#!/usr/bin/env bash
# shellcheck disable=SC1091,SC2050,SC2170

# using options from https://github.com/paboyle/Grid/tree/develop/systems/Tursa

#SBATCH -J power-loc32-16A-840
#SBATCH -A dp207
#SBATCH -t 48:00:00
#SBATCH --nodes=16
#SBATCH --ntasks=64
#SBATCH --ntasks-per-node=4
#SBATCH --cpus-per-task=8
#SBATCH --partition=gpu
#SBATCH --gres=gpu:4
#SBATCH --output=%x.%j.out
#SBATCH --error=%x.%j.err
#SBATCH --reservation=dc-port1_61
#SBATCH --qos=reservation
#SBATCH --no-requeue

set -e

# OpenMP/OpenMPI/UCX environment ###############################################
export OMP_NUM_THREADS=4
export OMPI_MCA_btl=^uct,openib
export OMPI_MCA_pml=ucx
export UCX_TLS=gdr_copy,rc,rc_x,sm,cuda_copy,cuda_ipc
export UCX_RNDV_SCHEME=put_zcopy
export UCX_RNDV_THRESH=16384
export UCX_IB_GPU_DIRECT_RDMA=yes
export UCX_MEMTYPE_CACHE=n

# IO environment ###############################################################

if [ 16 -eq 1 ]; then
	export OMPI_MCA_io=ompio
else
	export OMPI_MCA_io=romio321
fi
export OMPI_MCA_btl_openib_allow_ib=true
export OMPI_MCA_btl_openib_device_type=infiniband
export OMPI_MCA_btl_openib_if_exclude=mlx5_1,mlx5_2,mlx5_3

# load environment #############################################################
env_dir="$(readlink -f /mnt/lustre/tursafs1/home/dp207/dp207/shared/env/versions/220428)"
source "${env_dir}/env-base.sh"
if [ "${SLURM_JOB_PARTITION}" = 'gpu' ]; then
	source "${env_dir}/env-gpu.sh"
else
	echo "error: partition ${SLURM_JOB_PARTITION} not supported for this template" 1>&2
  exit 1
fi
spack load sshpass

# application and parameters ###################################################
app='/mnt/lustre/tursafs1/home/dp207/dp207/dc-port1/power-bench/2-racks/size-loc32/Benchmark_dwf_fp32'
opt=('--comms-overlap' '--comms-concurrent')
par=''

# collect job information ######################################################
job_info_dir=job/${SLURM_JOB_NAME}.${SLURM_JOB_ID}
mkdir -p "${job_info_dir}"

date                         > "${job_info_dir}/start-date"
echo "epoch $(date '+%s')"   >> "${job_info_dir}/start-date"
set                          > "${job_info_dir}/env"
ldd ${app}                   > "${job_info_dir}/ldd"
md5sum ${app}                > "${job_info_dir}/app-hash"
readelf -a ${app}            > "${job_info_dir}/elf"
echo "${SLURM_JOB_NODELIST}" > "${job_info_dir}/nodes"
cp "${BASH_SOURCE[0]}"       "${job_info_dir}/script"
if [ -n "${par}" ]; then cp "${par}" "${job_info_dir}/par"; fi

# GPU frequency control ########################################################
power_dir='/mnt/lustre/tursafs1/home/dp207/dp207/dc-port1/power-bench/2-racks/size-loc32'
freq=840

# set frequency
for h in $(scontrol show hostnames "${SLURM_JOB_NODELIST}"); do 
	${power_dir}/remote-sudo.sh "$h" "nvidia-smi -ac 1215,${freq}"
done
# start NVIDIA SMI monitoring
tmp=$(mktemp)
sleep 1
coproc nvidia-smi dmon -o DT &> "${tmp}"

# run! #########################################################################
mpirun -np "${SLURM_NTASKS}" -x LD_LIBRARY_PATH --bind-to none \
	./gpu-mpi-wrapper.sh \
  ${app} "${par}" "${opt[@]}" \
	--mpi 2.2.2.8 \
  --accelerator-threads 8 \
	--grid 64.64.64.256 \
	--shm 2048 &> "${job_info_dir}/log"

# if we reach that point the application exited successfully ###################
touch "${job_info_dir}/success"
date > "${job_info_dir}/end-date"
echo "epoch $(date '+%s')" >> "${job_info_dir}/end-date"

# reset GPUS ###################################################################
# stop monitoring
kill -INT "${COPROC_PID}"

# make monitoring DB
${power_dir}/dmon-to-db.sh "${tmp}" smi-dmon-16A.db "clock_limit_${freq}"

# reset clocks
for h in $(scontrol show hostnames "${SLURM_JOB_NODELIST}"); do 
	${power_dir}/remote-sudo.sh "$h" 'nvidia-smi -ac 1215,1410' 
done
################################################################################