Broken external PRRTE compile

[rhc54]

Thank you for taking the time to submit an issue!

Background information

What version of Open MPI are you using? (e.g., v4.1.6, v5.0.1, git branch name and hash, etc.)

Head of main branch

Describe how Open MPI was installed (e.g., from a source/distribution tarball, from a git clone, from an operating system distribution package, etc.)

Git clone

If you are building/installing from a git clone, please copy-n-paste the output from git submodule status.

Please describe the system on which you are running

• Operating system/version: Ubuntu
• Computer hardware: container
• Network type: TCP

---

Details of the problem

Building OMPI with an external PRRTE, I'm hitting the following error:

schizo-ompi-rst-content/schizo-ompi-cli.rstxt:387: CRITICAL: Problems with "include" directive path:
InputError: [Errno 2] No such file or directory: 'prrte-rst-content/cli-no-app-prefix.rst'. [docutils]
make[1]: *** [Makefile:2743: _build/man/ompi-wrapper-compiler.1] Error 1
make: *** [Makefile:1514: all-recursive] Error 1

Note that the internal PRRTE submodule does not have the referenced file. However, neither does the internal copy of schizo-ompi-cli.rxtxt contain that reference.

However, the external PRRTE has both the reference in the rxtxt file, and the cli snippet.

So it appears that your man page build script is getting confused and trying to combine the external schizo-ompi-cli.rstxt with the internal docs directory content. This is going to fail more often than not as PRRTE continues to evolve.

[jsquyres]

I'm unable to reproduce. Can you send me a recipe?

Here's what I tried:

• Get a fresh openpmix clone
• Make a python venv, pip install -r docs/requirements.txt (in pmix clone)
• ./autogen.pl |& tee auto.out
• ./configure --enable-sphinx --with-hwloc=/opt/homebrew --with-libevent=/opt/homebrew --prefix=$bogus |& tee config.out
• make -j 32 |& tee make.out
• make install > /dev/null

Repeat essentially the same thing for prrte:

• Get a fresh prrte clone
• ./autogen.pl |& tee auto.out
• ./configure --enable-sphinx --with-hwloc=/opt/homebrew --with-libevent=/opt/homebrew --with-pmix=$bogus --prefix=$bogus |& tee config.out
• make -j 32 |& tee make.out
• make install > /dev/null

Repeat essentially the same thing for ompi:

• Get a fresh ompi clone
• ./autogen.pl |& tee auto.out
• ./configure --enable-sphinx --with-hwloc=/opt/homebrew --with-libevent=/opt/homebrew --with-pmix=$bogus --with-prrte=$bogus --prefix=$bogus |& tee config.out

I did not do a full ompi build, though, because a compile error occurs in runtime/prte_init.c. Instead, I just did

• cd docs
• make

This completed successfully.

Should I be doing something else?

Or, I wonder if there's some kind of dependency mismatch that's occurring in a non-clean tree that can cause something to not get built or generated properly...?

[rhc54]

Well, I got a brand new OMPI clone just now (using my Ubuntu-latest container). Have my external versions of PMIx and PRRTE (both head of current master branches) installed, which means PRRTE has the updated rstxt file and accompanying snippets. I then just do:

./autogen.pl > /dev/null && \
./configure --prefix=${MPI_ROOT} \
            --with-hwloc=${HWLOC_INSTALL_PATH} \
            --with-libevent=${LIBEVENT_INSTALL_PATH} \
            --with-pmix=${PMIX_ROOT} \
            --with-prrte=${PRRTE_ROOT} \
            --enable-mpirun-prefix-by-default \
	    --enable-debug > /dev/null && \
make -j 10 > /dev/null && \
make -j 10 install > /dev/null

which is a script I've been using to build OMPI with external support for a long time.

FWIW: Seeing a bunch of these:

configure.ac:114: warning: AC_ARG_PROGRAM was called before AC_CANONICAL_TARGET
../../lib/autoconf/general.m4:1844: AC_CANONICAL_TARGET is expanded from...
configure.ac:114: the top level

during the configure stage. Probably unrelated - just noting them.

When it tries to run the make step, I hit:

.....
../../../../ompi/mca/part/persist/part_persist.h: In function ‘mca_part_persist_psend_init’:
../../../../ompi/include/mpi.h:555:32: warning: operand of ‘?:’ changes signedness from ‘int’ to ‘unsigned int’ due to unsignedness of other operand [-Wsign-compare]
  555 | #define MPI_UNDEFINED          -32766                  /* undefined stuff */
      |                                ^~~~~~
../../../../ompi/mca/part/persist/part_persist.h:445:48: note: in expansion of macro ‘MPI_UNDEFINED’
  445 |     dt_size = (dt_size_ > (size_t) UINT_MAX) ? MPI_UNDEFINED : (uint32_t) dt_size_;
      |                                                ^~~~~~~~~~~~~

schizo-ompi-rst-content/schizo-ompi-cli.rstxt:387: CRITICAL: Problems with "include" directive path:
InputError: [Errno 2] No such file or directory: 'prrte-rst-content/cli-no-app-prefix.rst'. [docutils]
make[1]: *** [Makefile:2745: _build/man/ompi-wrapper-compiler.1] Error 1
make: *** [Makefile:1516: all-recursive] Error 1

which comes out after the MPI layer finishes compiling.

[rhc54]

I did not do a full ompi build, though, because a compile error occurs in runtime/prte_init.c

This is disturbing - it means that you are trying to compile the internal PRRTE using the external PMIx, which cannot work because (as we have discussed) your internal PRRTE is pointing to your OMPI fork, which is known to be broken and cannot support PMIx master.

Question is: why are you trying to compile the internal PRRTE when we are using the external one? I grok the need to run configure through it - but why the compile??

[jsquyres]

This is disturbing - it means that you are trying to compile the internal PRRTE using the external PMIx, which cannot work because (as we have discussed) your internal PRRTE is pointing to your OMPI fork, which is known to be broken and cannot support PMIx master.

Turns out that this was because I was using --with-prte instead of --with-prrte 🤬

[jsquyres]

Fixed by openpmix/prrte#2209.

[rhc54]

Sorry, but that isn't the correct fix - the referenced file does indeed exist in upstream PRRTE master and v3.0 release branch.

[jsquyres]

Just to close the loop on this issue: the above-cited PR was updated with the correct fix and merged. Nothing needed over here in OMPI.