Sun, 23 Oct 2011 23:06:06 -0700
7096030: G1: PrintGCDetails enhancements
7102445: G1: Unnecessary Resource allocations during RSet scanning
Summary: Add a new per-worker thread line in the PrintGCDetails output. GC Worker Other is the difference between the elapsed time for the parallel phase of the evacuation pause and the sum of the times of the sub-phases (external root scanning, mark stack scanning, RSet updating, RSet scanning, object copying, and termination) for that worker. During RSet scanning, stack allocate DirtyCardToOopClosure objects; allocating these in a resource area was causing abnormally high GC Worker Other times while the worker thread freed ResourceArea chunks.
Reviewed-by: tonyp, jwilhelm, brutisso
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 class G1CollectorPolicy: public CollectorPolicy {
87 private:
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 CollectionSetChooser* _collectionSetChooser;
111 double _cur_collection_start_sec;
112 size_t _cur_collection_pause_used_at_start_bytes;
113 size_t _cur_collection_pause_used_regions_at_start;
114 size_t _prev_collection_pause_used_at_end_bytes;
115 double _cur_collection_par_time_ms;
116 double _cur_satb_drain_time_ms;
117 double _cur_clear_ct_time_ms;
118 double _cur_ref_proc_time_ms;
119 double _cur_ref_enq_time_ms;
121 #ifndef PRODUCT
122 // Card Table Count Cache stats
123 double _min_clear_cc_time_ms; // min
124 double _max_clear_cc_time_ms; // max
125 double _cur_clear_cc_time_ms; // clearing time during current pause
126 double _cum_clear_cc_time_ms; // cummulative clearing time
127 jlong _num_cc_clears; // number of times the card count cache has been cleared
128 #endif
130 // Statistics for recent GC pauses. See below for how indexed.
131 TruncatedSeq* _recent_rs_scan_times_ms;
133 // These exclude marking times.
134 TruncatedSeq* _recent_pause_times_ms;
135 TruncatedSeq* _recent_gc_times_ms;
137 TruncatedSeq* _recent_CS_bytes_used_before;
138 TruncatedSeq* _recent_CS_bytes_surviving;
140 TruncatedSeq* _recent_rs_sizes;
142 TruncatedSeq* _concurrent_mark_remark_times_ms;
143 TruncatedSeq* _concurrent_mark_cleanup_times_ms;
145 Summary* _summary;
147 NumberSeq* _all_pause_times_ms;
148 NumberSeq* _all_full_gc_times_ms;
149 double _stop_world_start;
150 NumberSeq* _all_stop_world_times_ms;
151 NumberSeq* _all_yield_times_ms;
153 size_t _region_num_young;
154 size_t _region_num_tenured;
155 size_t _prev_region_num_young;
156 size_t _prev_region_num_tenured;
158 NumberSeq* _all_mod_union_times_ms;
160 int _aux_num;
161 NumberSeq* _all_aux_times_ms;
162 double* _cur_aux_start_times_ms;
163 double* _cur_aux_times_ms;
164 bool* _cur_aux_times_set;
166 double* _par_last_gc_worker_start_times_ms;
167 double* _par_last_ext_root_scan_times_ms;
168 double* _par_last_mark_stack_scan_times_ms;
169 double* _par_last_update_rs_times_ms;
170 double* _par_last_update_rs_processed_buffers;
171 double* _par_last_scan_rs_times_ms;
172 double* _par_last_obj_copy_times_ms;
173 double* _par_last_termination_times_ms;
174 double* _par_last_termination_attempts;
175 double* _par_last_gc_worker_end_times_ms;
176 double* _par_last_gc_worker_times_ms;
178 // Each workers 'other' time i.e. the elapsed time of the parallel
179 // phase of the pause minus the sum of the individual sub-phase
180 // times for a given worker thread.
