portelli/Grid
1
0
Fork 0
mirror of https://github.com/paboyle/Grid.git synced 2025-06-17 15:27:06 +01:00
Code Releases Activity

Significantly better performance on Aurora without using pipeline mode

Browse Source
This commit is contained in:
Peter Boyle
2025-01-30 16:36:46 +00:00
parent d6b2727f86
commit 94019a922e
12 changed files with 306 additions and 101 deletions
Download Patch File Download Diff File Expand all files Collapse all files

16
systems/Aurora/benchmarks/bench2.pbs
View File

@ -27,10 +27,22 @@ export MPICH_OFI_NIC_POLICY=GPU
#export MPIR_CVAR_CH4_OFI_GPU_PIPELINE_NUM_BUFFERS_PER_CHUNK=16
#export MPIR_CVAR_CH4_OFI_GPU_PIPELINE_MAX_NUM_BUFFERS=16
#
# Local vol 16.16.16.32
#
#VOL=32.64.64.96
for VOL in 32.32.32.96 32.64.64.96
do
for AT in 32
do
CMD="mpiexec -np 24 -ppn 12 -envall \
./gpu_tile.sh ./Benchmark_dwf_fp32 --mpi 2.2.2.3 --grid 32.64.64.96 \
--shm-mpi 0 --shm 2048 --device-mem 32000 --accelerator-threads 8 "
./gpu_tile.sh ./Benchmark_dwf_fp32 --mpi 2.2.2.3 --grid $VOL \
--shm-mpi 0 --shm 2048 --device-mem 32000 --accelerator-threads $AT --comms-overlap "
echo $CMD
$CMD
done
done