Add reuse client

src/tool/hpcrun/hpcrun_stats.c

@@ -102,8 +102,10 @@ static atomic_long num_falseWWIns = ATOMIC_VAR_INIT(0);
 static atomic_long num_falseRWIns = ATOMIC_VAR_INIT(0);
 static atomic_long num_falseWRIns = ATOMIC_VAR_INIT(0);
 
-static atomic_long num_reuse = ATOMIC_VAR_INIT(0);
+static atomic_long num_reuseSpatial = ATOMIC_VAR_INIT(0);
+static atomic_long num_reuseTemporal =  ATOMIC_VAR_INIT(0);
 static atomic_long num_latency = ATOMIC_VAR_INIT(0);
+static atomic_long num_corrected_reuse_distance = ATOMIC_VAR_INIT(0);
 
 static atomic_long num_unwind_intervals_total = ATOMIC_VAR_INIT(0);
 static atomic_long num_unwind_intervals_suspicious = ATOMIC_VAR_INIT(0);
@@ -155,6 +157,11 @@ hpcrun_stats_reinit(void)
   atomic_store_explicit(&num_trueWWIns, 0, memory_order_relaxed);
   atomic_store_explicit(&num_trueRWIns, 0, memory_order_relaxed);
   atomic_store_explicit(&num_trueWRIns, 0, memory_order_relaxed);
+
+  atomic_store_explicit(&num_reuseSpatial, 0, memory_order_relaxed);
+  atomic_store_explicit(&num_reuseTemporal, 0, memory_order_relaxed);
+  atomic_store_explicit(&num_latency, 0, memory_order_relaxed);
+  atomic_store_explicit(&num_corrected_reuse_distance, 0, memory_order_relaxed);
 }
 
 
@@ -274,95 +281,107 @@ void
 hpcrun_stats_num_insane_ip_inc(long val)
 {
   atomic_fetch_add_explicit(&num_insane_ip, val, memory_order_relaxed);
-}                
-                 
-                 
-long             
+}
+
+
+long
 hpcrun_stats_num_insane_ip(void)
-{                
+{
   return atomic_load_explicit(&num_insane_ip, memory_order_relaxed);
-}                
+}
 
 
 void
 hpcrun_stats_num_writtenBytes_inc(long val)
 {
   atomic_fetch_add_explicit(&num_writtenBytes, val, memory_order_relaxed);
-}                
-                 
+}
+
 
 void
 hpcrun_stats_num_usedBytes_inc(long val)
 {
   atomic_fetch_add_explicit(&num_usedBytes, val, memory_order_relaxed);
-}                
+}
 
 void
 hpcrun_stats_num_deadBytes_inc(long val)
 {
   atomic_fetch_add_explicit(&num_deadBytes, val, memory_order_relaxed);
-}                
+}
 
 void
 hpcrun_stats_num_newBytes_inc(long val)
 {
   atomic_fetch_add_explicit(&num_newBytes, val, memory_order_relaxed);
-}                
+}
 
 void
 hpcrun_stats_num_oldAppxBytes_inc(long val)
 {
   atomic_fetch_add_explicit(&num_oldAppxBytes, val, memory_order_relaxed);
-}   
+}
 
 void
 hpcrun_stats_num_oldBytes_inc(long val)
 {
   atomic_fetch_add_explicit(&num_oldBytes, val, memory_order_relaxed);
-}                
-             
+}
+
 void
 hpcrun_stats_num_loadedBytes_inc(long val)
 {
   atomic_fetch_add_explicit(&num_loadedBytes, val, memory_order_relaxed);
-}                
+}
 
 void
 hpcrun_stats_num_accessedIns_inc(long val)
 {
   atomic_fetch_add_explicit(&num_accessedIns, val, memory_order_relaxed);
-}            
+}
+
+void
+hpcrun_stats_num_reuseTemporal_inc(long val)
+{
+  atomic_fetch_add_explicit(&num_reuseTemporal, val, memory_order_relaxed);
+}
 
 void
-hpcrun_stats_num_reuse_inc(long val)
+hpcrun_stats_num_reuseSpatial_inc(long val)
 {
-  atomic_fetch_add_explicit(&num_reuse, val, memory_order_relaxed);
-}            
+  atomic_fetch_add_explicit(&num_reuseSpatial, val, memory_order_relaxed);
+}
 
 void
 hpcrun_stats_num_latency_inc(long val)
 {
   atomic_fetch_add_explicit(&num_latency, val, memory_order_relaxed);
-}            
+}
+
+void
+hpcrun_stats_num_corrected_reuse_distance_inc(long val)
+{
+  atomic_fetch_add_explicit(&num_corrected_reuse_distance, val, memory_order_relaxed);
+}
 
 void
 hpcrun_stats_num_falseWWIns_inc(long val)
 {
   atomic_fetch_add_explicit(&num_falseWWIns, val, memory_order_relaxed);
-}            
+}
 
 
 void
 hpcrun_stats_num_falseRWIns_inc(long val)
 {
   atomic_fetch_add_explicit(&num_falseRWIns, val, memory_order_relaxed);
-}            
+}
 
 void
 hpcrun_stats_num_falseWRIns_inc(long val)
 {
   atomic_fetch_add_explicit(&num_falseWRIns, val, memory_order_relaxed);
-}            
+}
 
 void
 hpcrun_stats_num_trueWWIns_inc(long val)
@@ -384,7 +403,7 @@ hpcrun_stats_num_trueWRIns_inc(long val)
 
 
 //-----------------------------
-// samples total 
+// samples total
 //-----------------------------
 
 void
@@ -403,7 +422,7 @@ hpcrun_stats_num_samples_total(void)
 
 
 //-----------------------------
-// samples attempted 
+// samples attempted
 //-----------------------------
 
 void
@@ -422,7 +441,7 @@ hpcrun_stats_num_samples_attempted(void)
 
 
 //-----------------------------
-// samples blocked async 
+// samples blocked async
 //-----------------------------
 
 // The async blocks happen in the signal handlers, without getting to
@@ -444,7 +463,7 @@ hpcrun_stats_num_samples_blocked_async(void)
 
 
 //-----------------------------
-// samples blocked dlopen 
+// samples blocked dlopen
 //-----------------------------
 
 void
@@ -617,7 +636,6 @@ hpcrun_stats_num_samples_yielded(void)
 //-----------------------------
 // print summary
 //-----------------------------
-
 void
 hpcrun_stats_print_summary(void)
 {
@@ -637,9 +655,9 @@ hpcrun_stats_print_summary(void)
   getrusage(RUSAGE_SELF, &rusage);
 
   //AMSG("WATCHPOINT ANOMALIES: samples:%ld, SM_imprecise:%ld, WP_Set:%ld, WP_triggered:%ld, WP_SampleTriggering:%ld, WP_ImpreciseIP:%ld, WP_InsaneIP:%ld, WP_Off8Addr:%ld, WP_ImpreciseAddr:%ld, WP_Dropped:%ld", num_samples_total, num_samples_imprecise, num_watchpoints_set, num_watchpoints_triggered, num_sample_triggering_watchpoints,  num_watchpoints_imprecise, num_insane_ip, num_watchpoints_imprecise_address_8_byte, num_watchpoints_imprecise_address, num_watchpoints_dropped);
-  AMSG("WATCHPOINT ANOMALIES: samples:%.2e, SM_imprecise:%.2e, WP_Set:%.2e, WP_triggered:%.2e, WP_SampleTriggering:%.2e, WP_ImpreciseIP:%.2e, WP_InsaneIP:%.2e, WP_Off8Addr:%.2e, WP_ImpreciseAddr:%.2e, WP_Dropped:%.2e", (double)atomic_load(&num_samples_total), (double)atomic_load(&num_samples_imprecise), (double)atomic_load(&num_watchpoints_set), (double)atomic_load(&num_watchpoints_triggered), (double)atomic_load(&num_sample_triggering_watchpoints),  (double)atomic_load(&num_watchpoints_imprecise), (double)atomic_load(&num_insane_ip), (double)atomic_load(&num_watchpoints_imprecise_address_8_byte), (double)atomic_load(&num_watchpoints_imprecise_address), (double)atomic_load(&num_watchpoints_dropped));
+  AMSG("WATCHPOINT ANOMALIES: samples:%.2e, SM_imprecise:%.2e, WP_Set:%.2e, WP_triggered:%.2e, WP_SampleTriggering:%.2e, WP_ImpreciseIP:%.2e, WP_InsaneIP:%.2e, WP_Off8Addr:%.2e, WP_ImpreciseAddr:%.2e, WP_Dropped:%.2e, CORRECTED_REUSE_DISTANCE:%.2e", (double)atomic_load(&num_samples_total), (double)atomic_load(&num_samples_imprecise), (double)atomic_load(&num_watchpoints_set), (double)atomic_load(&num_watchpoints_triggered), (double)atomic_load(&num_sample_triggering_watchpoints),  (double)atomic_load(&num_watchpoints_imprecise), (double)atomic_load(&num_insane_ip), (double)atomic_load(&num_watchpoints_imprecise_address_8_byte), (double)atomic_load(&num_watchpoints_imprecise_address), (double)atomic_load(&num_watchpoints_dropped), (double)atomic_load(&num_corrected_reuse_distance));
 
-  AMSG("WATCHPOINT STATS: writtenBytes:%ld, usedBytes:%ld, deadBytes:%ld, newBytes:%ld, oldBytes:%ld, oldAppxBytes:%ld, loadedBytes:%ld, accessedIns:%ld, falseWWIns:%ld, falseRWIns:%ld, falseWRIns:%ld, trueWWIns:%ld, trueRWIns:%ld, trueWRIns:%ld, RSS:%ld, reuse:%ld, latency:%ld", num_writtenBytes, num_usedBytes, num_deadBytes, num_newBytes, num_oldBytes, num_oldAppxBytes, num_loadedBytes, num_accessedIns, num_falseWWIns, num_falseRWIns, num_falseWRIns, num_trueWWIns, num_trueRWIns, num_trueWRIns,  (size_t)(rusage.ru_maxrss), num_reuse, num_latency);
+  AMSG("WATCHPOINT STATS: writtenBytes:%ld, usedBytes:%ld, deadBytes:%ld, newBytes:%ld, oldBytes:%ld, oldAppxBytes:%ld, loadedBytes:%ld, accessedIns:%ld, falseWWIns:%ld, falseRWIns:%ld, falseWRIns:%ld, trueWWIns:%ld, trueRWIns:%ld, trueWRIns:%ld, RSS:%ld, reuseTemporal:%ld, reuseSpatial:%ldlatency:%ld", num_writtenBytes, num_usedBytes, num_deadBytes, num_newBytes, num_oldBytes, num_oldAppxBytes, num_loadedBytes, num_accessedIns, num_falseWWIns, num_falseRWIns, num_falseWRIns, num_trueWWIns, num_trueRWIns, num_trueWRIns,  (size_t)(rusage.ru_maxrss), num_reuseTemporal, num_reuseSpatial, num_latency);
 
   AMSG("SAMPLE ANOMALIES: blocks: %ld (async: %ld, dlopen: %ld), "
        "errors: %ld (segv: %ld, soft: %ld)",

src/tool/hpcrun/hpcrun_stats.h

@@ -52,7 +52,7 @@
 void hpcrun_stats_reinit(void);
 
 //-----------------------------
-// watchpoint 
+// watchpoint
 //-----------------------------
 void hpcrun_stats_num_samples_imprecise_inc(long val);
 long hpcrun_stats_num_samples_imprecise(void);
@@ -85,36 +85,37 @@ void hpcrun_stats_num_deadBytes_inc(long val);
 void hpcrun_stats_num_newBytes_inc(long val);
 void hpcrun_stats_num_oldBytes_inc(long val);
 void hpcrun_stats_num_oldAppxBytes_inc(long val);
-void hpcrun_stats_num_reuse_inc(long val);
+void hpcrun_stats_num_reuseTemporal_inc(long val);
+void hpcrun_stats_num_reuseSpatial_inc(long val);
 void hpcrun_stats_num_loadedBytes_inc(long val);
 
 
 //-----------------------------
-// samples total 
+// samples total
 //-----------------------------
 
 void hpcrun_stats_num_samples_total_inc(void);
 long hpcrun_stats_num_samples_total(void);
 
 
 //-----------------------------
-// samples attempted 
+// samples attempted
 //-----------------------------
 
 void hpcrun_stats_num_samples_attempted_inc(void);
 long hpcrun_stats_num_samples_attempted(void);
 
 
 //-----------------------------
-// samples blocked async 
+// samples blocked async
 //-----------------------------
 
 void hpcrun_stats_num_samples_blocked_async_inc(void);
 long hpcrun_stats_num_samples_blocked_async(void);
 
 
 //-----------------------------
-// samples blocked dlopen 
+// samples blocked dlopen
 //-----------------------------
 
 void hpcrun_stats_num_samples_blocked_dlopen_inc(void);

src/tool/hpcrun/metrics.c

@@ -211,7 +211,7 @@ hpcrun_get_num_metrics()
       id2metric[l->id] = &(l->val);
     }
     metric_proc_tbl = (metric_upd_proc_t**) hpcrun_malloc(n_metrics * sizeof(metric_upd_proc_t*));
-    
+
     for(metric_proc_map_t* l = proc_map; l; l = l->next) {
     //    for(metric_proc_map_t* l = proc_map; l; l = l->next) {
       TMSG(METRICS_FINALIZE, "metric_proc[%d] = %p", l->id, l->proc);
@@ -233,15 +233,15 @@ hpcrun_get_num_metrics()
 
 // Finalize metrics
 
-void hpcrun_finalize_metrics() 
+void hpcrun_finalize_metrics()
 {
   hpcrun_get_num_metrics();
 }
 
 metric_desc_t*
 hpcrun_id2metric(int metric_id)
 {
-  hpcrun_get_num_metrics(); 
+  hpcrun_get_num_metrics();
   if ((0 <= metric_id) && (metric_id < n_metrics)) {
     return id2metric[metric_id];
   }
@@ -307,7 +307,7 @@ hpcrun_new_metric_of_kind(kind_info_t* kind)
   kind->idx++;
 
   n_metrics++;
-  
+
   //
   // No preallocation for metric_proc tbl
   //
@@ -316,7 +316,7 @@ hpcrun_new_metric_of_kind(kind_info_t* kind)
   m->id   = metric_data->id;
   m->proc = (metric_upd_proc_t*) NULL;
   proc_map = m;
-  
+
   return metric_data->id;
 }
 
@@ -326,7 +326,7 @@ hpcrun_new_metric(void)
   return hpcrun_new_metric_of_kind(current_kind);
 }
 
-metric_desc_t* 
+metric_desc_t*
 hpcrun_set_metric_info_w_fn(int metric_id, const char* name,
 			    MetricFlags_ValFmt_t valFmt, size_t period,
 			    metric_upd_proc_t upd_fn, metric_desc_properties_t prop)
@@ -377,7 +377,7 @@ hpcrun_set_metric_info_w_fn(int metric_id, const char* name,
 }
 
 
-metric_desc_t* 
+metric_desc_t*
 hpcrun_set_metric_info_and_period(int metric_id, const char* name,
 				  MetricFlags_ValFmt_t valFmt, size_t period, metric_desc_properties_t prop)
 {
@@ -454,7 +454,7 @@ hpcrun_metric_std(int metric_id, metric_set_t* set,
   switch (minfo->flags.fields.valFmt) {
     case MetricFlags_ValFmt_Int:
       if (operation == '+')
-        loc->i += val.i; 
+        loc->i += val.i;
       else if (operation == '=')
         loc->i = val.i;
       break;
@@ -479,13 +479,13 @@ hpcrun_metric_std_set(int metric_id, metric_set_t* set,
 }
 
 //
-// Given two metrics, metric_id1 and metric_id2, 
-// bump up metric_id2 to reach metric_id1 and return 
+// Given two metrics, metric_id1 and metric_id2,
+// bump up metric_id2 to reach metric_id1 and return
 // the difference between them multiplied by the period.
 //
 int
-hpcrun_get_weighted_metric_diff(int metric_id1, int metric_id2, 
-                                metric_set_t* set, cct_metric_data_t * diff, 
+hpcrun_get_weighted_metric_diff(int metric_id1, int metric_id2,
+                                metric_set_t* set, cct_metric_data_t * diff,
                                 cct_metric_data_t * diffWithPeriod)
 {
     metric_desc_t* minfo1 = hpcrun_id2metric(metric_id1);
@@ -509,8 +509,12 @@ hpcrun_get_weighted_metric_diff(int metric_id1, int metric_id2,
             diff->i = (loc1->i - loc2->i);
             break;
         case MetricFlags_ValFmt_Real:
-            assert(loc1->r >= loc2->r);
-            diff->r = (loc1->r - loc2->r);
+            if (loc1->r < loc2->r){
+                diff->r = 0;
+            }
+            else {
+                diff->r = (loc1->r - loc2->r);
+            }
             diffWithPeriod->r = (loc1->r - loc2->r) * minfo1->period;
             break;
         default: