

Usage

1. Run all benchmarks

   dpbench -a run

2. Generate report

   dpbench report

3. Device Customization

   If a framework is SYCL based, an extra configuration option sycl_device may be set in the framework config file or by passing --sycl-device argument to dpbench run to control what device the framework uses for execution. The sycl_device value should be a legal SYCL device filter string. The dpcpp, dpnp, and numba_dpex frameworks support the sycl_device option.

   Here is an example:

   dpbench -b black_scholes -i dpnp run --sycl-device=level_zero:gpu:0

4. All available options are available using dpbench --help and dpbench <command> --help:

   usage: dpbench [-h] [-b [BENCHMARKS]] [-i [IMPLEMENTATIONS]] [-a | --all-implementations | --no-all-implementations] [--version] [-r [RUN_ID]] [--last-run | --no-last-run] [-d [RESULTS_DB]]
              [--log-level [{critical,fatal,error,warning,info,debug}]]
              {run,report,config} ...
   
   positional arguments:
   {run,report,config}
   
   options:
   -h, --help            show this help message and exit
   -b [BENCHMARKS], --benchmarks [BENCHMARKS]
                           Comma separated list of benchmarks. Leave empty to load all benchmarks.
   -i [IMPLEMENTATIONS], --implementations [IMPLEMENTATIONS]
                           Comma separated list of implementations. Use --all-implementations to load all available implementations.
   -a, --all-implementations, --no-all-implementations
                           If set, all available implementations will be loaded.
   --version             show program's version number and exit
   -r [RUN_ID], --run-id [RUN_ID]
                           run_id to perform actions on. Use --last-run to use latest available run, or leave empty to create new one.
   --last-run, --no-last-run
                           Sets run_id to the latest run_id from the database.
   -d [RESULTS_DB], --results-db [RESULTS_DB]
                           Path to a database to store results.
   --log-level [{critical,fatal,error,warning,info,debug}]
                           Log level.
   

   usage: dpbench run [-h] [-p [{S,M16Gb,M,L}]] [-s | --validate | --no-validate] [--dpbench | --no-dpbench] [--experimental-npbench | --no-experimental-npbench] [--experimental-polybench | --no-experimental-polybench]
                  [--experimental-rodinia | --no-experimental-rodinia] [-r [REPEAT]] [-t [TIMEOUT]] [--precision [{single,double}]] [--print-results | --no-print-results] [--save | --no-save] [--sycl-device [SYCL_DEVICE]]
                  [--skip-expected-failures | --no-skip-expected-failures]
   
   Subcommand to run benchmark executions.
   
   options:
   -h, --help            show this help message and exit
   -p [{S,M16Gb,M,L}], --preset [{S,M16Gb,M,L}]
                           Preset to use for benchmark execution.
   -s, --validate, --no-validate
                           Set if the validation will be run for each benchmark.
   --dpbench, --no-dpbench
                           Set if run dpbench benchmarks.
   --experimental-npbench, --no-experimental-npbench
                           Set if run npbench benchmarks.
   --experimental-polybench, --no-experimental-polybench
                           Set if run polybench benchmarks.
   --experimental-rodinia, --no-experimental-rodinia
                           Set if run rodinia benchmarks.
   -r [REPEAT], --repeat [REPEAT]
                           Number of repeats for each benchmark.
   -t [TIMEOUT], --timeout [TIMEOUT]
                           Timeout time in seconds for each benchmark execution.
   --precision [{single,double}]
                           Data precision to use for array initialization.
   --print-results, --no-print-results
                           Show the result summary or not
   --save, --no-save     Either to save execution into database.
   --sycl-device [SYCL_DEVICE]
                           Sycl device to overwrite for framework configurations.
   --skip-expected-failures, --no-skip-expected-failures
                           Either to save execution into database.
   

   usage: dpbench report [--comparisons [COMPARISON_PAIRS]] [--csv]
   
   Subcommand to generate a summary report from the local DB
   
   options:
   -c, --comparisons [COMPARISON_PAIRS]
                           Comma separated list of implementation pairs to be compared
   --csv
                           Sets the general summary report to output in CSV format (default: False)
   

Performance Measurement

For each benchmark, we measure the execution time of the computationally intesive part, but not the intialization or shutdown. We provide three inputs (a.k.a presets) for each benchmark.

• S - Minimal input to verify that programs are executable
• M - Medium-sized input for performance measurements on client devices
• L - Large-sized input for performance measurements on servers

As a rough guideline for selecting input sizes, S inputs need to be small enough for python and numpy implementations to execute in <100ms. M and L inputs need to be large enough to obtain useful performance insights on client and servers devices, respectively. Also, note that the python and numpy implementations are not expected to work with M and L inputs.