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

Optional host buffer bounce for no CUDA aware MPI

Browse Source
This commit is contained in:
Peter Boyle
2025-01-28 15:22:46 +00:00
parent 1caf8b0f86
commit 74a4f43946
7 changed files with 84 additions and 53 deletions
Download Patch File Download Diff File Expand all files Collapse all files

32
systems/Aurora/benchmarks/bench1.pbs
View File

@ -1,6 +1,7 @@
#!/bin/bash
#PBS -q EarlyAppAccess
##PBS -q EarlyAppAccess
#PBS -q debug
#PBS -l select=1
#PBS -l walltime=00:20:00
#PBS -A LatticeQCD_aesp_CNDA
@ -12,27 +13,24 @@ source ../sourceme.sh
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 MPICH_OFI_NIC_POLICY=GPU
#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
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_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
CMD="mpiexec -np 12 -ppn 12 -envall \
./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"
./gpu_tile.sh ./Benchmark_dwf_fp32 --mpi 2.1.2.3 --grid 32.32.64.96 \
--shm-mpi 1 --shm 2048 --device-mem 32000 --accelerator-threads 8 "
#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
$CMD

34
systems/Aurora/benchmarks/gpu_tile_compact.sh
View File

@ -1,34 +0,0 @@
#!/bin/bash
#export NUMA_MAP=(2 2 2 3 3 3 2 2 2 3 3 3 )
#export NUMA_MAP=(0 0 1 1 0 0 1 1 0 0 1 1);
#export GPU_MAP=(0.0 0.1 3.0 3.1 1.0 1.1 4.0 4.1 2.0 2.1 5.0 5.1)
export NUMA_MAP=(0 0 0 0 0 0 1 1 1 1 1 1 );
export GPU_MAP=(0.0 1.0 2.0 3.0 4.0 5.0 0.1 1.1 2.1 3.1 4.1 5.1 )
export NUMA=${NUMA_MAP[$PALS_LOCAL_RANKID]}
export gpu_id=${GPU_MAP[$PALS_LOCAL_RANKID]}
unset EnableWalkerPartition
export EnableImplicitScaling=0
export ZE_AFFINITY_MASK=$gpu_id
export ONEAPI_DEVICE_FILTER=gpu,level_zero
export SYCL_PI_LEVEL_ZERO_DEVICE_SCOPE_EVENTS=0
export SYCL_PI_LEVEL_ZERO_USE_IMMEDIATE_COMMANDLISTS=1
export SYCL_PI_LEVEL_ZERO_USE_COPY_ENGINE=0:5
#export SYCL_PI_LEVEL_ZERO_USE_COPY_ENGINE=0:2
export SYCL_PI_LEVEL_ZERO_USE_COPY_ENGINE_FOR_D2D_COPY=1
#export SYCL_PI_LEVEL_ZERO_USM_RESIDENT=1
echo "rank $PALS_RANKID ; local rank $PALS_LOCAL_RANKID ; ZE_AFFINITY_MASK=$ZE_AFFINITY_MASK ; NUMA $NUMA "
if [ $PALS_RANKID = "0" ]
then
# numactl -m $NUMA -N $NUMA onetrace --chrome-device-timeline "$@"
# numactl -m $NUMA -N $NUMA unitrace --chrome-kernel-logging --chrome-mpi-logging --chrome-sycl-logging --demangle "$@"
numactl -m $NUMA -N $NUMA "$@"
else
numactl -m $NUMA -N $NUMA "$@"
fi