CPU-DPU microbenchmarks: explicitly log number of instructions

4 files changed, 12 insertions, 8 deletions

diff --git a/Microbenchmarks/CPU-DPU/host/app.c b/Microbenchmarks/CPU-DPU/host/app.c
index bb86d09..ba63d89 100644
--- a/Microbenchmarks/CPU-DPU/host/app.c
+++ b/Microbenchmarks/CPU-DPU/host/app.c
@@ -103,9 +103,9 @@ int main(int argc, char **argv) {
     read_input(A, B, input_size);
 
     //printf("NR_TASKLETS\t%d\tBL\t%d\n", NR_TASKLETS, BL);
-    printf("[::] NMC reconfiguration | n_dpus=%d n_ranks=%d n_tasklets=%d n_nops=%d e_type=%s n_elements=%u e_mode=%s"
+    printf("[::] NMC reconfiguration | n_dpus=%d n_ranks=%d n_tasklets=%d n_nops=%d n_instr=%d e_type=%s n_elements=%u e_mode=%s"
         " | latency_dpu_alloc_ns=%lu latency_dpu_load_ns=%lu latency_dpu_get_ns=%lu\n",
-        nr_of_dpus, nr_of_ranks, NR_TASKLETS, p.n_nops, XSTR(T), transfer_size, transfer_mode,
+        nr_of_dpus, nr_of_ranks, NR_TASKLETS, p.n_nops, p.n_instr, XSTR(T), transfer_size, transfer_mode,
         timer.nanoseconds[4], timer.nanoseconds[5], timer.nanoseconds[6]);
 
     // Loop over main kernel
@@ -170,9 +170,9 @@ int main(int argc, char **argv) {
             stop(&timer, 3);
 
         if (rep >= p.n_warmup) {
-            printf("[::] transfer UPMEM | n_dpus=%d n_ranks=%d n_tasklets=%d n_nops=%d e_type=%s n_elements=%u e_mode=%s"
+            printf("[::] transfer UPMEM | n_dpus=%d n_ranks=%d n_tasklets=%d n_nops=%d n_instr=%d e_type=%s n_elements=%u e_mode=%s"
                 " | latency_dram_mram_ns=%lu latency_mram_dram_ns=%lu throughput_dram_mram_Bps=%f throughput_mram_dram_Bps=%f",
-                nr_of_dpus, nr_of_ranks, NR_TASKLETS, p.n_nops, XSTR(T), transfer_size, transfer_mode,
+                nr_of_dpus, nr_of_ranks, NR_TASKLETS, p.n_nops, p.n_instr, XSTR(T), transfer_size, transfer_mode,
                 timer.nanoseconds[1], timer.nanoseconds[3],
                 transfer_size * sizeof(T) * 1e9 / timer.nanoseconds[1],
                 transfer_size * sizeof(T) * 1e9 / timer.nanoseconds[3]);
diff --git a/Microbenchmarks/CPU-DPU/run-alloc-rank-stress.sh b/Microbenchmarks/CPU-DPU/run-alloc-rank-stress.sh
index da136c0..e166a4e 100755
--- a/Microbenchmarks/CPU-DPU/run-alloc-rank-stress.sh
+++ b/Microbenchmarks/CPU-DPU/run-alloc-rank-stress.sh
@@ -5,6 +5,8 @@ mkdir -p "$(hostname)-alloc"
 NCORES=$(grep -c '^processor' /proc/cpuinfo)
 cleanexit() {
 	pkill -f "stress -c ${NCORES}"
+	xz -f -v -9 -M 800M "$(hostname)-alloc/rank-stress-c${NCORES}.txt"
+	exit 0
 }
 
 trap cleanexit TERM INT
@@ -14,5 +16,3 @@ stress -c ${NCORES} &
 ./run-alloc-rank.sh | tee "$(hostname)-alloc/rank-stress-c${NCORES}.txt"
 
 cleanexit
-
-xz -f -v -9 -M 800M "$(hostname)-alloc/rank-stress-c${NCORES}.txt"
diff --git a/Microbenchmarks/CPU-DPU/run-alloc-rank.sh b/Microbenchmarks/CPU-DPU/run-alloc-rank.sh
index a6907fe..2e20e52 100755
--- a/Microbenchmarks/CPU-DPU/run-alloc-rank.sh
+++ b/Microbenchmarks/CPU-DPU/run-alloc-rank.sh
@@ -16,7 +16,7 @@ for i in 1 4 8 16 32 48 64; do
 		n_nops=$((j * 256))
 		if make -B NR_DPUS=$i NR_TASKLETS=1 BL=10 DPU_BINARY=\'\"bin/dpu_size\"\'; then
 			for l in $(seq 1 100); do
-				bin/host_code -w 1 -e 0 -x 1 -i 65536 -N $n_nops || true
+				bin/host_code -w 1 -e 0 -x 1 -i 65536 -N $n_nops -I $(size -A bin/dpu_size | awk '($1 == ".text") {print $2/8}') || true
 			done
 		fi
 	done
diff --git a/Microbenchmarks/CPU-DPU/support/params.h b/Microbenchmarks/CPU-DPU/support/params.h
index 1ecf71d..3bb1535 100644
--- a/Microbenchmarks/CPU-DPU/support/params.h
+++ b/Microbenchmarks/CPU-DPU/support/params.h
@@ -7,6 +7,7 @@ typedef struct Params {
     unsigned int   input_size;
     unsigned int   n_threads;
     unsigned int   n_nops;
+    unsigned int   n_instr;
     int   n_warmup;
     int   n_reps;
     int  exp;
@@ -26,6 +27,7 @@ static void usage() {
         "\n    -i <I>    input size (default=8K elements)"
         "\n    -n <N>    number of threads per pool (default=8)"
         "\n    -N <N>    number of nops in dpu task (default=0)"
+        "\n    -I <N>    number of instructions in dpu binary (default=0)"
         "\n");
 }
 
@@ -34,12 +36,13 @@ struct Params input_params(int argc, char **argv) {
     p.input_size    = 8 << 10;
     p.n_threads     = 8;
     p.n_nops        = 0;
+    p.n_instr       = 0;
     p.n_warmup      = 1;
     p.n_reps        = 3;
     p.exp           = 0;
 
     int opt;
-    while((opt = getopt(argc, argv, "hi:n:w:e:x:N:")) >= 0) {
+    while((opt = getopt(argc, argv, "hi:n:w:e:x:N:I:")) >= 0) {
         switch(opt) {
         case 'h':
         usage();
@@ -48,6 +51,7 @@ struct Params input_params(int argc, char **argv) {
         case 'i': p.input_size    = atoi(optarg); break;
         case 'n': p.n_threads     = atoi(optarg); break;
         case 'N': p.n_nops        = atoi(optarg); break;
+        case 'I': p.n_instr       = atoi(optarg); break;
         case 'w': p.n_warmup      = atoi(optarg); break;
         case 'e': p.n_reps        = atoi(optarg); break;
         case 'x': p.exp           = atoi(optarg); break;