portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2025-06-12 20:27:06 +01:00
Code Releases Activity

Scaling benchmarks, verbosity and MPICH aware in acceleratorInit()

Browse Source 
For some reason Dirichlet benchmark fails on several nodes; need to
debug this.
This commit is contained in:
Peter Boyle
2024-02-13 19:48:03 +00:00
parent 5ef4da3f29
commit 303b83cdb8
3 changed files with 106 additions and 13 deletions
Download Patch File Download Diff File Expand all files Collapse all files

80
systems/Aurora/benchmarks/bench_scaling.pbs Normal file
View File

@ -0,0 +1,80 @@
#!/bin/bash
## qsub -q EarlyAppAccess -A Aurora_Deployment -I -l select=1 -l walltime=60:00
#PBS -q EarlyAppAccess
#PBS -l select=32
#PBS -l walltime=01:00:00
#PBS -A LatticeQCD_aesp_CNDA
#export OMP_PROC_BIND=spread
#unset OMP_PLACES
cd $PBS_O_WORKDIR
source ../sourceme.sh
cat $PBS_NODEFILE
export OMP_NUM_THREADS=3
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
# 12 ppn, 32 nodes, 384 ranks
#
CMD="mpiexec -np 384 -ppn 12 -envall \
./gpu_tile_compact.sh \
./Benchmark_comms_host_device --mpi 4.6.4.4 --grid 32.24.32.192 \
--shm-mpi 1 --shm 2048 --device-mem 32000 --accelerator-threads 32"
$CMD
CMD="mpiexec -np 12 -ppn 12 -envall \
./gpu_tile_compact.sh \
./Benchmark_dwf_fp32 --mpi 1.2.2.3 --grid 16.64.64.96 \
--shm-mpi 1 --shm 2048 --device-mem 32000 --accelerator-threads 32 --comms-overlap"
$CMD | tee 1node.dwf
CMD="mpiexec -np 24 -ppn 12 -envall \
./gpu_tile_compact.sh \
./Benchmark_dwf_fp32 --mpi 2.2.2.3 --grid 32.64.64.96 \
--shm-mpi 1 --shm 2048 --device-mem 32000 --accelerator-threads 32 --comms-overlap"
$CMD | tee 2node.dwf
CMD="mpiexec -np 48 -ppn 12 -envall \
./gpu_tile_compact.sh \
./Benchmark_dwf_fp32 --mpi 2.2.2.6 --grid 32.64.64.192 \
--shm-mpi 1 --shm 2048 --device-mem 32000 --accelerator-threads 32 --comms-overlap"
$CMD | tee 4node.dwf
CMD="mpiexec -np 96 -ppn 12 -envall \
./gpu_tile_compact.sh \
./Benchmark_dwf_fp32 --mpi 2.2.4.6 --grid 32.64.128.192 \
--shm-mpi 1 --shm 2048 --device-mem 32000 --accelerator-threads 32 --comms-overlap"
$CMD | tee 8node.dwf
CMD="mpiexec -np 192 -ppn 12 -envall \
./gpu_tile_compact.sh \
./Benchmark_dwf_fp32 --mpi 2.4.4.6 --grid 32.128.128.192 \
--shm-mpi 1 --shm 2048 --device-mem 32000 --accelerator-threads 32 --comms-overlap"
$CMD | tee 16node.dwf
CMD="mpiexec -np 384 -ppn 12 -envall \
./gpu_tile_compact.sh \
./Benchmark_dwf_fp32 --mpi 4.4.4.6 --grid 64.128.128.192 \
--shm-mpi 1 --shm 2048 --device-mem 32000 --accelerator-threads 32 --comms-overlap"
$CMD | tee 32node.dwf