src/share/vm/gc_implementation/g1/g1CollectorPolicy.hpp

Fri, 12 Aug 2011 11:31:06 -0400

author
tonyp
date
Fri, 12 Aug 2011 11:31:06 -0400
changeset 3028
f44782f04dd4
parent 3021
14a2fd14c0db
child 3065
ff53346271fe
permissions
-rw-r--r--

7039627: G1: avoid BOT updates for survivor allocations and dirty survivor regions incrementally
Summary: Refactor the allocation code during GC to use the G1AllocRegion abstraction. Use separate subclasses of G1AllocRegion for survivor and old regions. Avoid BOT updates and dirty survivor cards incrementally for the former.
Reviewed-by: brutisso, johnc, ysr

     1 /*
     2  * Copyright (c) 2001, 2011, Oracle and/or its affiliates. All rights reserved.
     3  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
     4  *
     5  * This code is free software; you can redistribute it and/or modify it
     6  * under the terms of the GNU General Public License version 2 only, as
     7  * published by the Free Software Foundation.
     8  *
     9  * This code is distributed in the hope that it will be useful, but WITHOUT
    10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
    11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
    12  * version 2 for more details (a copy is included in the LICENSE file that
    13  * accompanied this code).
    14  *
    15  * You should have received a copy of the GNU General Public License version
    16  * 2 along with this work; if not, write to the Free Software Foundation,
    17  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
    18  *
    19  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
    20  * or visit www.oracle.com if you need additional information or have any
    21  * questions.
    22  *
    23  */
    25 #ifndef SHARE_VM_GC_IMPLEMENTATION_G1_G1COLLECTORPOLICY_HPP
    26 #define SHARE_VM_GC_IMPLEMENTATION_G1_G1COLLECTORPOLICY_HPP
    28 #include "gc_implementation/g1/collectionSetChooser.hpp"
    29 #include "gc_implementation/g1/g1MMUTracker.hpp"
    30 #include "memory/collectorPolicy.hpp"
    32 // A G1CollectorPolicy makes policy decisions that determine the
    33 // characteristics of the collector.  Examples include:
    34 //   * choice of collection set.
    35 //   * when to collect.
    37 class HeapRegion;
    38 class CollectionSetChooser;
    40 // Yes, this is a bit unpleasant... but it saves replicating the same thing
    41 // over and over again and introducing subtle problems through small typos and
    42 // cutting and pasting mistakes. The macros below introduces a number
    43 // sequnce into the following two classes and the methods that access it.
    45 #define define_num_seq(name)                                                  \
    46 private:                                                                      \
    47   NumberSeq _all_##name##_times_ms;                                           \
    48 public:                                                                       \
    49   void record_##name##_time_ms(double ms) {                                   \
    50     _all_##name##_times_ms.add(ms);                                           \
    51   }                                                                           \
    52   NumberSeq* get_##name##_seq() {                                             \
    53     return &_all_##name##_times_ms;                                           \
    54   }
    56 class MainBodySummary;
    58 class PauseSummary: public CHeapObj {
    59   define_num_seq(total)
    60     define_num_seq(other)
    62 public:
    63   virtual MainBodySummary*    main_body_summary()    { return NULL; }
    64 };
    66 class MainBodySummary: public CHeapObj {
    67   define_num_seq(satb_drain) // optional
    68   define_num_seq(parallel) // parallel only
    69     define_num_seq(ext_root_scan)
    70     define_num_seq(mark_stack_scan)
    71     define_num_seq(update_rs)
    72     define_num_seq(scan_rs)
    73     define_num_seq(obj_copy)
    74     define_num_seq(termination) // parallel only
    75     define_num_seq(parallel_other) // parallel only
    76   define_num_seq(mark_closure)
    77   define_num_seq(clear_ct)  // parallel only
    78 };
    80 class Summary: public PauseSummary,
    81                public MainBodySummary {
    82 public:
    83   virtual MainBodySummary*    main_body_summary()    { return this; }
    84 };
    86 class G1CollectorPolicy: public CollectorPolicy {
    87 protected:
    88   // The number of pauses during the execution.
    89   long _n_pauses;
    91   // either equal to the number of parallel threads, if ParallelGCThreads
    92   // has been set, or 1 otherwise
    93   int _parallel_gc_threads;
    95   enum SomePrivateConstants {
    96     NumPrevPausesForHeuristics = 10
    97   };
    99   G1MMUTracker* _mmu_tracker;
   101   void initialize_flags();
   103   void initialize_all() {
   104     initialize_flags();
   105     initialize_size_info();
   106     initialize_perm_generation(PermGen::MarkSweepCompact);
   107   }
   109   virtual size_t default_init_heap_size() {
   110     // Pick some reasonable default.
   111     return 8*M;
   112   }
   114   double _cur_collection_start_sec;
   115   size_t _cur_collection_pause_used_at_start_bytes;
   116   size_t _cur_collection_pause_used_regions_at_start;
   117   size_t _prev_collection_pause_used_at_end_bytes;
   118   double _cur_collection_par_time_ms;
   119   double _cur_satb_drain_time_ms;
   120   double _cur_clear_ct_time_ms;
   121   bool   _satb_drain_time_set;
   123 #ifndef PRODUCT
   124   // Card Table Count Cache stats
   125   double _min_clear_cc_time_ms;         // min
   126   double _max_clear_cc_time_ms;         // max
   127   double _cur_clear_cc_time_ms;         // clearing time during current pause
   128   double _cum_clear_cc_time_ms;         // cummulative clearing time
   129   jlong  _num_cc_clears;                // number of times the card count cache has been cleared
   130 #endif
   132   // Statistics for recent GC pauses.  See below for how indexed.
   133   TruncatedSeq* _recent_rs_scan_times_ms;
   135   // These exclude marking times.
   136   TruncatedSeq* _recent_pause_times_ms;
   137   TruncatedSeq* _recent_gc_times_ms;
   139   TruncatedSeq* _recent_CS_bytes_used_before;
   140   TruncatedSeq* _recent_CS_bytes_surviving;
   142   TruncatedSeq* _recent_rs_sizes;
   144   TruncatedSeq* _concurrent_mark_init_times_ms;
   145   TruncatedSeq* _concurrent_mark_remark_times_ms;
   146   TruncatedSeq* _concurrent_mark_cleanup_times_ms;
   148   Summary*           _summary;
   150   NumberSeq* _all_pause_times_ms;
   151   NumberSeq* _all_full_gc_times_ms;
   152   double _stop_world_start;
   153   NumberSeq* _all_stop_world_times_ms;
   154   NumberSeq* _all_yield_times_ms;
   156   size_t     _region_num_young;
   157   size_t     _region_num_tenured;
   158   size_t     _prev_region_num_young;
   159   size_t     _prev_region_num_tenured;
   161   NumberSeq* _all_mod_union_times_ms;
   163   int        _aux_num;
   164   NumberSeq* _all_aux_times_ms;
   165   double*    _cur_aux_start_times_ms;
   166   double*    _cur_aux_times_ms;
   167   bool*      _cur_aux_times_set;
   169   double* _par_last_gc_worker_start_times_ms;
   170   double* _par_last_ext_root_scan_times_ms;
   171   double* _par_last_mark_stack_scan_times_ms;
   172   double* _par_last_update_rs_times_ms;
   173   double* _par_last_update_rs_processed_buffers;
   174   double* _par_last_scan_rs_times_ms;
   175   double* _par_last_obj_copy_times_ms;
   176   double* _par_last_termination_times_ms;
   177   double* _par_last_termination_attempts;
   178   double* _par_last_gc_worker_end_times_ms;
   179   double* _par_last_gc_worker_times_ms;
   181   // indicates that we are in young GC mode
   182   bool _in_young_gc_mode;
   184   // indicates whether we are in full young or partially young GC mode
   185   bool _full_young_gcs;
   187   // if true, then it tries to dynamically adjust the length of the
   188   // young list
   189   bool _adaptive_young_list_length;
   190   size_t _young_list_min_length;
   191   size_t _young_list_target_length;
   192   size_t _young_list_fixed_length;
   194   // The max number of regions we can extend the eden by while the GC
   195   // locker is active. This should be >= _young_list_target_length;
   196   size_t _young_list_max_length;
   198   size_t _young_cset_length;
   199   bool   _last_young_gc_full;
   201   unsigned              _full_young_pause_num;
   202   unsigned              _partial_young_pause_num;
   204   bool                  _during_marking;
   205   bool                  _in_marking_window;
   206   bool                  _in_marking_window_im;
   208   SurvRateGroup*        _short_lived_surv_rate_group;
   209   SurvRateGroup*        _survivor_surv_rate_group;
   210   // add here any more surv rate groups
   212   double                _gc_overhead_perc;
   214   bool during_marking() {
   215     return _during_marking;
   216   }
   218   // <NEW PREDICTION>
   220 private:
   221   enum PredictionConstants {
   222     TruncatedSeqLength = 10
   223   };
   225   TruncatedSeq* _alloc_rate_ms_seq;
   226   double        _prev_collection_pause_end_ms;
   228   TruncatedSeq* _pending_card_diff_seq;
   229   TruncatedSeq* _rs_length_diff_seq;
   230   TruncatedSeq* _cost_per_card_ms_seq;
   231   TruncatedSeq* _fully_young_cards_per_entry_ratio_seq;
   232   TruncatedSeq* _partially_young_cards_per_entry_ratio_seq;
   233   TruncatedSeq* _cost_per_entry_ms_seq;
   234   TruncatedSeq* _partially_young_cost_per_entry_ms_seq;
   235   TruncatedSeq* _cost_per_byte_ms_seq;
   236   TruncatedSeq* _constant_other_time_ms_seq;
   237   TruncatedSeq* _young_other_cost_per_region_ms_seq;
   238   TruncatedSeq* _non_young_other_cost_per_region_ms_seq;
   240   TruncatedSeq* _pending_cards_seq;
   241   TruncatedSeq* _scanned_cards_seq;
   242   TruncatedSeq* _rs_lengths_seq;
   244   TruncatedSeq* _cost_per_byte_ms_during_cm_seq;
   246   TruncatedSeq* _young_gc_eff_seq;
   248   TruncatedSeq* _max_conc_overhead_seq;
   250   size_t _recorded_young_regions;
   251   size_t _recorded_non_young_regions;
   252   size_t _recorded_region_num;
   254   size_t _free_regions_at_end_of_collection;
   256   size_t _recorded_rs_lengths;
   257   size_t _max_rs_lengths;
   259   size_t _recorded_marked_bytes;
   260   size_t _recorded_young_bytes;
   262   size_t _predicted_pending_cards;
   263   size_t _predicted_cards_scanned;
   264   size_t _predicted_rs_lengths;
   265   size_t _predicted_bytes_to_copy;
   267   double _predicted_survival_ratio;
   268   double _predicted_rs_update_time_ms;
   269   double _predicted_rs_scan_time_ms;
   270   double _predicted_object_copy_time_ms;
   271   double _predicted_constant_other_time_ms;
   272   double _predicted_young_other_time_ms;
   273   double _predicted_non_young_other_time_ms;
   274   double _predicted_pause_time_ms;
   276   double _vtime_diff_ms;
   278   double _recorded_young_free_cset_time_ms;
   279   double _recorded_non_young_free_cset_time_ms;
   281   double _sigma;
   282   double _expensive_region_limit_ms;
   284   size_t _rs_lengths_prediction;
   286   size_t _known_garbage_bytes;
   287   double _known_garbage_ratio;
   289   double sigma() {
   290     return _sigma;
   291   }
   293   // A function that prevents us putting too much stock in small sample
   294   // sets.  Returns a number between 2.0 and 1.0, depending on the number
   295   // of samples.  5 or more samples yields one; fewer scales linearly from
   296   // 2.0 at 1 sample to 1.0 at 5.
   297   double confidence_factor(int samples) {
   298     if (samples > 4) return 1.0;
   299     else return  1.0 + sigma() * ((double)(5 - samples))/2.0;
   300   }
   302   double get_new_neg_prediction(TruncatedSeq* seq) {
   303     return seq->davg() - sigma() * seq->dsd();
   304   }
   306 #ifndef PRODUCT
   307   bool verify_young_ages(HeapRegion* head, SurvRateGroup *surv_rate_group);
   308 #endif // PRODUCT
   310   void adjust_concurrent_refinement(double update_rs_time,
   311                                     double update_rs_processed_buffers,
   312                                     double goal_ms);
   314 protected:
   315   double _pause_time_target_ms;
   316   double _recorded_young_cset_choice_time_ms;
   317   double _recorded_non_young_cset_choice_time_ms;
   318   bool   _within_target;
   319   size_t _pending_cards;
   320   size_t _max_pending_cards;
   322 public:
   324   void set_region_short_lived(HeapRegion* hr) {
   325     hr->install_surv_rate_group(_short_lived_surv_rate_group);
   326   }
   328   void set_region_survivors(HeapRegion* hr) {
   329     hr->install_surv_rate_group(_survivor_surv_rate_group);
   330   }
   332 #ifndef PRODUCT
   333   bool verify_young_ages();
   334 #endif // PRODUCT
   336   double get_new_prediction(TruncatedSeq* seq) {
   337     return MAX2(seq->davg() + sigma() * seq->dsd(),
   338                 seq->davg() * confidence_factor(seq->num()));
   339   }
   341   size_t young_cset_length() {
   342     return _young_cset_length;
   343   }
   345   void record_max_rs_lengths(size_t rs_lengths) {
   346     _max_rs_lengths = rs_lengths;
   347   }
   349   size_t predict_pending_card_diff() {
   350     double prediction = get_new_neg_prediction(_pending_card_diff_seq);
   351     if (prediction < 0.00001)
   352       return 0;
   353     else
   354       return (size_t) prediction;
   355   }
   357   size_t predict_pending_cards() {
   358     size_t max_pending_card_num = _g1->max_pending_card_num();
   359     size_t diff = predict_pending_card_diff();
   360     size_t prediction;
   361     if (diff > max_pending_card_num)
   362       prediction = max_pending_card_num;
   363     else
   364       prediction = max_pending_card_num - diff;
   366     return prediction;
   367   }
   369   size_t predict_rs_length_diff() {
   370     return (size_t) get_new_prediction(_rs_length_diff_seq);
   371   }
   373   double predict_alloc_rate_ms() {
   374     return get_new_prediction(_alloc_rate_ms_seq);
   375   }
   377   double predict_cost_per_card_ms() {
   378     return get_new_prediction(_cost_per_card_ms_seq);
   379   }
   381   double predict_rs_update_time_ms(size_t pending_cards) {
   382     return (double) pending_cards * predict_cost_per_card_ms();
   383   }
   385   double predict_fully_young_cards_per_entry_ratio() {
   386     return get_new_prediction(_fully_young_cards_per_entry_ratio_seq);
   387   }
   389   double predict_partially_young_cards_per_entry_ratio() {
   390     if (_partially_young_cards_per_entry_ratio_seq->num() < 2)
   391       return predict_fully_young_cards_per_entry_ratio();
   392     else
   393       return get_new_prediction(_partially_young_cards_per_entry_ratio_seq);
   394   }
   396   size_t predict_young_card_num(size_t rs_length) {
   397     return (size_t) ((double) rs_length *
   398                      predict_fully_young_cards_per_entry_ratio());
   399   }
   401   size_t predict_non_young_card_num(size_t rs_length) {
   402     return (size_t) ((double) rs_length *
   403                      predict_partially_young_cards_per_entry_ratio());
   404   }
   406   double predict_rs_scan_time_ms(size_t card_num) {
   407     if (full_young_gcs())
   408       return (double) card_num * get_new_prediction(_cost_per_entry_ms_seq);
   409     else
   410       return predict_partially_young_rs_scan_time_ms(card_num);
   411   }
   413   double predict_partially_young_rs_scan_time_ms(size_t card_num) {
   414     if (_partially_young_cost_per_entry_ms_seq->num() < 3)
   415       return (double) card_num * get_new_prediction(_cost_per_entry_ms_seq);
   416     else
   417       return (double) card_num *
   418         get_new_prediction(_partially_young_cost_per_entry_ms_seq);
   419   }
   421   double predict_object_copy_time_ms_during_cm(size_t bytes_to_copy) {
   422     if (_cost_per_byte_ms_during_cm_seq->num() < 3)
   423       return 1.1 * (double) bytes_to_copy *
   424         get_new_prediction(_cost_per_byte_ms_seq);
   425     else
   426       return (double) bytes_to_copy *
   427         get_new_prediction(_cost_per_byte_ms_during_cm_seq);
   428   }
   430   double predict_object_copy_time_ms(size_t bytes_to_copy) {
   431     if (_in_marking_window && !_in_marking_window_im)
   432       return predict_object_copy_time_ms_during_cm(bytes_to_copy);
   433     else
   434       return (double) bytes_to_copy *
   435         get_new_prediction(_cost_per_byte_ms_seq);
   436   }
   438   double predict_constant_other_time_ms() {
   439     return get_new_prediction(_constant_other_time_ms_seq);
   440   }
   442   double predict_young_other_time_ms(size_t young_num) {
   443     return
   444       (double) young_num *
   445       get_new_prediction(_young_other_cost_per_region_ms_seq);
   446   }
   448   double predict_non_young_other_time_ms(size_t non_young_num) {
   449     return
   450       (double) non_young_num *
   451       get_new_prediction(_non_young_other_cost_per_region_ms_seq);
   452   }
   454   void check_if_region_is_too_expensive(double predicted_time_ms);
   456   double predict_young_collection_elapsed_time_ms(size_t adjustment);
   457   double predict_base_elapsed_time_ms(size_t pending_cards);
   458   double predict_base_elapsed_time_ms(size_t pending_cards,
   459                                       size_t scanned_cards);
   460   size_t predict_bytes_to_copy(HeapRegion* hr);
   461   double predict_region_elapsed_time_ms(HeapRegion* hr, bool young);
   463     // for use by: calculate_young_list_target_length(rs_length)
   464   bool predict_will_fit(size_t young_region_num,
   465                         double base_time_ms,
   466                         size_t init_free_regions,
   467                         double target_pause_time_ms);
   469   void start_recording_regions();
   470   void record_cset_region_info(HeapRegion* hr, bool young);
   471   void record_non_young_cset_region(HeapRegion* hr);
   473   void set_recorded_young_regions(size_t n_regions);
   474   void set_recorded_young_bytes(size_t bytes);
