Add OpenMPI host injection script #53
Conversation
|
This needs CI testing for two scenarios:
Both cases should verify that the expected libraries are being picked up and run an OSU benchmark to verify that executables do not crash |
Co-authored-by: ocaisa <[email protected]>
|
@agimenog That test passes, well done! Now add an additional test that injects OpenMPI from the host operating system (installed via |
| if [ -n "$(ls -A ${host_injection_mpi_path})" ]; then | ||
| echo "MPI was already injected" | ||
| if ${FORCE}; then | ||
| echo "Forcing new MPI injection" |
There was a problem hiding this comment.
I would move the existing directory to a time stamped version mv ${host_injection_mpi_path} ${host_injection_mpi_path}_<timestamp> in this situation
There was a problem hiding this comment.
Changes for this already commited
|
Hi @ocaisa, I will work on that. Just to clarify, should it be on a different .yml file or in the same one? |
Same one, just another step |
| mkdir ${temp_inject_path} | ||
|
|
||
| # Get all library files from openmpi dir | ||
| find ${MPI_PATH} -type f -name "*.so*" -exec cp {} ${temp_inject_path} \; |
There was a problem hiding this comment.
Hmm, this is probably not going to play nice with an OS installed version of OpenMPI as the libraries may well be in the same place as all the other libraries on the system. We should only search for the libraries that make up the compatibility libraries (see https://docs.open-mpi.org/en/v5.0.x/version-numbering.html#shared-library-version-number):
libmpi
libmpi_mpifh
libmpi_usempi_tkr
libmpi_usempi_ignore_tkr
libmpi_usempif08
libmpi_cxx
libmpi_java
liboshmem
There was a problem hiding this comment.
I also wonder what will happen when it tries to load a provider...
There was a problem hiding this comment.
I think we also need libmca*.so* and mca_*.so (see https://docs.open-mpi.org/en/main/mca.html)
There was a problem hiding this comment.
Thanks for the documentation! I also think this will fail as it was originally thinked to inject all libs from one unique directory.
I will take a look on that.
There was a problem hiding this comment.
The provider thing is very tricky, I expect that they would need to be in the same relative location to the libraries
There was a problem hiding this comment.
I have no idea how to inspect or test that
Hi,
Reubicating this PR following the previous one EESSI/software-layer#1085
Tested the functionality and checked that the MPI injection now also works for OpenMPI/5.0.7-GCC-14.2.0.
We added some warnings and instructions to follow as newer MPI implementations include libcuda.so which is difficult to locate depending on the system.
Also, where would be the best place to document how to use the script?
Regards,
Arturo.