portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2025-06-13 20:57:06 +01:00
Code Releases Activity

Almost working on Aurora

Browse Source
This commit is contained in:
Peter Boyle
2024-09-23 09:43:50 +00:00
parent e637fbacae
commit 02c8178f16
7 changed files with 118 additions and 98 deletions
Download Patch File Download Diff File Expand all files Collapse all files

61
systems/Aurora/benchmarks/bench1.pbs
View File

@ -5,63 +5,34 @@
#PBS -l walltime=00:20:00
#PBS -A LatticeQCD_aesp_CNDA
#export OMP_PROC_BIND=spread
#unset OMP_PLACES
cd $PBS_O_WORKDIR
source ../sourceme.sh
module load pti-gpu
#cat $PBS_NODEFILE
cp $PBS_NODEFILE nodefile
export OMP_NUM_THREADS=4
export MPIR_CVAR_CH4_OFI_ENABLE_GPU_PIPELINE=1
#unset MPIR_CVAR_CH4_OFI_GPU_PIPELINE_D2H_ENGINE_TYPE
#unset MPIR_CVAR_CH4_OFI_GPU_PIPELINE_H2D_ENGINE_TYPE
#unset MPIR_CVAR_GPU_USE_IMMEDIATE_COMMAND_LIST
unset MPIR_CVAR_CH4_OFI_GPU_PIPELINE_D2H_ENGINE_TYPE
unset MPIR_CVAR_CH4_OFI_GPU_PIPELINE_H2D_ENGINE_TYPE
unset MPIR_CVAR_GPU_USE_IMMEDIATE_COMMAND_LIST
#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_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
# 12 ppn, 2 nodes, 24 ranks
#
CMD="mpiexec -np 1 -ppn 1 -envall \
./gpu_tile_compact.sh \
./Benchmark_usqcd --mpi 1.1.1.1 --grid 24.32.32.24 \
--shm-mpi 0 --shm 2048 --device-mem 32000 --accelerator-threads 32"
$CMD | tee usqcd.log
CMD="mpiexec -np 1 -ppn 1 -envall \
./gpu_tile_compact.sh \
./Benchmark_dwf_fp32 --mpi 1.1.1.1 --grid 16.32.32.32 \
--shm-mpi 0 --shm 2048 --device-mem 32000 --accelerator-threads 32 "
$CMD | tee 1tile.dwf
CMD="mpiexec -np 12 -ppn 12 -envall \
./gpu_tile_compact.sh \
./Benchmark_dwf_fp32 --mpi 2.2.1.3 --grid 32.32.32.48 \
--shm-mpi 0 --shm 2048 --device-mem 32000 --accelerator-threads 32 --comms-overlap"
#$CMD | tee 1node.32.32.32.48.dwf
./Benchmark_dwf_fp32 --mpi 2.1.2.3 --grid 32.32.64.48 \
--shm-mpi 1 --shm 2048 --device-mem 32000 --accelerator-threads 32 --debug-signals"
CMD="mpiexec -np 12 -ppn 12 -envall \
./gpu_tile_compact.sh \
./Benchmark_dwf_fp32 --mpi 2.2.1.3 --grid 64.64.32.96 \
--shm-mpi 0 --shm 2048 --device-mem 32000 --accelerator-threads 32 --comms-overlap"
#$CMD | tee 1node.64.64.32.96.dwf
CMD="mpiexec -np 12 -ppn 12 -envall \
./gpu_tile_compact.sh \
./Benchmark_dwf_fp32 --mpi 2.2.1.3 --grid 64.32.32.48 \
--shm-mpi 0 --shm 2048 --device-mem 32000 --accelerator-threads 32 --comms-overlap"
#$CMD | tee 1node.64.32.32.48.dwf
#for f in 1 2 3 4 5 6 7 8
for f in 1
do
echo $CMD
$CMD | tee 1node.32.32.64.48.dwf.hbm.$f
done

