[Darshan-users] Darshan Intel MPI, SLURM, Fortran

[Harms, Kevin]

Alain,

  Can you reproduce the crash by just running the binary on the login node? If so, you can run the test under GDB and see if you can get a backtrace. How are you setting the LD_PRELOAD variable? srun --export=LD_PRELOAD=<path> ?

kevin


-----Original Message-----
From: <darshan-users-bounces at lists.mcs.anl.gov> on behalf of Alain REFLOCH <alain.refloch at onera.fr>
Date: Friday, April 28, 2017 at 8:03 AM
To: "darshan-users at lists.mcs.anl.gov" <darshan-users at lists.mcs.anl.gov>
Subject: [Darshan-users] Darshan Intel MPI, SLURM, Fortran

    
    
    
    I try my first test with Darshan (version darshan-3.1.4).
    No problem for install 
    I use Intel compiler v 15 and Intel MPI, only a message at the compilation :
    lib/darshan-mpiio.c(230): warning #147: declaration is incompatible with "int MPI_File_open(MPI_Comm={int}, const char *, int, MPI_Info={int}, MPI_File *)" (declared at line 192 of "/opt/software/common/intel/impi/5.1.3.258/include64/mpio.h")
    Resolve by put option : -DHAVE_MPIIO_CONST
    
    I
     test with the file io-sample.c in the directory darshan-3.1.4/darshan-test
    all is ok 
    
    For install I have do this :
    ./configure  --prefix=$HOME/Darshan --with-mem-align=8 --with-log-path=$SCRATCHDIR/Darshan --with-jobid-env=SLURM_JOB_ID --with-zlib=/opt/software/occigen/libraries/zlib/1.2.8/intel/17.0/nompi CC='mpiicc -DHAVE_MPIIO_CONST'
    
    I am on a Bull supercomputer with Haswell processor, with slum for the batch. I have use SLURM_JOB_ID or NONE
    It's OK, I have a file in the repository ....../year/months/days.
    
    After I try on my big program. All it's ok with no LD_PRELOAD, but with PRELOAD
    I have No message but the job is kill by SLURM without message. 
    The main is in Fortran,  call to mpi_init and finalize is in Fortran. 
    But in the code there is a part in C/C++ for the mesh file with MPI_IO (all parts
    open, write, read, close for this file are in C/C++), I am interested by the Darshan informations
    
    on this file (no on the ascii files open in Fortran). There is in the code a parallel partitionner
    for domains decompostion (ParMetis), there is no recompilation of this part.
    
    My question is : is Darshan ok for this configuration (main in Fortran, partial recompilation
    of the MPI_IO part write in C)
    
    bests regards
    Alain