summaryrefslogtreecommitdiff
path: root/TS
diff options
context:
space:
mode:
Diffstat (limited to 'TS')
-rw-r--r--TS/Makefile2
-rwxr-xr-xTS/benchmark-scripts/ccmcc25-sim.sh25
-rwxr-xr-xTS/benchmark-scripts/ccmcc25.sh23
-rw-r--r--TS/host/app.c30
-rwxr-xr-xTS/run-paper-strong-full.sh29
-rwxr-xr-xTS/run-paper-strong-rank.sh28
-rwxr-xr-xTS/run-paper-weak.sh29
7 files changed, 53 insertions, 113 deletions
diff --git a/TS/Makefile b/TS/Makefile
index 8eedaf9..2fce611 100644
--- a/TS/Makefile
+++ b/TS/Makefile
@@ -12,6 +12,8 @@ aspectc ?= 0
aspectc_timing ?= 0
dfatool_timing ?= 1
+HOST_CC := ${CC}
+
COMMON_FLAGS := -Wall -Wextra -g -Iinclude -DBL=${BL}
HOST_FLAGS := ${COMMON_FLAGS} -O3 `dpu-pkg-config --cflags --libs dpu` -DNR_TASKLETS=${NR_TASKLETS} -DNR_DPUS=${NR_DPUS} -DWITH_ALLOC_OVERHEAD=${WITH_ALLOC_OVERHEAD} -DWITH_LOAD_OVERHEAD=${WITH_LOAD_OVERHEAD} -DWITH_FREE_OVERHEAD=${WITH_FREE_OVERHEAD} -DDFATOOL_TIMING=${dfatool_timing} -DASPECTC=${aspectc} -lm
DPU_FLAGS := ${COMMON_FLAGS} -O2 -DNR_TASKLETS=${NR_TASKLETS}
diff --git a/TS/benchmark-scripts/ccmcc25-sim.sh b/TS/benchmark-scripts/ccmcc25-sim.sh
new file mode 100755
index 0000000..0df03d9
--- /dev/null
+++ b/TS/benchmark-scripts/ccmcc25-sim.sh
@@ -0,0 +1,25 @@
+#!/bin/bash
+
+mkdir -p log/$(hostname)
+
+run_benchmark_nmc() {
+ local "$@"
+ set -e
+ make -B NR_DPUS=${nr_dpus} NR_TASKLETS=${nr_tasklets} BL=8 \
+ aspectc=1 aspectc_timing=1 dfatool_timing=0
+ bin/ts_host -w 0 -e 5 -n ${ts_size} 2>&1
+}
+
+export -f run_benchmark_nmc
+
+fn=log/$(hostname)/ccmcc25-sim
+
+source ~/lib/local/upmem/upmem-2025.1.0-Linux-x86_64/upmem_env.sh simulator
+
+echo "prim-benchmarks TS $(git describe --all --long) $(git rev-parse HEAD) $(date -R)" >> ${fn}.txt
+
+parallel -j1 --eta --joblog ${fn}.joblog --resume --header : \
+ run_benchmark_nmc nr_dpus={nr_dpus} nr_tasklets=8 ts_size={ts_size} \
+ ::: nr_dpus 1 2 4 8 16 32 48 64 \
+ ::: ts_size 2048 4096 8192 16384 32768 \
+>> ${fn}.txt
diff --git a/TS/benchmark-scripts/ccmcc25.sh b/TS/benchmark-scripts/ccmcc25.sh
index c7f93c3..74c8371 100755
--- a/TS/benchmark-scripts/ccmcc25.sh
+++ b/TS/benchmark-scripts/ccmcc25.sh
@@ -1,9 +1,6 @@
#!/bin/bash
mkdir -p log/$(hostname)
-fn=log/$(hostname)/ccmcc25
-
-source /opt/upmem/upmem-2025.1.0-Linux-x86_64/upmem_env.sh
run_benchmark_nmc() {
local "$@"
@@ -16,10 +13,18 @@ run_benchmark_nmc() {
export -f run_benchmark_nmc
-echo "prim-benchmarks TS $(git describe --all --long) $(git rev-parse HEAD) $(date -R)" >> ${fn}.txt
+for sdk in 2023.2.0 2024.1.0 2024.2.0 2025.1.0; do
+
+ fn=log/$(hostname)/ccmcc25-sdk${sdk}
+
+ source /opt/upmem/upmem-${sdk}-Linux-x86_64/upmem_env.sh
+
+ echo "prim-benchmarks TS $(git describe --all --long) $(git rev-parse HEAD) $(date -R)" >> ${fn}.txt
+
