Wed, 21 Dec 2011 22:13:31 +0100
7113021: G1: automatically enable young gen size auto-tuning when -Xms==-Xmx
Summary: Use a percentage of -Xms as min and another percentage of -Xmx as max for the young gen size
Reviewed-by: tonyp, johnc
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)
78 };
80 class Summary: public PauseSummary,
81 public MainBodySummary {
82 public:
83 virtual MainBodySummary* main_body_summary() { return this; }
84 };
86 // There are three command line options related to the young gen size:
87 // NewSize, MaxNewSize and NewRatio (There is also -Xmn, but that is
88 // just a short form for NewSize==MaxNewSize). G1 will use its internal
89 // heuristics to calculate the actual young gen size, so these options
90 // basically only limit the range within which G1 can pick a young gen
91 // size. Also, these are general options taking byte sizes. G1 will
92 // internally work with a number of regions instead. So, some rounding
93 // will occur.
94 //
95 // If nothing related to the the young gen size is set on the command
96 // line we should allow the young gen to be between
97 // G1DefaultMinNewGenPercent and G1DefaultMaxNewGenPercent of the
98 // heap size. This means that every time the heap size changes the
99 // limits for the young gen size will be updated.
100 //
101 // If only -XX:NewSize is set we should use the specified value as the
102 // minimum size for young gen. Still using G1DefaultMaxNewGenPercent
103 // of the heap as maximum.
104 //
105 // If only -XX:MaxNewSize is set we should use the specified value as the
106 // maximum size for young gen. Still using G1DefaultMinNewGenPercent
107 // of the heap as minimum.
108 //
109 // If -XX:NewSize and -XX:MaxNewSize are both specified we use these values.
110 // No updates when the heap size changes. There is a special case when
111 // NewSize==MaxNewSize. This is interpreted as "fixed" and will use a
112 // different heuristic for calculating the collection set when we do mixed
113 // collection.
114 //
115 // If only -XX:NewRatio is set we should use the specified ratio of the heap
116 // as both min and max. This will be interpreted as "fixed" just like the
117 // NewSize==MaxNewSize case above. But we will update the min and max
118 // everytime the heap size changes.
119 //
120 // NewSize and MaxNewSize override NewRatio. So, NewRatio is ignored if it is
121 // combined with either NewSize or MaxNewSize. (A warning message is printed.)
122 class G1YoungGenSizer : public CHeapObj {
123 private:
124 enum SizerKind {
125 SizerDefaults,
126 SizerNewSizeOnly,
127 SizerMaxNewSizeOnly,
128 SizerMaxAndNewSize,
129 SizerNewRatio
130 };
131 SizerKind _sizer_kind;
132 size_t _min_desired_young_length;
133 size_t _max_desired_young_length;
134 bool _adaptive_size;
135 size_t calculate_default_min_length(size_t new_number_of_heap_regions);
136 size_t calculate_default_max_length(size_t new_number_of_heap_regions);
138 public:
139 G1YoungGenSizer();
140 void heap_size_changed(size_t new_number_of_heap_regions);
141 size_t min_desired_young_length() {
142 return _min_desired_young_length;
143 }
144 size_t max_desired_young_length() {
145 return _max_desired_young_length;
146 }
147 bool adaptive_young_list_length() {
148 return _adaptive_size;
149 }
150 };
152 class G1CollectorPolicy: public CollectorPolicy {
153 private:
154 // either equal to the number of parallel threads, if ParallelGCThreads
155 // has been set, or 1 otherwise
156 int _parallel_gc_threads;
158 // The number of GC threads currently active.
159 uintx _no_of_gc_threads;
161 enum SomePrivateConstants {
162 NumPrevPausesForHeuristics = 10
163 };
165 G1MMUTracker* _mmu_tracker;
167 void initialize_flags();
169 void initialize_all() {
170 initialize_flags();
171 initialize_size_info();
172 initialize_perm_generation(PermGen::MarkSweepCompact);
173 }
175 CollectionSetChooser* _collectionSetChooser;
177 double _cur_collection_start_sec;
178 size_t _cur_collection_pause_used_at_start_bytes;
179 size_t _cur_collection_pause_used_regions_at_start;
180 size_t _prev_collection_pause_used_at_end_bytes;
181 double _cur_collection_par_time_ms;
182 double _cur_satb_drain_time_ms;
183 double _cur_clear_ct_time_ms;
184 double _cur_ref_proc_time_ms;
185 double _cur_ref_enq_time_ms;
187 #ifndef PRODUCT
188 // Card Table Count Cache stats
189 double _min_clear_cc_time_ms; // min
190 double _max_clear_cc_time_ms; // max
191 double _cur_clear_cc_time_ms; // clearing time during current pause
192 double _cum_clear_cc_time_ms; // cummulative clearing time
193 jlong _num_cc_clears; // number of times the card count cache has been cleared
194 #endif
196 // These exclude marking times.
197 TruncatedSeq* _recent_gc_times_ms;
199 TruncatedSeq* _concurrent_mark_remark_times_ms;
200 TruncatedSeq* _concurrent_mark_cleanup_times_ms;
202 Summary* _summary;
204 NumberSeq* _all_pause_times_ms;
205 NumberSeq* _all_full_gc_times_ms;
206 double _stop_world_start;
207 NumberSeq* _all_stop_world_times_ms;
208 NumberSeq* _all_yield_times_ms;
210 int _aux_num;
211 NumberSeq* _all_aux_times_ms;
212 double* _cur_aux_start_times_ms;
213 double* _cur_aux_times_ms;
214 bool* _cur_aux_times_set;
216 double* _par_last_gc_worker_start_times_ms;
217 double* _par_last_ext_root_scan_times_ms;
218 double* _par_last_mark_stack_scan_times_ms;
219 double* _par_last_update_rs_times_ms;
220 double* _par_last_update_rs_processed_buffers;
221 double* _par_last_scan_rs_times_ms;
222 double* _par_last_obj_copy_times_ms;
223 double* _par_last_termination_times_ms;
224 double* _par_last_termination_attempts;
225 double* _par_last_gc_worker_end_times_ms;
226 double* _par_last_gc_worker_times_ms;
228 // Each workers 'other' time i.e. the elapsed time of the parallel
229 // phase of the pause minus the sum of the individual sub-phase
230 // times for a given worker thread.
231 double* _par_last_gc_worker_other_times_ms;
233 // indicates whether we are in young or mixed GC mode
234 bool _gcs_are_young;
236 size_t _young_list_target_length;
237 size_t _young_list_fixed_length;
238 size_t _prev_eden_capacity; // used for logging
240 // The max number of regions we can extend the eden by while the GC
241 // locker is active. This should be >= _young_list_target_length;
242 size_t _young_list_max_length;
244 bool _last_gc_was_young;
246 unsigned _young_pause_num;
247 unsigned _mixed_pause_num;
249 bool _during_marking;
250 bool _in_marking_window;
251 bool _in_marking_window_im;
253 SurvRateGroup* _short_lived_surv_rate_group;
254 SurvRateGroup* _survivor_surv_rate_group;
255 // add here any more surv rate groups
257 double _gc_overhead_perc;
259 double _reserve_factor;
260 size_t _reserve_regions;
262 bool during_marking() {
263 return _during_marking;
264 }
266 private:
267 enum PredictionConstants {
268 TruncatedSeqLength = 10
269 };
271 TruncatedSeq* _alloc_rate_ms_seq;
272 double _prev_collection_pause_end_ms;
274 TruncatedSeq* _pending_card_diff_seq;
275 TruncatedSeq* _rs_length_diff_seq;
276 TruncatedSeq* _cost_per_card_ms_seq;
277 TruncatedSeq* _young_cards_per_entry_ratio_seq;
278 TruncatedSeq* _mixed_cards_per_entry_ratio_seq;
279 TruncatedSeq* _cost_per_entry_ms_seq;
280 TruncatedSeq* _mixed_cost_per_entry_ms_seq;
281 TruncatedSeq* _cost_per_byte_ms_seq;
282 TruncatedSeq* _constant_other_time_ms_seq;
283 TruncatedSeq* _young_other_cost_per_region_ms_seq;
284 TruncatedSeq* _non_young_other_cost_per_region_ms_seq;
286 TruncatedSeq* _pending_cards_seq;
287 TruncatedSeq* _rs_lengths_seq;
289 TruncatedSeq* _cost_per_byte_ms_during_cm_seq;
291 TruncatedSeq* _young_gc_eff_seq;
293 G1YoungGenSizer* _young_gen_sizer;
295 size_t _eden_cset_region_length;
296 size_t _survivor_cset_region_length;
297 size_t _old_cset_region_length;
299 void init_cset_region_lengths(size_t eden_cset_region_length,
300 size_t survivor_cset_region_length);
302 size_t eden_cset_region_length() { return _eden_cset_region_length; }
303 size_t survivor_cset_region_length() { return _survivor_cset_region_length; }
304 size_t old_cset_region_length() { return _old_cset_region_length; }
306 size_t _free_regions_at_end_of_collection;
308 size_t _recorded_rs_lengths;
309 size_t _max_rs_lengths;
311 double _recorded_young_free_cset_time_ms;
312 double _recorded_non_young_free_cset_time_ms;
314 double _sigma;
315 double _expensive_region_limit_ms;
317 size_t _rs_lengths_prediction;
319 size_t _known_garbage_bytes;
320 double _known_garbage_ratio;
322 double sigma() {
323 return _sigma;
324 }
326 // A function that prevents us putting too much stock in small sample
327 // sets. Returns a number between 2.0 and 1.0, depending on the number
328 // of samples. 5 or more samples yields one; fewer scales linearly from
329 // 2.0 at 1 sample to 1.0 at 5.
330 double confidence_factor(int samples) {
331 if (samples > 4) return 1.0;
332 else return 1.0 + sigma() * ((double)(5 - samples))/2.0;
333 }
335 double get_new_neg_prediction(TruncatedSeq* seq) {
336 return seq->davg() - sigma() * seq->dsd();
337 }
339 #ifndef PRODUCT
340 bool verify_young_ages(HeapRegion* head, SurvRateGroup *surv_rate_group);
341 #endif // PRODUCT
343 void adjust_concurrent_refinement(double update_rs_time,
344 double update_rs_processed_buffers,
345 double goal_ms);
347 uintx no_of_gc_threads() { return _no_of_gc_threads; }
348 void set_no_of_gc_threads(uintx v) { _no_of_gc_threads = v; }
350 double _pause_time_target_ms;
351 double _recorded_young_cset_choice_time_ms;
352 double _recorded_non_young_cset_choice_time_ms;
353 size_t _pending_cards;
354 size_t _max_pending_cards;
356 public:
357 // Accessors
359 void set_region_eden(HeapRegion* hr, int young_index_in_cset) {
360 hr->set_young();
361 hr->install_surv_rate_group(_short_lived_surv_rate_group);
362 hr->set_young_index_in_cset(young_index_in_cset);
363 }
365 void set_region_survivor(HeapRegion* hr, int young_index_in_cset) {
366 assert(hr->is_young() && hr->is_survivor(), "pre-condition");
367 hr->install_surv_rate_group(_survivor_surv_rate_group);
368 hr->set_young_index_in_cset(young_index_in_cset);
369 }
371 #ifndef PRODUCT
372 bool verify_young_ages();
373 #endif // PRODUCT
375 double get_new_prediction(TruncatedSeq* seq) {
376 return MAX2(seq->davg() + sigma() * seq->dsd(),
377 seq->davg() * confidence_factor(seq->num()));
378 }
380 void record_max_rs_lengths(size_t rs_lengths) {
381 _max_rs_lengths = rs_lengths;
382 }
384 size_t predict_pending_card_diff() {
385 double prediction = get_new_neg_prediction(_pending_card_diff_seq);
386 if (prediction < 0.00001) {
387 return 0;
388 } else {
389 return (size_t) prediction;
390 }
391 }
393 size_t predict_pending_cards() {
394 size_t max_pending_card_num = _g1->max_pending_card_num();
395 size_t diff = predict_pending_card_diff();
396 size_t prediction;
397 if (diff > max_pending_card_num) {
398 prediction = max_pending_card_num;
399 } else {
400 prediction = max_pending_card_num - diff;
401 }
403 return prediction;
404 }
406 size_t predict_rs_length_diff() {
407 return (size_t) get_new_prediction(_rs_length_diff_seq);
408 }
410 double predict_alloc_rate_ms() {
411 return get_new_prediction(_alloc_rate_ms_seq);
412 }
414 double predict_cost_per_card_ms() {
415 return get_new_prediction(_cost_per_card_ms_seq);
416 }
418 double predict_rs_update_time_ms(size_t pending_cards) {
419 return (double) pending_cards * predict_cost_per_card_ms();
420 }
422 double predict_young_cards_per_entry_ratio() {
423 return get_new_prediction(_young_cards_per_entry_ratio_seq);
424 }
426 double predict_mixed_cards_per_entry_ratio() {
427 if (_mixed_cards_per_entry_ratio_seq->num() < 2) {
428 return predict_young_cards_per_entry_ratio();
429 } else {
430 return get_new_prediction(_mixed_cards_per_entry_ratio_seq);
431 }
432 }
434 size_t predict_young_card_num(size_t rs_length) {
435 return (size_t) ((double) rs_length *
436 predict_young_cards_per_entry_ratio());
437 }
439 size_t predict_non_young_card_num(size_t rs_length) {
440 return (size_t) ((double) rs_length *
441 predict_mixed_cards_per_entry_ratio());
442 }
444 double predict_rs_scan_time_ms(size_t card_num) {
445 if (gcs_are_young()) {
446 return (double) card_num * get_new_prediction(_cost_per_entry_ms_seq);
447 } else {
448 return predict_mixed_rs_scan_time_ms(card_num);
449 }
450 }
452 double predict_mixed_rs_scan_time_ms(size_t card_num) {
453 if (_mixed_cost_per_entry_ms_seq->num() < 3) {
454 return (double) card_num * get_new_prediction(_cost_per_entry_ms_seq);
455 } else {
456 return (double) (card_num *
457 get_new_prediction(_mixed_cost_per_entry_ms_seq));
458 }
459 }
461 double predict_object_copy_time_ms_during_cm(size_t bytes_to_copy) {
462 if (_cost_per_byte_ms_during_cm_seq->num() < 3) {
463 return (1.1 * (double) bytes_to_copy) *
464 get_new_prediction(_cost_per_byte_ms_seq);
465 } else {
466 return (double) bytes_to_copy *
467 get_new_prediction(_cost_per_byte_ms_during_cm_seq);
468 }
469 }
471 double predict_object_copy_time_ms(size_t bytes_to_copy) {
472 if (_in_marking_window && !_in_marking_window_im) {
473 return predict_object_copy_time_ms_during_cm(bytes_to_copy);
474 } else {
475 return (double) bytes_to_copy *
476 get_new_prediction(_cost_per_byte_ms_seq);
477 }
478 }
480 double predict_constant_other_time_ms() {
481 return get_new_prediction(_constant_other_time_ms_seq);
482 }
484 double predict_young_other_time_ms(size_t young_num) {
485 return (double) young_num *
486 get_new_prediction(_young_other_cost_per_region_ms_seq);
487 }
