Compiling with MPI and Intel 19.0

frodo · Post by **frodo** » Sun Mar 03, 2019 11:13 pm

Hello,

I'm having trouble running the abinit (8.10.2) executable after compiling with the Intel 19.0 compilers and with MPI enabled (64 bit intel).

If I compile with either the gnu tools (gcc, gfortran 7.3.0) or the Intel tools (icc, ifort), and without MPI enabled, make check shows all fast tests succeed.

If I compile with the gnu tools (gcc, gfortran) and MPI enabled, make check runs abinit without mpirun and all fast tests succeed.

If I compile with the Intel tools (mpiicc, mpiifort) and MPI enabled, make check still runs abinit without mpirun and all fast tests fail with the following error:

forrtl: severe (24): end-of-file during read, unit 5, file /proc/19230/fd/0
Image PC Routine Line Source
libifcoremt.so.5 000014A9FFCA97B6 for__io_return Unknown Unknown
libifcoremt.so.5 000014A9FFCE7C00 for_read_seq_fmt Unknown Unknown
abinit 00000000015B9312 Unknown Unknown Unknown
abinit 0000000000409DEF Unknown Unknown Unknown
abinit 0000000000409B22 Unknown Unknown Unknown
libc-2.27.so 000014A9FD6E1B97 __libc_start_main Unknown Unknown
abinit 0000000000409A0A Unknown Unknown Unknown

If I compile with the Intel tools and MPI enabled and run "runtests.py fast --force-mpirun", then abinit is run with "mpirun -np 1" and all tests succeed.

My understanding is that executables compiled with mpiifort must be run with mpirun even if np=1.

It seems that runtests.py tries to run serial tests without mpirun. This seems to work when abinit is compiled with the gnu tools, but not when compiled with the intel tools.

Is this a known difference in behavior for mpi executables compiled with the gnu tools vs the intel tools? If so, why doesn't runtests.py use mpirun for the intel compiled executable even for serial tests. Or am I doing something wrong?

Also, when compiling with Intel and MPI, setting with_mpi_incs and with_mpi_libs has no effect. They are not used in the compilation. I assume this is because mpiifort is a wrapper that is supplying these. Is that correct?

I am using Intel Parallel Studio XE and source psxevars.sh to set the environment before compiling/running with intel.

Thanks for any suggestions.

ebousquet · Post by **ebousquet** » Thu Mar 07, 2019 5:24 pm

Dear Frodo,
You can type "./runtests.py -h" to see all the available options.
For the parallel tests, you have the option runtests.py paral -n XX with XX the number of CPU on which you want to run the mpi. See the other options to force some executable, etc.
Best wishes,
Eric

frodo · Post by **frodo** » Thu Mar 07, 2019 7:52 pm

Hi Eric,

Thank you for the reply.

Yes, I know about runtests.py and use it all the time.

My question was why running "make check" fails with the indicated error when abinit is compiled with Intel 19.0 (mpiifort), where "make check" does not fail when abinit is compiled with gnu (mpif90). In both cases, "make check" invokes runtests.py to run abinit directly, i.e., without mpirun. This works for the gnu compiled version but fails with the error I indicated for the Intel compiled version.

Let me try to make what I am saying more clear:

Compiling with GNU:

I compile abinit with the following config.ac: