summaryrefslogtreecommitdiff
diff options
context:
space:
mode:
-rw-r--r--HST-S/baselines/cpu/app_baseline.c2
-rwxr-xr-xHST-S/dimes-hetsim-hbm.sh51
-rwxr-xr-xHST-S/dimes-hetsim-nmc.sh103
-rwxr-xr-xHST-S/dimes-hetsim.sh20
4 files changed, 155 insertions, 21 deletions
diff --git a/HST-S/baselines/cpu/app_baseline.c b/HST-S/baselines/cpu/app_baseline.c
index 0c766e0..0e16944 100644
--- a/HST-S/baselines/cpu/app_baseline.c
+++ b/HST-S/baselines/cpu/app_baseline.c
@@ -288,7 +288,7 @@ int main(int argc, char **argv) {
#pragma omp atomic
nr_threads++;
- printf("[::] HST-S CPU | n_threads=%d e_type=%s n_elements=%d n_bins=%d"
+ printf("[::] HST-S-CPU | n_threads=%d e_type=%s n_elements=%d n_bins=%d"
#if NUMA
" numa_node_in=%d numa_node_out=%d numa_node_cpu=%d numa_distance_in_cpu=%d numa_distance_cpu_out=%d"
#endif
diff --git a/HST-S/dimes-hetsim-hbm.sh b/HST-S/dimes-hetsim-hbm.sh
new file mode 100755
index 0000000..08842b7
--- /dev/null
+++ b/HST-S/dimes-hetsim-hbm.sh
@@ -0,0 +1,51 @@
+#!/bin/bash
+
+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))
+
+# Here: 2 GiB
+input_size_dpu=$((2**29))
+
+(
+
+echo "single-node execution, upstream ref (1/4)" >&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 \
+ ::: nr_threads 1 2 4 8 12 16 \
+ ::: cpu $(seq 0 7) \
+ ::: ram $(seq 0 15)
+
+echo "single-node execution, DPU ref (2/4)" >&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 \
+ ::: nr_threads 1 2 4 8 12 16 \
+ ::: cpu $(seq 0 7) \
+ ::: ram $(seq 0 15)
+
+echo "multi-node execution, upstream ref (3/4)" >&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 \
+ ::: 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
+
+xz -f -v -9 -M 800M ${fn}.txt
diff --git a/HST-S/dimes-hetsim-nmc.sh b/HST-S/dimes-hetsim-nmc.sh
new file mode 100755
index 0000000..b3f0669
--- /dev/null
+++ b/HST-S/dimes-hetsim-nmc.sh
@@ -0,0 +1,103 @@
+#!/bin/bash
+
+mkdir -p log/$(hostname) baselines/cpu/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))
+
+# Here: 2 GiB
+input_size_dpu=$((2**29))
+
+run_benchmark_nmc() {
+ local "$@"
+ sudo limit_ranks_to_numa_node ${numa_rank}
+ if make -B NR_DPUS=${nr_dpus} NR_TASKLETS=${nr_tasklets} BL=10 WITH_ALLOC_OVERHEAD=1 WITH_LOAD_OVERHEAD=1 WITH_FREE_OVERHEAD=1; then
+ bin/host_code -w 0 -e 100 -b 256 -x 1 -i ${input_size}
+ fi
+ return $?
+}
+
+export -f run_benchmark_nmc
+
+(
+
+echo "NMC single-node upstream-ref (1/4)" >&2
+
+parallel -j1 --eta --joblog ${fn}.1.joblog --resume --header : \
+ run_benchmark_nmc nr_dpus={nr_dpus} nr_tasklets=16 numa_rank={numa_rank} \
+ input_size=${input_size_upstream} \
+ ::: numa_rank 0 1 \
+ ::: nr_dpus 64 128 256 512 768 1024
+
+echo "NMC multi-node upstream-ref (2/4)" >&2
+
+parallel -j1 --eta --joblog ${fn}.2.joblog --resume --header : \
+ run_benchmark_nmc nr_dpus={nr_dpus} nr_tasklets=16 numa_rank={numa_rank} \
+ input_size=${input_size_upstream} \
+ ::: numa_rank -1 \
+ ::: nr_dpus 1536 2048 2304
+
+echo "NMC single-node DPU-ref (3/4)" >&2
+
+parallel -j1 --eta --joblog ${fn}.3.joblog --resume --header : \
+ run_benchmark_nmc nr_dpus={nr_dpus} nr_tasklets=16 numa_rank={numa_rank} \
+ input_size=${input_size_dpu} \
+ ::: numa_rank 0 1 \
+ ::: nr_dpus 64 128 256 512 768 1024
+
+echo "NMC multi-node DPU-ref (4/4)" >&2
+
+parallel -j1 --eta --joblog ${fn}.4.joblog --resume --header : \
+ run_benchmark_nmc nr_dpus={nr_dpus} nr_tasklets=16 numa_rank={numa_rank} \
+ input_size=${input_size_dpu} \
+ ::: 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 upstream-ref (1/4)" >&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 20 -x 1 \
+ ::: cpu 0 1 \
+ ::: ram 0 1 \
+ ::: nr_threads 1 2 4 8 12 16 32
+
+echo "CPU single-node DPU-ref (2/4)" >&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 20 -x 1 \
+ ::: i $(seq 1 20) \
+ ::: cpu 0 1 \
+ ::: ram 0 1 \
+ ::: nr_threads 1 2 4 8 12 16 32
+
+echo "CPU multi-node upstream-ref (3/4)" >&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 20 -x 1 \
+ ::: i $(seq 1 20) \
+ ::: cpu -1 \
+ ::: ram 0 1 \
+ ::: nr_threads 48 64
+
+echo "CPU multi-node 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 20 -x 1 \
+ ::: i $(seq 1 20) \
+ ::: cpu -1 \
+ ::: ram 0 1 \
+ ::: nr_threads 48 64
+
+) > ${fn}.txt
+
+xz -f -v -9 -M 800M ${fn}.txt
diff --git a/HST-S/dimes-hetsim.sh b/HST-S/dimes-hetsim.sh
deleted file mode 100755
index 54b660d..0000000
--- a/HST-S/dimes-hetsim.sh
+++ /dev/null
@@ -1,20 +0,0 @@
-#!/bin/sh
-
-cd baselines/cpu
-make -B NUMA=1
-
-# upstream uses 1006632960 uint32 elements == 3.75 GiB
-# This is way larger than the input file...
-
-for i in `seq 1 10`; do
- for nr_threads in 1 2 4 8 12 16; do
- for cpu in 0 1 2 3 4 5 6 7; do
- for ram in 0 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15; do
- for j in `seq 1 10`; do
- ./hist -A $ram -B $ram -C $cpu -t $nr_threads -w 0 -e 100 -x 2
- done
- ./hist -i 1006632960 -A $ram -B $ram -C $cpu -t $nr_threads -w 0 -e 20 -x 2
- done
- done
- done
-done