Track different perf counters. Colorbench: random in the entire arena.

2 files changed, 88 insertions, 31 deletions

diff --git a/bin/colorbench.c b/bin/colorbench.c
index de7189a..c91b731 100644
--- a/bin/colorbench.c
+++ b/bin/colorbench.c
@@ -16,7 +16,7 @@
 #include "litmus.h"
 
 #define DEBUG           1
-#define NR_LOOPS        100
+#define NR_LOOPS        10
 
 /* Pound */
 #define NR_CPUS         4
@@ -76,7 +76,8 @@ static pthread_barrier_t barrier;
 static int nr_threads;
 static int arena_size;
 static int color_shift;
-static int *page_line_order;
+//static int *page_line_order;
+static int *arena_line_order;
 static struct perf_counter perf_counters[NR_CPUS * NR_PERF_COUNTERS];
 
 #ifdef DEBUG
@@ -137,6 +138,7 @@ void sattolo(int *items, const int len)
  * Starting at position 0 in the page_line_order means the cycle ends with 0.
  * We use 0 in the arena to signify that we are done reading.
  */
+#if 0
 static void init_arena_page_line_order(int *arena)
 {
         int cur_page;
@@ -166,6 +168,22 @@ static void init_arena_page_line_order(int *arena)
                 }
         }
 }
+#endif
+static void init_arena_line_order(int *arena)
+{
+        int cur_line;
+        for (cur_line = 0; cur_line < ARENA_LINES; cur_line++) {
+                const int idx = LINE_INTS * cur_line;
+                const int next_line = arena_line_order[cur_line];
+                int next_idx = LINE_INTS * next_line;
+
+                if (0 == next_line) {
+                        /* special case: cycle end */
+                        next_idx = 0;
+                }
+                arena[idx] = next_idx;
+        }
+}
 
 static void setup_colors(struct pthread_state *state)
 {
@@ -194,6 +212,30 @@ static int loop_once(struct pthread_state *state)
         return j;
 }
 
+static void print_perf_counters(void)
+{
+        uint64_t val;
+        int err, cpu, i;
+        const char *name;
+
+        for (cpu = 0; cpu < NR_CPUS; cpu++) {
+                printf("CPU %d\n", cpu);
+                for (i = 0; i < NR_PERF_COUNTERS; i++) {
+                        const int idx = cpu * NR_PERF_COUNTERS + i;
+                        name = get_perf_name(&perf_counters[idx]);
+                        err = read_perf_counter(&perf_counters[idx], &val);
+                        if (err)
+                                printf("%50s: ERROR\n", name);
+                        else
+#if 0
+                                printf("%50s: %10.3f\n", name,
+                                                ((double)val) / NR_LOOPS);
+#endif
+                                printf("%50s: %lu\n", name, val);
+                }
+        }
+}
+
 int thread_init(struct pthread_state *state)
 {
         const int cpu = THREAD_CPU(state->tid);
@@ -220,7 +262,8 @@ int thread_init(struct pthread_state *state)
                 goto out;
         }
 
-        init_arena_page_line_order(state->arena);
+        //init_arena_page_line_order(state->arena);
+        init_arena_line_order(state->arena);
 
         err = mlockall(MCL_CURRENT|MCL_FUTURE);
         if (err)
@@ -241,6 +284,17 @@ static void change_counters(const int request)
         }
 }
 
+static void reset_counters(void)
+{
+        int cpu, counter;
+        for (cpu = 0; cpu < NR_CPUS; cpu++) {
+                for (counter = 0; counter < NR_PERF_COUNTERS; counter++) {
+                        const int ctr_idx = cpu * NR_PERF_COUNTERS + counter;
+                        ioctl(perf_counters[ctr_idx].fd, PERF_EVENT_IOC_RESET);
+                }
+        }
+}
+
 void * thread_start(void *data)
 {
         struct pthread_state *state = (struct pthread_state*) data;
@@ -252,14 +306,20 @@ void * thread_start(void *data)
 
         pthread_barrier_wait(&barrier);
 
