Collective message failure with PSM2 #60
Description
Hi,
for several weeks I try to understand a problem (wrong behavior) with fortran MPI_ALLTOALLW calls. The problem only occur on a Debian supercomputer using this opa-psm2 library for it's omni-path architecture. I, and 2 OpenMPI developpers, have tested many other achitectures (intel or amd cpu, with ethernet, omni-path or infiniband network and running RedHat or Suse OS. The problem do not occur in any of these tests. More over, if on the debian computer I build OpenMPI using --without-psm2 flag the problem do not occur but omni-path performances are not reached.
I'm building OpenMPI 4.0.5 with gcc 6.3 or gcc 10.2 (same behavior)
Please find in attachement a really small test case showing the problem. If all runs fine it prints " Test pass!" else it shows the wrong values and calls mpi_abort().
To run this test:
- make
- mpirun -np 4 ./test_layout_array
Patrick