1
0
mirror of https://github.com/paboyle/Grid.git synced 2024-09-19 16:55:37 +01:00

This appears working now on Sunspot

This commit is contained in:
Peter Boyle 2024-03-22 15:43:30 +00:00
parent d1e9fe50d2
commit 503dec34ef

View File

@ -1,6 +1,6 @@
#!/bin/bash
#PBS -l select=16:system=sunspot,place=scatter
#PBS -l select=32:system=sunspot,place=scatter
#PBS -A LatticeQCD_aesp_CNDA
#PBS -l walltime=02:00:00
#PBS -N reproN
@ -15,13 +15,22 @@ 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 MPIR_CVAR_CH4_OFI_GPU_PIPELINE_D2H_ENGINE_TYPE=0
#export MPIR_CVAR_CH4_OFI_GPU_PIPELINE_H2D_ENGINE_TYPE=0
#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 MPIR_CVAR_GPU_USE_IMMEDIATE_COMMAND_LIST=1
export SYCL_PI_LEVEL_ZERO_USE_IMMEDIATE_COMMANDLISTS=1
export SYCL_PI_LEVEL_ZERO_USE_COPY_ENGINE=1
export SYCL_PI_LEVEL_ZERO_USE_COPY_ENGINE_FOR_D2D_COPY=1
export GRID_PRINT_ENTIRE_LOG=0
export GRID_CHECKSUM_RECV_BUF=1
export GRID_CHECKSUM_SEND_BUF=0
export MPICH_OFI_NIC_POLICY=GPU
export MPIR_CVAR_ALLREDUCE_DEVICE_COLLECTIVE=0
@ -46,29 +55,39 @@ cd $PBS_O_WORKDIR
THIS_NODE=`head -n$n $PBS_NODEFILE | tail -n1 `
echo Node $n is $THIS_NODE
DIR=repro.$PBS_JOBID/node-$n-$THIS_NODE
DIR=reproN.$PBS_JOBID/node-$n-$THIS_NODE
mkdir -p $DIR
cd $DIR
echo $THIS_NODE > nodefile
#CMD="mpiexec -np 12 -ppn 12 -envall --hostfile nodefile \
# ../../gpu_tile_compact.sh \
# ../../Test_dwf_mixedcg_prec --mpi 1.2.2.3 --grid 32.64.64.96 \
# --shm-mpi 0 --shm 4096 --device-mem 32000 --accelerator-threads 32 --seconds 6000 --debug-stdout --log Message --comms-overlap"
CMD="mpiexec -np 12 -ppn 12 -envall --hostfile nodefile \
../../gpu_tile_compact.sh \
../../Test_dwf_mixedcg_prec --mpi 1.2.2.3 --grid 32.64.64.96 \
--shm-mpi 0 --shm 4096 --device-mem 32000 --accelerator-threads 32 --seconds 6000 --debug-stdout --log Message --comms-overlap"
--shm-mpi 1 --shm 4096 --device-mem 32000 --accelerator-threads 32 --seconds 6000 --debug-stdout --log Message --comms-overlap"
echo $CMD > command-line
env > environment
$CMD &
done
wait
# Suspicious wait is allowing jobs to collide and knock out
#wait
sleep 6500
for n in ` eval echo {1..$NN} `
do
THIS_NODE=`head -n$n $PBS_NODEFILE | tail -n1 `
DIR=repro.$PBS_JOBID/node-$n-$THIS_NODE
DIR=reproN.$PBS_JOBID/node-$n-$THIS_NODE
cd $DIR