181 double* _par_last_gc_worker_other_times_ms;
183 // indicates whether we are in full young or partially young GC mode
184 bool _full_young_gcs;
186 // if true, then it tries to dynamically adjust the length of the
187 // young list
188 bool _adaptive_young_list_length;
189 size_t _young_list_target_length;
190 size_t _young_list_fixed_length;
191 size_t _prev_eden_capacity; // used for logging
193 // The max number of regions we can extend the eden by while the GC
194 // locker is active. This should be >= _young_list_target_length;
195 size_t _young_list_max_length;
197 size_t _young_cset_length;
198 bool _last_young_gc_full;
200 unsigned _full_young_pause_num;
201 unsigned _partial_young_pause_num;
203 bool _during_marking;
204 bool _in_marking_window;
205 bool _in_marking_window_im;
207 SurvRateGroup* _short_lived_surv_rate_group;
208 SurvRateGroup* _survivor_surv_rate_group;
209 // add here any more surv rate groups
211 double _gc_overhead_perc;
213 double _reserve_factor;
214 size_t _reserve_regions;
216 bool during_marking() {
217 return _during_marking;
218 }
220 // <NEW PREDICTION>
222 private:
223 enum PredictionConstants {
224 TruncatedSeqLength = 10
225 };
227 TruncatedSeq* _alloc_rate_ms_seq;
228 double _prev_collection_pause_end_ms;
230 TruncatedSeq* _pending_card_diff_seq;
231 TruncatedSeq* _rs_length_diff_seq;
232 TruncatedSeq* _cost_per_card_ms_seq;
233 TruncatedSeq* _fully_young_cards_per_entry_ratio_seq;
234 TruncatedSeq* _partially_young_cards_per_entry_ratio_seq;
235 TruncatedSeq* _cost_per_entry_ms_seq;
236 TruncatedSeq* _partially_young_cost_per_entry_ms_seq;
237 TruncatedSeq* _cost_per_byte_ms_seq;
238 TruncatedSeq* _constant_other_time_ms_seq;
239 TruncatedSeq* _young_other_cost_per_region_ms_seq;
240 TruncatedSeq* _non_young_other_cost_per_region_ms_seq;
242 TruncatedSeq* _pending_cards_seq;
243 TruncatedSeq* _scanned_cards_seq;
244 TruncatedSeq* _rs_lengths_seq;
246 TruncatedSeq* _cost_per_byte_ms_during_cm_seq;
248 TruncatedSeq* _young_gc_eff_seq;
250 TruncatedSeq* _max_conc_overhead_seq;
252 bool _using_new_ratio_calculations;
253 size_t _min_desired_young_length; // as set on the command line or default calculations
254 size_t _max_desired_young_length; // as set on the command line or default calculations
256 size_t _recorded_young_regions;
257 size_t _recorded_non_young_regions;
258 size_t _recorded_region_num;
260 size_t _free_regions_at_end_of_collection;
262 size_t _recorded_rs_lengths;
263 size_t _max_rs_lengths;
265 size_t _recorded_marked_bytes;
266 size_t _recorded_young_bytes;
268 size_t _predicted_pending_cards;
269 size_t _predicted_cards_scanned;
270 size_t _predicted_rs_lengths;
271 size_t _predicted_bytes_to_copy;
273 double _predicted_survival_ratio;
274 double _predicted_rs_update_time_ms;
275 double _predicted_rs_scan_time_ms;
276 double _predicted_object_copy_time_ms;
277 double _predicted_constant_other_time_ms;
278 double _predicted_young_other_time_ms;
279 double _predicted_non_young_other_time_ms;
280 double _predicted_pause_time_ms;
282 double _vtime_diff_ms;
284 double _recorded_young_free_cset_time_ms;
285 double _recorded_non_young_free_cset_time_ms;
287 double _sigma;
288 double _expensive_region_limit_ms;
290 size_t _rs_lengths_prediction;
292 size_t _known_garbage_bytes;
293 double _known_garbage_ratio;
295 double sigma() {
296 return _sigma;
297 }
299 // A function that prevents us putting too much stock in small sample
300 // sets. Returns a number between 2.0 and 1.0, depending on the number
301 // of samples. 5 or more samples yields one; fewer scales linearly from
302 // 2.0 at 1 sample to 1.0 at 5.
303 double confidence_factor(int samples) {
304 if (samples > 4) return 1.0;
305 else return 1.0 + sigma() * ((double)(5 - samples))/2.0;
306 }
308 double get_new_neg_prediction(TruncatedSeq* seq) {
309 return seq->davg() - sigma() * seq->dsd();
310 }
312 #ifndef PRODUCT
313 bool verify_young_ages(HeapRegion* head, SurvRateGroup *surv_rate_group);
314 #endif // PRODUCT
316 void adjust_concurrent_refinement(double update_rs_time,
317 double update_rs_processed_buffers,
318 double goal_ms);
320 double _pause_time_target_ms;
321 double _recorded_young_cset_choice_time_ms;
322 double _recorded_non_young_cset_choice_time_ms;
323 bool _within_target;
324 size_t _pending_cards;
325 size_t _max_pending_cards;
327 public:
329 void set_region_short_lived(HeapRegion* hr) {
330 hr->install_surv_rate_group(_short_lived_surv_rate_group);
331 }
333 void set_region_survivors(HeapRegion* hr) {
334 hr->install_surv_rate_group(_survivor_surv_rate_group);
335 }
337 #ifndef PRODUCT
338 bool verify_young_ages();
339 #endif // PRODUCT
341 double get_new_prediction(TruncatedSeq* seq) {
342 return MAX2(seq->davg() + sigma() * seq->dsd(),
343 seq->davg() * confidence_factor(seq->num()));
344 }
346 size_t young_cset_length() {
347 return _young_cset_length;
348 }
350 void record_max_rs_lengths(size_t rs_lengths) {
351 _max_rs_lengths = rs_lengths;
352 }
354 size_t predict_pending_card_diff() {
355 double prediction = get_new_neg_prediction(_pending_card_diff_seq);
356 if (prediction < 0.00001)
357 return 0;
358 else
359 return (size_t) prediction;
360 }
362 size_t predict_pending_cards() {
363 size_t max_pending_card_num = _g1->max_pending_card_num();
364 size_t diff = predict_pending_card_diff();
365 size_t prediction;
366 if (diff > max_pending_card_num)
367 prediction = max_pending_card_num;
368 else
369 prediction = max_pending_card_num - diff;
371 return prediction;
372 }
374 size_t predict_rs_length_diff() {
375 return (size_t) get_new_prediction(_rs_length_diff_seq);
376 }
378 double predict_alloc_rate_ms() {
379 return get_new_prediction(_alloc_rate_ms_seq);
380 }
382 double predict_cost_per_card_ms() {
383 return get_new_prediction(_cost_per_card_ms_seq);
384 }
386 double predict_rs_update_time_ms(size_t pending_cards) {
387 return (double) pending_cards * predict_cost_per_card_ms();
388 }
390 double predict_fully_young_cards_per_entry_ratio() {
391 return get_new_prediction(_fully_young_cards_per_entry_ratio_seq);
392 }
394 double predict_partially_young_cards_per_entry_ratio() {
395 if (_partially_young_cards_per_entry_ratio_seq->num() < 2)
396 return predict_fully_young_cards_per_entry_ratio();
397 else
398 return get_new_prediction(_partially_young_cards_per_entry_ratio_seq);
399 }
401 size_t predict_young_card_num(size_t rs_length) {
402 return (size_t) ((double) rs_length *
403 predict_fully_young_cards_per_entry_ratio());
404 }
406 size_t predict_non_young_card_num(size_t rs_length) {
407 return (size_t) ((double) rs_length *
408 predict_partially_young_cards_per_entry_ratio());
409 }
411 double predict_rs_scan_time_ms(size_t card_num) {
412 if (full_young_gcs())
413 return (double) card_num * get_new_prediction(_cost_per_entry_ms_seq);
414 else
415 return predict_partially_young_rs_scan_time_ms(card_num);
416 }
418 double predict_partially_young_rs_scan_time_ms(size_t card_num) {
419 if (_partially_young_cost_per_entry_ms_seq->num() < 3)
420 return (double) card_num * get_new_prediction(_cost_per_entry_ms_seq);
421 else
422 return (double) card_num *
423 get_new_prediction(_partially_young_cost_per_entry_ms_seq);
424 }
426 double predict_object_copy_time_ms_during_cm(size_t bytes_to_copy) {
427 if (_cost_per_byte_ms_during_cm_seq->num() < 3)
428 return 1.1 * (double) bytes_to_copy *
429 get_new_prediction(_cost_per_byte_ms_seq);
430 else
431 return (double) bytes_to_copy *
432 get_new_prediction(_cost_per_byte_ms_during_cm_seq);
433 }
435 double predict_object_copy_time_ms(size_t bytes_to_copy) {
436 if (_in_marking_window && !_in_marking_window_im)
437 return predict_object_copy_time_ms_during_cm(bytes_to_copy);
438 else
439 return (double) bytes_to_copy *
440 get_new_prediction(_cost_per_byte_ms_seq);
441 }
443 double predict_constant_other_time_ms() {
444 return get_new_prediction(_constant_other_time_ms_seq);
445 }
447 double predict_young_other_time_ms(size_t young_num) {
448 return
449 (double) young_num *
450 get_new_prediction(_young_other_cost_per_region_ms_seq);
451 }
453 double predict_non_young_other_time_ms(size_t non_young_num) {
454 return
455 (double) non_young_num *
456 get_new_prediction(_non_young_other_cost_per_region_ms_seq);
457 }
459 void check_if_region_is_too_expensive(double predicted_time_ms);
461 double predict_young_collection_elapsed_time_ms(size_t adjustment);
462 double predict_base_elapsed_time_ms(size_t pending_cards);
463 double predict_base_elapsed_time_ms(size_t pending_cards,
464 size_t scanned_cards);
465 size_t predict_bytes_to_copy(HeapRegion* hr);
466 double predict_region_elapsed_time_ms(HeapRegion* hr, bool young);
468 void start_recording_regions();
469 void record_cset_region_info(HeapRegion* hr, bool young);
470 void record_non_young_cset_region(HeapRegion* hr);
472 void set_recorded_young_regions(size_t n_regions);
473 void set_recorded_young_bytes(size_t bytes);
474 void set_recorded_rs_lengths(size_t rs_lengths);
475 void set_predicted_bytes_to_copy(size_t bytes);
477 void end_recording_regions();
479 void record_vtime_diff_ms(double vtime_diff_ms) {
480 _vtime_diff_ms = vtime_diff_ms;
481 }
483 void record_young_free_cset_time_ms(double time_ms) {
484 _recorded_young_free_cset_time_ms = time_ms;
485 }
487 void record_non_young_free_cset_time_ms(double time_ms) {
488 _recorded_non_young_free_cset_time_ms = time_ms;
489 }
491 double predict_young_gc_eff() {
492 return get_new_neg_prediction(_young_gc_eff_seq);
493 }
495 double predict_survivor_regions_evac_time();
497 // </NEW PREDICTION>
499 void cset_regions_freed() {
500 bool propagate = _last_young_gc_full && !_in_marking_window;
501 _short_lived_surv_rate_group->all_surviving_words_recorded(propagate);
502 _survivor_surv_rate_group->all_surviving_words_recorded(propagate);
503 // also call it on any more surv rate groups
504 }
506 void set_known_garbage_bytes(size_t known_garbage_bytes) {
507 _known_garbage_bytes = known_garbage_bytes;
508 size_t heap_bytes = _g1->capacity();
509 _known_garbage_ratio = (double) _known_garbage_bytes / (double) heap_bytes;
510 }
512 void decrease_known_garbage_bytes(size_t known_garbage_bytes) {
513 guarantee( _known_garbage_bytes >= known_garbage_bytes, "invariant" );
515 _known_garbage_bytes -= known_garbage_bytes;
516 size_t heap_bytes = _g1->capacity();
517 _known_garbage_ratio = (double) _known_garbage_bytes / (double) heap_bytes;
518 }
520 G1MMUTracker* mmu_tracker() {
521 return _mmu_tracker;
522 }
524 double max_pause_time_ms() {
525 return _mmu_tracker->max_gc_time() * 1000.0;
526 }
528 double predict_remark_time_ms() {
529 return get_new_prediction(_concurrent_mark_remark_times_ms);
530 }
532 double predict_cleanup_time_ms() {
533 return get_new_prediction(_concurrent_mark_cleanup_times_ms);
534 }
536 // Returns an estimate of the survival rate of the region at yg-age
537 // "yg_age".
538 double predict_yg_surv_rate(int age, SurvRateGroup* surv_rate_group) {
539 TruncatedSeq* seq = surv_rate_group->get_seq(age);
540 if (seq->num() == 0)
541 gclog_or_tty->print("BARF! age is %d", age);
542 guarantee( seq->num() > 0, "invariant" );
543 double pred = get_new_prediction(seq);
544 if (pred > 1.0)
545 pred = 1.0;
546 return pred;
547 }
549 double predict_yg_surv_rate(int age) {
550 return predict_yg_surv_rate(age, _short_lived_surv_rate_group);
551 }
553 double accum_yg_surv_rate_pred(int age) {
554 return _short_lived_surv_rate_group->accum_surv_rate_pred(age);
555 }
557 private:
558 void print_stats(int level, const char* str, double value);
559 void print_stats(int level, const char* str, int value);
561 void print_par_stats(int level, const char* str, double* data);
562 void print_par_sizes(int level, const char* str, double* data);
564 void check_other_times(int level,
565 NumberSeq* other_times_ms,
566 NumberSeq* calc_other_times_ms) const;
568 void print_summary (PauseSummary* stats) const;
570 void print_summary (int level, const char* str, NumberSeq* seq) const;
571 void print_summary_sd (int level, const char* str, NumberSeq* seq) const;
573 double avg_value (double* data);
574 double max_value (double* data);
575 double sum_of_values (double* data);
576 double max_sum (double* data1, double* data2);
578 int _last_satb_drain_processed_buffers;
579 int _last_update_rs_processed_buffers;
580 double _last_pause_time_ms;
582 size_t _bytes_in_collection_set_before_gc;
583 size_t _bytes_copied_during_gc;
585 // Used to count used bytes in CS.
586 friend class CountCSClosure;
588 // Statistics kept per GC stoppage, pause or full.
589 TruncatedSeq* _recent_prev_end_times_for_all_gcs_sec;
591 // Add a new GC of the given duration and end time to the record.
592 void update_recent_gc_times(double end_time_sec, double elapsed_ms);
594 // The head of the list (via "next_in_collection_set()") representing the
595 // current collection set. Set from the incrementally built collection
596 // set at the start of the pause.
597 HeapRegion* _collection_set;
599 // The number of regions in the collection set. Set from the incrementally
600 // built collection set at the start of an evacuation pause.
601 size_t _collection_set_size;
603 // The number of bytes in the collection set before the pause. Set from
604 // the incrementally built collection set at the start of an evacuation
605 // pause.
606 size_t _collection_set_bytes_used_before;
608 // The associated information that is maintained while the incremental
609 // collection set is being built with young regions. Used to populate
610 // the recorded info for the evacuation pause.
612 enum CSetBuildType {
613 Active, // We are actively building the collection set
614 Inactive // We are not actively building the collection set
615 };
617 CSetBuildType _inc_cset_build_state;
619 // The head of the incrementally built collection set.
620 HeapRegion* _inc_cset_head;
622 // The tail of the incrementally built collection set.
623 HeapRegion* _inc_cset_tail;
625 // The number of regions in the incrementally built collection set.
626 // Used to set _collection_set_size at the start of an evacuation
627 // pause.
628 size_t _inc_cset_size;
630 // Used as the index in the surving young words structure
631 // which tracks the amount of space, for each young region,
632 // that survives the pause.
633 size_t _inc_cset_young_index;
635 // The number of bytes in the incrementally built collection set.
636 // Used to set _collection_set_bytes_used_before at the start of
637 // an evacuation pause.
638 size_t _inc_cset_bytes_used_before;
640 // Used to record the highest end of heap region in collection set
641 HeapWord* _inc_cset_max_finger;
643 // The number of recorded used bytes in the young regions
644 // of the collection set. This is the sum of the used() bytes
645 // of retired young regions in the collection set.
646 size_t _inc_cset_recorded_young_bytes;
648 // The RSet lengths recorded for regions in the collection set
649 // (updated by the periodic sampling of the regions in the
650 // young list/collection set).
651 size_t _inc_cset_recorded_rs_lengths;
653 // The predicted elapsed time it will take to collect the regions
654 // in the collection set (updated by the periodic sampling of the
655 // regions in the young list/collection set).
656 double _inc_cset_predicted_elapsed_time_ms;
658 // The predicted bytes to copy for the regions in the collection
659 // set (updated by the periodic sampling of the regions in the
660 // young list/collection set).
661 size_t _inc_cset_predicted_bytes_to_copy;
663 // Stash a pointer to the g1 heap.
664 G1CollectedHeap* _g1;
666 // The average time in ms per collection pause, averaged over recent pauses.
667 double recent_avg_time_for_pauses_ms();
669 // The average time in ms for RS scanning, per pause, averaged
670 // over recent pauses. (Note the RS scanning time for a pause
671 // is itself an average of the RS scanning time for each worker
672 // thread.)
673 double recent_avg_time_for_rs_scan_ms();
675 // The number of "recent" GCs recorded in the number sequences
676 int number_of_recent_gcs();
678 // The average survival ratio, computed by the total number of bytes
679 // suriviving / total number of bytes before collection over the last
680 // several recent pauses.
681 double recent_avg_survival_fraction();
682 // The survival fraction of the most recent pause; if there have been no
683 // pauses, returns 1.0.
684 double last_survival_fraction();
686 // Returns a "conservative" estimate of the recent survival rate, i.e.,
687 // one that may be higher than "recent_avg_survival_fraction".
688 // This is conservative in several ways:
689 // If there have been few pauses, it will assume a potential high
690 // variance, and err on the side of caution.
691 // It puts a lower bound (currently 0.1) on the value it will return.
692 // To try to detect phase changes, if the most recent pause ("latest") has a
693 // higher-than average ("avg") survival rate, it returns that rate.
694 // "work" version is a utility function; young is restricted to young regions.
695 double conservative_avg_survival_fraction_work(double avg,
696 double latest);
698 // The arguments are the two sequences that keep track of the number of bytes
699 // surviving and the total number of bytes before collection, resp.,
700 // over the last evereal recent pauses
701 // Returns the survival rate for the category in the most recent pause.
702 // If there have been no pauses, returns 1.0.
703 double last_survival_fraction_work(TruncatedSeq* surviving,
704 TruncatedSeq* before);
706 // The arguments are the two sequences that keep track of the number of bytes
707 // surviving and the total number of bytes before collection, resp.,
708 // over the last several recent pauses
709 // Returns the average survival ration over the last several recent pauses
710 // If there have been no pauses, return 1.0
711 double recent_avg_survival_fraction_work(TruncatedSeq* surviving,
712 TruncatedSeq* before);
714 double conservative_avg_survival_fraction() {
715 double avg = recent_avg_survival_fraction();
716 double latest = last_survival_fraction();
717 return conservative_avg_survival_fraction_work(avg, latest);
718 }
720 // The ratio of gc time to elapsed time, computed over recent pauses.
721 double _recent_avg_pause_time_ratio;
723 double recent_avg_pause_time_ratio() {
724 return _recent_avg_pause_time_ratio;
725 }
727 // Number of pauses between concurrent marking.
728 size_t _pauses_btwn_concurrent_mark;
730 // At the end of a pause we check the heap occupancy and we decide
731 // whether we will start a marking cycle during the next pause. If
732 // we decide that we want to do that, we will set this parameter to
733 // true. So, this parameter will stay true between the end of a
734 // pause and the beginning of a subsequent pause (not necessarily
735 // the next one, see the comments on the next field) when we decide
736 // that we will indeed start a marking cycle and do the initial-mark
737 // work.
738 volatile bool _initiate_conc_mark_if_possible;
740 // If initiate_conc_mark_if_possible() is set at the beginning of a
741 // pause, it is a suggestion that the pause should start a marking
742 // cycle by doing the initial-mark work. However, it is possible
743 // that the concurrent marking thread is still finishing up the
744 // previous marking cycle (e.g., clearing the next marking
745 // bitmap). If that is the case we cannot start a new cycle and
746 // we'll have to wait for the concurrent marking thread to finish
747 // what it is doing. In this case we will postpone the marking cycle
748 // initiation decision for the next pause. When we eventually decide
749 // to start a cycle, we will set _during_initial_mark_pause which
750 // will stay true until the end of the initial-mark pause and it's
751 // the condition that indicates that a pause is doing the
752 // initial-mark work.
753 volatile bool _during_initial_mark_pause;
755 bool _should_revert_to_full_young_gcs;
756 bool _last_full_young_gc;
758 // This set of variables tracks the collector efficiency, in order to
759 // determine whether we should initiate a new marking.
760 double _cur_mark_stop_world_time_ms;
761 double _mark_remark_start_sec;
762 double _mark_cleanup_start_sec;
763 double _mark_closure_time_ms;
765 // Update the young list target length either by setting it to the
766 // desired fixed value or by calculating it using G1's pause
767 // prediction model. If no rs_lengths parameter is passed, predict
768 // the RS lengths using the prediction model, otherwise use the
769 // given rs_lengths as the prediction.
770 void update_young_list_target_length(size_t rs_lengths = (size_t) -1);
772 // Calculate and return the minimum desired young list target
773 // length. This is the minimum desired young list length according
774 // to the user's inputs.
775 size_t calculate_young_list_desired_min_length(size_t base_min_length);
777 // Calculate and return the maximum desired young list target
778 // length. This is the maximum desired young list length according
779 // to the user's inputs.
780 size_t calculate_young_list_desired_max_length();
782 // Calculate and return the maximum young list target length that
783 // can fit into the pause time goal. The parameters are: rs_lengths
784 // represent the prediction of how large the young RSet lengths will
785 // be, base_min_length is the alreay existing number of regions in
786 // the young list, min_length and max_length are the desired min and
787 // max young list length according to the user's inputs.
788 size_t calculate_young_list_target_length(size_t rs_lengths,
789 size_t base_min_length,
790 size_t desired_min_length,
791 size_t desired_max_length);
793 // Check whether a given young length (young_length) fits into the
794 // given target pause time and whether the prediction for the amount
795 // of objects to be copied for the given length will fit into the
796 // given free space (expressed by base_free_regions). It is used by
797 // calculate_young_list_target_length().
798 bool predict_will_fit(size_t young_length, double base_time_ms,
799 size_t base_free_regions, double target_pause_time_ms);
801 // Count the number of bytes used in the CS.
802 void count_CS_bytes_used();
804 void update_young_list_size_using_newratio(size_t number_of_heap_regions);
806 public:
808 G1CollectorPolicy();
810 virtual G1CollectorPolicy* as_g1_policy() { return this; }
812 virtual CollectorPolicy::Name kind() {
813 return CollectorPolicy::G1CollectorPolicyKind;
814 }
816 // Check the current value of the young list RSet lengths and
817 // compare it against the last prediction. If the current value is
818 // higher, recalculate the young list target length prediction.
819 void revise_young_list_target_length_if_necessary();
821 size_t bytes_in_collection_set() {
822 return _bytes_in_collection_set_before_gc;
823 }
825 unsigned calc_gc_alloc_time_stamp() {
826 return _all_pause_times_ms->num() + 1;
827 }
829 // This should be called after the heap is resized.
830 void record_new_heap_size(size_t new_number_of_regions);
832 public:
834 void init();
836 // Create jstat counters for the policy.
837 virtual void initialize_gc_policy_counters();
839 virtual HeapWord* mem_allocate_work(size_t size,
840 bool is_tlab,
841 bool* gc_overhead_limit_was_exceeded);
843 // This method controls how a collector handles one or more
844 // of its generations being fully allocated.
845 virtual HeapWord* satisfy_failed_allocation(size_t size,
846 bool is_tlab);
848 BarrierSet::Name barrier_set_name() { return BarrierSet::G1SATBCTLogging; }
850 GenRemSet::Name rem_set_name() { return GenRemSet::CardTable; }
852 // The number of collection pauses so far.
853 long n_pauses() const { return _n_pauses; }
855 // Update the heuristic info to record a collection pause of the given
856 // start time, where the given number of bytes were used at the start.
857 // This may involve changing the desired size of a collection set.
859 void record_stop_world_start();
861 void record_collection_pause_start(double start_time_sec, size_t start_used);
863 // Must currently be called while the world is stopped.
864 void record_concurrent_mark_init_end(double
865 mark_init_elapsed_time_ms);
867 void record_mark_closure_time(double mark_closure_time_ms);
869 void record_concurrent_mark_remark_start();
870 void record_concurrent_mark_remark_end();
872 void record_concurrent_mark_cleanup_start();
873 void record_concurrent_mark_cleanup_end();
874 void record_concurrent_mark_cleanup_completed();
876 void record_concurrent_pause();
877 void record_concurrent_pause_end();
879 void record_collection_pause_end();
880 void print_heap_transition();
882 // Record the fact that a full collection occurred.
883 void record_full_collection_start();
884 void record_full_collection_end();
886 void record_gc_worker_start_time(int worker_i, double ms) {
887 _par_last_gc_worker_start_times_ms[worker_i] = ms;
888 }
890 void record_ext_root_scan_time(int worker_i, double ms) {
891 _par_last_ext_root_scan_times_ms[worker_i] = ms;
892 }
894 void record_mark_stack_scan_time(int worker_i, double ms) {
895 _par_last_mark_stack_scan_times_ms[worker_i] = ms;
896 }
898 void record_satb_drain_time(double ms) {
899 assert(_g1->mark_in_progress(), "shouldn't be here otherwise");
900 _cur_satb_drain_time_ms = ms;
901 }
903 void record_satb_drain_processed_buffers(int processed_buffers) {
904 assert(_g1->mark_in_progress(), "shouldn't be here otherwise");
905 _last_satb_drain_processed_buffers = processed_buffers;
906 }
908 void record_mod_union_time(double ms) {
909 _all_mod_union_times_ms->add(ms);
910 }
912 void record_update_rs_time(int thread, double ms) {
913 _par_last_update_rs_times_ms[thread] = ms;
914 }
916 void record_update_rs_processed_buffers (int thread,
917 double processed_buffers) {
918 _par_last_update_rs_processed_buffers[thread] = processed_buffers;
919 }
921 void record_scan_rs_time(int thread, double ms) {
922 _par_last_scan_rs_times_ms[thread] = ms;
923 }
925 void reset_obj_copy_time(int thread) {
