diff options
author | Birte Kristina Friesel <birte.friesel@uos.de> | 2024-07-05 11:34:36 +0200 |
---|---|---|
committer | Birte Kristina Friesel <birte.friesel@uos.de> | 2024-07-05 11:34:36 +0200 |
commit | f4e5bca49a26f13788d275c352a1268a6e65b1c0 (patch) | |
tree | 448949acf49731d88569ae68ae9de26206764ad8 /GEMV | |
parent | f5377a304ba5390400ebff633f162212d3c7ab5d (diff) |
GEMV: add HetSim scripts
Diffstat (limited to 'GEMV')
-rwxr-xr-x | GEMV/dimes-hetsim-hbm.sh | 43 | ||||
-rwxr-xr-x | GEMV/dimes-hetsim-nmc.sh | 74 |
2 files changed, 117 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 ${fn}.txt diff --git a/GEMV/dimes-hetsim-nmc.sh b/GEMV/dimes-hetsim-nmc.sh new file mode 100755 index 0000000..a4d2973 --- /dev/null +++ b/GEMV/dimes-hetsim-nmc.sh @@ -0,0 +1,74 @@ +#!/bin/bash + +mkdir -p log/$(hostname) baselines/cpu/log/$(hostname) +fn=log/$(hostname)/$(date +%Y%m%d) + +# upstream DPU version uses -m 163840 -n 4096 +# → (163840 * 4096 + 4096) * uint32 ≈ 2.5 GiB + +# upstream baseline uses 20480 rows and 8192 cols, allocating 20480 * double + 8192 * double + 20480 * double + 20480 * 8192 * double +# → ≈ 1.25 GiB + +run_benchmark_nmc() { + local "$@" + sudo limit_ranks_to_numa_node ${numa_rank} + if make -B NR_DPUS=${nr_dpus} NR_TASKLETS=${nr_tasklets} WITH_ALLOC_OVERHEAD=1 WITH_LOAD_OVERHEAD=1 WITH_FREE_OVERHEAD=1; then + bin/gemv_host -w 0 -e 100 -m 163840 -n 4096 + fi + return $? +} + +export -f run_benchmark_nmc + +run_benchmark_baseline() { + local "$@" + OMP_NUM_THREADS=$nr_threads ./gemv $ram $ram $cpu + return $? +} + +export -f run_benchmark_baseline + +( + +echo "NMC single-node (1/2)" >&2 + +parallel -j1 --eta --joblog ${fn}.1.joblog --resume --header : \ + run_benchmark_nmc nr_dpus={nr_dpus} nr_tasklets=16 numa_rank={numa_rank} \ + ::: numa_rank 0 1 \ + ::: nr_dpus 64 128 256 512 768 1024 + +echo "NMC multi-node (2/2)" >&2 + +parallel -j1 --eta --joblog ${fn}.2.joblog --resume --header : \ + run_benchmark_nmc nr_dpus={nr_dpus} nr_tasklets=16 numa_rank={numa_rank} \ + ::: numa_rank -1 \ + ::: nr_dpus 1536 2048 2304 + +) > ${fn}.txt + +xz -f -v -9 -M 800M ${fn}.txt + +cd baselines/cpu +make -B NUMA=1 + +( + +echo "CPU single-node (1/2)" >&2 + +parallel -j1 --eta --joblog ${fn}.1.joblog --resume --header : \ + run_benchmark_baseline nr_threads={nr_threads} ram={ram} cpu={cpu} \ + ::: cpu 0 1 \ + ::: ram 0 1 \ + ::: nr_threads 1 2 4 8 12 16 32 + +echo "CPU multi-node (2/2)" >&2 + +parallel -j1 --eta --joblog ${fn}.2.joblog --resume --header : \ + run_benchmark_baseline nr_threads={nr_threads} ram={ram} cpu={cpu} \ + ::: cpu -1 \ + ::: ram 0 1 \ + ::: nr_threads 48 64 + +) > ${fn}.txt + +xz -f -v -9 -M 800M ${fn}.txt |