Hi, Apologies if this has been resolved but I couldn't spot a definitive fix in this thread:
"Fortran and OpenMPI 1.8.3 compiled with Intel-15 does nothing silently" http://www.open-mpi.org/community/lists/users/2014/11/25823.php and the related thread on Intel's forums: https://software.intel.com/en-us/forums/topic/540673 I recently compiled OpenMPI 1.8.7 with default options and Intel 15.0.3.187: $ ./configure --prefix=/common/debian/8.1/Compiler/intel/2015.3/openmpi/openmpi-1.8.7 CC=icc CXX=icpc F77=ifort FC=ifort $ make all $ make install A simple test program: program test use mpi integer :: ierr real(8) :: a call mpi_init(ierr) call random_number(a) write(*,*)"hello" call mpi_finalize(ierr) end program test fails if compiled with OpenMP (even if OpenMP is not used): $ mpif90 test.f90 -openmp $ ./a.out Segmentation fault Running with more cores causes a silent failure: $ mpirun -np 4 ./a.out <no output> though an exit code of 139 is returned. The Intel forum thread alleges that this is (at least for MVAPICH2) because incorrect Intel runtime sources are included in an MPI library, which is backed up by looking at the OpenMPI and Intel libraries: Intel: $ nm libifcore.so.5 | grep for_rtl_init 0000000000050550 T for_rtl_init_ $ nm libifcoremt.so.5 | grep for_rtl_init 0000000000050ed0 T for_rtl_init_ OpenMPI: $ nm libmpi_usempif08.so | grep for_rtl_init 00000000000303f0 T for_rtl_init_ I appear to be able to work around this by forcing the use of shared Intel libraries: $ mpif90 test.f90 -openmp -shared-intel $ ./a.out hello $ mpirun -np 4 ./a.out hello hello hello hello Is there a way to resolve this in the OpenMPI build itself instead? Thanks, --James -- James Spencer http://www.cmth.ph.ic.ac.uk/people/j.spencer/ http://www.github.com/jsspencer
