Tue, 24 Aug 2010 17:24:33 -0400
6974966: G1: unnecessary direct-to-old allocations
Summary: This change revamps the slow allocation path of G1. Improvements include the following: a) Allocations directly to old regions are now totally banned. G1 now only allows allocations out of young regions (with the only exception being humongous regions). b) The thread that allocates a new region (which is now guaranteed to be young) does not dirty all its cards. Each thread that successfully allocates out of a young region is now responsible for dirtying the cards that corresponding to the "block" that just got allocated. c) allocate_new_tlab() and mem_allocate() are now implemented differently and TLAB allocations are only done by allocate_new_tlab(). d) If a thread schedules an evacuation pause in order to satisfy an allocation request, it will perform the allocation at the end of the safepoint so that the thread that initiated the GC also gets "first pick" of any space made available by the GC. e) If a thread is unable to allocate a humongous object it will schedule an evacuation pause in case it reclaims enough regions so that the humongous allocation can be satisfied aftewards. f) The G1 policy is more careful to set the young list target length to be the survivor number +1. g) Lots of code tidy up, removal, refactoring to make future changes easier.
Reviewed-by: johnc, ysr
1 /*
2 * Copyright (c) 2001, 2010, 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 double _cur_CH_strong_roots_end_sec;
133 double _cur_CH_strong_roots_dur_ms;
134 double _cur_G1_strong_roots_end_sec;
135 double _cur_G1_strong_roots_dur_ms;
137 // Statistics for recent GC pauses. See below for how indexed.
138 TruncatedSeq* _recent_CH_strong_roots_times_ms;
139 TruncatedSeq* _recent_G1_strong_roots_times_ms;
140 TruncatedSeq* _recent_evac_times_ms;
141 // These exclude marking times.
142 TruncatedSeq* _recent_pause_times_ms;
143 TruncatedSeq* _recent_gc_times_ms;
145 TruncatedSeq* _recent_CS_bytes_used_before;
146 TruncatedSeq* _recent_CS_bytes_surviving;
148 TruncatedSeq* _recent_rs_sizes;
150 TruncatedSeq* _concurrent_mark_init_times_ms;
151 TruncatedSeq* _concurrent_mark_remark_times_ms;
152 TruncatedSeq* _concurrent_mark_cleanup_times_ms;
154 Summary* _summary;
156 NumberSeq* _all_pause_times_ms;
157 NumberSeq* _all_full_gc_times_ms;
158 double _stop_world_start;
159 NumberSeq* _all_stop_world_times_ms;
160 NumberSeq* _all_yield_times_ms;
162 size_t _region_num_young;
163 size_t _region_num_tenured;
164 size_t _prev_region_num_young;
165 size_t _prev_region_num_tenured;
167 NumberSeq* _all_mod_union_times_ms;
169 int _aux_num;
170 NumberSeq* _all_aux_times_ms;
171 double* _cur_aux_start_times_ms;
172 double* _cur_aux_times_ms;
173 bool* _cur_aux_times_set;
175 double* _par_last_gc_worker_start_times_ms;
176 double* _par_last_ext_root_scan_times_ms;
177 double* _par_last_mark_stack_scan_times_ms;
178 double* _par_last_update_rs_times_ms;
179 double* _par_last_update_rs_processed_buffers;
180 double* _par_last_scan_rs_times_ms;
181 double* _par_last_obj_copy_times_ms;
182 double* _par_last_termination_times_ms;
183 double* _par_last_termination_attempts;
184 double* _par_last_gc_worker_end_times_ms;
186 // indicates that we are in young GC mode
187 bool _in_young_gc_mode;
189 // indicates whether we are in full young or partially young GC mode
190 bool _full_young_gcs;
192 // if true, then it tries to dynamically adjust the length of the
193 // young list
194 bool _adaptive_young_list_length;
195 size_t _young_list_min_length;
196 size_t _young_list_target_length;
197 size_t _young_list_fixed_length;
199 size_t _young_cset_length;
200 bool _last_young_gc_full;
202 unsigned _full_young_pause_num;
203 unsigned _partial_young_pause_num;
205 bool _during_marking;
206 bool _in_marking_window;
207 bool _in_marking_window_im;
209 SurvRateGroup* _short_lived_surv_rate_group;
210 SurvRateGroup* _survivor_surv_rate_group;
211 // add here any more surv rate groups
213 double _gc_overhead_perc;
215 bool during_marking() {
216 return _during_marking;
217 }
219 // <NEW PREDICTION>
221 private:
222 enum PredictionConstants {
223 TruncatedSeqLength = 10
224 };
226 TruncatedSeq* _alloc_rate_ms_seq;
227 double _prev_collection_pause_end_ms;
229 TruncatedSeq* _pending_card_diff_seq;
230 TruncatedSeq* _rs_length_diff_seq;
231 TruncatedSeq* _cost_per_card_ms_seq;
232 TruncatedSeq* _fully_young_cards_per_entry_ratio_seq;
233 TruncatedSeq* _partially_young_cards_per_entry_ratio_seq;
234 TruncatedSeq* _cost_per_entry_ms_seq;
235 TruncatedSeq* _partially_young_cost_per_entry_ms_seq;
236 TruncatedSeq* _cost_per_byte_ms_seq;
237 TruncatedSeq* _constant_other_time_ms_seq;
238 TruncatedSeq* _young_other_cost_per_region_ms_seq;
239 TruncatedSeq* _non_young_other_cost_per_region_ms_seq;
241 TruncatedSeq* _pending_cards_seq;
242 TruncatedSeq* _scanned_cards_seq;
243 TruncatedSeq* _rs_lengths_seq;
245 TruncatedSeq* _cost_per_byte_ms_during_cm_seq;
247 TruncatedSeq* _young_gc_eff_seq;
249 TruncatedSeq* _max_conc_overhead_seq;
251 size_t _recorded_young_regions;
252 size_t _recorded_non_young_regions;
253 size_t _recorded_region_num;
255 size_t _free_regions_at_end_of_collection;
257 size_t _recorded_rs_lengths;
258 size_t _max_rs_lengths;
260 size_t _recorded_marked_bytes;
261 size_t _recorded_young_bytes;
263 size_t _predicted_pending_cards;
264 size_t _predicted_cards_scanned;
265 size_t _predicted_rs_lengths;
266 size_t _predicted_bytes_to_copy;
268 double _predicted_survival_ratio;
269 double _predicted_rs_update_time_ms;
270 double _predicted_rs_scan_time_ms;
271 double _predicted_object_copy_time_ms;
272 double _predicted_constant_other_time_ms;
273 double _predicted_young_other_time_ms;
274 double _predicted_non_young_other_time_ms;
275 double _predicted_pause_time_ms;
277 double _vtime_diff_ms;
279 double _recorded_young_free_cset_time_ms;
280 double _recorded_non_young_free_cset_time_ms;
282 double _sigma;
283 double _expensive_region_limit_ms;
285 size_t _rs_lengths_prediction;
287 size_t _known_garbage_bytes;
288 double _known_garbage_ratio;
290 double sigma() {
291 return _sigma;
292 }
294 // A function that prevents us putting too much stock in small sample
295 // sets. Returns a number between 2.0 and 1.0, depending on the number
296 // of samples. 5 or more samples yields one; fewer scales linearly from
297 // 2.0 at 1 sample to 1.0 at 5.
298 double confidence_factor(int samples) {
299 if (samples > 4) return 1.0;
300 else return 1.0 + sigma() * ((double)(5 - samples))/2.0;
301 }
303 double get_new_neg_prediction(TruncatedSeq* seq) {
304 return seq->davg() - sigma() * seq->dsd();
305 }
307 #ifndef PRODUCT
308 bool verify_young_ages(HeapRegion* head, SurvRateGroup *surv_rate_group);
309 #endif // PRODUCT
311 void adjust_concurrent_refinement(double update_rs_time,
312 double update_rs_processed_buffers,
313 double goal_ms);
315 protected:
316 double _pause_time_target_ms;
317 double _recorded_young_cset_choice_time_ms;
318 double _recorded_non_young_cset_choice_time_ms;
319 bool _within_target;
320 size_t _pending_cards;
321 size_t _max_pending_cards;
323 public:
325 void set_region_short_lived(HeapRegion* hr) {
326 hr->install_surv_rate_group(_short_lived_surv_rate_group);
327 }
329 void set_region_survivors(HeapRegion* hr) {
330 hr->install_surv_rate_group(_survivor_surv_rate_group);
331 }
333 #ifndef PRODUCT
334 bool verify_young_ages();
335 #endif // PRODUCT
337 double get_new_prediction(TruncatedSeq* seq) {
338 return MAX2(seq->davg() + sigma() * seq->dsd(),
339 seq->davg() * confidence_factor(seq->num()));
340 }
342 size_t young_cset_length() {
343 return _young_cset_length;
344 }
346 void record_max_rs_lengths(size_t rs_lengths) {
347 _max_rs_lengths = rs_lengths;
348 }
350 size_t predict_pending_card_diff() {
351 double prediction = get_new_neg_prediction(_pending_card_diff_seq);
352 if (prediction < 0.00001)
353 return 0;
354 else
355 return (size_t) prediction;
356 }
358 size_t predict_pending_cards() {
359 size_t max_pending_card_num = _g1->max_pending_card_num();
360 size_t diff = predict_pending_card_diff();
361 size_t prediction;
362 if (diff > max_pending_card_num)
363 prediction = max_pending_card_num;
364 else
365 prediction = max_pending_card_num - diff;
367 return prediction;
368 }
370 size_t predict_rs_length_diff() {
371 return (size_t) get_new_prediction(_rs_length_diff_seq);
372 }
374 double predict_alloc_rate_ms() {
375 return get_new_prediction(_alloc_rate_ms_seq);
376 }
378 double predict_cost_per_card_ms() {
379 return get_new_prediction(_cost_per_card_ms_seq);
380 }
382 double predict_rs_update_time_ms(size_t pending_cards) {
383 return (double) pending_cards * predict_cost_per_card_ms();
384 }
386 double predict_fully_young_cards_per_entry_ratio() {
387 return get_new_prediction(_fully_young_cards_per_entry_ratio_seq);
388 }
390 double predict_partially_young_cards_per_entry_ratio() {
391 if (_partially_young_cards_per_entry_ratio_seq->num() < 2)
392 return predict_fully_young_cards_per_entry_ratio();
393 else
394 return get_new_prediction(_partially_young_cards_per_entry_ratio_seq);
395 }
397 size_t predict_young_card_num(size_t rs_length) {
398 return (size_t) ((double) rs_length *
399 predict_fully_young_cards_per_entry_ratio());
400 }
402 size_t predict_non_young_card_num(size_t rs_length) {
403 return (size_t) ((double) rs_length *
404 predict_partially_young_cards_per_entry_ratio());
405 }
407 double predict_rs_scan_time_ms(size_t card_num) {
408 if (full_young_gcs())
409 return (double) card_num * get_new_prediction(_cost_per_entry_ms_seq);
410 else
411 return predict_partially_young_rs_scan_time_ms(card_num);
412 }
414 double predict_partially_young_rs_scan_time_ms(size_t card_num) {
415 if (_partially_young_cost_per_entry_ms_seq->num() < 3)
416 return (double) card_num * get_new_prediction(_cost_per_entry_ms_seq);
417 else
418 return (double) card_num *
419 get_new_prediction(_partially_young_cost_per_entry_ms_seq);
420 }
422 double predict_object_copy_time_ms_during_cm(size_t bytes_to_copy) {
423 if (_cost_per_byte_ms_during_cm_seq->num() < 3)
424 return 1.1 * (double) bytes_to_copy *
425 get_new_prediction(_cost_per_byte_ms_seq);
426 else
427 return (double) bytes_to_copy *
428 get_new_prediction(_cost_per_byte_ms_during_cm_seq);
429 }
431 double predict_object_copy_time_ms(size_t bytes_to_copy) {
432 if (_in_marking_window && !_in_marking_window_im)
433 return predict_object_copy_time_ms_during_cm(bytes_to_copy);
434 else
435 return (double) bytes_to_copy *
436 get_new_prediction(_cost_per_byte_ms_seq);
437 }
439 double predict_constant_other_time_ms() {
440 return get_new_prediction(_constant_other_time_ms_seq);
441 }
443 double predict_young_other_time_ms(size_t young_num) {
444 return
445 (double) young_num *
446 get_new_prediction(_young_other_cost_per_region_ms_seq);
447 }
449 double predict_non_young_other_time_ms(size_t non_young_num) {
450 return
451 (double) non_young_num *
452 get_new_prediction(_non_young_other_cost_per_region_ms_seq);
453 }
455 void check_if_region_is_too_expensive(double predicted_time_ms);
457 double predict_young_collection_elapsed_time_ms(size_t adjustment);
458 double predict_base_elapsed_time_ms(size_t pending_cards);
459 double predict_base_elapsed_time_ms(size_t pending_cards,
460 size_t scanned_cards);
461 size_t predict_bytes_to_copy(HeapRegion* hr);
462 double predict_region_elapsed_time_ms(HeapRegion* hr, bool young);
464 // for use by: calculate_young_list_target_length(rs_length)
465 bool predict_will_fit(size_t young_region_num,
466 double base_time_ms,
467 size_t init_free_regions,
468 double target_pause_time_ms);
470 void start_recording_regions();
471 void record_cset_region_info(HeapRegion* hr, bool young);
472 void record_non_young_cset_region(HeapRegion* hr);
474 void set_recorded_young_regions(size_t n_regions);
475 void set_recorded_young_bytes(size_t bytes);
476 void set_recorded_rs_lengths(size_t rs_lengths);
477 void set_predicted_bytes_to_copy(size_t bytes);
479 void end_recording_regions();
481 void record_vtime_diff_ms(double vtime_diff_ms) {
482 _vtime_diff_ms = vtime_diff_ms;
483 }
485 void record_young_free_cset_time_ms(double time_ms) {
486 _recorded_young_free_cset_time_ms = time_ms;
487 }
489 void record_non_young_free_cset_time_ms(double time_ms) {
490 _recorded_non_young_free_cset_time_ms = time_ms;
491 }
493 double predict_young_gc_eff() {
494 return get_new_neg_prediction(_young_gc_eff_seq);
495 }
497 double predict_survivor_regions_evac_time();
499 // </NEW PREDICTION>
501 public:
502 void cset_regions_freed() {
503 bool propagate = _last_young_gc_full && !_in_marking_window;
504 _short_lived_surv_rate_group->all_surviving_words_recorded(propagate);
505 _survivor_surv_rate_group->all_surviving_words_recorded(propagate);
506 // also call it on any more surv rate groups
507 }
509 void set_known_garbage_bytes(size_t known_garbage_bytes) {
510 _known_garbage_bytes = known_garbage_bytes;
511 size_t heap_bytes = _g1->capacity();
512 _known_garbage_ratio = (double) _known_garbage_bytes / (double) heap_bytes;
513 }
515 void decrease_known_garbage_bytes(size_t known_garbage_bytes) {
516 guarantee( _known_garbage_bytes >= known_garbage_bytes, "invariant" );
518 _known_garbage_bytes -= known_garbage_bytes;
519 size_t heap_bytes = _g1->capacity();
520 _known_garbage_ratio = (double) _known_garbage_bytes / (double) heap_bytes;
521 }
523 G1MMUTracker* mmu_tracker() {
524 return _mmu_tracker;
525 }
527 double max_pause_time_ms() {
528 return _mmu_tracker->max_gc_time() * 1000.0;
529 }
531 double predict_init_time_ms() {
532 return get_new_prediction(_concurrent_mark_init_times_ms);
533 }
535 double predict_remark_time_ms() {
536 return get_new_prediction(_concurrent_mark_remark_times_ms);
537 }
539 double predict_cleanup_time_ms() {
540 return get_new_prediction(_concurrent_mark_cleanup_times_ms);
541 }
543 // Returns an estimate of the survival rate of the region at yg-age
544 // "yg_age".
545 double predict_yg_surv_rate(int age, SurvRateGroup* surv_rate_group) {
546 TruncatedSeq* seq = surv_rate_group->get_seq(age);
547 if (seq->num() == 0)
548 gclog_or_tty->print("BARF! age is %d", age);
549 guarantee( seq->num() > 0, "invariant" );
550 double pred = get_new_prediction(seq);
551 if (pred > 1.0)
552 pred = 1.0;
553 return pred;
554 }
556 double predict_yg_surv_rate(int age) {
557 return predict_yg_surv_rate(age, _short_lived_surv_rate_group);
558 }
560 double accum_yg_surv_rate_pred(int age) {
561 return _short_lived_surv_rate_group->accum_surv_rate_pred(age);
562 }
564 protected:
565 void print_stats(int level, const char* str, double value);
566 void print_stats(int level, const char* str, int value);
568 void print_par_stats(int level, const char* str, double* data) {
569 print_par_stats(level, str, data, true);
570 }
571 void print_par_stats(int level, const char* str, double* data, bool summary);
572 void print_par_sizes(int level, const char* str, double* data, bool summary);
574 void check_other_times(int level,
575 NumberSeq* other_times_ms,
576 NumberSeq* calc_other_times_ms) const;
578 void print_summary (PauseSummary* stats) const;
580 void print_summary (int level, const char* str, NumberSeq* seq) const;
581 void print_summary_sd (int level, const char* str, NumberSeq* seq) const;
583 double avg_value (double* data);
584 double max_value (double* data);
585 double sum_of_values (double* data);
586 double max_sum (double* data1, double* data2);
588 int _last_satb_drain_processed_buffers;
589 int _last_update_rs_processed_buffers;
590 double _last_pause_time_ms;
592 size_t _bytes_in_to_space_before_gc;
593 size_t _bytes_in_to_space_after_gc;
594 size_t bytes_in_to_space_during_gc() {
595 return
596 _bytes_in_to_space_after_gc - _bytes_in_to_space_before_gc;
597 }
598 size_t _bytes_in_collection_set_before_gc;
599 // Used to count used bytes in CS.
600 friend class CountCSClosure;
602 // Statistics kept per GC stoppage, pause or full.
603 TruncatedSeq* _recent_prev_end_times_for_all_gcs_sec;
605 // We track markings.
606 int _num_markings;
607 double _mark_thread_startup_sec; // Time at startup of marking thread
609 // Add a new GC of the given duration and end time to the record.
610 void update_recent_gc_times(double end_time_sec, double elapsed_ms);
612 // The head of the list (via "next_in_collection_set()") representing the
613 // current collection set. Set from the incrementally built collection
614 // set at the start of the pause.
615 HeapRegion* _collection_set;
617 // The number of regions in the collection set. Set from the incrementally
618 // built collection set at the start of an evacuation pause.
619 size_t _collection_set_size;
621 // The number of bytes in the collection set before the pause. Set from
622 // the incrementally built collection set at the start of an evacuation
623 // pause.
624 size_t _collection_set_bytes_used_before;
626 // The associated information that is maintained while the incremental
627 // collection set is being built with young regions. Used to populate
628 // the recorded info for the evacuation pause.
630 enum CSetBuildType {
631 Active, // We are actively building the collection set
632 Inactive // We are not actively building the collection set
633 };
635 CSetBuildType _inc_cset_build_state;
637 // The head of the incrementally built collection set.
638 HeapRegion* _inc_cset_head;
640 // The tail of the incrementally built collection set.
641 HeapRegion* _inc_cset_tail;
643 // The number of regions in the incrementally built collection set.
644 // Used to set _collection_set_size at the start of an evacuation
645 // pause.
646 size_t _inc_cset_size;
648 // Used as the index in the surving young words structure
649 // which tracks the amount of space, for each young region,
650 // that survives the pause.
651 size_t _inc_cset_young_index;
653 // The number of bytes in the incrementally built collection set.
654 // Used to set _collection_set_bytes_used_before at the start of
655 // an evacuation pause.
656 size_t _inc_cset_bytes_used_before;
658 // Used to record the highest end of heap region in collection set
659 HeapWord* _inc_cset_max_finger;
661 // The number of recorded used bytes in the young regions
662 // of the collection set. This is the sum of the used() bytes
663 // of retired young regions in the collection set.
664 size_t _inc_cset_recorded_young_bytes;
666 // The RSet lengths recorded for regions in the collection set
667 // (updated by the periodic sampling of the regions in the
668 // young list/collection set).
669 size_t _inc_cset_recorded_rs_lengths;
671 // The predicted elapsed time it will take to collect the regions
672 // in the collection set (updated by the periodic sampling of the
673 // regions in the young list/collection set).
674 double _inc_cset_predicted_elapsed_time_ms;
676 // The predicted bytes to copy for the regions in the collection
677 // set (updated by the periodic sampling of the regions in the
678 // young list/collection set).
679 size_t _inc_cset_predicted_bytes_to_copy;
681 // Info about marking.
682 int _n_marks; // Sticky at 2, so we know when we've done at least 2.
684 // The number of collection pauses at the end of the last mark.
685 size_t _n_pauses_at_mark_end;
687 // Stash a pointer to the g1 heap.
688 G1CollectedHeap* _g1;
690 // The average time in ms per collection pause, averaged over recent pauses.
691 double recent_avg_time_for_pauses_ms();
693 // The average time in ms for processing CollectedHeap strong roots, per
694 // collection pause, averaged over recent pauses.
695 double recent_avg_time_for_CH_strong_ms();
697 // The average time in ms for processing the G1 remembered set, per
698 // pause, averaged over recent pauses.
699 double recent_avg_time_for_G1_strong_ms();
701 // The average time in ms for "evacuating followers", per pause, averaged
702 // over recent pauses.
703 double recent_avg_time_for_evac_ms();
705 // The number of "recent" GCs recorded in the number sequences
706 int number_of_recent_gcs();
708 // The average survival ratio, computed by the total number of bytes
709 // suriviving / total number of bytes before collection over the last
710 // several recent pauses.
711 double recent_avg_survival_fraction();
712 // The survival fraction of the most recent pause; if there have been no
713 // pauses, returns 1.0.
714 double last_survival_fraction();
716 // Returns a "conservative" estimate of the recent survival rate, i.e.,
717 // one that may be higher than "recent_avg_survival_fraction".
718 // This is conservative in several ways:
719 // If there have been few pauses, it will assume a potential high
720 // variance, and err on the side of caution.
721 // It puts a lower bound (currently 0.1) on the value it will return.
722 // To try to detect phase changes, if the most recent pause ("latest") has a
723 // higher-than average ("avg") survival rate, it returns that rate.
724 // "work" version is a utility function; young is restricted to young regions.
725 double conservative_avg_survival_fraction_work(double avg,
726 double latest);
728 // The arguments are the two sequences that keep track of the number of bytes
729 // surviving and the total number of bytes before collection, resp.,
730 // over the last evereal recent pauses
731 // Returns the survival rate for the category in the most recent pause.
732 // If there have been no pauses, returns 1.0.
733 double last_survival_fraction_work(TruncatedSeq* surviving,
734 TruncatedSeq* before);
736 // The arguments are the two sequences that keep track of the number of bytes
737 // surviving and the total number of bytes before collection, resp.,
738 // over the last several recent pauses
739 // Returns the average survival ration over the last several recent pauses
740 // If there have been no pauses, return 1.0
741 double recent_avg_survival_fraction_work(TruncatedSeq* surviving,
742 TruncatedSeq* before);
744 double conservative_avg_survival_fraction() {
745 double avg = recent_avg_survival_fraction();
746 double latest = last_survival_fraction();
747 return conservative_avg_survival_fraction_work(avg, latest);
748 }
750 // The ratio of gc time to elapsed time, computed over recent pauses.
751 double _recent_avg_pause_time_ratio;
753 double recent_avg_pause_time_ratio() {
754 return _recent_avg_pause_time_ratio;
755 }
757 // Number of pauses between concurrent marking.
758 size_t _pauses_btwn_concurrent_mark;
760 size_t _n_marks_since_last_pause;
762 // At the end of a pause we check the heap occupancy and we decide
763 // whether we will start a marking cycle during the next pause. If
764 // we decide that we want to do that, we will set this parameter to
765 // true. So, this parameter will stay true between the end of a
766 // pause and the beginning of a subsequent pause (not necessarily
767 // the next one, see the comments on the next field) when we decide
768 // that we will indeed start a marking cycle and do the initial-mark
769 // work.
770 volatile bool _initiate_conc_mark_if_possible;
772 // If initiate_conc_mark_if_possible() is set at the beginning of a
773 // pause, it is a suggestion that the pause should start a marking
774 // cycle by doing the initial-mark work. However, it is possible
775 // that the concurrent marking thread is still finishing up the
776 // previous marking cycle (e.g., clearing the next marking
777 // bitmap). If that is the case we cannot start a new cycle and
778 // we'll have to wait for the concurrent marking thread to finish
779 // what it is doing. In this case we will postpone the marking cycle
780 // initiation decision for the next pause. When we eventually decide
781 // to start a cycle, we will set _during_initial_mark_pause which
782 // will stay true until the end of the initial-mark pause and it's
783 // the condition that indicates that a pause is doing the
784 // initial-mark work.
785 volatile bool _during_initial_mark_pause;
787 bool _should_revert_to_full_young_gcs;
788 bool _last_full_young_gc;
790 // This set of variables tracks the collector efficiency, in order to
791 // determine whether we should initiate a new marking.
792 double _cur_mark_stop_world_time_ms;
793 double _mark_init_start_sec;
794 double _mark_remark_start_sec;
795 double _mark_cleanup_start_sec;
796 double _mark_closure_time_ms;
798 void calculate_young_list_min_length();
799 void calculate_young_list_target_length();
800 void calculate_young_list_target_length(size_t rs_lengths);
802 public:
804 G1CollectorPolicy();
806 virtual G1CollectorPolicy* as_g1_policy() { return this; }
808 virtual CollectorPolicy::Name kind() {
809 return CollectorPolicy::G1CollectorPolicyKind;
810 }
812 void check_prediction_validity();
814 size_t bytes_in_collection_set() {
815 return _bytes_in_collection_set_before_gc;
816 }
818 size_t bytes_in_to_space() {
819 return bytes_in_to_space_during_gc();
820 }
822 unsigned calc_gc_alloc_time_stamp() {
823 return _all_pause_times_ms->num() + 1;
824 }
826 protected:
828 // Count the number of bytes used in the CS.
829 void count_CS_bytes_used();
831 // Together these do the base cleanup-recording work. Subclasses might
832 // want to put something between them.
833 void record_concurrent_mark_cleanup_end_work1(size_t freed_bytes,
834 size_t max_live_bytes);
835 void record_concurrent_mark_cleanup_end_work2();
837 public:
839 virtual void init();
841 // Create jstat counters for the policy.
842 virtual void initialize_gc_policy_counters();
844 virtual HeapWord* mem_allocate_work(size_t size,
845 bool is_tlab,
846 bool* gc_overhead_limit_was_exceeded);
848 // This method controls how a collector handles one or more
849 // of its generations being fully allocated.
850 virtual HeapWord* satisfy_failed_allocation(size_t size,
851 bool is_tlab);
853 BarrierSet::Name barrier_set_name() { return BarrierSet::G1SATBCTLogging; }
855 GenRemSet::Name rem_set_name() { return GenRemSet::CardTable; }
857 // The number of collection pauses so far.
858 long n_pauses() const { return _n_pauses; }
860 // Update the heuristic info to record a collection pause of the given
861 // start time, where the given number of bytes were used at the start.
862 // This may involve changing the desired size of a collection set.
864 virtual void record_stop_world_start();
866 virtual void record_collection_pause_start(double start_time_sec,
867 size_t start_used);
869 // Must currently be called while the world is stopped.
870 virtual void record_concurrent_mark_init_start();
871 virtual void record_concurrent_mark_init_end();
872 void record_concurrent_mark_init_end_pre(double
873 mark_init_elapsed_time_ms);
875 void record_mark_closure_time(double mark_closure_time_ms);
877 virtual void record_concurrent_mark_remark_start();
878 virtual void record_concurrent_mark_remark_end();
880 virtual void record_concurrent_mark_cleanup_start();
881 virtual void record_concurrent_mark_cleanup_end(size_t freed_bytes,
882 size_t max_live_bytes);
883 virtual void record_concurrent_mark_cleanup_completed();
885 virtual void record_concurrent_pause();
886 virtual void record_concurrent_pause_end();
888 virtual void record_collection_pause_end_CH_strong_roots();
889 virtual void record_collection_pause_end_G1_strong_roots();
891 virtual void record_collection_pause_end();
893 // Record the fact that a full collection occurred.
894 virtual void record_full_collection_start();
895 virtual void record_full_collection_end();
897 void record_gc_worker_start_time(int worker_i, double ms) {
898 _par_last_gc_worker_start_times_ms[worker_i] = ms;
899 }
901 void record_ext_root_scan_time(int worker_i, double ms) {
902 _par_last_ext_root_scan_times_ms[worker_i] = ms;
903 }
905 void record_mark_stack_scan_time(int worker_i, double ms) {
906 _par_last_mark_stack_scan_times_ms[worker_i] = ms;
907 }
909 void record_satb_drain_time(double ms) {
910 _cur_satb_drain_time_ms = ms;
911 _satb_drain_time_set = true;
912 }
914 void record_satb_drain_processed_buffers (int processed_buffers) {
915 _last_satb_drain_processed_buffers = processed_buffers;
916 }
918 void record_mod_union_time(double ms) {
919 _all_mod_union_times_ms->add(ms);
920 }
922 void record_update_rs_time(int thread, double ms) {
923 _par_last_update_rs_times_ms[thread] = ms;
924 }
926 void record_update_rs_processed_buffers (int thread,
927 double processed_buffers) {
928 _par_last_update_rs_processed_buffers[thread] = processed_buffers;
929 }
931 void record_scan_rs_time(int thread, double ms) {
932 _par_last_scan_rs_times_ms[thread] = ms;
933 }
935 void reset_obj_copy_time(int thread) {
936 _par_last_obj_copy_times_ms[thread] = 0.0;
937 }
939 void reset_obj_copy_time() {
940 reset_obj_copy_time(0);
941 }
943 void record_obj_copy_time(int thread, double ms) {
944 _par_last_obj_copy_times_ms[thread] += ms;
945 }
947 void record_termination(int thread, double ms, size_t attempts) {
948 _par_last_termination_times_ms[thread] = ms;
949 _par_last_termination_attempts[thread] = (double) attempts;
950 }
952 void record_gc_worker_end_time(int worker_i, double ms) {
953 _par_last_gc_worker_end_times_ms[worker_i] = ms;
954 }
956 void record_pause_time_ms(double ms) {
957 _last_pause_time_ms = ms;
958 }
960 void record_clear_ct_time(double ms) {
961 _cur_clear_ct_time_ms = ms;
962 }
964 void record_par_time(double ms) {
965 _cur_collection_par_time_ms = ms;
966 }
968 void record_aux_start_time(int i) {
969 guarantee(i < _aux_num, "should be within range");
970 _cur_aux_start_times_ms[i] = os::elapsedTime() * 1000.0;
971 }
973 void record_aux_end_time(int i) {
974 guarantee(i < _aux_num, "should be within range");
975 double ms = os::elapsedTime() * 1000.0 - _cur_aux_start_times_ms[i];
976 _cur_aux_times_set[i] = true;
977 _cur_aux_times_ms[i] += ms;
978 }
980 #ifndef PRODUCT
981 void record_cc_clear_time(double ms) {
982 if (_min_clear_cc_time_ms < 0.0 || ms <= _min_clear_cc_time_ms)
983 _min_clear_cc_time_ms = ms;
984 if (_max_clear_cc_time_ms < 0.0 || ms >= _max_clear_cc_time_ms)
985 _max_clear_cc_time_ms = ms;
986 _cur_clear_cc_time_ms = ms;
987 _cum_clear_cc_time_ms += ms;
988 _num_cc_clears++;
989 }
990 #endif
992 // Record the fact that "bytes" bytes allocated in a region.
993 void record_before_bytes(size_t bytes);
994 void record_after_bytes(size_t bytes);
996 // Choose a new collection set. Marks the chosen regions as being
997 // "in_collection_set", and links them together. The head and number of
998 // the collection set are available via access methods.
999 virtual void choose_collection_set(double target_pause_time_ms) = 0;
1001 // The head of the list (via "next_in_collection_set()") representing the
1002 // current collection set.
1003 HeapRegion* collection_set() { return _collection_set; }
1005 void clear_collection_set() { _collection_set = NULL; }
1007 // The number of elements in the current collection set.
1008 size_t collection_set_size() { return _collection_set_size; }
1010 // Add "hr" to the CS.
1011 void add_to_collection_set(HeapRegion* hr);
1013 // Incremental CSet Support
1015 // The head of the incrementally built collection set.
1016 HeapRegion* inc_cset_head() { return _inc_cset_head; }
1018 // The tail of the incrementally built collection set.
1019 HeapRegion* inc_set_tail() { return _inc_cset_tail; }
1021 // The number of elements in the incrementally built collection set.
1022 size_t inc_cset_size() { return _inc_cset_size; }
1024 // Initialize incremental collection set info.
1025 void start_incremental_cset_building();
1027 void clear_incremental_cset() {
1028 _inc_cset_head = NULL;
1029 _inc_cset_tail = NULL;
1030 }
1032 // Stop adding regions to the incremental collection set
1033 void stop_incremental_cset_building() { _inc_cset_build_state = Inactive; }
1035 // Add/remove information about hr to the aggregated information
1036 // for the incrementally built collection set.
1037 void add_to_incremental_cset_info(HeapRegion* hr, size_t rs_length);
1038 void remove_from_incremental_cset_info(HeapRegion* hr);
1040 // Update information about hr in the aggregated information for
1041 // the incrementally built collection set.
1042 void update_incremental_cset_info(HeapRegion* hr, size_t new_rs_length);
1044 private:
1045 // Update the incremental cset information when adding a region
1046 // (should not be called directly).
1047 void add_region_to_incremental_cset_common(HeapRegion* hr);
1049 public:
1050 // Add hr to the LHS of the incremental collection set.
1051 void add_region_to_incremental_cset_lhs(HeapRegion* hr);
1053 // Add hr to the RHS of the incremental collection set.
1054 void add_region_to_incremental_cset_rhs(HeapRegion* hr);
1056 #ifndef PRODUCT
1057 void print_collection_set(HeapRegion* list_head, outputStream* st);
1058 #endif // !PRODUCT
1060 bool initiate_conc_mark_if_possible() { return _initiate_conc_mark_if_possible; }
1061 void set_initiate_conc_mark_if_possible() { _initiate_conc_mark_if_possible = true; }
1062 void clear_initiate_conc_mark_if_possible() { _initiate_conc_mark_if_possible = false; }
1064 bool during_initial_mark_pause() { return _during_initial_mark_pause; }
1065 void set_during_initial_mark_pause() { _during_initial_mark_pause = true; }
1066 void clear_during_initial_mark_pause(){ _during_initial_mark_pause = false; }
1068 // This sets the initiate_conc_mark_if_possible() flag to start a
1069 // new cycle, as long as we are not already in one. It's best if it
1070 // is called during a safepoint when the test whether a cycle is in
1071 // progress or not is stable.
1072 bool force_initial_mark_if_outside_cycle();
1074 // This is called at the very beginning of an evacuation pause (it
1075 // has to be the first thing that the pause does). If
1076 // initiate_conc_mark_if_possible() is true, and the concurrent
1077 // marking thread has completed its work during the previous cycle,
1078 // it will set during_initial_mark_pause() to so that the pause does
1079 // the initial-mark work and start a marking cycle.
1080 void decide_on_conc_mark_initiation();
1082 // If an expansion would be appropriate, because recent GC overhead had
1083 // exceeded the desired limit, return an amount to expand by.
1084 virtual size_t expansion_amount();
1086 // note start of mark thread
1087 void note_start_of_mark_thread();
1089 // The marked bytes of the "r" has changed; reclassify it's desirability
1090 // for marking. Also asserts that "r" is eligible for a CS.
1091 virtual void note_change_in_marked_bytes(HeapRegion* r) = 0;
1093 #ifndef PRODUCT
1094 // Check any appropriate marked bytes info, asserting false if
1095 // something's wrong, else returning "true".
1096 virtual bool assertMarkedBytesDataOK() = 0;
1097 #endif
1099 // Print tracing information.
1100 void print_tracing_info() const;
1102 // Print stats on young survival ratio
1103 void print_yg_surv_rate_info() const;
1105 void finished_recalculating_age_indexes(bool is_survivors) {
1106 if (is_survivors) {
1107 _survivor_surv_rate_group->finished_recalculating_age_indexes();
1108 } else {
1109 _short_lived_surv_rate_group->finished_recalculating_age_indexes();
1110 }
1111 // do that for any other surv rate groups
1112 }
1114 bool is_young_list_full() {
1115 size_t young_list_length = _g1->young_list()->length();
1116 size_t young_list_max_length = _young_list_target_length;
1117 if (G1FixedEdenSize) {
1118 young_list_max_length -= _max_survivor_regions;
1119 }
1121 return young_list_length >= young_list_max_length;
1122 }
1123 void update_region_num(bool young);
1125 bool in_young_gc_mode() {
1126 return _in_young_gc_mode;
1127 }
1128 void set_in_young_gc_mode(bool in_young_gc_mode) {
1129 _in_young_gc_mode = in_young_gc_mode;
1130 }
1132 bool full_young_gcs() {
1133 return _full_young_gcs;
1134 }
1135 void set_full_young_gcs(bool full_young_gcs) {
1136 _full_young_gcs = full_young_gcs;
1137 }
1139 bool adaptive_young_list_length() {
1140 return _adaptive_young_list_length;
1141 }
1142 void set_adaptive_young_list_length(bool adaptive_young_list_length) {
1143 _adaptive_young_list_length = adaptive_young_list_length;
1144 }
1146 inline double get_gc_eff_factor() {
1147 double ratio = _known_garbage_ratio;
1149 double square = ratio * ratio;
1150 // square = square * square;
1151 double ret = square * 9.0 + 1.0;
1152 #if 0
1153 gclog_or_tty->print_cr("ratio = %1.2lf, ret = %1.2lf", ratio, ret);
1154 #endif // 0
1155 guarantee(0.0 <= ret && ret < 10.0, "invariant!");
1156 return ret;
1157 }
1159 //
1160 // Survivor regions policy.
1161 //
1162 protected:
1164 // Current tenuring threshold, set to 0 if the collector reaches the
1165 // maximum amount of suvivors regions.
1166 int _tenuring_threshold;
1168 // The limit on the number of regions allocated for survivors.
1169 size_t _max_survivor_regions;
1171 // The amount of survor regions after a collection.
1172 size_t _recorded_survivor_regions;
1173 // List of survivor regions.
1174 HeapRegion* _recorded_survivor_head;
1175 HeapRegion* _recorded_survivor_tail;
1177 ageTable _survivors_age_table;
1179 public:
1181 inline GCAllocPurpose
1182 evacuation_destination(HeapRegion* src_region, int age, size_t word_sz) {
1183 if (age < _tenuring_threshold && src_region->is_young()) {
1184 return GCAllocForSurvived;
1185 } else {
1186 return GCAllocForTenured;
1187 }
1188 }
1190 inline bool track_object_age(GCAllocPurpose purpose) {
1191 return purpose == GCAllocForSurvived;
1192 }
1194 inline GCAllocPurpose alternative_purpose(int purpose) {
1195 return GCAllocForTenured;
1196 }
1198 static const size_t REGIONS_UNLIMITED = ~(size_t)0;
1200 size_t max_regions(int purpose);
1202 // The limit on regions for a particular purpose is reached.
1203 void note_alloc_region_limit_reached(int purpose) {
1204 if (purpose == GCAllocForSurvived) {
1205 _tenuring_threshold = 0;
1206 }
1207 }
1209 void note_start_adding_survivor_regions() {
1210 _survivor_surv_rate_group->start_adding_regions();
1211 }
1213 void note_stop_adding_survivor_regions() {
1214 _survivor_surv_rate_group->stop_adding_regions();
1215 }
1217 void record_survivor_regions(size_t regions,
1218 HeapRegion* head,
1219 HeapRegion* tail) {
1220 _recorded_survivor_regions = regions;
1221 _recorded_survivor_head = head;
1222 _recorded_survivor_tail = tail;
1223 }
1225 size_t recorded_survivor_regions() {
1226 return _recorded_survivor_regions;
1227 }
1229 void record_thread_age_table(ageTable* age_table)
1230 {
1231 _survivors_age_table.merge_par(age_table);
1232 }
1234 // Calculates survivor space parameters.
1235 void calculate_survivors_policy();
1237 };
1239 // This encapsulates a particular strategy for a g1 Collector.
1240 //
1241 // Start a concurrent mark when our heap size is n bytes
1242 // greater then our heap size was at the last concurrent
1243 // mark. Where n is a function of the CMSTriggerRatio
1244 // and the MinHeapFreeRatio.
1245 //
1246 // Start a g1 collection pause when we have allocated the
1247 // average number of bytes currently being freed in
1248 // a collection, but only if it is at least one region
1249 // full
1250 //
1251 // Resize Heap based on desired
1252 // allocation space, where desired allocation space is
1253 // a function of survival rate and desired future to size.
1254 //
1255 // Choose collection set by first picking all older regions
1256 // which have a survival rate which beats our projected young
1257 // survival rate. Then fill out the number of needed regions
1258 // with young regions.
1260 class G1CollectorPolicy_BestRegionsFirst: public G1CollectorPolicy {
1261 CollectionSetChooser* _collectionSetChooser;
1262 // If the estimated is less then desirable, resize if possible.
1263 void expand_if_possible(size_t numRegions);
1265 virtual void choose_collection_set(double target_pause_time_ms);
1266 virtual void record_collection_pause_start(double start_time_sec,
1267 size_t start_used);
1268 virtual void record_concurrent_mark_cleanup_end(size_t freed_bytes,
1269 size_t max_live_bytes);
1270 virtual void record_full_collection_end();
1272 public:
1273 G1CollectorPolicy_BestRegionsFirst() {
1274 _collectionSetChooser = new CollectionSetChooser();
1275 }
1276 void record_collection_pause_end();
1277 // This is not needed any more, after the CSet choosing code was
1278 // changed to use the pause prediction work. But let's leave the
1279 // hook in just in case.
1280 void note_change_in_marked_bytes(HeapRegion* r) { }
1281 #ifndef PRODUCT
1282 bool assertMarkedBytesDataOK();
1283 #endif
1284 };
1286 // This should move to some place more general...
1288 // If we have "n" measurements, and we've kept track of their "sum" and the
1289 // "sum_of_squares" of the measurements, this returns the variance of the
1290 // sequence.
1291 inline double variance(int n, double sum_of_squares, double sum) {
1292 double n_d = (double)n;
1293 double avg = sum/n_d;
1294 return (sum_of_squares - 2.0 * avg * sum + n_d * avg * avg) / n_d;
1295 }
1297 // Local Variables: ***
1298 // c-indentation-style: gnu ***
1299 // End: ***
1301 #endif // SHARE_VM_GC_IMPLEMENTATION_G1_G1COLLECTORPOLICY_HPP