489 double predict_non_young_other_time_ms(size_t non_young_num) {
490 return (double) non_young_num *
491 get_new_prediction(_non_young_other_cost_per_region_ms_seq);
492 }
494 void check_if_region_is_too_expensive(double predicted_time_ms);
496 double predict_young_collection_elapsed_time_ms(size_t adjustment);
497 double predict_base_elapsed_time_ms(size_t pending_cards);
498 double predict_base_elapsed_time_ms(size_t pending_cards,
499 size_t scanned_cards);
500 size_t predict_bytes_to_copy(HeapRegion* hr);
501 double predict_region_elapsed_time_ms(HeapRegion* hr, bool young);
503 void set_recorded_rs_lengths(size_t rs_lengths);
505 size_t cset_region_length() { return young_cset_region_length() +
506 old_cset_region_length(); }
507 size_t young_cset_region_length() { return eden_cset_region_length() +
508 survivor_cset_region_length(); }
510 void record_young_free_cset_time_ms(double time_ms) {
511 _recorded_young_free_cset_time_ms = time_ms;
512 }
514 void record_non_young_free_cset_time_ms(double time_ms) {
515 _recorded_non_young_free_cset_time_ms = time_ms;
516 }
518 double predict_young_gc_eff() {
519 return get_new_neg_prediction(_young_gc_eff_seq);
520 }
522 double predict_survivor_regions_evac_time();
524 void cset_regions_freed() {
525 bool propagate = _last_gc_was_young && !_in_marking_window;
526 _short_lived_surv_rate_group->all_surviving_words_recorded(propagate);
527 _survivor_surv_rate_group->all_surviving_words_recorded(propagate);
528 // also call it on any more surv rate groups
529 }
531 void set_known_garbage_bytes(size_t known_garbage_bytes) {
532 _known_garbage_bytes = known_garbage_bytes;
533 size_t heap_bytes = _g1->capacity();
534 _known_garbage_ratio = (double) _known_garbage_bytes / (double) heap_bytes;
535 }
537 void decrease_known_garbage_bytes(size_t known_garbage_bytes) {
538 guarantee( _known_garbage_bytes >= known_garbage_bytes, "invariant" );
540 _known_garbage_bytes -= known_garbage_bytes;
541 size_t heap_bytes = _g1->capacity();
542 _known_garbage_ratio = (double) _known_garbage_bytes / (double) heap_bytes;
543 }
545 G1MMUTracker* mmu_tracker() {
546 return _mmu_tracker;
547 }
549 double max_pause_time_ms() {
550 return _mmu_tracker->max_gc_time() * 1000.0;
551 }
553 double predict_remark_time_ms() {
554 return get_new_prediction(_concurrent_mark_remark_times_ms);
555 }
557 double predict_cleanup_time_ms() {
558 return get_new_prediction(_concurrent_mark_cleanup_times_ms);
559 }
561 // Returns an estimate of the survival rate of the region at yg-age
562 // "yg_age".
563 double predict_yg_surv_rate(int age, SurvRateGroup* surv_rate_group) {
564 TruncatedSeq* seq = surv_rate_group->get_seq(age);
565 if (seq->num() == 0)
566 gclog_or_tty->print("BARF! age is %d", age);
567 guarantee( seq->num() > 0, "invariant" );
568 double pred = get_new_prediction(seq);
569 if (pred > 1.0)
570 pred = 1.0;
571 return pred;
572 }
574 double predict_yg_surv_rate(int age) {
575 return predict_yg_surv_rate(age, _short_lived_surv_rate_group);
576 }
578 double accum_yg_surv_rate_pred(int age) {
579 return _short_lived_surv_rate_group->accum_surv_rate_pred(age);
580 }
582 private:
583 void print_stats(int level, const char* str, double value);
584 void print_stats(int level, const char* str, int value);
586 void print_par_stats(int level, const char* str, double* data);
587 void print_par_sizes(int level, const char* str, double* data);
589 void check_other_times(int level,
590 NumberSeq* other_times_ms,
591 NumberSeq* calc_other_times_ms) const;
593 void print_summary (PauseSummary* stats) const;
595 void print_summary (int level, const char* str, NumberSeq* seq) const;
596 void print_summary_sd (int level, const char* str, NumberSeq* seq) const;
598 double avg_value (double* data);
599 double max_value (double* data);
600 double sum_of_values (double* data);
601 double max_sum (double* data1, double* data2);
603 double _last_pause_time_ms;
605 size_t _bytes_in_collection_set_before_gc;
606 size_t _bytes_copied_during_gc;
608 // Used to count used bytes in CS.
609 friend class CountCSClosure;
611 // Statistics kept per GC stoppage, pause or full.
612 TruncatedSeq* _recent_prev_end_times_for_all_gcs_sec;
614 // Add a new GC of the given duration and end time to the record.
615 void update_recent_gc_times(double end_time_sec, double elapsed_ms);
617 // The head of the list (via "next_in_collection_set()") representing the
618 // current collection set. Set from the incrementally built collection
619 // set at the start of the pause.
620 HeapRegion* _collection_set;
622 // The number of bytes in the collection set before the pause. Set from
623 // the incrementally built collection set at the start of an evacuation
624 // pause.
625 size_t _collection_set_bytes_used_before;
627 // The associated information that is maintained while the incremental
628 // collection set is being built with young regions. Used to populate
629 // the recorded info for the evacuation pause.
631 enum CSetBuildType {
632 Active, // We are actively building the collection set
633 Inactive // We are not actively building the collection set
634 };
636 CSetBuildType _inc_cset_build_state;
638 // The head of the incrementally built collection set.
639 HeapRegion* _inc_cset_head;
641 // The tail of the incrementally built collection set.
642 HeapRegion* _inc_cset_tail;
644 // The number of bytes in the incrementally built collection set.
645 // Used to set _collection_set_bytes_used_before at the start of
646 // an evacuation pause.
647 size_t _inc_cset_bytes_used_before;
649 // Used to record the highest end of heap region in collection set
650 HeapWord* _inc_cset_max_finger;
652 // The RSet lengths recorded for regions in the CSet. It is updated
653 // by the thread that adds a new region to the CSet. We assume that
654 // only one thread can be allocating a new CSet region (currently,
655 // it does so after taking the Heap_lock) hence no need to
656 // synchronize updates to this field.
657 size_t _inc_cset_recorded_rs_lengths;
659 // A concurrent refinement thread periodcially samples the young
660 // region RSets and needs to update _inc_cset_recorded_rs_lengths as
661 // the RSets grow. Instead of having to syncronize updates to that
662 // field we accumulate them in this field and add it to
663 // _inc_cset_recorded_rs_lengths_diffs at the start of a GC.
664 ssize_t _inc_cset_recorded_rs_lengths_diffs;
666 // The predicted elapsed time it will take to collect the regions in
667 // the CSet. This is updated by the thread that adds a new region to
668 // the CSet. See the comment for _inc_cset_recorded_rs_lengths about
669 // MT-safety assumptions.
670 double _inc_cset_predicted_elapsed_time_ms;
672 // See the comment for _inc_cset_recorded_rs_lengths_diffs.
673 double _inc_cset_predicted_elapsed_time_ms_diffs;
675 // Stash a pointer to the g1 heap.
676 G1CollectedHeap* _g1;
678 // The ratio of gc time to elapsed time, computed over recent pauses.
679 double _recent_avg_pause_time_ratio;
681 double recent_avg_pause_time_ratio() {
682 return _recent_avg_pause_time_ratio;
683 }
685 // At the end of a pause we check the heap occupancy and we decide
686 // whether we will start a marking cycle during the next pause. If
687 // we decide that we want to do that, we will set this parameter to
688 // true. So, this parameter will stay true between the end of a
689 // pause and the beginning of a subsequent pause (not necessarily
690 // the next one, see the comments on the next field) when we decide
691 // that we will indeed start a marking cycle and do the initial-mark
692 // work.
693 volatile bool _initiate_conc_mark_if_possible;
695 // If initiate_conc_mark_if_possible() is set at the beginning of a
696 // pause, it is a suggestion that the pause should start a marking
697 // cycle by doing the initial-mark work. However, it is possible
698 // that the concurrent marking thread is still finishing up the
699 // previous marking cycle (e.g., clearing the next marking
700 // bitmap). If that is the case we cannot start a new cycle and
701 // we'll have to wait for the concurrent marking thread to finish
702 // what it is doing. In this case we will postpone the marking cycle
703 // initiation decision for the next pause. When we eventually decide
704 // to start a cycle, we will set _during_initial_mark_pause which
705 // will stay true until the end of the initial-mark pause and it's
706 // the condition that indicates that a pause is doing the
707 // initial-mark work.
708 volatile bool _during_initial_mark_pause;
710 bool _should_revert_to_young_gcs;
711 bool _last_young_gc;
713 // This set of variables tracks the collector efficiency, in order to
714 // determine whether we should initiate a new marking.
715 double _cur_mark_stop_world_time_ms;
716 double _mark_remark_start_sec;
717 double _mark_cleanup_start_sec;
718 double _mark_closure_time_ms;
720 // Update the young list target length either by setting it to the
721 // desired fixed value or by calculating it using G1's pause
722 // prediction model. If no rs_lengths parameter is passed, predict
723 // the RS lengths using the prediction model, otherwise use the
724 // given rs_lengths as the prediction.
725 void update_young_list_target_length(size_t rs_lengths = (size_t) -1);
727 // Calculate and return the minimum desired young list target
728 // length. This is the minimum desired young list length according
729 // to the user's inputs.
730 size_t calculate_young_list_desired_min_length(size_t base_min_length);
732 // Calculate and return the maximum desired young list target
733 // length. This is the maximum desired young list length according
734 // to the user's inputs.
735 size_t calculate_young_list_desired_max_length();
737 // Calculate and return the maximum young list target length that
738 // can fit into the pause time goal. The parameters are: rs_lengths
739 // represent the prediction of how large the young RSet lengths will
740 // be, base_min_length is the alreay existing number of regions in
741 // the young list, min_length and max_length are the desired min and
742 // max young list length according to the user's inputs.
743 size_t calculate_young_list_target_length(size_t rs_lengths,
744 size_t base_min_length,
745 size_t desired_min_length,
746 size_t desired_max_length);
748 // Check whether a given young length (young_length) fits into the
749 // given target pause time and whether the prediction for the amount
750 // of objects to be copied for the given length will fit into the
751 // given free space (expressed by base_free_regions). It is used by
752 // calculate_young_list_target_length().
753 bool predict_will_fit(size_t young_length, double base_time_ms,
754 size_t base_free_regions, double target_pause_time_ms);
756 // Count the number of bytes used in the CS.
757 void count_CS_bytes_used();
759 public:
761 G1CollectorPolicy();
763 virtual G1CollectorPolicy* as_g1_policy() { return this; }
765 virtual CollectorPolicy::Name kind() {
766 return CollectorPolicy::G1CollectorPolicyKind;
767 }
769 // Check the current value of the young list RSet lengths and
770 // compare it against the last prediction. If the current value is
771 // higher, recalculate the young list target length prediction.
772 void revise_young_list_target_length_if_necessary();
774 size_t bytes_in_collection_set() {
775 return _bytes_in_collection_set_before_gc;
776 }
778 unsigned calc_gc_alloc_time_stamp() {
779 return _all_pause_times_ms->num() + 1;
780 }
782 // This should be called after the heap is resized.
783 void record_new_heap_size(size_t new_number_of_regions);
785 void init();
787 // Create jstat counters for the policy.
788 virtual void initialize_gc_policy_counters();
790 virtual HeapWord* mem_allocate_work(size_t size,
791 bool is_tlab,
792 bool* gc_overhead_limit_was_exceeded);
794 // This method controls how a collector handles one or more
795 // of its generations being fully allocated.
796 virtual HeapWord* satisfy_failed_allocation(size_t size,
797 bool is_tlab);
799 BarrierSet::Name barrier_set_name() { return BarrierSet::G1SATBCTLogging; }
801 GenRemSet::Name rem_set_name() { return GenRemSet::CardTable; }
803 // Update the heuristic info to record a collection pause of the given
804 // start time, where the given number of bytes were used at the start.
805 // This may involve changing the desired size of a collection set.
807 void record_stop_world_start();
809 void record_collection_pause_start(double start_time_sec, size_t start_used);
811 // Must currently be called while the world is stopped.
812 void record_concurrent_mark_init_end(double
813 mark_init_elapsed_time_ms);
815 void record_mark_closure_time(double mark_closure_time_ms) {
816 _mark_closure_time_ms = mark_closure_time_ms;
817 }
819 void record_concurrent_mark_remark_start();
820 void record_concurrent_mark_remark_end();
822 void record_concurrent_mark_cleanup_start();
823 void record_concurrent_mark_cleanup_end(int no_of_gc_threads);
824 void record_concurrent_mark_cleanup_completed();
826 void record_concurrent_pause();
827 void record_concurrent_pause_end();
829 void record_collection_pause_end(int no_of_gc_threads);
830 void print_heap_transition();
832 // Record the fact that a full collection occurred.
833 void record_full_collection_start();
834 void record_full_collection_end();
836 void record_gc_worker_start_time(int worker_i, double ms) {
837 _par_last_gc_worker_start_times_ms[worker_i] = ms;
838 }
840 void record_ext_root_scan_time(int worker_i, double ms) {
841 _par_last_ext_root_scan_times_ms[worker_i] = ms;
842 }
844 void record_mark_stack_scan_time(int worker_i, double ms) {
845 _par_last_mark_stack_scan_times_ms[worker_i] = ms;
846 }
848 void record_satb_drain_time(double ms) {
849 assert(_g1->mark_in_progress(), "shouldn't be here otherwise");
850 _cur_satb_drain_time_ms = ms;
851 }
853 void record_update_rs_time(int thread, double ms) {
854 _par_last_update_rs_times_ms[thread] = ms;
855 }
857 void record_update_rs_processed_buffers (int thread,
858 double processed_buffers) {
859 _par_last_update_rs_processed_buffers[thread] = processed_buffers;
860 }
862 void record_scan_rs_time(int thread, double ms) {
863 _par_last_scan_rs_times_ms[thread] = ms;
864 }
866 void reset_obj_copy_time(int thread) {
867 _par_last_obj_copy_times_ms[thread] = 0.0;
868 }
870 void reset_obj_copy_time() {
871 reset_obj_copy_time(0);
872 }
874 void record_obj_copy_time(int thread, double ms) {
875 _par_last_obj_copy_times_ms[thread] += ms;
876 }
878 void record_termination(int thread, double ms, size_t attempts) {
879 _par_last_termination_times_ms[thread] = ms;
880 _par_last_termination_attempts[thread] = (double) attempts;
881 }
883 void record_gc_worker_end_time(int worker_i, double ms) {
884 _par_last_gc_worker_end_times_ms[worker_i] = ms;
885 }
887 void record_pause_time_ms(double ms) {
888 _last_pause_time_ms = ms;
889 }
891 void record_clear_ct_time(double ms) {
892 _cur_clear_ct_time_ms = ms;
893 }
895 void record_par_time(double ms) {
896 _cur_collection_par_time_ms = ms;
897 }
899 void record_aux_start_time(int i) {
900 guarantee(i < _aux_num, "should be within range");
901 _cur_aux_start_times_ms[i] = os::elapsedTime() * 1000.0;
902 }
904 void record_aux_end_time(int i) {
905 guarantee(i < _aux_num, "should be within range");
906 double ms = os::elapsedTime() * 1000.0 - _cur_aux_start_times_ms[i];
907 _cur_aux_times_set[i] = true;
908 _cur_aux_times_ms[i] += ms;
909 }
911 void record_ref_proc_time(double ms) {
912 _cur_ref_proc_time_ms = ms;
913 }
915 void record_ref_enq_time(double ms) {
916 _cur_ref_enq_time_ms = ms;
917 }
919 #ifndef PRODUCT
920 void record_cc_clear_time(double ms) {
921 if (_min_clear_cc_time_ms < 0.0 || ms <= _min_clear_cc_time_ms)
922 _min_clear_cc_time_ms = ms;
923 if (_max_clear_cc_time_ms < 0.0 || ms >= _max_clear_cc_time_ms)
924 _max_clear_cc_time_ms = ms;
925 _cur_clear_cc_time_ms = ms;
926 _cum_clear_cc_time_ms += ms;
927 _num_cc_clears++;
928 }
929 #endif
931 // Record how much space we copied during a GC. This is typically
932 // called when a GC alloc region is being retired.
933 void record_bytes_copied_during_gc(size_t bytes) {
934 _bytes_copied_during_gc += bytes;
935 }
937 // The amount of space we copied during a GC.
938 size_t bytes_copied_during_gc() {
939 return _bytes_copied_during_gc;
940 }
942 // Choose a new collection set. Marks the chosen regions as being
943 // "in_collection_set", and links them together. The head and number of
944 // the collection set are available via access methods.
945 void choose_collection_set(double target_pause_time_ms);
947 // The head of the list (via "next_in_collection_set()") representing the
948 // current collection set.
949 HeapRegion* collection_set() { return _collection_set; }
951 void clear_collection_set() { _collection_set = NULL; }
953 // Add old region "hr" to the CSet.
954 void add_old_region_to_cset(HeapRegion* hr);
956 // Incremental CSet Support
958 // The head of the incrementally built collection set.
959 HeapRegion* inc_cset_head() { return _inc_cset_head; }
961 // The tail of the incrementally built collection set.
962 HeapRegion* inc_set_tail() { return _inc_cset_tail; }
964 // Initialize incremental collection set info.
965 void start_incremental_cset_building();
967 // Perform any final calculations on the incremental CSet fields
968 // before we can use them.
969 void finalize_incremental_cset_building();
971 void clear_incremental_cset() {
972 _inc_cset_head = NULL;
973 _inc_cset_tail = NULL;
974 }
976 // Stop adding regions to the incremental collection set
977 void stop_incremental_cset_building() { _inc_cset_build_state = Inactive; }
979 // Add information about hr to the aggregated information for the
980 // incrementally built collection set.
981 void add_to_incremental_cset_info(HeapRegion* hr, size_t rs_length);
983 // Update information about hr in the aggregated information for
984 // the incrementally built collection set.
985 void update_incremental_cset_info(HeapRegion* hr, size_t new_rs_length);
987 private:
988 // Update the incremental cset information when adding a region
989 // (should not be called directly).
990 void add_region_to_incremental_cset_common(HeapRegion* hr);
992 public:
993 // Add hr to the LHS of the incremental collection set.
994 void add_region_to_incremental_cset_lhs(HeapRegion* hr);
996 // Add hr to the RHS of the incremental collection set.
997 void add_region_to_incremental_cset_rhs(HeapRegion* hr);
999 #ifndef PRODUCT
1000 void print_collection_set(HeapRegion* list_head, outputStream* st);
1001 #endif // !PRODUCT
1003 bool initiate_conc_mark_if_possible() { return _initiate_conc_mark_if_possible; }
1004 void set_initiate_conc_mark_if_possible() { _initiate_conc_mark_if_possible = true; }
1005 void clear_initiate_conc_mark_if_possible() { _initiate_conc_mark_if_possible = false; }
1007 bool during_initial_mark_pause() { return _during_initial_mark_pause; }
1008 void set_during_initial_mark_pause() { _during_initial_mark_pause = true; }
1009 void clear_during_initial_mark_pause(){ _during_initial_mark_pause = false; }
1011 // This sets the initiate_conc_mark_if_possible() flag to start a
1012 // new cycle, as long as we are not already in one. It's best if it
1013 // is called during a safepoint when the test whether a cycle is in
1014 // progress or not is stable.
1015 bool force_initial_mark_if_outside_cycle(GCCause::Cause gc_cause);
1017 // This is called at the very beginning of an evacuation pause (it
1018 // has to be the first thing that the pause does). If
1019 // initiate_conc_mark_if_possible() is true, and the concurrent
1020 // marking thread has completed its work during the previous cycle,
1021 // it will set during_initial_mark_pause() to so that the pause does
1022 // the initial-mark work and start a marking cycle.
1023 void decide_on_conc_mark_initiation();
1025 // If an expansion would be appropriate, because recent GC overhead had
1026 // exceeded the desired limit, return an amount to expand by.
1027 size_t expansion_amount();
1029 #ifndef PRODUCT
1030 // Check any appropriate marked bytes info, asserting false if
1031 // something's wrong, else returning "true".
1032 bool assertMarkedBytesDataOK();
1033 #endif
1035 // Print tracing information.
1036 void print_tracing_info() const;
1038 // Print stats on young survival ratio
1039 void print_yg_surv_rate_info() const;
1041 void finished_recalculating_age_indexes(bool is_survivors) {
1042 if (is_survivors) {
1043 _survivor_surv_rate_group->finished_recalculating_age_indexes();
1044 } else {
1045 _short_lived_surv_rate_group->finished_recalculating_age_indexes();
1046 }
1047 // do that for any other surv rate groups
1048 }
1050 bool is_young_list_full() {
1051 size_t young_list_length = _g1->young_list()->length();
1052 size_t young_list_target_length = _young_list_target_length;
1053 return young_list_length >= young_list_target_length;
1054 }
1056 bool can_expand_young_list() {
1057 size_t young_list_length = _g1->young_list()->length();
1058 size_t young_list_max_length = _young_list_max_length;
1059 return young_list_length < young_list_max_length;
1060 }
1062 size_t young_list_max_length() {
1063 return _young_list_max_length;
1064 }
1066 bool gcs_are_young() {
1067 return _gcs_are_young;
1068 }
1069 void set_gcs_are_young(bool gcs_are_young) {
1070 _gcs_are_young = gcs_are_young;
1071 }
1073 bool adaptive_young_list_length() {
1074 return _young_gen_sizer->adaptive_young_list_length();
1075 }
1077 inline double get_gc_eff_factor() {
1078 double ratio = _known_garbage_ratio;
1080 double square = ratio * ratio;
1081 // square = square * square;
1082 double ret = square * 9.0 + 1.0;
1083 #if 0
1084 gclog_or_tty->print_cr("ratio = %1.2lf, ret = %1.2lf", ratio, ret);
1085 #endif // 0
1086 guarantee(0.0 <= ret && ret < 10.0, "invariant!");
1087 return ret;
1088 }
1090 private:
1091 //
1092 // Survivor regions policy.
1093 //
1095 // Current tenuring threshold, set to 0 if the collector reaches the
1096 // maximum amount of suvivors regions.
1097 int _tenuring_threshold;
1099 // The limit on the number of regions allocated for survivors.
1100 size_t _max_survivor_regions;
1102 // For reporting purposes.
1103 size_t _eden_bytes_before_gc;
1104 size_t _survivor_bytes_before_gc;
1105 size_t _capacity_before_gc;
1107 // The amount of survor regions after a collection.
1108 size_t _recorded_survivor_regions;
1109 // List of survivor regions.
1110 HeapRegion* _recorded_survivor_head;
1111 HeapRegion* _recorded_survivor_tail;
1113 ageTable _survivors_age_table;
1115 public:
1117 inline GCAllocPurpose
1118 evacuation_destination(HeapRegion* src_region, int age, size_t word_sz) {
1119 if (age < _tenuring_threshold && src_region->is_young()) {
1120 return GCAllocForSurvived;
1121 } else {
1122 return GCAllocForTenured;
1123 }
1124 }
1126 inline bool track_object_age(GCAllocPurpose purpose) {
1127 return purpose == GCAllocForSurvived;
1128 }
1130 static const size_t REGIONS_UNLIMITED = ~(size_t)0;
1132 size_t max_regions(int purpose);
1134 // The limit on regions for a particular purpose is reached.
1135 void note_alloc_region_limit_reached(int purpose) {
1136 if (purpose == GCAllocForSurvived) {
1137 _tenuring_threshold = 0;
1138 }
1139 }
1141 void note_start_adding_survivor_regions() {
1142 _survivor_surv_rate_group->start_adding_regions();
1143 }
1145 void note_stop_adding_survivor_regions() {
1146 _survivor_surv_rate_group->stop_adding_regions();
1147 }
1149 void record_survivor_regions(size_t regions,
1150 HeapRegion* head,
1151 HeapRegion* tail) {
1152 _recorded_survivor_regions = regions;
1153 _recorded_survivor_head = head;
1154 _recorded_survivor_tail = tail;
1155 }
1157 size_t recorded_survivor_regions() {
1158 return _recorded_survivor_regions;
1159 }
1161 void record_thread_age_table(ageTable* age_table)
1162 {
1163 _survivors_age_table.merge_par(age_table);
1164 }
1166 void update_max_gc_locker_expansion();
1168 // Calculates survivor space parameters.
1169 void update_survivors_policy();
1171 };
1173 // This should move to some place more general...
1175 // If we have "n" measurements, and we've kept track of their "sum" and the
1176 // "sum_of_squares" of the measurements, this returns the variance of the
1177 // sequence.
1178 inline double variance(int n, double sum_of_squares, double sum) {
1179 double n_d = (double)n;
1180 double avg = sum/n_d;
1181 return (sum_of_squares - 2.0 * avg * sum + n_d * avg * avg) / n_d;
1182 }
1184 #endif // SHARE_VM_GC_IMPLEMENTATION_G1_G1COLLECTORPOLICY_HPP