#!/usr/bin/env bash

#SBATCH -J sette
#SBATCH -o sette.%j.out
#SBATCH -e sette.%j.err
#SBATCH --export=ALL
#SBATCH --parsable
#SBATCH --exclusive
#SBATCH -N 1

##SBATCH --time=00:30:00
#SBATCH --time=00:15:00

##SBATCH -p debug64
##SBATCH -A smer
##SBATCH --qos=debug

#SBATCH -p normal256
##SBATCH -A cmems
##SBATCH --qos=coper


# Test specific settings. Do not hand edit these lines; the fcm_job.sh script will set these
# (via sed operating on this template job file). 
#
  echo " ";
  export XIO_HOME="/home/ext/mr/smer/samsong/SRC/XIOS/xios-2.5/BEL_INTEL18_r1900"
  O_PER_NODE=32
  X_PER_NODE=8
  OCORES=NPROCS
  XCORES=NXIOPROCS
  if [ $OCORES -le 32 ] ; then O_PER_NODE=$OCORES; fi
  export SETTE_DIR=DEF_SETTE_DIR

###############################################################
#
#
# load sette functions (only post_test_tidyup needed)
#
  . ${SETTE_DIR}/all_functions.sh
###############################################################
#
# modules to load
module purge
module load intel/2018.5.274
module load intelmpi/2018.5.274
module load phdf5/1.10.5
module load netcdf_par/4.7.1

# Don't remove neither change the following line
# BODY
#
# Test specific settings. Do not hand edit these lines; the fcm_job.sh script will set these
# (via sed operating on this template job file). Note that the number of compute nodes required
# is also set by the fcm_job.sh on the PBS select header line above.
#
# These variables are needed by post_test_tidyup function in all_functions.sh
#
  export INPUT_DIR=DEF_INPUT_DIR
  export CONFIG_DIR=DEF_CONFIG_DIR
  export TOOLS_DIR=DEF_TOOLS_DIR
  export NEMO_VALIDATION_DIR=DEF_NEMO_VALIDATION
  export NEW_CONF=DEF_NEW_CONF
  export CMP_NAM=DEF_CMP_NAM
  export TEST_NAME=DEF_TEST_NAME
  export EXE_DIR=DEF_EXE_DIR
  ulimit -a
  ulimit -s unlimited
#
# end of set up
###############################################################
#
# change to the working directory 
#
  cd $EXE_DIR

  echo Running on host `hostname`
  echo Time is `date`
  echo Directory is `pwd`
# 
#  Run the parallel MPI executable 
#
  if [ $XCORES -gt 0 ]; then
#
#  Run MPMD case
#
     #XIOS will run on a separate node so will run in parallel queue
     if [ ! -f ./xios_server.exe ] && [ -f ${XIO_HOME}/bin/xios_server.exe ]; then
        cp ${XIO_HOME}/bin/xios_server.exe .
     fi
     if [ ! -f ./xios_server.exe ]; then
        echo "./xios_server.exe not found"
        echo "run aborted"
        exit
     fi

#    cat > mpmd.conf <<EOF
#0-$((OCORES-1)) ./nemo
#${OCORES}-39 ./xios_server.exe
#EOF
    cat > mpmd.conf <<EOF
-n ${OCORES} ./nemo
-n ${XCORES} ./xios_server.exe
EOF

#     echo time srun --cpu_bind=cores --mpi=pmi2 -m cyclic -n 40 --multi-prog ./mpmd.conf
#          time srun --cpu_bind=cores --mpi=pmi2 -m cyclic -n 40 --multi-prog ./mpmd.conf

      # Comm/Fabric
      # -----------
      export DAPL_ACK_RETRY=7
      export DAPL_ACK_TIMER=20
      export DAPL_IB_SL=0
      export DAPL_UCM_CQ_SIZE=8192
      export DAPL_UCM_DREQ_RETRY=4
      export DAPL_UCM_QP_SIZE=8192
      export DAPL_UCM_REP_TIME=8000
      export DAPL_UCM_RTU_TIME=8000
      export DAPL_UCM_WAIT_TIME=10000
      export I_MPI_CHECK_DAPL_PROVIDER_COMPATIBILITY=0
      export I_MPI_CHECK_DAPL_PROVIDER_MISMATCH=none
      export I_MPI_DAPL_RDMA_MIXED=enable
      export I_MPI_DAPL_SCALABLE_PROGRESS=1
      export I_MPI_DAPL_TRANSLATION_CACHE=1
      export I_MPI_DAPL_UD_DIRECT_COPY_THRESHOLD=65536
      export I_MPI_DAPL_UD=on
      export I_MPI_FABRICS=shm:dapl
      export I_MPI_DAPL_PROVIDER=ofa-v2-mlx5_0-1u
      export I_MPI_FALLBACK=disable
      export I_MPI_FALLBACK_DEVICE=disable
      export I_MPI_DYNAMIC_CONNECTION=1
      export I_MPI_FAST_COLLECTIVES=1
      export I_MPI_LARGE_SCALE_THRESHOLD=8192
      # File system
      # -----------
      export I_MPI_EXTRA_FILESYSTEM_LIST=lustre
      export I_MPI_EXTRA_FILESYSTEM=on
      # Slurm
      # -----
      export I_MPI_HYDRA_BOOTSTRAP=slurm
      export I_MPI_SLURM_EXT=0
      # Force kill job
      # --------------
      export I_MPI_JOB_SIGNAL_PROPAGATION=on
      export I_MPI_JOB_ABORT_SIGNAL=9
      # Extra
      # -----
      export I_MPI_LIBRARY_KIND=release_mt
      export EC_MPI_ATEXIT=0
      export EC_PROFILE_HEAP=0
      # Process placement (cyclic)
      # --------------------------
      export I_MPI_JOB_RESPECT_PROCESS_PLACEMENT=off
      export I_MPI_PERHOST=1
      # Process pinning
      # ---------------
      export I_MPI_PIN=enable
      export I_MPI_PIN_PROCESSOR_LIST="allcores:map=scatter" # map=spread

     echo time mpiexec.hydra -configfile ./mpmd.conf
          time mpiexec.hydra -configfile ./mpmd.conf
#
  else
#
# Run SPMD case
#
    echo time srun --cpu_bind=cores --mpi=pmi2 -m cyclic -n ${OCORES} ./nemo
         time srun --cpu_bind=cores --mpi=pmi2 -m cyclic -n ${OCORES} ./nemo
  fi
#

#
  post_test_tidyup
# END_BODY
# Don't remove neither change the previous line
  exit