1
0
mirror of https://github.com/paboyle/Grid.git synced 2024-11-15 02:05:37 +00:00
Grid/systems/Aurora/tests/repro1gpu.pbs
2024-09-10 18:11:52 +00:00

87 lines
2.2 KiB
Bash

#!/bin/bash
#PBS -l select=16
#PBS -q EarlyAppAccess
#PBS -A LatticeQCD_aesp_CNDA
#PBS -l walltime=02:00:00
#PBS -N repro1gpu
#PBS -k doe
#export OMP_PROC_BIND=spread
#unset OMP_PLACES
#module load oneapi/eng-compiler/2023.05.15.003
#module load mpich/51.2/icc-all-deterministic-pmix-gpu
# 56 cores / 6 threads ~9
export OMP_NUM_THREADS=6
export MPIR_CVAR_CH4_OFI_ENABLE_GPU_PIPELINE=1
export MPIR_CVAR_CH4_OFI_GPU_PIPELINE_D2H_ENGINE_TYPE=0
export MPIR_CVAR_CH4_OFI_GPU_PIPELINE_H2D_ENGINE_TYPE=0
export MPIR_CVAR_GPU_USE_IMMEDIATE_COMMAND_LIST=1
export MPIR_CVAR_CH4_OFI_GPU_PIPELINE_BUFFER_SZ=1048576
export MPIR_CVAR_CH4_OFI_GPU_PIPELINE_THRESHOLD=131072
export MPIR_CVAR_CH4_OFI_GPU_PIPELINE_NUM_BUFFERS_PER_CHUNK=16
export MPIR_CVAR_CH4_OFI_GPU_PIPELINE_MAX_NUM_BUFFERS=16
export MPICH_OFI_NIC_POLICY=GPU
export MPIR_CVAR_ALLREDUCE_DEVICE_COLLECTIVE=0
export MPIR_CVAR_REDUCE_DEVICE_COLLECTIVE=0
export MPIR_CVAR_ALLREDUCE_INTRA_ALGORITHM=recursive_doubling
unset MPIR_CVAR_CH4_COLL_SELECTION_TUNING_JSON_FILE
unset MPIR_CVAR_COLL_SELECTION_TUNING_JSON_FILE
unset MPIR_CVAR_CH4_POSIX_COLL_SELECTION_TUNING_JSON_FILE
export SYCL_PROGRAM_COMPILE_OPTIONS="-ze-opt-large-register-file"
cd $PBS_O_WORKDIR
source ../sourceme.sh
NN=`cat $PBS_NODEFILE | wc -l`
echo $PBS_NODEFILE
cat $PBS_NODEFILE
echo $NN nodes in node file
for n in `eval echo {1..$NN}`
do
THIS_NODE=`head -n$n $PBS_NODEFILE | tail -n1 `
echo Node $n is $THIS_NODE
for g in {0..11}
do
export NUMA_MAP=(0 0 0 1 1 1 0 0 0 1 1 1 )
export TILE_MAP=(0 0 0 0 0 0 1 1 1 1 1 1 )
export GPU_MAP=(0 1 2 3 4 5 0 1 2 3 4 5 )
export numa=${NUMA_MAP[$g]}
export gpu_id=${GPU_MAP[$g]}
export tile_id=${TILE_MAP[$g]}
export gpu=$gpu_id.$tile_id
cd $PBS_O_WORKDIR
DIR=repro.1gpu.$PBS_JOBID/node-$n-$THIS_NODE-GPU-$gpu
mkdir -p $DIR
cd $DIR
echo $THIS_NODE > nodefile
echo $gpu > gpu
export ZE_AFFINITY_MASK=$gpu
export ONEAPI_DEVICE_FILTER=gpu,level_zero
CMD="mpiexec -np 1 -ppn 1 -envall --hostfile nodefile \
numactl -N $numa -m $numa ../../Test_dwf_mixedcg_prec --mpi 1.1.1.1 --grid 16.16.32.32 \
--shm-mpi 0 --shm 4096 --device-mem 32000 --accelerator-threads 32 --seconds 6000 --debug-stdout --log Message"
echo $CMD
$CMD &
done
done
wait