Using a fresh installation of OpenMPI on an up-to-date Fedora 40 machine gives mpirun version 5.0.2
which is less tolerant than version 4.1.7a1 (from the Nvidia HPC SDK 2024) towards being run without command line options.
It seems that an binary compiled by mpicc which takes no options will work the same under both versions, e.g., mpirun ./binary.out.
But if the binary takes optional input, and none is provided to mpirun, then problems can occur -- namely,
mpirun ./compiled_instrument.out --seed=1 ...
raises an error because mpirun does not recognize the option --seed.
If the binary options are separated from the mpirun call, e.g.,
mpirun ./compiled_instrument.out -- --seed=1 ...
then everything works as expected, but inserting the separator does not fit easily into the command-building logic:
An alternative solution may be generally workable: any* mpirun option provided before the name of the binary allows for normal operation.
This has only been tested with -n (equivalent to -np for OpenMPI, at least); but may work with other flags.
Interestingly, the explicit setting of -n 0 appears to be the same behavior as not providing -n -- so this could be an appropriate flag to always set for mccode_antlr.compiler.c:run_compiled_instrument
Using a fresh installation of
OpenMPIon an up-to-date Fedora 40 machine givesmpirunversion5.0.2which is less tolerant than version4.1.7a1(from the Nvidia HPC SDK 2024) towards being run without command line options.It seems that an binary compiled by mpicc which takes no options will work the same under both versions, e.g.,
mpirun ./binary.out. But if the binary takes optional input, and none is provided tompirun, then problems can occur -- namely,raises an error because
mpirundoes not recognize the option--seed.If the binary options are separated from the
mpiruncall, e.g.,then everything works as expected, but inserting the separator does not fit easily into the command-building logic:
https://github.com/McStasMcXtrace/mccode-antlr/blob/20979a3708e5c7dab088e53b9791f9f0929625aa/mccode_antlr/compiler/c.py#L170-L177
An alternative solution may be generally workable: any*
mpirunoption provided before the name of the binary allows for normal operation. This has only been tested with-n(equivalent to-npfor OpenMPI, at least); but may work with other flags.Interestingly, the explicit setting of
-n 0appears to be the same behavior as not providing-n-- so this could be an appropriate flag to always set formccode_antlr.compiler.c:run_compiled_instrument