Wrong parsing of command line arguments with mpirun

[angainor]

In 3.1.1 and 3.1.3 I get the following warning starting MPI applications that take the --am argument:

$ mpirun -H c11-3 ls --am aa bb
--------------------------------------------------------------------------
Process 130051 Unable to locate the variable file "aa" in the following search path:
   /cluster/software/OpenMPI/3.1.1-GCC-7.3.0-2.30/share/openmpi/amca-param-sets:/cluster/projects/nn9999k/
--------------------------------------------------------------------------
--------------------------------------------------------------------------
You're using the --am option. Please be advised that this option is deprecated;
you should use --tune instead. The --tune option allows one to specify mca
parameters as well as environment variables from within a file using the same
command line syntax (e.g. -mca var val -mca var "val" -x var=val -x var).
The --am option will be removed in a future release.
--------------------------------------------------------------------------

Of course ls doesn't take this argument, but that is only to illustrate the problem. It seems that mpirun is trying to analyze the application's arguments, and beyond the application name. Note that the above message is only displayed when the -H switch is present, and when there are at least 2 arguments after --am. Pretty funny, and only annoying: I do not see any other side effects except for the noise.

[jsquyres]

Hah, that's a fun bug!

Yes, it looks like something in the command line parsing is reading --am before it discards everything to the right (from OMPI's processing, anyway) of the command to be executed. I simplified your example down to:

$ mpirun echo --am a b

A similar error happens on master HEAD, so this is likely behavior that has escaped attention for quite a while.

[SimonR1]

Dear everyone,
I have the same problem and would be very grateful for any help or advice.

[awlauria]

I can take a look and see what's going on.

[rhc54]

FWIW: this is fixed in the RTE for OMPI v5. Problem is in not processing the cmd line before searching for mca params (including --am), thereby not knowing where the executable starts.

[awlauria]

Thanks @rhc54 .

@SimonR1 this will be fixed in the next major release - OMPI v5.0 via prrte.

[jsquyres]

I confirm: this is fixed on Open MPI master / what will eventually become v5.0.

I'm kinda guessing that this is not going to be fixed in prior releases, mainly due to lack of resources (i.e., no one has submitted a fix in 20+ months).

[loganharbour]

For those that have also scrambled across this like myself, it appears this is still not fully resolved:

[harblh@br006: ~ ]
$ mpiexec hostname
br006
br006
br010
br010
[harblh@br006: ~ ]
$ mpiexec hostname --help
mpiexec (Open MPI) 5.0.5

Usage: mpiexec [OPTION]...

See the mpirun(1) man page or HTML help for a detailed list of command
line options that are available.

Report bugs to https://www.open-mpi.org/community/help/

In the second case, --help is being passed as an argument to mpiexec

[rhc54]

It is (and has been for some time) fixed on the main branch, but there is a missing cherry-pick into the release branch. Have that in-process now.

[loganharbour]

It is (and has been for some time) fixed on the main branch, but there is a missing cherry-pick into the release branch. Have that in-process now.

Great. Thanks for addressing this quickly!

[everythingfunctional]

I also just encountered this.

[rhc54]

Ok....but as stated, it has been fixed and just awaits another release.

[loganharbour]

To be clear - the fix that @rhc54 was referencing did work for me. Just hasn't made it into a release.