+#if 0
         if (0 == state->tid)
                 change_counters(PERF_EVENT_IOC_ENABLE);
 
         pthread_barrier_wait(&barrier);
+#endif
+
+        change_counters(PERF_EVENT_IOC_ENABLE);
 
         for (i = 0; i < NR_LOOPS; ++i) {
                 loop_once(state);
-                pthread_barrier_wait(&barrier);
+                //pthread_barrier_wait(&barrier);
+                print_perf_counters();
+                reset_counters();
         }
 
         if (0 == state->tid)
@@ -304,27 +364,6 @@ out:
         return ret;
 }
 
-static void print_perf_counters(void)
-{
-        uint64_t val;
-        int err, cpu, i;
-        const char *name;
-
-        for (cpu = 0; cpu < NR_CPUS; cpu++) {
-                printf("CPU %d\n", cpu);
-                for (i = 0; i < NR_PERF_COUNTERS; i++) {
-                        const int idx = cpu * NR_PERF_COUNTERS + i;
-                        name = get_perf_name(&perf_counters[idx]);
-                        err = read_perf_counter(&perf_counters[idx], &val);
-                        if (err)
-                                printf("%50s: ERROR\n", name);
-                        else
-                                printf("%50s: %10.3f\n", name,
-                                                ((double)val) / NR_LOOPS);
-                }
-        }
-}
-
 #define CHECK(fun, ...) { \
         int err = fun(__VA_ARGS__); \
         if (err) { \
@@ -366,8 +405,10 @@ int main(int argc, char **argv)
         }
 
         pthread_state   = malloc(nr_threads * sizeof(*pthread_state));
-        page_line_order = malloc(PAGE_LINES * sizeof(*page_line_order));
-        if (!pthread_state || !page_line_order) {
+        //page_line_order = malloc(PAGE_LINES * sizeof(*page_line_order));
+        arena_line_order = malloc(ARENA_LINES * sizeof(*arena_line_order));
+        //if (!pthread_state || !page_line_order) {
+        if (!pthread_state || !arena_line_order) {
                 debug_print("could not malloc\n");
                 ret = 1;
                 goto out;
@@ -379,7 +420,8 @@ int main(int argc, char **argv)
         CHECK(pthread_attr_setdetachstate, &attr, PTHREAD_CREATE_JOINABLE);
         CHECK(pthread_barrier_init, &barrier, NULL, nr_threads);
 
-        sattolo(page_line_order, PAGE_LINES);
+        //sattolo(page_line_order, PAGE_LINES);
+        sattolo(arena_line_order, ARENA_LINES);
         //sequential(page_line_order, PAGE_LINES);
 
         for (i = 0; i < nr_threads; i++) {
@@ -405,8 +447,13 @@ int main(int argc, char **argv)
         CHECK(pthread_barrier_destroy, &barrier);
         CHECK(pthread_attr_destroy, &attr);
 
-        printf("arena pages: %d\n", ARENA_PAGES);
-        print_perf_counters();
+        printf("arena size:        %d\n", arena_size);
+        printf("color shift:       %d\n", color_shift);
+        printf("arena pages:       %d\n", ARENA_PAGES);
+        printf("arena lines:       %d\n", ARENA_LINES);
+        printf("use colors:        %d\n", USE_COLORS);
+        printf("contiguous colors: %d\n", CONTIG_COLORS);
+        //print_perf_counters();
 
 out:
         return ret;
diff --git a/bin/perfcounters.c b/bin/perfcounters.c
index 402e047..6302164 100644
--- a/bin/perfcounters.c
+++ b/bin/perfcounters.c
@@ -68,7 +68,6 @@ static struct perf_counter_setup perf_setup[NR_PERF_COUNTERS] = {
                 .type   = PERF_TYPE_RAW,
                 .config = ATTR_CONFIG(0x0f, 0x20),
         },
-#endif
         {
                 .name   = "L2_RQSTS.PREFETCH_HIT",
                 .type   = PERF_TYPE_RAW,
@@ -79,6 +78,17 @@ static struct perf_counter_setup perf_setup[NR_PERF_COUNTERS] = {
                 .type   = PERF_TYPE_RAW,
                 .config = ATTR_CONFIG(0x24, 0x80),
         },
+#endif
+        {
+                .name   = "MEM_LOAD_RETIRED.OTHER_CORE_L2_HIT_HITM",
+                .type   = PERF_TYPE_RAW,
+                .config = ATTR_CONFIG(0xcb, 0x08),
+        },
+        {
+                .name   = "MEM_LOAD_RETIRED.L3_UNSHARED_HIT",
+                .type   = PERF_TYPE_RAW,
+                .config = ATTR_CONFIG(0xcb, 0x04),
+        },
         {
                 .name   = "MEM_LOAD_RETIRED.L3_MISS",
                 .type   = PERF_TYPE_RAW,