27
systems/Aurora/benchmarks/bench2.pbs
View File

@ -11,17 +11,16 @@
cd $PBS_O_WORKDIR
source ../sourceme.sh
module load pti-gpu
#module load pti-gpu
#cat $PBS_NODEFILE
cp $PBS_NODEFILE nodefile
export OMP_NUM_THREADS=4
export MPIR_CVAR_CH4_OFI_ENABLE_GPU_PIPELINE=1
#unset MPIR_CVAR_CH4_OFI_GPU_PIPELINE_D2H_ENGINE_TYPE
#unset MPIR_CVAR_CH4_OFI_GPU_PIPELINE_H2D_ENGINE_TYPE
#unset MPIR_CVAR_GPU_USE_IMMEDIATE_COMMAND_LIST
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
@ -34,22 +33,26 @@ export MPICH_OFI_NIC_POLICY=GPU
# 12 ppn, 2 nodes, 24 ranks
#
CMD="mpiexec -np 24 -ppn 12 -envall \
./gpu_tile_compact.sh \
./gpu_tile.sh \
./Benchmark_comms_host_device --mpi 2.2.2.3 --grid 24.32.32.24 \
--shm-mpi 0 --shm 2048 --device-mem 32000 --accelerator-threads 32"
$CMD | tee 2node.comms
#$CMD | tee 2node.comms.hbm
CMD="mpiexec -np 24 -ppn 12 -envall \
./gpu_tile_compact.sh \
./Benchmark_dwf_fp32 --mpi 2.2.2.3 --grid 32.32.64.48 \
--shm-mpi 1 --shm 2048 --device-mem 32000 --accelerator-threads 32 "
$CMD | tee 2node.32.32.64.48.dwf
--shm-mpi 1 --shm 2048 --device-mem 32000 --accelerator-threads 32 --comms-overlap --debug-signals"
#for f in 1 2 3 4 5 6 7 8
for f in 1
do
echo $CMD
$CMD | tee 2node.32.32.64.48.dwf.hbm.$f
done
CMD="mpiexec -np 24 -ppn 12 -envall \
./gpu_tile_compact.sh \
./gpu_tile.sh \
./Benchmark_dwf_fp32 --mpi 2.2.2.3 --grid 64.64.64.96 \
--shm-mpi 1 --shm 2048 --device-mem 32000 --accelerator-threads 32 "
$CMD | tee 2node.64.64.64.96.dwf
--shm-mpi 0 --shm 2048 --device-mem 32000 --accelerator-threads 32 --comms-overlap"
#$CMD | tee 2node.64.64.64.96.dwf.hbm

2
systems/Aurora/sourceme.sh
View File

@ -1,4 +1,6 @@
module load oneapi/release/2023.12.15.001
#module load mpich/icc-all-debug-pmix-gpu/52.2
#module load mpich-config/mode/deterministic
#module load intel_compute_runtime/release/821.35
source ~/spack/share/spack/setup-env.sh
spack load c-lime

45
systems/Aurora/tests/reproBigJob.pbs
View File

@ -15,13 +15,13 @@
# 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_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_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_CH4_OFI_GPU_PIPELINE_BUFFER_SZ=10485760
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
@ -30,20 +30,22 @@ export SYCL_PI_LEVEL_ZERO_USE_COPY_ENGINE_FOR_D2D_COPY=1
export SYCL_PROGRAM_COMPILE_OPTIONS="-ze-opt-large-register-file"
export GRID_PRINT_ENTIRE_LOG=0
export GRID_CHECKSUM_RECV_BUF=1
export GRID_CHECKSUM_SEND_BUF=1
export GRID_CHECKSUM_RECV_BUF=0
export GRID_CHECKSUM_SEND_BUF=0
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 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
cd $PBS_O_WORKDIR
cp $PBS_NODEFILE nodefile
DIR=reproBigJob.$PBS_JOBID
mkdir -p $DIR
@ -51,10 +53,19 @@ cd $DIR
cp $PBS_NODEFILE nodefile
BINARY=../Test_dwf_mixedcg_prec
echo > pingjob <<EOF
while read node ;
do
echo ssh $node killall -s USR1 -- ../Test_dwf_mixedcg_prec
done < nodefile
EOF
CMD="mpiexec -np 384 -ppn 12 -envall --hostfile nodefile \
../gpu_tile_compact.sh \
../Test_dwf_mixedcg_prec --mpi 4.4.4.6 --grid 128.128.128.96 \
--shm-mpi 1 --comms-overlap --shm 4096 --device-mem 32000 --accelerator-threads 32 --seconds 6000 --debug-stdout --log Message --debug-signals"
$BINARY --mpi 4.4.4.6 --grid 128.128.128.96 \
--shm-mpi 1 --shm 4096 --device-mem 32000 --accelerator-threads 32 --seconds 6000 --debug-stdout --log Message --debug-signals"
echo $CMD > command-line
env > environment