VA baseline: configurable memcpy NUMA binding

3 files changed, 50 insertions, 18 deletions

diff --git a/VA/baselines/cpu/app_baseline.c b/VA/baselines/cpu/app_baseline.c
index 0451079..4c8610a 100644
--- a/VA/baselines/cpu/app_baseline.c
+++ b/VA/baselines/cpu/app_baseline.c
@@ -39,6 +39,7 @@ static T *B;
 static T *C;
 
 #if NUMA_MEMCPY
+int numa_node_cpu_memcpy = -1;
 int numa_node_local = -1;
 int numa_node_in_is_local = 0;
 static T *A_local;
@@ -73,6 +74,7 @@ typedef struct Params {
     int numa_node_cpu;
 #endif
 #if NUMA_MEMCPY
+    int numa_node_cpu_memcpy;
     struct bitmask* bitmask_cpu;
 #endif
 }Params;
@@ -106,11 +108,12 @@ struct Params input_params(int argc, char **argv) {
     p.numa_node_cpu  = -1;
 #endif
 #if NUMA_MEMCPY
+    p.numa_node_cpu_memcpy  = -1;
     p.bitmask_cpu    = NULL;
 #endif
 
     int opt;
-    while((opt = getopt(argc, argv, "hi:w:e:x:t:a:b:c:C:")) >= 0) {
+    while((opt = getopt(argc, argv, "hi:w:e:x:t:a:b:c:C:M:")) >= 0) {
         switch(opt) {
         case 'h':
         usage();
@@ -127,6 +130,7 @@ struct Params input_params(int argc, char **argv) {
         case 'c': p.numa_node_cpu = atoi(optarg); break;
 #if NUMA_MEMCPY
         case 'C': p.bitmask_cpu   = numa_parse_nodestring(optarg); break;
+        case 'M': p.numa_node_cpu_memcpy = atoi(optarg); break;
 #endif // NUMA_MEMCPY
 #endif // NUMA
         default:
@@ -222,8 +226,8 @@ int main(int argc, char **argv) {
     }
 
     numa_node_cpu = p.numa_node_cpu;
-    if (numa_node_cpu != -1) {
-        if (numa_run_on_node(numa_node_cpu) == -1) {
+    if (p.numa_node_cpu != -1) {
+        if (numa_run_on_node(p.numa_node_cpu) == -1) {
             perror("numa_run_on_node");
             numa_node_cpu = -1;
         }
@@ -239,12 +243,21 @@ int main(int argc, char **argv) {
     for(int rep = 0; rep < p.n_warmup + p.n_reps; rep++) {
 
 #if NUMA_MEMCPY
+        numa_node_cpu_memcpy = p.numa_node_cpu_memcpy;
         start(&timer, 1, 0);
         if (!numa_node_in_is_local) {
             A_local = (T*) numa_alloc(input_size * sizeof(T));
             B_local = (T*) numa_alloc(input_size * sizeof(T));
         }
         stop(&timer, 1);
+        if (!numa_node_in_is_local) {
+            if (p.numa_node_cpu_memcpy != -1) {
+                if (numa_run_on_node(p.numa_node_cpu_memcpy) == -1) {
+                    perror("numa_run_on_node");
+                    numa_node_cpu_memcpy = -1;
+                }
+            }
+        }
         start(&timer, 2, 0);
         if (!numa_node_in_is_local) {
             memcpy(A_local, A, input_size * sizeof(T));
@@ -254,6 +267,12 @@ int main(int argc, char **argv) {
             B_local = B;
         }
         stop(&timer, 2);
+        if (p.numa_node_cpu != -1) {
+            if (numa_run_on_node(p.numa_node_cpu) == -1) {
+                perror("numa_run_on_node");
+                numa_node_cpu = -1;
+            }
+        }
         mp_pages[0] = A_local;
         if (move_pages(0, 1, mp_pages, NULL, mp_status, 0) == -1) {
             perror("move_pages(A_local)");
@@ -287,10 +306,10 @@ int main(int argc, char **argv) {
         if (rep >= p.n_warmup) {
 #if NUMA_MEMCPY
             printf("[::] VA-CPU-MEMCPY | n_threads=%d e_type=%s n_elements=%d"
-                " numa_node_in=%d numa_node_local=%d numa_node_out=%d numa_node_cpu=%d numa_distance_in_cpu=%d numa_distance_cpu_out=%d"
+                " numa_node_in=%d numa_node_local=%d numa_node_out=%d numa_node_cpu=%d numa_node_cpu_memcpy=%d numa_distance_in_cpu=%d numa_distance_cpu_out=%d"
                 " | throughput_MBps=%f",
                 nr_threads, XSTR(T), input_size,
-                numa_node_in, numa_node_local, numa_node_out, numa_node_cpu, numa_distance(numa_node_in, numa_node_cpu), numa_distance(numa_node_cpu, numa_node_out),
+                numa_node_in, numa_node_local, numa_node_out, numa_node_cpu, numa_node_cpu_memcpy, numa_distance(numa_node_in, numa_node_cpu), numa_distance(numa_node_cpu, numa_node_out),
                 input_size * 3 * sizeof(T) / timer.time[0]);
             printf(" throughput_MOpps=%f",
                 input_size / timer.time[0]);
diff --git a/VA/dimes-hetsim-hbm.sh b/VA/dimes-hetsim-hbm.sh
index 7c0df73..d260c1e 100755
--- a/VA/dimes-hetsim-hbm.sh
+++ b/VA/dimes-hetsim-hbm.sh
@@ -11,15 +11,16 @@ fn=log/$(hostname)/dimes-hetsim-hbm
 
