v4.0.0: Check for ABI compatibility with v3.1.x

[jsquyres]

Per https://www.mail-archive.com/[email protected]/msg00015.html and webex discussion on 2018-07-17, @hppritcha and @gpaulsen will be conducting an ABI compatibility check between the v4.0.x branch (which will be created tomorrow) and the v3.1.x branch.

If possible, we would like to not change libmpi/liboshmem's .so major version for all the reasons cited in https://www.mail-archive.com/[email protected]/msg00015.html. We're pretty sure that this means increasing both c and a by 10.

Howard and Geoff will need to check the branches manually, and probably do a bunch of testing, such as:

1. Compile and install Open MPI v3.1.x into /path
2. Compile a bunch of tests against that Open MPI installation
3. rm -rf /path
4. Compile and install Open MPI v4.0.x (with appropriate c:r:a values and possibly --enable-mpi1-compat) into /path
5. Run the tests (without recompiling/relinking) and make sure they all work properly

[jsquyres]

FYI: @amckinstry

[amckinstry]

Thanks for the pointer; this plan makes sense to me. Postponing symbol versioning and adding it in v5.0 would be best.

Since symbols will be dropped in v5.0, a new major SOVERSION would be required.
From a Debian perspective, the next release is Q2/Q3 2019, with a transition freeze in January, so we'd go with V4.0, or if a new major SOVERSION is required, probably 3.x, as its unlikely to get a full transition through before January (transitions basically cause traffic jams in our release planning).
As soon as a test release of v4.0 is available to work with, I'm going to do an experimental package in Debian and test-build all our MPI packages against it. This will test to see if anyone is using the MPI-1 features or the CXX interface ; i'll disable them in the test build and only enable them if necessary, in prep for v5.0.

[gpaulsen]

I have completed this testing. It is important to acknowledge that testing alone can never fully guarantee ABI compatibility. This tested was done in combination with discussion in the weekly web-ex discussions with our core developers and deemed the gating factor if we can keep our .so versions compatible so as to NOT require users' to rebuild/link their mpi applications.

I built and ran the open-mpi ibm test suite (generally considered good coverage of the MPI API interface) with top of v3.1.x branch. I then deleted that OpenMPI v3.1.x install, and built Open MPI v4.0.x two time (I repeated this process both with and without the new configure flag: -enable-mpi1-compatibility). I then updated the variables:

OPAL_PREFIX
OPAL_LIBDIR
PATH
LD_LIBRARY_PATH

I was then able to RUN the binaries under each subdirectory of the ibm test suite.
A few tests failed both test runs due to not enough ranks (I used --np 4, along with vader btl). I also killed the int_overflow test in collectives subdir as that test is a resource HOG!. These failures ensured that I corectly saw what a failure looked like. I then checked return codes for all of the other tests, and they all correctly returned 0 return codes.

PASSED.

I will close this issue and create a PR of the VERSION file to v4.0.x that reflects NO need to require customers who built with v3.1.x to recompile/relink with v4.0.0.

[jsquyres]

Per 2018-09-11 webex, @gpaulsen says this is now complete / ok to close.