dlr-sc / googletest_mpi Goto Github PK

If we disable MPI support (with -DGTEST_HAS_MPI=0) then the new *_MPI macros are not defined causing compiler errors in code using them. Instead, they should be defined to their non-mpi counterparts to make code using the *_MPI macros behave like expected (as serial code) if compiled without MPI support.

The new github citation feature offers a simple way to cite software repositories.
We should add a citation file to this repo.

Describe the bug

internal tests are not working properly with MPI support.

Steps to reproduce the bug

cmake with

 BUILD_GMOCK                      ON
 BUILD_SHARED_LIBS                OFF
 CMAKE_BUILD_TYPE
 CMAKE_INSTALL_PREFIX             /usr/local
 INSTALL_GTEST                    ON
 gmock_build_tests                OFF
 gtest_build_samples              ON
 gtest_build_tests                ON
 gtest_disable_mpi                OFF
 gtest_disable_pthreads           OFF
 gtest_force_shared_crt           OFF
 gtest_hide_internal_symbols      OFF

make
make test

Does the bug persist in the most recent commit?

We recommend using the latest commit in the master branch in your projects.

What operating system and version are you using?

Ubuntu

What compiler and version are you using?

mpicc -v
mpicc for MPICH version 4.0.2
gcc version 12.1.0 (Spack GCC) 

gcc -v
gcc version 12.1.0 (Spack GCC) 

What build system are you using?

cmake --version
cmake version 3.23.1

Additional context

LastTest.log
LastTestsFailed.log
CTestCostData.txt

If an error occurs on rank k!=0, the message is lost because there is only a reduction on the test result to verify the test failed. On Jenkins we therefore see only the file and line where the test failed, not the details.

See also the corresponding PHIST issue:

https://bitbucket.org/essex/phist/issues/174/missing-test-output-in-xml-files

If no --random_seed is given, it is created from the current time, which is typically different for each MPI process, hence each one picks a different order of tests.
This was already fixed in essex/phist, commit 3d8c3ce0f9e52

The MPI extensions are based on an old version of Google Test (v. 1.8.0).

We should rebase it to the current version before integrating it into master.

1.10. introduced the functions TestPartSkipped and TestResult::Skipped to gtest.cc.
Do we need to add anything here to support MPI?
Such as: only skip a test if all processes skip the test?

The stress test possibly uses multiple pthreads. MPI collectives are possibly not correctly matched. Not sure if this is fixable...

One idea would be to use an own communicator for all threads, but I don't know if this is possible...

After doing a fresh clone from the repository I tried to build the package with CMake.
The configure step works fine but the build crashes due to compilation errors.
I am guessing it is treating warnings as errors.

I tried with main branch and v1.10.x branch.

I also tried with the latest tarball release but it still does not work.

The bug can be reproduced by executing the 'make' command.

I am working Ubuntu with gcc 11.4.0

Do I need to use any special flags for compiling ?

I am attaching a log file for reference.

log.txt

Some users may want to call the new collective macros EXPECT/ASSERT_*_MPI on a subset of COMM_WORLD and it would thus be useful to include an option for this.

Why did we not include it from the start?

Because this would alter the number of arguments of the macros and it would be tedious to allways carry the mpi communicator around.

Maybe there was another reason as well that we could not include this feature, but i cannot remember. We should nevertheless investigate this, if it becomes an issue.

Update to gtest version 1.10

• Merge with gtest 1.10 (in progress)
• Adapt code to new 1.10
  • XML representation of a UnitTest (gtest.cc:3787 gtest_xml_output_unittest.py)
  • any other issues
• Test whether everything is working