Hi, I'm trying to deploy OpenMPI 4.0.5 on the university's supercomputer:
* Debian GNU/Linux 9 (stretch) * Intel Corporation Omni-Path HFI Silicon 100 Series [discrete] (rev 11) and for several days I have a bug (wrong results using MPI_AllToAllW) on this server when using OmniPath. Running 4 threads on a single node, using OpenMPI 4.0.5 built without omnipath support, the code is working: CC=$(which gcc) CXX=$(which g++) FC=$(which gfortran) ../configure --with-hwloc --enable-mpirun-prefix-by-default \ --prefix=/bettik/begou/OpenMPI405-noib --enable-mpi1-compatibility \ --enable-mpi-cxx --enable-cxx-exceptions --without-verbs --without-ofi --without-psm --without-psm2 --without-openib \ --without-slurm If I use omnipath, still with 4 threads on one node, the test-case does not work (incorrect results): CFLAGS="-O3 -march=native -mtune=native" CXXFLAGS="-O3 -march=native -mtune=native" FCFLAGS="-O3 -march=native -mtune=native" \ CC=$(which gcc) CXX=$(which g++) FC=$(which gfortran) ../configure --with-hwloc --enable-mpirun-prefix-by-default \ --prefix=/bettik/begou/OpenMPI405 --enable-mpi1-compatibility \ --enable-mpi-cxx --enable-cxx-exceptions --without-verbs I do not undestand what could be wrong as the code is running on many architecture with various interconnect and openMPI versions. Thanks for your suggestions. Patrick