 make -B NUMA=1 NUMA_MEMCPY=1
 
-echo "CPU single-node operation with setup cost, cpu/out on same node (1/4)" >&2
+echo "CPU single-node operation with setup cost, memcpy node == input node, cpu node == output node (1/4)" >&2
 
 parallel -j1 --eta --joblog ${fn}.1.joblog --resume --header : \
-	./va -i {input_size} -a {ram_in} -b {ram_out} -c {cpu} -C {ram_local} -t {nr_threads} -w 0 -e 40 \
+	./va -i {input_size} -a {ram_in} -b {ram_out} -c {cpu} -C {ram_local} -M {cpu_memcpy} -t {nr_threads} -w 0 -e 20 \
 	::: nr_threads 1 2 4 8 12 16 \
-	::: ram_in $(seq 0 15) \
-	:::        cpu $(seq 0 7) $(seq 0 7) \
-	:::+ ram_local $(seq 0 15) \
-	:::+   ram_out $(seq 0 15) \
+	:::      ram_in $(seq 0 15) \
+	:::+ cpu_memcpy $(seq 0 7) $(seq 0 7) \
+	:::   ram_local $(seq 0 15) \
+	:::+        cpu $(seq 0 7) $(seq 0 7) \
+	:::+    ram_out $(seq 0 15) \
 	::: input_size 167772160
 
 make -B NUMA=1
@@ -47,7 +48,7 @@ parallel -j1 --eta --joblog ${fn}.3.joblog --resume --header : \
 echo "multi-node execution (4/4)" >&2
 
 parallel -j1 --eta --joblog ${fn}.4.joblog --resume --header : \
-	./va -i {input_size} -a {ram_in} -b {ram_out} -c {cpu} -t {nr_threads} -w 0 -e 40 \
+	./va -i {input_size} -a {ram_in} -b {ram_out} -c {cpu} -t {nr_threads} -w 0 -e 20 \
 	::: nr_threads 32 48 64 96 128 \
 	::: cpu -1 \
 	::: ram_in $(seq 0 15) \
diff --git a/VA/dimes-hetsim-nmc.sh b/VA/dimes-hetsim-nmc.sh
index cabae57..a176027 100755
--- a/VA/dimes-hetsim-nmc.sh
+++ b/VA/dimes-hetsim-nmc.sh
@@ -40,20 +40,20 @@ make -B NUMA=1
 
 (
 
-echo "CPU single-node operation (1/3)" >&2
+echo "CPU single-node operation (1/4)" >&2
 
 parallel -j1 --eta --joblog ${fn}.1.joblog --resume --header : \
-	./va -i {input_size} -a {ram_in} -b {ram_out} -c {cpu} -t {nr_threads} -w 0 -e 40 \
+	./va -i {input_size} -a {ram_in} -b {ram_out} -c {cpu} -t {nr_threads} -w 0 -e 20 \
 	::: ram_in 0 1 \
 	::: ram_out 0 1 \
 	::: cpu 0 1 \
 	::: nr_threads 1 2 4 8 12 16 \
 	::: input_size 167772160
 
-echo "CPU multi-node operation (2/3)" >&2
+echo "CPU multi-node operation (2/4)" >&2
 
 parallel -j1 --eta --joblog ${fn}.2.joblog --resume --header : \
-	./va -i {input_size} -a {ram_in} -b {ram_out} -c {cpu} -t {nr_threads} -w 0 -e 40 \
+	./va -i {input_size} -a {ram_in} -b {ram_out} -c {cpu} -t {nr_threads} -w 0 -e 20 \
 	::: ram_in 0 1 \
 	::: ram_out 0 1 \
 	::: cpu -1 \
@@ -66,11 +66,23 @@ make -B NUMA=1 NUMA_MEMCPY=1
 
 (
 
-echo "CPU single-node operation with setup cost, cpu/out on same node (3/3)" >&2
+echo "CPU single-node operation with setup cost, memcpy node == input node, cpu node == output node (3/4)" >&2
 
 parallel -j1 --eta --joblog ${fn}.3.joblog --resume --header : \
-	./va -i {input_size} -a {ram_in} -b {ram_out} -c {cpu} -C {cpu} -t {nr_threads} -w 0 -e 40 \
+	./va -i {input_size} -a {ram_in} -b {ram_out} -c {cpu} -C {cpu} -M {cpu_memcpy} -t {nr_threads} -w 0 -e 20 \
+	:::      ram_in 0 1 \
+	:::+ cpu_memcpy 0 1 \
+	:::      cpu 0 1 \
+	:::+ ram_out 0 1 \
+	::: nr_threads 1 2 4 8 12 16 \
+	::: input_size 167772160
+
+echo "CPU single-node operation with setup cost, memcpy node == 0, cpu node == output node (4/4)" >&2
+
+parallel -j1 --eta --joblog ${fn}.4.joblog --resume --header : \
+	./va -i {input_size} -a {ram_in} -b {ram_out} -c {cpu} -C {cpu} -M {cpu_memcpy} -t {nr_threads} -w 0 -e 20 \
 	::: ram_in 0 1 \
+	::: cpu_memcpy 0 \
 	:::      cpu 0 1 \
 	:::+ ram_out 0 1 \
 	::: nr_threads 1 2 4 8 12 16 \