diff options
Diffstat (limited to 'GEMV/dimes-hetsim-hbm.sh')
-rwxr-xr-x | GEMV/dimes-hetsim-hbm.sh | 43 |
1 files changed, 43 insertions, 0 deletions
diff --git a/GEMV/dimes-hetsim-hbm.sh b/GEMV/dimes-hetsim-hbm.sh new file mode 100755 index 0000000..d771269 --- /dev/null +++ b/GEMV/dimes-hetsim-hbm.sh @@ -0,0 +1,43 @@ +#!/bin/bash + +cd baselines/cpu +make -B NUMA=1 + +mkdir -p log/$(hostname) +fn=log/$(hostname)/$(date +%Y%m%d)-baseline + +# upstream DPU version uses -m 163840 -n 4096 +# → (163840 * 4096 + 4096) * uint32 ≈ 2.5 GiB + +# upstream baseline uses 20480 rows and 8192 cols, allocating 20480 * double + 8192 * double + 20480 * double + 20480 * 8192 * double +# → ≈ 1.25 GiB + +run_benchmark() { + local "$@" + OMP_NUM_THREADS=$nr_threads ./gemv $ram $ram $cpu + return $? +} + +export -f run_benchmark + +( + +echo "CPU single-node (1/2)" >&2 + +parallel -j1 --eta --joblog ${fn}.1.joblog --resume --header : \ + run_benchmark nr_threads={nr_threads} ram={ram} cpu={cpu} \ + ::: cpu $(seq 0 7) \ + ::: ram $(seq 0 15) \ + ::: nr_threads 1 2 4 8 12 16 + +echo "CPU multi-node (2/2)" >&2 + +parallel -j1 --eta --joblog ${fn}.2.joblog --resume --header : \ + run_benchmark nr_threads={nr_threads} ram={ram} cpu={cpu} \ + ::: cpu -1 \ + ::: ram $(seq 0 15) \ + ::: nr_threads 32 48 64 96 128 + +) > ${fn}.txt + +xz -f -v -9 -M 800M ${fn}.txt |