+ parallel -j1 --eta --joblog ${fn}.joblog --resume --header : \
+ run_benchmark_nmc nr_dpus={nr_dpus} nr_tasklets=8 numa_rank=any ts_size={ts_size} \
+ ::: nr_dpus 64 128 256 512 768 1024 1536 2048 2304 \
+ ::: ts_size 8388608 16777216 33554432 67108864 \
+ >> ${fn}.txt
-parallel -j1 --eta --joblog ${fn}.joblog --resume --header : \
- run_benchmark_nmc nr_dpus={nr_dpus} nr_tasklets=8 numa_rank=any ts_size={ts_size} \
- ::: nr_dpus 64 128 256 512 768 1024 1536 2048 2304 \
- ::: ts_size 8388608 16777216 33554432 67108864 \
->> ${fn}.txt
+done
diff --git a/TS/host/app.c b/TS/host/app.c
index a75cfe9..bfa14df 100644
--- a/TS/host/app.c
+++ b/TS/host/app.c
@@ -224,6 +224,15 @@ int main(int argc, char **argv)
for (int rep = 0; rep < p.n_warmup + p.n_reps; rep++) {
+ if (rep >= p.n_warmup) {
+ start(&timer, 6, 0);
+ }
+ streamp(tSeries, AMean, ASigma, ts_size - query_length - 1,
+ query, query_length, query_mean, query_std);
+ if (rep >= p.n_warmup) {
+ stop(&timer, 6);
+ }
+
#if WITH_ALLOC_OVERHEAD
if (rep >= p.n_warmup) {
start(&timer, 0, 0);
@@ -250,16 +259,10 @@ int main(int argc, char **argv)
start(&timer, 2, 0);
}
uint32_t i = 0;
-
- DPU_FOREACH(dpu_set, dpu) {
- input_arguments.exclusion_zone = 0;
-
- DPU_ASSERT(dpu_copy_to
- (dpu, "DPU_INPUT_ARGUMENTS", 0,
- (const void *)&input_arguments,
- sizeof(input_arguments)));
- i++;
+ DPU_FOREACH(dpu_set, dpu, i) {
+ DPU_ASSERT(dpu_prepare_xfer(dpu, &input_arguments));
}
+ DPU_ASSERT(dpu_push_xfer(dpu_set, DPU_XFER_TO_DPU, "DPU_INPUT_ARGUMENTS", 0, sizeof(input_arguments), DPU_XFER_DEFAULT));
i = 0;
mem_offset = 0;
@@ -402,15 +405,6 @@ int main(int argc, char **argv)
#endif
#endif
- if (rep >= p.n_warmup) {
- start(&timer, 6, 0);
- }
- streamp(tSeries, AMean, ASigma, ts_size - query_length - 1,
- query, query_length, query_mean, query_std);
- if (rep >= p.n_warmup) {
- stop(&timer, 6);
- }
-
int status = (minHost == result.minValue);
if (status) {
printf("[" ANSI_COLOR_GREEN "OK" ANSI_COLOR_RESET
diff --git a/TS/run-paper-strong-full.sh b/TS/run-paper-strong-full.sh
deleted file mode 100755
index 5b7656d..0000000
--- a/TS/run-paper-strong-full.sh
+++ /dev/null
@@ -1,29 +0,0 @@
-#!/bin/bash
-
-set -e
-
-# BL: use 2^(BL) B blocks for MRAM <-> WRAM transfers on PIM module
-# T: data type
-# -w: number of un-timed warmup iterations
-# -e: number of timed iterations
-# -i; ignored, always uses 262144 elements
-
-(
-
-echo "prim-benchmarks TS strong-full (dfatool edition)"
-echo "Started at $(date)"
-echo "Revision $(git describe --always)"
-
-# >2048 is not part of upstream
-# 12 tasklets are not part of upstream (code does not work with 16…)
-for nr_dpus in 2543 2304 256 512 1024 2048; do
- for nr_tasklets in 1 2 4 8 12 16; do
- echo
- # upstream code did not respect $BL in the makefile and used 256B (BL=8) instead.
- # This appears to be faster than BL=10.
- if make -B NR_DPUS=${nr_dpus} NR_TASKLETS=${nr_tasklets} BL=8; then
- timeout --foreground -k 1m 30m bin/ts_host -w 0 -e 100 -n 33554432 || true
- fi
- done
-done
-) | tee log-paper-strong-full.txt
diff --git a/TS/run-paper-strong-rank.sh b/TS/run-paper-strong-rank.sh
deleted file mode 100755
index 58ad641..0000000
--- a/TS/run-paper-strong-rank.sh
+++ /dev/null
@@ -1,28 +0,0 @@
-#!/bin/bash
-
-set -e
-
-# BL: use 2^(BL) B blocks for MRAM <-> WRAM transfers on PIM module
-# T: data type
-# -w: number of un-timed warmup iterations
-# -e: number of timed iterations
-# -i; ignored, always uses 262144 elements
-
-(
-
-echo "prim-benchmarks TS strong-rank (dfatool edition)"
-echo "Started at $(date)"
-echo "Revision $(git describe --always)"
-
-# >64 are not part of upstream config space
-for nr_dpus in 128 1 4 16 64; do
- for nr_tasklets in 1 2 4 8 16; do
- echo
- # upstream code did not respect $BL in the makefile and used 256B (BL=8) instead.
- # BL=10 appears to be slightly faster.
- if make -B NR_DPUS=${nr_dpus} NR_TASKLETS=${nr_tasklets} BL=10; then
- timeout --foreground -k 1m 60m bin/ts_host -w 0 -e 50 -n 524288 || true
- fi
- done
-done
-) | tee log-paper-strong-rank.txt
diff --git a/TS/run-paper-weak.sh b/TS/run-paper-weak.sh
deleted file mode 100755
index 64892f4..0000000
--- a/TS/run-paper-weak.sh
+++ /dev/null
@@ -1,29 +0,0 @@
-#!/bin/bash
-
-set -e
-
-# BL: use 2^(BL) B blocks for MRAM <-> WRAM transfers on PIM module
-# T: data type
-# -w: number of un-timed warmup iterations
-# -e: number of timed iterations
-# -i; ignored, always uses 262144 elements
-
-(
-
-echo "prim-benchmarks TS weak (dfatool edition)"
-echo "Started at $(date)"
-echo "Revision $(git describe --always)"
-
-# 256 and 512 are not part of upstream
-for nr_dpus in 1 4 16 64; do
- for nr_tasklets in 1 2 4 8 16; do
- echo
- # upstream code did not respect $BL in the makefile and used 256B (BL=8) instead.
- # BL=10 appears to be slightly faster.
- if make -B NR_DPUS=${nr_dpus} NR_TASKLETS=${nr_tasklets} BL=8; then
- i=$(( nr_dpus * 524288 ))
- timeout --foreground -k 1m 30m bin/ts_host -w 0 -e 100 -n $i || true
- fi
- done
-done
-) | tee log-paper-weak.txt
generated by cgit v1.2.3 (git 2.46.0) at 2025-12-22 02:16:53 +0000