diff options
Diffstat (limited to 'HST-S/dimes-hetsim-hbm.sh')
-rwxr-xr-x | HST-S/dimes-hetsim-hbm.sh | 60 |
1 files changed, 30 insertions, 30 deletions
diff --git a/HST-S/dimes-hetsim-hbm.sh b/HST-S/dimes-hetsim-hbm.sh index 08842b7..44c191d 100755 --- a/HST-S/dimes-hetsim-hbm.sh +++ b/HST-S/dimes-hetsim-hbm.sh @@ -1,51 +1,51 @@ -#!/bin/bash +#!/bin/sh cd baselines/cpu -make -B NUMA=1 mkdir -p log/$(hostname) -fn=log/$(hostname)/$(date +%Y%m%d) - -# Upstream DPU version uses 256 bins and 1536 * 1024 * 64 uint32 elements == 384 MiB total (-x 2 with implicit -z 64) -input_size_upstream=$((1536 * 1024 * 64)) +fn=log/$(hostname)/dimes-hetsim-hbm -# Here: 2 GiB -input_size_dpu=$((2**29)) +# Input: (2^29 == 536870912) * int32 == 2 GiB ( -echo "single-node execution, upstream ref (1/4)" >&2 +make -B NUMA=1 NUMA_MEMCPY=1 + +echo "CPU single-node operation with setup cost, memcpy node == input node, cpu node == output node (1/3)" >&2 parallel -j1 --eta --joblog ${fn}.1.joblog --resume --header : \ - ./hist -i ${input_size_upstream} -A {ram} -B {ram} -C {cpu} -t {nr_threads} -w 0 -e 5 -x 1 \ + ./hist -i {input_size} -b 256 -A {ram_in} -B {ram_out} -C {cpu} -D {ram_local} -M {cpu-memcpy} -t {nr_threads} \ + ::: i $(seq 1 5) \ ::: nr_threads 1 2 4 8 12 16 \ - ::: cpu $(seq 0 7) \ - ::: ram $(seq 0 15) + ::: ram_in $(seq 0 15) \ + :::+ cpu_memcpy $(seq 0 7) $(seq 0 7) \ + ::: ram_local $(seq 0 15) \ + :::+ cpu $(seq 0 7) $(seq 0 7) \ + :::+ ram_out $(seq 0 15) \ + ::: input_size 536870912 + +make -B NUMA=1 -echo "single-node execution, DPU ref (2/4)" >&2 +echo "single-node execution, cpu/out on same node (2/3)" >&2 parallel -j1 --eta --joblog ${fn}.2.joblog --resume --header : \ - ./hist -i ${input_size_dpu} -A {ram} -B {ram} -C {cpu} -t {nr_threads} -w 0 -e 5 -x 1 \ + ./hist -i {input_size} -b 256 -A {ram_in} -B {ram_out} -C {cpu} -t {nr_threads} \ + ::: i $(seq 1 5) \ ::: nr_threads 1 2 4 8 12 16 \ - ::: cpu $(seq 0 7) \ - ::: ram $(seq 0 15) + ::: ram_in $(seq 0 15) \ + ::: cpu $(seq 0 7) $(seq 0 7) \ + :::+ ram_out $(seq 0 15) \ + ::: input_size 536870912 -echo "multi-node execution, upstream ref (3/4)" >&2 +echo "multi-node execution (3/3)" >&2 parallel -j1 --eta --joblog ${fn}.3.joblog --resume --header : \ - ./hist -i ${input_size_upstream} -A {ram} -B {ram} -C {cpu} -t {nr_threads} -w 0 -e 40 -x 1 \ + ./hist -i {input_size} -b 256 -A {ram_in} -B {ram_out} -C {cpu} -t {nr_threads} \ + ::: i $(seq 1 20) \ ::: nr_threads 32 48 64 96 128 \ ::: cpu -1 \ - ::: ram $(seq 0 15) - -echo "multi-node execution, DPU ref (4/4)" >&2 - -parallel -j1 --eta --joblog ${fn}.4.joblog --resume --header : \ - ./hist -i ${input_size_dpu} -A {ram} -B {ram} -C {cpu} -t {nr_threads} -w 0 -e 40 -x 1 \ - ::: nr_threads 32 48 64 96 128 \ - ::: cpu -1 \ - ::: ram $(seq 0 15) - -) > ${fn}.txt + ::: ram_in $(seq 0 15) \ + ::: ram_out $(seq 0 15) \ + ::: input_size 536870912 -xz -f -v -9 -M 800M ${fn}.txt +) >> ${fn}.txt |