   475   void set_recorded_rs_lengths(size_t rs_lengths);
   476   void set_predicted_bytes_to_copy(size_t bytes);
   478   void end_recording_regions();
   480   void record_vtime_diff_ms(double vtime_diff_ms) {
   481     _vtime_diff_ms = vtime_diff_ms;
   482   }
   484   void record_young_free_cset_time_ms(double time_ms) {
   485     _recorded_young_free_cset_time_ms = time_ms;
   486   }
   488   void record_non_young_free_cset_time_ms(double time_ms) {
   489     _recorded_non_young_free_cset_time_ms = time_ms;
   490   }
   492   double predict_young_gc_eff() {
   493     return get_new_neg_prediction(_young_gc_eff_seq);
   494   }
   496   double predict_survivor_regions_evac_time();
   498   // </NEW PREDICTION>
   500 public:
   501   void cset_regions_freed() {
   502     bool propagate = _last_young_gc_full && !_in_marking_window;
   503     _short_lived_surv_rate_group->all_surviving_words_recorded(propagate);
   504     _survivor_surv_rate_group->all_surviving_words_recorded(propagate);
   505     // also call it on any more surv rate groups
   506   }
   508   void set_known_garbage_bytes(size_t known_garbage_bytes) {
   509     _known_garbage_bytes = known_garbage_bytes;
   510     size_t heap_bytes = _g1->capacity();
   511     _known_garbage_ratio = (double) _known_garbage_bytes / (double) heap_bytes;
   512   }
   514   void decrease_known_garbage_bytes(size_t known_garbage_bytes) {
   515     guarantee( _known_garbage_bytes >= known_garbage_bytes, "invariant" );
   517     _known_garbage_bytes -= known_garbage_bytes;
   518     size_t heap_bytes = _g1->capacity();
   519     _known_garbage_ratio = (double) _known_garbage_bytes / (double) heap_bytes;
   520   }
   522   G1MMUTracker* mmu_tracker() {
   523     return _mmu_tracker;
   524   }
   526   double max_pause_time_ms() {
   527     return _mmu_tracker->max_gc_time() * 1000.0;
   528   }
   530   double predict_init_time_ms() {
   531     return get_new_prediction(_concurrent_mark_init_times_ms);
   532   }
   534   double predict_remark_time_ms() {
   535     return get_new_prediction(_concurrent_mark_remark_times_ms);
   536   }
   538   double predict_cleanup_time_ms() {
   539     return get_new_prediction(_concurrent_mark_cleanup_times_ms);
   540   }
   542   // Returns an estimate of the survival rate of the region at yg-age
   543   // "yg_age".
   544   double predict_yg_surv_rate(int age, SurvRateGroup* surv_rate_group) {
   545     TruncatedSeq* seq = surv_rate_group->get_seq(age);
   546     if (seq->num() == 0)
   547       gclog_or_tty->print("BARF! age is %d", age);
   548     guarantee( seq->num() > 0, "invariant" );
   549     double pred = get_new_prediction(seq);
   550     if (pred > 1.0)
   551       pred = 1.0;
   552     return pred;
   553   }
   555   double predict_yg_surv_rate(int age) {
   556     return predict_yg_surv_rate(age, _short_lived_surv_rate_group);
   557   }
   559   double accum_yg_surv_rate_pred(int age) {
   560     return _short_lived_surv_rate_group->accum_surv_rate_pred(age);
   561   }
   563 protected:
   564   void print_stats(int level, const char* str, double value);
   565   void print_stats(int level, const char* str, int value);
   567   void print_par_stats(int level, const char* str, double* data);
   568   void print_par_sizes(int level, const char* str, double* data);
   570   void check_other_times(int level,
   571                          NumberSeq* other_times_ms,
   572                          NumberSeq* calc_other_times_ms) const;
   574   void print_summary (PauseSummary* stats) const;
   576   void print_summary (int level, const char* str, NumberSeq* seq) const;
   577   void print_summary_sd (int level, const char* str, NumberSeq* seq) const;
   579   double avg_value (double* data);
   580   double max_value (double* data);
   581   double sum_of_values (double* data);
   582   double max_sum (double* data1, double* data2);
   584   int _last_satb_drain_processed_buffers;
   585   int _last_update_rs_processed_buffers;
   586   double _last_pause_time_ms;
   588   size_t _bytes_in_collection_set_before_gc;
   589   size_t _bytes_copied_during_gc;
   591   // Used to count used bytes in CS.
   592   friend class CountCSClosure;
   594   // Statistics kept per GC stoppage, pause or full.
   595   TruncatedSeq* _recent_prev_end_times_for_all_gcs_sec;
   597   // We track markings.
   598   int _num_markings;
   599   double _mark_thread_startup_sec;       // Time at startup of marking thread
   601   // Add a new GC of the given duration and end time to the record.
   602   void update_recent_gc_times(double end_time_sec, double elapsed_ms);
   604   // The head of the list (via "next_in_collection_set()") representing the
   605   // current collection set. Set from the incrementally built collection
   606   // set at the start of the pause.
   607   HeapRegion* _collection_set;
   609   // The number of regions in the collection set. Set from the incrementally
   610   // built collection set at the start of an evacuation pause.
   611   size_t _collection_set_size;
   613   // The number of bytes in the collection set before the pause. Set from
   614   // the incrementally built collection set at the start of an evacuation
   615   // pause.
   616   size_t _collection_set_bytes_used_before;
   618   // The associated information that is maintained while the incremental
   619   // collection set is being built with young regions. Used to populate
   620   // the recorded info for the evacuation pause.
   622   enum CSetBuildType {
   623     Active,             // We are actively building the collection set
   624     Inactive            // We are not actively building the collection set
   625   };
   627   CSetBuildType _inc_cset_build_state;
   629   // The head of the incrementally built collection set.
   630   HeapRegion* _inc_cset_head;
   632   // The tail of the incrementally built collection set.
   633   HeapRegion* _inc_cset_tail;
   635   // The number of regions in the incrementally built collection set.
   636   // Used to set _collection_set_size at the start of an evacuation
   637   // pause.
   638   size_t _inc_cset_size;
   640   // Used as the index in the surving young words structure
   641   // which tracks the amount of space, for each young region,
   642   // that survives the pause.
   643   size_t _inc_cset_young_index;
   645   // The number of bytes in the incrementally built collection set.
   646   // Used to set _collection_set_bytes_used_before at the start of
   647   // an evacuation pause.
   648   size_t _inc_cset_bytes_used_before;
   650   // Used to record the highest end of heap region in collection set
   651   HeapWord* _inc_cset_max_finger;
   653   // The number of recorded used bytes in the young regions
   654   // of the collection set. This is the sum of the used() bytes
   655   // of retired young regions in the collection set.
   656   size_t _inc_cset_recorded_young_bytes;
   658   // The RSet lengths recorded for regions in the collection set
   659   // (updated by the periodic sampling of the regions in the
   660   // young list/collection set).
   661   size_t _inc_cset_recorded_rs_lengths;
   663   // The predicted elapsed time it will take to collect the regions
   664   // in the collection set (updated by the periodic sampling of the
   665   // regions in the young list/collection set).
   666   double _inc_cset_predicted_elapsed_time_ms;
   668   // The predicted bytes to copy for the regions in the collection
   669   // set (updated by the periodic sampling of the regions in the
   670   // young list/collection set).
   671   size_t _inc_cset_predicted_bytes_to_copy;
   673   // Info about marking.
   674   int _n_marks; // Sticky at 2, so we know when we've done at least 2.
   676   // The number of collection pauses at the end of the last mark.
   677   size_t _n_pauses_at_mark_end;
   679   // Stash a pointer to the g1 heap.
   680   G1CollectedHeap* _g1;
   682   // The average time in ms per collection pause, averaged over recent pauses.
   683   double recent_avg_time_for_pauses_ms();
   685   // The average time in ms for RS scanning, per pause, averaged
   686   // over recent pauses. (Note the RS scanning time for a pause
   687   // is itself an average of the RS scanning time for each worker
   688   // thread.)
   689   double recent_avg_time_for_rs_scan_ms();
   691   // The number of "recent" GCs recorded in the number sequences
   692   int number_of_recent_gcs();
   694   // The average survival ratio, computed by the total number of bytes
   695   // suriviving / total number of bytes before collection over the last
   696   // several recent pauses.
   697   double recent_avg_survival_fraction();
   698   // The survival fraction of the most recent pause; if there have been no
   699   // pauses, returns 1.0.
   700   double last_survival_fraction();
   702   // Returns a "conservative" estimate of the recent survival rate, i.e.,
   703   // one that may be higher than "recent_avg_survival_fraction".
   704   // This is conservative in several ways:
   705   //   If there have been few pauses, it will assume a potential high
   706   //     variance, and err on the side of caution.
   707   //   It puts a lower bound (currently 0.1) on the value it will return.
   708   //   To try to detect phase changes, if the most recent pause ("latest") has a
   709   //     higher-than average ("avg") survival rate, it returns that rate.
   710   // "work" version is a utility function; young is restricted to young regions.
   711   double conservative_avg_survival_fraction_work(double avg,
   712                                                  double latest);
   714   // The arguments are the two sequences that keep track of the number of bytes
   715   //   surviving and the total number of bytes before collection, resp.,
   716   //   over the last evereal recent pauses
   717   // Returns the survival rate for the category in the most recent pause.
   718   // If there have been no pauses, returns 1.0.
   719   double last_survival_fraction_work(TruncatedSeq* surviving,
   720                                      TruncatedSeq* before);
   722   // The arguments are the two sequences that keep track of the number of bytes
   723   //   surviving and the total number of bytes before collection, resp.,
   724   //   over the last several recent pauses
   725   // Returns the average survival ration over the last several recent pauses
   726   // If there have been no pauses, return 1.0
   727   double recent_avg_survival_fraction_work(TruncatedSeq* surviving,
   728                                            TruncatedSeq* before);
   730   double conservative_avg_survival_fraction() {
   731     double avg = recent_avg_survival_fraction();
   732     double latest = last_survival_fraction();
   733     return conservative_avg_survival_fraction_work(avg, latest);
   734   }
   736   // The ratio of gc time to elapsed time, computed over recent pauses.
   737   double _recent_avg_pause_time_ratio;
   739   double recent_avg_pause_time_ratio() {
   740     return _recent_avg_pause_time_ratio;
   741   }
   743   // Number of pauses between concurrent marking.
   744   size_t _pauses_btwn_concurrent_mark;
   746   size_t _n_marks_since_last_pause;
   748   // At the end of a pause we check the heap occupancy and we decide
   749   // whether we will start a marking cycle during the next pause. If
   750   // we decide that we want to do that, we will set this parameter to
   751   // true. So, this parameter will stay true between the end of a
   752   // pause and the beginning of a subsequent pause (not necessarily
   753   // the next one, see the comments on the next field) when we decide
   754   // that we will indeed start a marking cycle and do the initial-mark
   755   // work.
   756   volatile bool _initiate_conc_mark_if_possible;
   758   // If initiate_conc_mark_if_possible() is set at the beginning of a
   759   // pause, it is a suggestion that the pause should start a marking
   760   // cycle by doing the initial-mark work. However, it is possible
   761   // that the concurrent marking thread is still finishing up the
   762   // previous marking cycle (e.g., clearing the next marking
   763   // bitmap). If that is the case we cannot start a new cycle and
   764   // we'll have to wait for the concurrent marking thread to finish
   765   // what it is doing. In this case we will postpone the marking cycle
   766   // initiation decision for the next pause. When we eventually decide
   767   // to start a cycle, we will set _during_initial_mark_pause which
   768   // will stay true until the end of the initial-mark pause and it's
   769   // the condition that indicates that a pause is doing the
   770   // initial-mark work.
   771   volatile bool _during_initial_mark_pause;
   773   bool _should_revert_to_full_young_gcs;
   774   bool _last_full_young_gc;
   776   // This set of variables tracks the collector efficiency, in order to
   777   // determine whether we should initiate a new marking.
   778   double _cur_mark_stop_world_time_ms;
   779   double _mark_init_start_sec;
   780   double _mark_remark_start_sec;
   781   double _mark_cleanup_start_sec;
   782   double _mark_closure_time_ms;
   784   void   calculate_young_list_min_length();
   785   void   calculate_young_list_target_length();
   786   void   calculate_young_list_target_length(size_t rs_lengths);
   788 public:
   790   G1CollectorPolicy();
   792   virtual G1CollectorPolicy* as_g1_policy() { return this; }
   794   virtual CollectorPolicy::Name kind() {
   795     return CollectorPolicy::G1CollectorPolicyKind;
   796   }
   798   void check_prediction_validity();
   800   size_t bytes_in_collection_set() {
   801     return _bytes_in_collection_set_before_gc;
   802   }
   804   unsigned calc_gc_alloc_time_stamp() {
   805     return _all_pause_times_ms->num() + 1;
   806   }
   808 protected:
   810   // Count the number of bytes used in the CS.
   811   void count_CS_bytes_used();
   813   // Together these do the base cleanup-recording work.  Subclasses might
   814   // want to put something between them.
   815   void record_concurrent_mark_cleanup_end_work1(size_t freed_bytes,
   816                                                 size_t max_live_bytes);
   817   void record_concurrent_mark_cleanup_end_work2();
   819 public:
   821   virtual void init();
   823   // Create jstat counters for the policy.
   824   virtual void initialize_gc_policy_counters();
   826   virtual HeapWord* mem_allocate_work(size_t size,
   827                                       bool is_tlab,
   828                                       bool* gc_overhead_limit_was_exceeded);
   830   // This method controls how a collector handles one or more
   831   // of its generations being fully allocated.
   832   virtual HeapWord* satisfy_failed_allocation(size_t size,
   833                                               bool is_tlab);
   835   BarrierSet::Name barrier_set_name() { return BarrierSet::G1SATBCTLogging; }
   837   GenRemSet::Name  rem_set_name()     { return GenRemSet::CardTable; }
   839   // The number of collection pauses so far.
   840   long n_pauses() const { return _n_pauses; }
   842   // Update the heuristic info to record a collection pause of the given
   843   // start time, where the given number of bytes were used at the start.
   844   // This may involve changing the desired size of a collection set.
   846   virtual void record_stop_world_start();
   848   virtual void record_collection_pause_start(double start_time_sec,
   849                                              size_t start_used);
   851   // Must currently be called while the world is stopped.
   852   virtual void record_concurrent_mark_init_start();
   853   virtual void record_concurrent_mark_init_end();
   854   void record_concurrent_mark_init_end_pre(double
   855                                            mark_init_elapsed_time_ms);
   857   void record_mark_closure_time(double mark_closure_time_ms);
   859   virtual void record_concurrent_mark_remark_start();
   860   virtual void record_concurrent_mark_remark_end();
   862   virtual void record_concurrent_mark_cleanup_start();
   863   virtual void record_concurrent_mark_cleanup_end(size_t freed_bytes,
   864                                                   size_t max_live_bytes);
   865   virtual void record_concurrent_mark_cleanup_completed();
   867   virtual void record_concurrent_pause();
   868   virtual void record_concurrent_pause_end();
   870   virtual void record_collection_pause_end();
   871   void print_heap_transition();
   873   // Record the fact that a full collection occurred.
   874   virtual void record_full_collection_start();
   875   virtual void record_full_collection_end();
   877   void record_gc_worker_start_time(int worker_i, double ms) {
   878     _par_last_gc_worker_start_times_ms[worker_i] = ms;
   879   }
   881   void record_ext_root_scan_time(int worker_i, double ms) {
   882     _par_last_ext_root_scan_times_ms[worker_i] = ms;
   883   }
   885   void record_mark_stack_scan_time(int worker_i, double ms) {
   886     _par_last_mark_stack_scan_times_ms[worker_i] = ms;
   887   }
   889   void record_satb_drain_time(double ms) {
   890     _cur_satb_drain_time_ms = ms;
   891     _satb_drain_time_set    = true;
   892   }
   894   void record_satb_drain_processed_buffers (int processed_buffers) {
   895     _last_satb_drain_processed_buffers = processed_buffers;
   896   }
   898   void record_mod_union_time(double ms) {
   899     _all_mod_union_times_ms->add(ms);
   900   }
   902   void record_update_rs_time(int thread, double ms) {
   903     _par_last_update_rs_times_ms[thread] = ms;
   904   }
   906   void record_update_rs_processed_buffers (int thread,
   907                                            double processed_buffers) {
   908     _par_last_update_rs_processed_buffers[thread] = processed_buffers;
   909   }
   911   void record_scan_rs_time(int thread, double ms) {
   912     _par_last_scan_rs_times_ms[thread] = ms;
   913   }
   915   void reset_obj_copy_time(int thread) {
   916     _par_last_obj_copy_times_ms[thread] = 0.0;
   917   }
   919   void reset_obj_copy_time() {
   920     reset_obj_copy_time(0);
   921   }
   923   void record_obj_copy_time(int thread, double ms) {
   924     _par_last_obj_copy_times_ms[thread] += ms;
   925   }
   927   void record_termination(int thread, double ms, size_t attempts) {
   928     _par_last_termination_times_ms[thread] = ms;
   929     _par_last_termination_attempts[thread] = (double) attempts;
   930   }
   932   void record_gc_worker_end_time(int worker_i, double ms) {
   933     _par_last_gc_worker_end_times_ms[worker_i] = ms;
   934   }
   936   void record_pause_time_ms(double ms) {
   937     _last_pause_time_ms = ms;
   938   }
   940   void record_clear_ct_time(double ms) {
   941     _cur_clear_ct_time_ms = ms;
   942   }
   944   void record_par_time(double ms) {
   945     _cur_collection_par_time_ms = ms;
   946   }
   948   void record_aux_start_time(int i) {
   949     guarantee(i < _aux_num, "should be within range");
   950     _cur_aux_start_times_ms[i] = os::elapsedTime() * 1000.0;
   951   }
   953   void record_aux_end_time(int i) {
   954     guarantee(i < _aux_num, "should be within range");
   955     double ms = os::elapsedTime() * 1000.0 - _cur_aux_start_times_ms[i];
   956     _cur_aux_times_set[i] = true;
   957     _cur_aux_times_ms[i] += ms;
   958   }
   960 #ifndef PRODUCT
   961   void record_cc_clear_time(double ms) {
   962     if (_min_clear_cc_time_ms < 0.0 || ms <= _min_clear_cc_time_ms)
   963       _min_clear_cc_time_ms = ms;
   964     if (_max_clear_cc_time_ms < 0.0 || ms >= _max_clear_cc_time_ms)
   965       _max_clear_cc_time_ms = ms;
   966     _cur_clear_cc_time_ms = ms;
   967     _cum_clear_cc_time_ms += ms;
   968     _num_cc_clears++;
   969   }
   970 #endif
   972   // Record how much space we copied during a GC. This is typically
   973   // called when a GC alloc region is being retired.
   974   void record_bytes_copied_during_gc(size_t bytes) {
   975     _bytes_copied_during_gc += bytes;
   976   }
   978   // The amount of space we copied during a GC.
   979   size_t bytes_copied_during_gc() {
   980     return _bytes_copied_during_gc;
   981   }
   983   // Choose a new collection set.  Marks the chosen regions as being
   984   // "in_collection_set", and links them together.  The head and number of
   985   // the collection set are available via access methods.
   986   virtual void choose_collection_set(double target_pause_time_ms) = 0;
   988   // The head of the list (via "next_in_collection_set()") representing the
   989   // current collection set.
   990   HeapRegion* collection_set() { return _collection_set; }
   992   void clear_collection_set() { _collection_set = NULL; }
   994   // The number of elements in the current collection set.
   995   size_t collection_set_size() { return _collection_set_size; }
   997   // Add "hr" to the CS.
   998   void add_to_collection_set(HeapRegion* hr);
  1000   // Incremental CSet Support
  1002   // The head of the incrementally built collection set.
  1003   HeapRegion* inc_cset_head() { return _inc_cset_head; }
  1005   // The tail of the incrementally built collection set.
  1006   HeapRegion* inc_set_tail() { return _inc_cset_tail; }
  1008   // The number of elements in the incrementally built collection set.
  1009   size_t inc_cset_size() { return _inc_cset_size; }
  1011   // Initialize incremental collection set info.
  1012   void start_incremental_cset_building();
  1014   void clear_incremental_cset() {
  1015     _inc_cset_head = NULL;
  1016     _inc_cset_tail = NULL;
  1019   // Stop adding regions to the incremental collection set
  1020   void stop_incremental_cset_building() { _inc_cset_build_state = Inactive; }
  1022   // Add/remove information about hr to the aggregated information
  1023   // for the incrementally built collection set.
  1024   void add_to_incremental_cset_info(HeapRegion* hr, size_t rs_length);
  1025   void remove_from_incremental_cset_info(HeapRegion* hr);
  1027   // Update information about hr in the aggregated information for
  1028   // the incrementally built collection set.
  1029   void update_incremental_cset_info(HeapRegion* hr, size_t new_rs_length);
  1031 private:
  1032   // Update the incremental cset information when adding a region
  1033   // (should not be called directly).
  1034   void add_region_to_incremental_cset_common(HeapRegion* hr);
  1036 public:
  1037   // Add hr to the LHS of the incremental collection set.
  1038   void add_region_to_incremental_cset_lhs(HeapRegion* hr);
  1040   // Add hr to the RHS of the incremental collection set.
  1041   void add_region_to_incremental_cset_rhs(HeapRegion* hr);
  1043 #ifndef PRODUCT
  1044   void print_collection_set(HeapRegion* list_head, outputStream* st);
  1045 #endif // !PRODUCT
  1047   bool initiate_conc_mark_if_possible()       { return _initiate_conc_mark_if_possible;  }
  1048   void set_initiate_conc_mark_if_possible()   { _initiate_conc_mark_if_possible = true;  }
  1049   void clear_initiate_conc_mark_if_possible() { _initiate_conc_mark_if_possible = false; }
  1051   bool during_initial_mark_pause()      { return _during_initial_mark_pause;  }
  1052   void set_during_initial_mark_pause()  { _during_initial_mark_pause = true;  }
  1053   void clear_during_initial_mark_pause(){ _during_initial_mark_pause = false; }
  1055   // This sets the initiate_conc_mark_if_possible() flag to start a
  1056   // new cycle, as long as we are not already in one. It's best if it
  1057   // is called during a safepoint when the test whether a cycle is in
  1058   // progress or not is stable.
  1059   bool force_initial_mark_if_outside_cycle();
  1061   // This is called at the very beginning of an evacuation pause (it
  1062   // has to be the first thing that the pause does). If
  1063   // initiate_conc_mark_if_possible() is true, and the concurrent
  1064   // marking thread has completed its work during the previous cycle,
  1065   // it will set during_initial_mark_pause() to so that the pause does
  1066   // the initial-mark work and start a marking cycle.
  1067   void decide_on_conc_mark_initiation();
  1069   // If an expansion would be appropriate, because recent GC overhead had
  1070   // exceeded the desired limit, return an amount to expand by.
  1071   virtual size_t expansion_amount();
  1073   // note start of mark thread
  1074   void note_start_of_mark_thread();
  1076   // The marked bytes of the "r" has changed; reclassify it's desirability
  1077   // for marking.  Also asserts that "r" is eligible for a CS.
  1078   virtual void note_change_in_marked_bytes(HeapRegion* r) = 0;
  1080 #ifndef PRODUCT
  1081   // Check any appropriate marked bytes info, asserting false if
  1082   // something's wrong, else returning "true".
  1083   virtual bool assertMarkedBytesDataOK() = 0;
  1084 #endif
  1086   // Print tracing information.
  1087   void print_tracing_info() const;
  1089   // Print stats on young survival ratio
  1090   void print_yg_surv_rate_info() const;
  1092   void finished_recalculating_age_indexes(bool is_survivors) {
  1093     if (is_survivors) {
  1094       _survivor_surv_rate_group->finished_recalculating_age_indexes();
  1095     } else {
  1096       _short_lived_surv_rate_group->finished_recalculating_age_indexes();
  1098     // do that for any other surv rate groups
  1101   bool is_young_list_full() {
  1102     size_t young_list_length = _g1->young_list()->length();
  1103     size_t young_list_target_length = _young_list_target_length;
  1104     if (G1FixedEdenSize) {
  1105       young_list_target_length -= _max_survivor_regions;
  1107     return young_list_length >= young_list_target_length;
  1110   bool can_expand_young_list() {
  1111     size_t young_list_length = _g1->young_list()->length();
  1112     size_t young_list_max_length = _young_list_max_length;
  1113     if (G1FixedEdenSize) {
  1114       young_list_max_length -= _max_survivor_regions;
  1116     return young_list_length < young_list_max_length;
  1119   void update_region_num(bool young);
  1121   bool in_young_gc_mode() {
  1122     return _in_young_gc_mode;
  1124   void set_in_young_gc_mode(bool in_young_gc_mode) {
  1125     _in_young_gc_mode = in_young_gc_mode;
  1128   bool full_young_gcs() {
  1129     return _full_young_gcs;
  1131   void set_full_young_gcs(bool full_young_gcs) {
  1132     _full_young_gcs = full_young_gcs;
  1135   bool adaptive_young_list_length() {
  1136     return _adaptive_young_list_length;
  1138   void set_adaptive_young_list_length(bool adaptive_young_list_length) {
  1139     _adaptive_young_list_length = adaptive_young_list_length;
  1142   inline double get_gc_eff_factor() {
  1143     double ratio = _known_garbage_ratio;
  1145     double square = ratio * ratio;
  1146     // square = square * square;
  1147     double ret = square * 9.0 + 1.0;
  1148 #if 0
  1149     gclog_or_tty->print_cr("ratio = %1.2lf, ret = %1.2lf", ratio, ret);
  1150 #endif // 0
  1151     guarantee(0.0 <= ret && ret < 10.0, "invariant!");
  1152     return ret;
  1155   //
  1156   // Survivor regions policy.
  1157   //
  1158 protected:
  1160   // Current tenuring threshold, set to 0 if the collector reaches the
  1161   // maximum amount of suvivors regions.
  1162   int _tenuring_threshold;
  1164   // The limit on the number of regions allocated for survivors.
  1165   size_t _max_survivor_regions;
  1167   // For reporting purposes.
  1168   size_t _eden_bytes_before_gc;
  1169   size_t _survivor_bytes_before_gc;
  1170   size_t _capacity_before_gc;
  1172   // The amount of survor regions after a collection.
  1173   size_t _recorded_survivor_regions;
  1174   // List of survivor regions.
  1175   HeapRegion* _recorded_survivor_head;
  1176   HeapRegion* _recorded_survivor_tail;
  1178   ageTable _survivors_age_table;
  1180 public:
  1182   inline GCAllocPurpose
  1183     evacuation_destination(HeapRegion* src_region, int age, size_t word_sz) {
  1184       if (age < _tenuring_threshold && src_region->is_young()) {
  1185         return GCAllocForSurvived;
  1186       } else {
  1187         return GCAllocForTenured;
  1191   inline bool track_object_age(GCAllocPurpose purpose) {
  1192     return purpose == GCAllocForSurvived;
  1195   static const size_t REGIONS_UNLIMITED = ~(size_t)0;
  1197   size_t max_regions(int purpose);
  1199   // The limit on regions for a particular purpose is reached.
  1200   void note_alloc_region_limit_reached(int purpose) {
  1201     if (purpose == GCAllocForSurvived) {
  1202       _tenuring_threshold = 0;
  1206   void note_start_adding_survivor_regions() {
  1207     _survivor_surv_rate_group->start_adding_regions();
  1210   void note_stop_adding_survivor_regions() {
  1211     _survivor_surv_rate_group->stop_adding_regions();
  1214   void record_survivor_regions(size_t      regions,
  1215                                HeapRegion* head,
  1216                                HeapRegion* tail) {
  1217     _recorded_survivor_regions = regions;
  1218     _recorded_survivor_head    = head;
  1219     _recorded_survivor_tail    = tail;
  1222   size_t recorded_survivor_regions() {
  1223     return _recorded_survivor_regions;
  1226   void record_thread_age_table(ageTable* age_table)
  1228     _survivors_age_table.merge_par(age_table);
  1231   void calculate_max_gc_locker_expansion();
  1233   // Calculates survivor space parameters.
  1234   void calculate_survivors_policy();
  1236 };
  1238 // This encapsulates a particular strategy for a g1 Collector.
  1239 //
  1240 //      Start a concurrent mark when our heap size is n bytes
  1241 //            greater then our heap size was at the last concurrent
  1242 //            mark.  Where n is a function of the CMSTriggerRatio
  1243 //            and the MinHeapFreeRatio.
  1244 //
  1245 //      Start a g1 collection pause when we have allocated the
  1246 //            average number of bytes currently being freed in
  1247 //            a collection, but only if it is at least one region
  1248 //            full
  1249 //
  1250 //      Resize Heap based on desired
  1251 //      allocation space, where desired allocation space is
  1252 //      a function of survival rate and desired future to size.
  1253 //
  1254 //      Choose collection set by first picking all older regions
  1255 //      which have a survival rate which beats our projected young
  1256 //      survival rate.  Then fill out the number of needed regions
  1257 //      with young regions.
  1259 class G1CollectorPolicy_BestRegionsFirst: public G1CollectorPolicy {
  1260   CollectionSetChooser* _collectionSetChooser;
  1261   // If the estimated is less then desirable, resize if possible.
  1262   void expand_if_possible(size_t numRegions);
  1264   virtual void choose_collection_set(double target_pause_time_ms);
  1265   virtual void record_collection_pause_start(double start_time_sec,
  1266                                              size_t start_used);
  1267   virtual void record_concurrent_mark_cleanup_end(size_t freed_bytes,
  1268                                                   size_t max_live_bytes);
  1269   virtual void record_full_collection_end();
  1271 public:
  1272   G1CollectorPolicy_BestRegionsFirst() {
  1273     _collectionSetChooser = new CollectionSetChooser();
  1275   void record_collection_pause_end();
  1276   // This is not needed any more, after the CSet choosing code was
  1277   // changed to use the pause prediction work. But let's leave the
  1278   // hook in just in case.
  1279   void note_change_in_marked_bytes(HeapRegion* r) { }
  1280 #ifndef PRODUCT
  1281   bool assertMarkedBytesDataOK();
  1282 #endif
  1283 };
  1285 // This should move to some place more general...
  1287 // If we have "n" measurements, and we've kept track of their "sum" and the
  1288 // "sum_of_squares" of the measurements, this returns the variance of the
  1289 // sequence.
  1290 inline double variance(int n, double sum_of_squares, double sum) {
  1291   double n_d = (double)n;
  1292   double avg = sum/n_d;
  1293   return (sum_of_squares - 2.0 * avg * sum + n_d * avg * avg) / n_d;
  1296 // Local Variables: ***
  1297 // c-indentation-style: gnu ***
  1298 // End: ***
  1300 #endif // SHARE_VM_GC_IMPLEMENTATION_G1_G1COLLECTORPOLICY_HPP

mercurial