Bring up a basic GitHub CI #40
Comments
anyzelman
pushed a commit
that referenced
this issue
Dec 2, 2024
This MR makes all LPF functional tests use GoogleTest. The main changes for CMake are: - reorganize CMake to use GTest CMake package for compilation, and gtest_add_tests clause to define Gtests; - remove all use of the run.sh script; - make explicit the list of tests and do not use file(GLOB, ...) clauses to look for all files in test directories, as recommended by CMakesource files: The main changes for source files are: - every single test file needed to be modified to not include internal Test.h but gtest.h, and to use the assert/equal clauses of GoogleTest - all death tests are slightly modified now to expect FAIL() after the expected fail condition - some modernization of C to C++ was needed for a number of files, incl. use of new/delete instead of malloc/free, or string manipulation instead of C char * manipulation A new Python wrapper script test_launcher.py is being employed now for more flexible checking of return codes, which is needed for death tests and normal tests alike. The expected return codes are still parsed from scanning the source files as before. For the complete history, see GitHub PR #26 . This MR also resolves bugs that have appeared on more modern cluster deployments compared to the previous main tag of LPF. This is hence a priority MR to base all further extensions and evolutions of LPF on. This MR was tested successfully for: - x86_64, CentOS Stream 9, GCC 11.5.0, MPICH 4.1.1 (caveat 2) - x86_64, CentOS Stream 9, GCC 11.5.0, OpenMPI 4.1.1 (caveats 1, 2, and 5) - x86_64, Fedora, GCC 9.3.1, MPICH 3.3.2 (caveat 2) - x86_64, Ubuntu 23.04, GCC 13.1.0, OpenMPI 4.1.6 (caveat 5) - x86_64, openEuler 20.03 LTS, GCC 7.3, MPICH 3.2.1 - x86_64, Ubuntu 22.04 LTS, GCC 11.4.0, MPICH 4.3.0b1 (caveat 3, transient) - ARM, Ubuntu 22.04 LTS, GCC 11.4.0, MPICH 4.3.0b1 (caveats 3 & 4) These are under different variations of calls to `bootstrap.sh` to test all the above-described changes. The following caveats apply: 1. with OpenMPI and default arguments to its `mpirun`, oversubscription to larger process counts may be limited, causing some tests to fail. Issue #37 has been raised to pass the appropriate flags to the OpenMPI `mpirun` during testing; 2. `make install` post-install checks fail (as they should) when: 1) tests are disabled, 2) ibverbs dev libraries are present, but 3) no IB card is present. On the machines with IB we tested on, all post-install checks succeed. 3. The hybrid backend on MPICH 4.3.0b1 does not pass the debug layer tests due to a segfault in MPICH's MPI_Abort. Issue #41 has been raised to track this. This error presently only is consistently reproducible on our cluster's ARM nodes and transient on our cluster's x86_64 nodes. They have not been observed outside our cluster. 4. With MPICH 4.3.0b1, we furthermore hit issues #43 and #44. Also these issues are only reproducible on ARM nodes. 5. With OpenMPI 4.1.6 and OpenMPI 4.1.1 on x86_64 we hit issue #45 6. At present, we have no CI flow that can run all build variations relevant to this MR. Issues #40 has been raised to (partially) address as well as to track this issue.
KADichev
added a commit
that referenced
this issue
Dec 19, 2024
This MR makes all LPF functional tests use GoogleTest. The main changes for CMake are: - reorganize CMake to use GTest CMake package for compilation, and gtest_add_tests clause to define Gtests; - remove all use of the run.sh script; - make explicit the list of tests and do not use file(GLOB, ...) clauses to look for all files in test directories, as recommended by CMakesource files: The main changes for source files are: - every single test file needed to be modified to not include internal Test.h but gtest.h, and to use the assert/equal clauses of GoogleTest - all death tests are slightly modified now to expect FAIL() after the expected fail condition - some modernization of C to C++ was needed for a number of files, incl. use of new/delete instead of malloc/free, or string manipulation instead of C char * manipulation A new Python wrapper script test_launcher.py is being employed now for more flexible checking of return codes, which is needed for death tests and normal tests alike. The expected return codes are still parsed from scanning the source files as before. For the complete history, see GitHub PR #26 . This MR also resolves bugs that have appeared on more modern cluster deployments compared to the previous main tag of LPF. This is hence a priority MR to base all further extensions and evolutions of LPF on. This MR was tested successfully for: - x86_64, CentOS Stream 9, GCC 11.5.0, MPICH 4.1.1 (caveat 2) - x86_64, CentOS Stream 9, GCC 11.5.0, OpenMPI 4.1.1 (caveats 1, 2, and 5) - x86_64, Fedora, GCC 9.3.1, MPICH 3.3.2 (caveat 2) - x86_64, Ubuntu 23.04, GCC 13.1.0, OpenMPI 4.1.6 (caveat 5) - x86_64, openEuler 20.03 LTS, GCC 7.3, MPICH 3.2.1 - x86_64, Ubuntu 22.04 LTS, GCC 11.4.0, MPICH 4.3.0b1 (caveat 3, transient) - ARM, Ubuntu 22.04 LTS, GCC 11.4.0, MPICH 4.3.0b1 (caveats 3 & 4) These are under different variations of calls to `bootstrap.sh` to test all the above-described changes. The following caveats apply: 1. with OpenMPI and default arguments to its `mpirun`, oversubscription to larger process counts may be limited, causing some tests to fail. Issue #37 has been raised to pass the appropriate flags to the OpenMPI `mpirun` during testing; 2. `make install` post-install checks fail (as they should) when: 1) tests are disabled, 2) ibverbs dev libraries are present, but 3) no IB card is present. On the machines with IB we tested on, all post-install checks succeed. 3. The hybrid backend on MPICH 4.3.0b1 does not pass the debug layer tests due to a segfault in MPICH's MPI_Abort. Issue #41 has been raised to track this. This error presently only is consistently reproducible on our cluster's ARM nodes and transient on our cluster's x86_64 nodes. They have not been observed outside our cluster. 4. With MPICH 4.3.0b1, we furthermore hit issues #43 and #44. Also these issues are only reproducible on ARM nodes. 5. With OpenMPI 4.1.6 and OpenMPI 4.1.1 on x86_64 we hit issue #45 6. At present, we have no CI flow that can run all build variations relevant to this MR. Issues #40 has been raised to (partially) address as well as to track this issue.
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
This CI would test primarily the pthread engine, and optionally can test if all other engines build. More exhaustive testing can only run on our internal infrastructure using our internal CI approach, which at present tests the mpimsg, mpirma, and ibverbs engines thoroughly using the functional test suite introduced with MR #26
The text was updated successfully, but these errors were encountered: