8030815: Code roots are not accounted for in region prediction

Reviewed-by: iwalulya, ayang

Author: Thomas Schatzl

src/hotspot/share/gc/g1/g1Analytics.cpp

@@ -80,9 +80,11 @@ G1Analytics::G1Analytics(const G1Predictions* predictor) :
     _card_scan_to_merge_ratio_seq(TruncatedSeqLength),
     _cost_per_card_scan_ms_seq(TruncatedSeqLength),
     _cost_per_card_merge_ms_seq(TruncatedSeqLength),
+    _cost_per_code_root_ms_seq(TruncatedSeqLength),
     _cost_per_byte_copied_ms_seq(TruncatedSeqLength),
     _pending_cards_seq(TruncatedSeqLength),
     _rs_length_seq(TruncatedSeqLength),
+    _code_root_rs_length_seq(TruncatedSeqLength),
     _constant_other_time_ms_seq(TruncatedSeqLength),
     _young_other_cost_per_region_ms_seq(TruncatedSeqLength),
     _non_young_other_cost_per_region_ms_seq(TruncatedSeqLength),
@@ -104,6 +106,7 @@ G1Analytics::G1Analytics(const G1Predictions* predictor) :
   _card_scan_to_merge_ratio_seq.set_initial(young_card_scan_to_merge_ratio_defaults[index]);
   _cost_per_card_scan_ms_seq.set_initial(young_only_cost_per_card_scan_ms_defaults[index]);
   _rs_length_seq.set_initial(0);
+  _code_root_rs_length_seq.set_initial(0);
   _cost_per_byte_copied_ms_seq.set_initial(cost_per_byte_ms_defaults[index]);
 
   _constant_other_time_ms_seq.add(constant_other_time_ms_defaults[index]);
@@ -186,6 +189,10 @@ void G1Analytics::report_cost_per_card_merge_ms(double cost_per_card_ms, bool fo
   _cost_per_card_merge_ms_seq.add(cost_per_card_ms, for_young_only_phase);
 }
 
+void G1Analytics::report_cost_per_code_root_scan_ms(double cost_per_code_root_ms, bool for_young_only_phase) {
+  _cost_per_code_root_ms_seq.add(cost_per_code_root_ms, for_young_only_phase);
+}
+
 void G1Analytics::report_card_scan_to_merge_ratio(double merge_to_scan_ratio, bool for_young_only_phase) {
   _card_scan_to_merge_ratio_seq.add(merge_to_scan_ratio, for_young_only_phase);
 }
@@ -214,6 +221,10 @@ void G1Analytics::report_rs_length(double rs_length, bool for_young_only_phase)
   _rs_length_seq.add(rs_length, for_young_only_phase);
 }
 
+void G1Analytics::report_code_root_rs_length(double code_root_rs_length, bool for_young_only_phase) {
+  _code_root_rs_length_seq.add(code_root_rs_length, for_young_only_phase);
+}
+
 double G1Analytics::predict_alloc_rate_ms() const {
   if (enough_samples_available(&_alloc_rate_ms_seq)) {
     return predict_zero_bounded(&_alloc_rate_ms_seq);
@@ -242,6 +253,10 @@ double G1Analytics::predict_card_merge_time_ms(size_t card_num, bool for_young_o
   return card_num * predict_zero_bounded(&_cost_per_card_merge_ms_seq, for_young_only_phase);
 }
 
+double G1Analytics::predict_code_root_scan_time_ms(size_t code_root_num, bool for_young_only_phase) const {
+  return code_root_num * predict_zero_bounded(&_cost_per_code_root_ms_seq, for_young_only_phase);
+}
+
 double G1Analytics::predict_card_scan_time_ms(size_t card_num, bool for_young_only_phase) const {
   return card_num * predict_zero_bounded(&_cost_per_card_scan_ms_seq, for_young_only_phase);
 }
@@ -274,6 +289,10 @@ size_t G1Analytics::predict_rs_length(bool for_young_only_phase) const {
   return predict_size(&_rs_length_seq, for_young_only_phase);
 }
 
+size_t G1Analytics::predict_code_root_rs_length(bool for_young_only_phase) const {
+  return predict_size(&_code_root_rs_length_seq, for_young_only_phase);
+}
+
 size_t G1Analytics::predict_pending_cards(bool for_young_only_phase) const {
   return predict_size(&_pending_cards_seq, for_young_only_phase);
 }

src/hotspot/share/gc/g1/g1Analytics.hpp

@@ -57,11 +57,14 @@ class G1Analytics: public CHeapObj<mtGC> {
   G1PhaseDependentSeq _cost_per_card_scan_ms_seq;
   // The cost to merge a card during young-only and mixed gcs in ms.
   G1PhaseDependentSeq _cost_per_card_merge_ms_seq;
+  // The cost to scan entries in the code root remembered set in ms.
+  G1PhaseDependentSeq _cost_per_code_root_ms_seq;
   // The cost to copy a byte in ms.
   G1PhaseDependentSeq _cost_per_byte_copied_ms_seq;
 
   G1PhaseDependentSeq _pending_cards_seq;
   G1PhaseDependentSeq _rs_length_seq;
+  G1PhaseDependentSeq _code_root_rs_length_seq;
 
   TruncatedSeq _constant_other_time_ms_seq;
   TruncatedSeq _young_other_cost_per_region_ms_seq;
@@ -127,6 +130,7 @@ class G1Analytics: public CHeapObj<mtGC> {
   void report_dirtied_cards_in_thread_buffers(size_t num_cards);
   void report_cost_per_card_scan_ms(double cost_per_remset_card_ms, bool for_young_only_phase);
   void report_cost_per_card_merge_ms(double cost_per_card_ms, bool for_young_only_phase);
+  void report_cost_per_code_root_scan_ms(double cost_per_code_root_ms, bool for_young_only_phase);
   void report_card_scan_to_merge_ratio(double cards_per_entry_ratio, bool for_young_only_phase);
   void report_rs_length_diff(double rs_length_diff, bool for_young_only_phase);
   void report_cost_per_byte_ms(double cost_per_byte_ms, bool for_young_only_phase);
@@ -135,6 +139,7 @@ class G1Analytics: public CHeapObj<mtGC> {
   void report_constant_other_time_ms(double constant_other_time_ms);
   void report_pending_cards(double pending_cards, bool for_young_only_phase);
   void report_rs_length(double rs_length, bool for_young_only_phase);
+  void report_code_root_rs_length(double code_root_rs_length, bool for_young_only_phase);
 
   double predict_alloc_rate_ms() const;
   int num_alloc_rate_ms() const;
@@ -150,6 +155,8 @@ class G1Analytics: public CHeapObj<mtGC> {
   double predict_card_merge_time_ms(size_t card_num, bool for_young_only_phase) const;
   double predict_card_scan_time_ms(size_t card_num, bool for_young_only_phase) const;
 
+  double predict_code_root_scan_time_ms(size_t code_root_num, bool for_young_only_phase) const;
+
   double predict_object_copy_time_ms(size_t bytes_to_copy, bool for_young_only_phase) const;
 
   double predict_constant_other_time_ms() const;
@@ -163,6 +170,7 @@ class G1Analytics: public CHeapObj<mtGC> {
   double predict_cleanup_time_ms() const;
 
   size_t predict_rs_length(bool for_young_only_phase) const;
+  size_t predict_code_root_rs_length(bool for_young_only_phase) const;
   size_t predict_pending_cards(bool for_young_only_phase) const;
 
   // Add a new GC of the given duration and end time to the record.

src/hotspot/share/gc/g1/g1ConcurrentRefine.cpp

@@ -280,18 +280,21 @@ uint64_t G1ConcurrentRefine::adjust_threads_wait_ms() const {
 class G1ConcurrentRefine::RemSetSamplingClosure : public HeapRegionClosure {
   G1CollectionSet* _cset;
   size_t _sampled_rs_length;
+  size_t _sampled_code_root_rs_length;
 
 public:
   explicit RemSetSamplingClosure(G1CollectionSet* cset) :
-    _cset(cset), _sampled_rs_length(0) {}
+    _cset(cset), _sampled_rs_length(0), _sampled_code_root_rs_length(0) {}
 
   bool do_heap_region(HeapRegion* r) override {
-    size_t rs_length = r->rem_set()->occupied();
-    _sampled_rs_length += rs_length;
+    HeapRegionRemSet* rem_set = r->rem_set();
+    _sampled_rs_length += rem_set->occupied();
+    _sampled_code_root_rs_length += rem_set->code_roots_list_length();
     return false;
   }
 
   size_t sampled_rs_length() const { return _sampled_rs_length; }
+  size_t sampled_code_root_rs_length() const { return _sampled_code_root_rs_length; }
 };
 
 // Adjust the target length (in regions) of the young gen, based on the the
@@ -311,7 +314,7 @@ void G1ConcurrentRefine::adjust_young_list_target_length() {
     G1CollectionSet* cset = G1CollectedHeap::heap()->collection_set();
     RemSetSamplingClosure cl{cset};
     cset->iterate(&cl);
-    _policy->revise_young_list_target_length(cl.sampled_rs_length());
+    _policy->revise_young_list_target_length(cl.sampled_rs_length(), cl.sampled_code_root_rs_length());
   }
 }
 

src/hotspot/share/gc/g1/g1Policy.cpp

@@ -189,13 +189,14 @@ void G1Policy::update_young_length_bounds() {
   assert(!Universe::is_fully_initialized() || SafepointSynchronize::is_at_safepoint(), "must be");
   bool for_young_only_phase = collector_state()->in_young_only_phase();
   update_young_length_bounds(_analytics->predict_pending_cards(for_young_only_phase),
-                             _analytics->predict_rs_length(for_young_only_phase));
+                             _analytics->predict_rs_length(for_young_only_phase),
+                             _analytics->predict_code_root_rs_length(for_young_only_phase));
 }
 
-void G1Policy::update_young_length_bounds(size_t pending_cards, size_t rs_length) {
+void G1Policy::update_young_length_bounds(size_t pending_cards, size_t rs_length, size_t code_root_rs_length) {
   uint old_young_list_target_length = young_list_target_length();
 
-  uint new_young_list_desired_length = calculate_young_desired_length(pending_cards, rs_length);
+  uint new_young_list_desired_length = calculate_young_desired_length(pending_cards, rs_length, code_root_rs_length);
   uint new_young_list_target_length = calculate_young_target_length(new_young_list_desired_length);
   uint new_young_list_max_length = calculate_young_max_length(new_young_list_target_length);
 
@@ -234,7 +235,9 @@ void G1Policy::update_young_length_bounds(size_t pending_cards, size_t rs_length
 // value smaller than what is already allocated or what can actually be allocated.
 // This return value is only an expectation.
 //
-uint G1Policy::calculate_young_desired_length(size_t pending_cards, size_t rs_length) const {
+uint G1Policy::calculate_young_desired_length(size_t pending_cards,
+                                              size_t rs_length,
+                                              size_t code_root_rs_length) const {
   uint min_young_length_by_sizer = _young_gen_sizer.min_desired_young_length();
   uint max_young_length_by_sizer = _young_gen_sizer.max_desired_young_length();
 
@@ -267,7 +270,7 @@ uint G1Policy::calculate_young_desired_length(size_t pending_cards, size_t rs_le
   if (use_adaptive_young_list_length()) {
     desired_eden_length_by_mmu = calculate_desired_eden_length_by_mmu();
 
-    double base_time_ms = predict_base_time_ms(pending_cards, rs_length);
+    double base_time_ms = predict_base_time_ms(pending_cards, rs_length, code_root_rs_length);
     double retained_time_ms = predict_retained_regions_evac_time();
     double total_time_ms = base_time_ms + retained_time_ms;
 
@@ -550,13 +553,13 @@ G1GCPhaseTimes* G1Policy::phase_times() const {
   return _phase_times;
 }
 
-void G1Policy::revise_young_list_target_length(size_t rs_length) {
+void G1Policy::revise_young_list_target_length(size_t rs_length, size_t code_root_rs_length) {
   guarantee(use_adaptive_young_list_length(), "should not call this otherwise" );
 
   size_t thread_buffer_cards = _analytics->predict_dirtied_cards_in_thread_buffers();
   G1DirtyCardQueueSet& dcqs = G1BarrierSet::dirty_card_queue_set();
   size_t pending_cards = dcqs.num_cards() + thread_buffer_cards;
-  update_young_length_bounds(pending_cards, rs_length);
+  update_young_length_bounds(pending_cards, rs_length, code_root_rs_length);
 }
 
 void G1Policy::record_full_collection_start() {
@@ -890,6 +893,17 @@ void G1Policy::record_young_collection_end(bool concurrent_operation_is_full_mar
     }
     _analytics->report_card_scan_to_merge_ratio(scan_to_merge_ratio, is_young_only_pause);
 
+    // Update prediction for code root scan
+    size_t const total_code_roots_scanned = p->sum_thread_work_items(G1GCPhaseTimes::CodeRoots, G1GCPhaseTimes::CodeRootsScannedNMethods) +
+                                            p->sum_thread_work_items(G1GCPhaseTimes::OptCodeRoots, G1GCPhaseTimes::CodeRootsScannedNMethods);
+
+    if (total_code_roots_scanned >= G1NumCodeRootsCostSampleThreshold) {
+      double avg_time_code_root_scan = average_time_ms(G1GCPhaseTimes::CodeRoots) +
+                                       average_time_ms(G1GCPhaseTimes::OptCodeRoots);
+
+      _analytics->report_cost_per_code_root_scan_ms(avg_time_code_root_scan / total_code_roots_scanned, is_young_only_pause);
+    }
+
     // Update prediction for copy cost per byte
     size_t copied_bytes = p->sum_thread_work_items(G1GCPhaseTimes::MergePSS, G1GCPhaseTimes::MergePSSCopiedBytes);
 
@@ -912,6 +926,7 @@ void G1Policy::record_young_collection_end(bool concurrent_operation_is_full_mar
 
     _analytics->report_pending_cards((double)pending_cards_at_gc_start(), is_young_only_pause);
     _analytics->report_rs_length((double)_rs_length, is_young_only_pause);
+    _analytics->report_code_root_rs_length((double)total_code_roots_scanned, is_young_only_pause);
   }
 
   assert(!(G1GCPauseTypeHelper::is_concurrent_start_pause(this_pause) && collector_state()->mark_or_rebuild_in_progress()),
@@ -1033,7 +1048,8 @@ void G1Policy::record_young_gc_pause_end(bool evacuation_failed) {
 }
 
 double G1Policy::predict_base_time_ms(size_t pending_cards,
-                                      size_t rs_length) const {
+                                      size_t rs_length,
+                                      size_t code_root_rs_length) const {
   bool in_young_only_phase = collector_state()->in_young_only_phase();
 
   size_t unique_cards_from_rs = _analytics->predict_scan_card_num(rs_length, in_young_only_phase);
@@ -1043,22 +1059,26 @@ double G1Policy::predict_base_time_ms(size_t pending_cards,
 
   double card_merge_time = _analytics->predict_card_merge_time_ms(pending_cards + rs_length, in_young_only_phase);
   double card_scan_time = _analytics->predict_card_scan_time_ms(effective_scanned_cards, in_young_only_phase);
+  double code_root_scan_time = _analytics->predict_code_root_scan_time_ms(code_root_rs_length, in_young_only_phase);
   double constant_other_time = _analytics->predict_constant_other_time_ms();
   double survivor_evac_time = predict_survivor_regions_evac_time();
 
-  double total_time = card_merge_time + card_scan_time + constant_other_time + survivor_evac_time;
+  double total_time = card_merge_time + card_scan_time + code_root_scan_time + constant_other_time + survivor_evac_time;
 
   log_trace(gc, ergo, heap)("Predicted base time: total %f lb_cards %zu rs_length %zu effective_scanned_cards %zu "
-                            "card_merge_time %f card_scan_time %f constant_other_time %f survivor_evac_time %f",
+                            "card_merge_time %f card_scan_time %f code_root_rs_length %zu code_root_scan_time %f "
+                            "constant_other_time %f survivor_evac_time %f",
                             total_time, pending_cards, rs_length, effective_scanned_cards,
-                            card_merge_time, card_scan_time, constant_other_time, survivor_evac_time);
+                            card_merge_time, card_scan_time, code_root_rs_length, code_root_scan_time,
+                            constant_other_time, survivor_evac_time);
   return total_time;
 }
 
 double G1Policy::predict_base_time_ms(size_t pending_cards) const {
   bool for_young_only_phase = collector_state()->in_young_only_phase();
   size_t rs_length = _analytics->predict_rs_length(for_young_only_phase);
-  return predict_base_time_ms(pending_cards, rs_length);
+  size_t code_root_rs_length = _analytics->predict_code_root_rs_length(for_young_only_phase);
+  return predict_base_time_ms(pending_cards, rs_length, code_root_rs_length);
 }
 
 size_t G1Policy::predict_bytes_to_copy(HeapRegion* hr) const {
@@ -1100,10 +1120,18 @@ double G1Policy::predict_region_merge_scan_time(HeapRegion* hr, bool for_young_o
     _analytics->predict_card_scan_time_ms(scan_card_num, for_young_only_phase);
 }
 
+double G1Policy::predict_region_code_root_scan_time(HeapRegion* hr, bool for_young_only_phase) const {
+  size_t code_root_length = hr->rem_set()->code_roots_list_length();
+
+  return
+    _analytics->predict_code_root_scan_time_ms(code_root_length, for_young_only_phase);
+}
+
 double G1Policy::predict_region_non_copy_time_ms(HeapRegion* hr,
                                                  bool for_young_only_phase) const {
 
-  double region_elapsed_time_ms = predict_region_merge_scan_time(hr, for_young_only_phase);
+  double region_elapsed_time_ms = predict_region_merge_scan_time(hr, for_young_only_phase) +
+                                  predict_region_code_root_scan_time(hr, for_young_only_phase);
   // The prediction of the "other" time for this region is based
   // upon the region type and NOT the GC type.
   if (hr->is_young()) {

src/hotspot/share/gc/g1/g1Policy.hpp

@@ -142,12 +142,15 @@ class G1Policy: public CHeapObj<mtGC> {
   // Base time contains handling remembered sets and constant other time of the
   // whole young gen, refinement buffers, and copying survivors.
   // Basically everything but copying eden regions.
-  double predict_base_time_ms(size_t pending_cards, size_t rs_length) const;
+  double predict_base_time_ms(size_t pending_cards, size_t rs_length, size_t code_root_length) const;
 
   // Copy time for a region is copying live data.
   double predict_region_copy_time_ms(HeapRegion* hr, bool for_young_only_phase) const;
-  // Merge-scan time for a region is handling remembered sets of that region (as a single unit).
+  // Merge-scan time for a region is handling card-based remembered sets of that region
+  // (as a single unit).
   double predict_region_merge_scan_time(HeapRegion* hr, bool for_young_only_phase) const;
+  // Code root scan time prediction for the given region.
+  double predict_region_code_root_scan_time(HeapRegion* hr, bool for_young_only_phase) const;
   // Non-copy time for a region is handling remembered sets and other time.
   double predict_region_non_copy_time_ms(HeapRegion* hr, bool for_young_only_phase) const;
 
@@ -207,10 +210,10 @@ class G1Policy: public CHeapObj<mtGC> {
   double _mark_cleanup_start_sec;
 
   // Updates the internal young gen maximum and target and desired lengths.
-  // If no parameters are passed, predict pending cards and the RS length using
-  // the prediction model.
+  // If no parameters are passed, predict pending cards, card set remset length and
+  // code root remset length using the prediction model.
   void update_young_length_bounds();
-  void update_young_length_bounds(size_t pending_cards, size_t rs_length);
+  void update_young_length_bounds(size_t pending_cards, size_t rs_length, size_t code_root_rs_length);
 
   // Calculate and return the minimum desired eden length based on the MMU target.
   uint calculate_desired_eden_length_by_mmu() const;
@@ -238,7 +241,7 @@ class G1Policy: public CHeapObj<mtGC> {
 
   // Calculate desired young length based on current situation without taking actually
   // available free regions into account.
-  uint calculate_young_desired_length(size_t pending_cards, size_t rs_length) const;
+  uint calculate_young_desired_length(size_t pending_cards, size_t rs_length, size_t code_root_rs_length) const;
   // Limit the given desired young length to available free regions.
   uint calculate_young_target_length(uint desired_young_length) const;
   // The GCLocker might cause us to need more regions than the target. Calculate
@@ -301,7 +304,7 @@ class G1Policy: public CHeapObj<mtGC> {
   // Check the current value of the young list RSet length and
   // compare it against the last prediction. If the current value is
   // higher, recalculate the young list target length prediction.
-  void revise_young_list_target_length(size_t rs_length);
+  void revise_young_list_target_length(size_t rs_length, size_t code_root_rs_length);
 
   // This should be called after the heap is resized.
   void record_new_heap_size(uint new_number_of_regions);

src/hotspot/share/gc/g1/g1_globals.hpp

@@ -325,9 +325,14 @@
           range(1, 256)                                                     \
                                                                             \
   product(uint, G1NumCardsCostSampleThreshold, 1000, DIAGNOSTIC,            \
-          "Threshold for the number of cards when reporting card cost "     \
-          "related prediction sample. That sample must involve the same or "\
-          "more than that number of cards to be used.")                     \
+          "Threshold for the number of cards when reporting remembered set "\
+          "card cost related prediction samples. A sample must involve "    \
+          "the same or more than that number of cards to be used.")         \
+                                                                            \
+  product(uint, G1NumCodeRootsCostSampleThreshold, 100, DIAGNOSTIC,         \
+          "Threshold for the number of code roots when reporting code root "\
+          "scan cost related prediction samples. A sample must involve "    \
+          "the same or more than this number of code roots to be used.")    \
                                                                             \
   GC_G1_EVACUATION_FAILURE_FLAGS(develop,                                   \
                     develop_pd,                                             \