6 files changed, 44 insertions, 140 deletions
diff --git a/BS/benchmark-scripts/milos-hbm-cxl.sh b/BS/benchmark-scripts/milos-hbm-cxl.sh
new file mode 100755
index 0000000..79d02c7
--- /dev/null
+++ b/BS/benchmark-scripts/milos-hbm-cxl.sh
@@ -0,0 +1,44 @@
+#!/bin/bash
+
+cd baselines/cpu
+make -B numa=1
+
+mkdir -p log/$(hostname)
+fn=log/$(hostname)/milos-hbm-cxl
+
+# * uint64 == 128 MiB
+num_queries_hbm=16777216
+
+run_benchmark() {
+	local "$@"
+	OMP_NUM_THREADS=${nr_threads} ./bs_omp ${input_size} ${num_queries} $ram $cpu 2>&1
+	return $?
+}
+
+export -f run_benchmark
+
+(
+
+echo "single-node execution, HBM ref (1/2)" >&2
+
+# 4 GiB
+parallel -j1 --eta --joblog ${fn}.1.joblog --resume --header : \
+	run_benchmark i={i} nr_threads={nr_threads} ram={ram} cpu={cpu} \
+	input_size=$(perl -E 'say 2 ** 29') num_queries=${num_queries_hbm} \
+	::: i $(seq 1 5) \
+	::: nr_threads 1 2 4 8 12 16 \
+	::: cpu $(seq 0 7) \
+	::: ram $(seq 0 16)
+
+echo "multi-node execution, HBM ref (2/2)" >&2
+
+# 8 GiB
+parallel -j1 --eta --joblog ${fn}.2.joblog --resume --header : \
+	run_benchmark i={i} nr_threads={nr_threads} ram={ram} cpu={cpu} \
+	input_size=$(perl -E 'say 2 ** 30') num_queries=${num_queries_hbm} \
+	::: i $(seq 1 40) \
+	::: nr_threads 32 48 64 96 128 \
+	::: cpu -1 \
+	::: ram $(seq 0 16)
+
+) >> ${fn}.txt
diff --git a/BS/run-fgbs24a.sh b/BS/run-fgbs24a.sh
deleted file mode 100755
index 06f8766..0000000
--- a/BS/run-fgbs24a.sh
+++ /dev/null
@@ -1,33 +0,0 @@
-#!/bin/bash
-
-set -e
-
-mkdir -p $(hostname)
-
-ts=$(date +%Y%m%d)
-
-# 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 BS (dfatool fgbs24a edition)"
-echo "Started at $(date)"
-echo "Revision $(git describe --always)"
-
-for nr_dpus in 2304 2048 2543; do
-	for nr_tasklets in 16; do
-		echo
-		if make -B NR_DPUS=${nr_dpus} NR_TASKLETS=${nr_tasklets} BL=10; then
-			timeout --foreground -k 1m 30m bin/bs_host -w 0 -e 100 -i 16777216 || true
-		fi
-		if make -B NR_DPUS=${nr_dpus} NR_TASKLETS=${nr_tasklets} BL=10 WITH_ALLOC_OVERHEAD=1 WITH_LOAD_OVERHEAD=1 WITH_FREE_OVERHEAD=1; then
-			timeout --foreground -k 1m 30m bin/bs_host -w 0 -e 100 -i 16777216 || true
-		fi
-	done
-done
-echo "Completed at $(date)"
-) | tee "$(hostname)/${ts}-fgbs24a.txt"
diff --git a/BS/run-paper-strong-full.sh b/BS/run-paper-strong-full.sh
deleted file mode 100755
index a6129aa..0000000
--- a/BS/run-paper-strong-full.sh
+++ /dev/null
@@ -1,26 +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 BS strong-full (dfatool edition)"
-echo "Started at $(date)"
-echo "Revision $(git describe --always)"
-
-# >2048 are not part of uptsream
-for nr_dpus in 2543 2304 256 512 1024 2048; do
-	for nr_tasklets in 1 2 4 8 16; do
-		echo
-		if make -B NR_DPUS=${nr_dpus} NR_TASKLETS=${nr_tasklets} BL=10 verbose=1; then
-			timeout --foreground -k 1m 30m bin/bs_host -w 0 -e 100 -i 16777216 || true
-		fi
-	done
-done
-) | tee log-paper-strong-full.txt
diff --git a/BS/run-paper-strong-rank.sh b/BS/run-paper-strong-rank.sh
deleted file mode 100755
index c2d4f36..0000000
--- a/BS/run-paper-strong-rank.sh
+++ /dev/null
@@ -1,25 +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 BS strong-rank (dfatool edition)"
-echo "Started at $(date)"
-echo "Revision $(git describe --always)"
-
-for nr_dpus in 1 4 16 64; do
-	for nr_tasklets in 1 2 4 8 16; do
-		echo
-		if make -B NR_DPUS=${nr_dpus} NR_TASKLETS=${nr_tasklets} BL=10 verbose=1; then
-			timeout --foreground -k 1m 30m bin/bs_host -w 0 -e 100 -i 262144 || true
-		fi
-	done
-done
-) | tee log-paper-strong-rank.txt
diff --git a/BS/run-paper-weak.sh b/BS/run-paper-weak.sh
deleted file mode 100755
index a27c547..0000000
--- a/BS/run-paper-weak.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
-# ... so the weak rank script might be bogus
-
-(
-
-echo "prim-benchmarks BS weak (dfatool edition)"
-echo "Started at $(date)"
-echo "Revision $(git describe --always)"
-
-for nr_dpus in 1 4 16 64; do
-	for nr_tasklets in 1 2 4 8 16; do
-		echo
-		# original Makefile sets PROBLEM_SIZE=2, for some reason.
-		if make -B NR_DPUS=${nr_dpus} NR_TASKLETS=${nr_tasklets} BL=10 verbose=1 PROBLEM_SIZE=2; then
-			i=$(( nr_dpus * 262144 ))
-			timeout --foreground -k 1m 30m bin/bs_host -w 0 -e 100 -i $i || true
-		fi
-	done
-done
-) | tee log-paper-weak.txt
diff --git a/BS/run.sh b/BS/run.sh
deleted file mode 100755
index 0c67c93..0000000
--- a/BS/run.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 BS (dfatool edition)"
-echo "Started at $(date)"
-echo "Revision $(git describe --always)"
-
-for i in 262144 16777216; do
-	for nr_dpus in 1 4 8 16 32 64 128 256 512 768 1024 1536 2048 2304 2542; do
-		for nr_tasklets in 8 12 16; do
-			echo
-			if make -B NR_DPUS=${nr_dpus} NR_TASKLETS=${nr_tasklets} BL=10; then
-				timeout --foreground -k 1m 30m bin/bs_host -w 0 -e 100 -i $i || true
-			fi
-		done
-	done
-done
-echo "Completed at $(date)"
-) | tee "log-$(hostname)-ndpus.txt"