ORCA

#!/bin/bash
###############################################################################
# orca_compute.sh  -  single-node ORCA 5.0.4 job template for duhpc
#
# COPY this, then edit INPUT and the #SBATCH lines. ORCA is run by FULL PATH and
# is told its core count via '%pal nprocs N end' INSIDE the .inp file. Do NOT
# wrap ORCA in mpirun/srun -- it spawns its own MPI ranks.
#
#   cp /scratch/apps/templates/orca_compute.sh myjob.sh
#   # edit INPUT=, --ntasks, and put a matching '%pal nprocs N end' in myjob.inp
#   sbatch myjob.sh
###############################################################################
#SBATCH --job-name=orca
#SBATCH --partition=compute
#SBATCH --nodes=1
#SBATCH --ntasks=16            # <-- MUST equal '%pal nprocs 16 end' in your .inp
#SBATCH --cpus-per-task=1      # ORCA parallelism is by MPI rank, not threads
#SBATCH --mem=32G             # total memory; set to a bit above (ntasks x %maxcore_MB). Adjust per job.
#SBATCH --time=2-00:00:00      # compute partition max is 5 days
#SBATCH --output=%x.%j.out
#SBATCH --error=%x.%j.err

#------------------------------------------------------------------ user edits
INPUT="myjob"                  # base name; expects ${INPUT}.inp, writes ${INPUT}.out
#----------------------------------------------------------------------------

module purge
module load orca/5.0.4

# InfiniBand/transport settings (belt-and-braces backup; also in the OpenMPI
# build's MCA config). Cluster uses native openib BTL, not UCX. Do not change.
export OMPI_MCA_btl_openib_allow_ib=1
export OMPI_MCA_pml=ob1
export OMPI_MCA_osc="^ucx"

# Belt-and-braces against any limits.conf drift (the SIGXCPU lesson).
ulimit -t unlimited 2>/dev/null || true
ulimit -u unlimited 2>/dev/null || true
ulimit -s unlimited 2>/dev/null || true   # ORCA likes a large stack

# Per-job node-local scratch. ORCA writes MANY temp files; keep them OFF Lustre
# during the run, then only the .out/.gbw/.xyz land in your project dir.
export ORCA_SCRATCH="/tmp/orca.${SLURM_JOB_ID}.${USER}"
mkdir -p "$ORCA_SCRATCH"
cp "${INPUT}.inp" "$ORCA_SCRATCH"/
# copy any restart/gbw/xyz dependencies your input references:
# cp ${INPUT}.gbw "$ORCA_SCRATCH"/ 2>/dev/null || true
cd "$ORCA_SCRATCH"

echo "Host          : $(hostname)"
echo "ORCA binary   : $(which orca)         (must be \$ORCA_DIR/orca)"
echo "mpirun        : $(which mpirun)       (must be the DEDICATED OpenMPI 4.1.1)"
echo "Scratch       : $ORCA_SCRATCH"
echo "Cores (SLURM) : ${SLURM_NTASKS}   <-- confirm '%pal nprocs ${SLURM_NTASKS} end' is in ${INPUT}.inp"

# FULL PATH is mandatory for ORCA parallel runs.
"$ORCA_DIR/orca" "${INPUT}.inp" > "${SLURM_SUBMIT_DIR}/${INPUT}.out" 2>&1
RC=$?

# Pull back the useful artefacts, then clean scratch.
cp -f "${INPUT}".gbw "${INPUT}".xyz "${INPUT}"*.txt "${SLURM_SUBMIT_DIR}/" 2>/dev/null || true
cd "${SLURM_SUBMIT_DIR}"
rm -rf "$ORCA_SCRATCH"

echo "ORCA exit code: $RC"
grep -q "ORCA TERMINATED NORMALLY" "${INPUT}.out" \
  && echo "==> NORMAL TERMINATION" \
  || echo "==> CHECK ${INPUT}.out -- did not terminate normally"
exit $RC

