Examples of parallel compiler speedup from Python asyncio-subprocess, ThreadPoolExecutor, and ProcessPoolExecutor. We observe asyncio is faster than ThreadPoolExecutor, which is faster than ProcessPoolExecutor.
These results may be useful for Meson maintainers and CMake maintainers to consider making some CMake internals parallelized.
The benchmarks below run compilers many times as a real-world example of performance improvements via different techniques in Python. These same relative speedups may hold similarly true in other coding languages.
The benchmark is of compiling ~ 15 small test Fortran source files as defined in fortran_test_generator() "Nrun" times.
For example, the benchmarks below use the respective compilers 8 * 15 = 120 times to help stabilize results.
The test Fortran source files are written once in a tempfile.TemporaryDirectory outside the benchmark loop.
Linux workstation:
Gfortran:
$ python compiler_checks.py -n 8 gfortran
Python 3.11.7 (main, Dec 15 2023, 18:12:31) [GCC 11.2.0] linux
0.174 seconds: asyncio: GNU Fortran (GCC) 11.4.1 20231218 (Red Hat 11.4.1-3)
0.427 seconds: ThreadPoolExecutor: GNU Fortran (GCC) 11.4.1 20231218 (Red Hat 11.4.1-3)
1.065 seconds: ProcessPoolExecutor: GNU Fortran (GCC) 11.4.1 20231218 (Red Hat 11.4.1-3)
3.630 seconds: serial: GNU Fortran (GCC) 11.4.1 20231218 (Red Hat 11.4.1-3)Intel oneAPI
$ python compiler_checks.py -n 8 ifx
Python 3.11.7 (main, Dec 15 2023, 18:12:31) [GCC 11.2.0] linux
0.306 seconds: asyncio: ifx (IFX) 2024.0.0 20231017
0.876 seconds: ThreadPoolExecutor: ifx (IFX) 2024.0.0 20231017
1.551 seconds: ProcessPoolExecutor: ifx (IFX) 2024.0.0 20231017
6.870 seconds: serial: ifx (IFX) 2024.0.0 20231017NVHPC
$ python compiler_checks.py -n 8 nvfortran
Python 3.11.7 (main, Dec 15 2023, 18:12:31) [GCC 11.2.0] linux
0.320 seconds: asyncio: nvfortran 23.11-0
1.056 seconds: ThreadPoolExecutor: nvfortran 23.11-0
1.653 seconds: ProcessPoolExecutor: nvfortran 23.11-0
7.036 seconds: serial: nvfortran 23.11-0AMD AOCC
$ python compiler_checks.py -n 8 flang
Python 3.11.7 (main, Dec 15 2023, 18:12:31) [GCC 11.2.0] linux
0.284 seconds: asyncio: flang AMD clang version 16.0.3 (CLANG: AOCC_4.1.0-Build#270 2023_07_10)
1.009 seconds: ThreadPoolExecutor: flang AMD clang version 16.0.3 (CLANG: AOCC_4.1.0-Build#270 2023_07_10)
1.638 seconds: ProcessPoolExecutor: flang AMD clang version 16.0.3 (CLANG: AOCC_4.1.0-Build#270 2023_07_10)
8.014 seconds: serial: flang AMD clang version 16.0.3 (CLANG: AOCC_4.1.0-Build#270 2023_07_10)macOS Apple Silicon:
% python compiler_checks.py -n 8
Python 3.11.7 (main, Dec 15 2023, 12:09:56) [Clang 14.0.6 ] darwin
1.282 seconds: asyncio: GNU Fortran (Homebrew GCC 13.2.0) 13.2.0
1.329 seconds: ThreadPoolExecutor: GNU Fortran (Homebrew GCC 13.2.0) 13.2.0
2.330 seconds: ProcessPoolExecutor: GNU Fortran (Homebrew GCC 13.2.0) 13.2.0
5.411 seconds: serial: GNU Fortran (Homebrew GCC 13.2.0) 13.2.0