STREAM: Add alloc/load overhead, n_ranks, and n_elements

1 files changed, 12 insertions, 8 deletions

diff --git a/Microbenchmarks/STREAM/run.sh b/Microbenchmarks/STREAM/run.sh
index 94b759b..778efa8 100755
--- a/Microbenchmarks/STREAM/run.sh
+++ b/Microbenchmarks/STREAM/run.sh
@@ -9,22 +9,26 @@ set -e
 # -i: input size (number of elements, not number of bytes!)
 # 2097152 B -> 2M is maximum for 64bit types (due to 16M per DPU)
 
+(
+
 echo "prim-benchmarks STREAM microbenchmark (dfatool edition)"
 echo "Started at $(date)"
 echo "Revision $(git describe --always)"
 
-for nr_dpus in 1 2 4 8 16 32 64 128 256 512; do
-	for mem in MRAM WRAM; do
-		for op in copy copyw add scale triad; do
-			for nr_tasklets in 1 2 3 4 6 8 10 12 16 20 24; do
-				for dt in uint8_t uint16_t uint32_t uint64_t float double; do
+for i in 4096 16384 131072 1048576 2097152; 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
+			for dt in uint8_t uint16_t uint32_t uint64_t float double; do
+				for op in copy copyw add scale triad; do
 					echo
-					if make -B MEM=${mem} OP=${op} NR_DPUS=${nr_dpus} NR_TASKLETS=${nr_tasklets} BL=10 T=${dt} UNROLL=1 \
-						|| make -B MEM=${mem} OP=${op} NR_DPUS=${nr_dpus} NR_TASKLETS=${nr_tasklets} BL=10 T=${dt} UNROLL=0; then
-						timeout -k 1m 30m bin/host_code -w 0 -e 20 -i 2097152 || true
+					if make -B OP=${op} NR_DPUS=${nr_dpus} NR_TASKLETS=${nr_tasklets} BL=10 T=${dt} UNROLL=1 WITH_ALLOC_OVERHEAD=1 WITH_LOAD_OVERHEAD=1 WITH_FREE_OVERHEAD=1 \
+						|| make -B OP=${op} NR_DPUS=${nr_dpus} NR_TASKLETS=${nr_tasklets} BL=10 T=${dt} UNROLL=0 WITH_ALLOC_OVERHEAD=1 WITH_LOAD_OVERHEAD=1 WITH_FREE_OVERHEAD=1; then
+						timeout --foreground -k 1m 30m bin/host_code -w 0 -e 40 -i $i || true
 					fi
 				done
 			done
 		done
 	done
 done
+echo "Completed at $(date)"
+) | tee "log-$(hostname)-ndpus.txt"