#!/bin/bash
###############################################################################
# orca_mpi.sh  -  MULTI-NODE ORCA 5.0.4 job template for duhpc (mpi partition)
#
# Multi-node ORCA differs from single-node in ONE important way: scratch must
# live on a SHARED filesystem (Lustre), not node-local /tmp, because ranks on
# different nodes need to see the same working files. We therefore stage into
# /scratch/home/$USER/orca-run-<jobid>/.
#
# As always: ORCA is invoked by FULL PATH, core count comes from
# '%pal nprocs N end' in the .inp, and you NEVER wrap ORCA in mpirun/srun --
# ORCA's own mpirun reads the SLURM allocation (our OpenMPI 4.1.1 is SLURM-aware).
#
# The mpi partition enforces MinNodes=2; for a single node use orca_compute.sh
# (the submit plugin will redirect you anyway).
###############################################################################
#SBATCH --job-name=orca_mpi
#SBATCH --partition=mpi
#SBATCH --nodes=2
#SBATCH --ntasks=32            # total ranks = '%pal nprocs 32 end' in your .inp
#SBATCH --ntasks-per-node=16
#SBATCH --cpus-per-task=1
#SBATCH --mem=0               # request all memory on each node (mpi nodes are exclusive)
#SBATCH --time=2-00:00:00
#SBATCH --output=%x.%j.out
#SBATCH --error=%x.%j.err

#------------------------------------------------------------------ user edits
INPUT="myjob"
#----------------------------------------------------------------------------

module purge
module load orca/5.0.4

# InfiniBand transport (belt-and-braces backup; also set in the OpenMPI build's
# etc/openmpi-mca-params.conf). This cluster uses the native openib BTL, NOT UCX
# (the available UCX has no InfiniBand support). Do not change these.
export OMPI_MCA_btl_openib_allow_ib=1
export OMPI_MCA_pml=ob1
export OMPI_MCA_osc="^ucx"
export OMPI_MCA_btl="openib,self,vader,tcp"

ulimit -t unlimited 2>/dev/null || true
ulimit -u unlimited 2>/dev/null || true
ulimit -s unlimited 2>/dev/null || true

# SHARED scratch on Lustre (visible to every node in the allocation).
export ORCA_SCRATCH="/scratch/home/${USER}/orca-run-${SLURM_JOB_ID}"
mkdir -p "$ORCA_SCRATCH"
cp "${INPUT}.inp" "$ORCA_SCRATCH"/
# cp ${INPUT}.gbw "$ORCA_SCRATCH"/ 2>/dev/null || true   # restart files if any
cd "$ORCA_SCRATCH"

echo "Nodes         : $SLURM_JOB_NODELIST"
echo "Total ranks   : $SLURM_NTASKS  <-- confirm '%pal nprocs $SLURM_NTASKS end' in ${INPUT}.inp"
echo "ORCA binary   : $(which orca)"
echo "mpirun        : $(which mpirun)  ($(mpirun --version | head -1))"
echo "Shared scratch: $ORCA_SCRATCH"

# ORCA reads the SLURM allocation through its bundled SLURM-aware OpenMPI.
# Full path is mandatory; do NOT prepend mpirun/srun.
"$ORCA_DIR/orca" "${INPUT}.inp" > "${SLURM_SUBMIT_DIR}/${INPUT}.out" 2>&1
RC=$?

cp -f "${INPUT}".gbw "${INPUT}".xyz "${INPUT}"*.txt "${SLURM_SUBMIT_DIR}/" 2>/dev/null || true
cd "${SLURM_SUBMIT_DIR}"
rm -rf "$ORCA_SCRATCH"

echo "ORCA exit code: $RC"
grep -q "ORCA TERMINATED NORMALLY" "${INPUT}.out" \
  && echo "==> NORMAL TERMINATION" \
  || echo "==> CHECK ${INPUT}.out -- did not terminate normally (multi-node MPI/IB issue?)"
exit $RC