926 _par_last_obj_copy_times_ms[thread] = 0.0;
927 }
929 void reset_obj_copy_time() {
930 reset_obj_copy_time(0);
931 }
933 void record_obj_copy_time(int thread, double ms) {
934 _par_last_obj_copy_times_ms[thread] += ms;
935 }
937 void record_termination(int thread, double ms, size_t attempts) {
938 _par_last_termination_times_ms[thread] = ms;
939 _par_last_termination_attempts[thread] = (double) attempts;
940 }
942 void record_gc_worker_end_time(int worker_i, double ms) {
943 _par_last_gc_worker_end_times_ms[worker_i] = ms;
944 }
946 void record_pause_time_ms(double ms) {
947 _last_pause_time_ms = ms;
948 }
950 void record_clear_ct_time(double ms) {
951 _cur_clear_ct_time_ms = ms;
952 }
954 void record_par_time(double ms) {
955 _cur_collection_par_time_ms = ms;
956 }
958 void record_aux_start_time(int i) {
959 guarantee(i < _aux_num, "should be within range");
960 _cur_aux_start_times_ms[i] = os::elapsedTime() * 1000.0;
961 }
963 void record_aux_end_time(int i) {
964 guarantee(i < _aux_num, "should be within range");
965 double ms = os::elapsedTime() * 1000.0 - _cur_aux_start_times_ms[i];
966 _cur_aux_times_set[i] = true;
967 _cur_aux_times_ms[i] += ms;
968 }
970 void record_ref_proc_time(double ms) {
971 _cur_ref_proc_time_ms = ms;
972 }
974 void record_ref_enq_time(double ms) {
975 _cur_ref_enq_time_ms = ms;
976 }
978 #ifndef PRODUCT
979 void record_cc_clear_time(double ms) {
980 if (_min_clear_cc_time_ms < 0.0 || ms <= _min_clear_cc_time_ms)
981 _min_clear_cc_time_ms = ms;
982 if (_max_clear_cc_time_ms < 0.0 || ms >= _max_clear_cc_time_ms)
983 _max_clear_cc_time_ms = ms;
984 _cur_clear_cc_time_ms = ms;
985 _cum_clear_cc_time_ms += ms;
986 _num_cc_clears++;
987 }
988 #endif
990 // Record how much space we copied during a GC. This is typically
991 // called when a GC alloc region is being retired.
992 void record_bytes_copied_during_gc(size_t bytes) {
993 _bytes_copied_during_gc += bytes;
994 }
996 // The amount of space we copied during a GC.
997 size_t bytes_copied_during_gc() {
998 return _bytes_copied_during_gc;
999 }
1001 // Choose a new collection set. Marks the chosen regions as being
1002 // "in_collection_set", and links them together. The head and number of
1003 // the collection set are available via access methods.
1004 void choose_collection_set(double target_pause_time_ms);
1006 // The head of the list (via "next_in_collection_set()") representing the
1007 // current collection set.
1008 HeapRegion* collection_set() { return _collection_set; }
1010 void clear_collection_set() { _collection_set = NULL; }
1012 // The number of elements in the current collection set.
1013 size_t collection_set_size() { return _collection_set_size; }
1015 // Add "hr" to the CS.
1016 void add_to_collection_set(HeapRegion* hr);
1018 // Incremental CSet Support
1020 // The head of the incrementally built collection set.
1021 HeapRegion* inc_cset_head() { return _inc_cset_head; }
1023 // The tail of the incrementally built collection set.
1024 HeapRegion* inc_set_tail() { return _inc_cset_tail; }
1026 // The number of elements in the incrementally built collection set.
1027 size_t inc_cset_size() { return _inc_cset_size; }
1029 // Initialize incremental collection set info.
1030 void start_incremental_cset_building();
1032 void clear_incremental_cset() {
1033 _inc_cset_head = NULL;
1034 _inc_cset_tail = NULL;
1035 }
1037 // Stop adding regions to the incremental collection set
1038 void stop_incremental_cset_building() { _inc_cset_build_state = Inactive; }
1040 // Add/remove information about hr to the aggregated information
1041 // for the incrementally built collection set.
1042 void add_to_incremental_cset_info(HeapRegion* hr, size_t rs_length);
1043 void remove_from_incremental_cset_info(HeapRegion* hr);
1045 // Update information about hr in the aggregated information for
1046 // the incrementally built collection set.
1047 void update_incremental_cset_info(HeapRegion* hr, size_t new_rs_length);
1049 private:
1050 // Update the incremental cset information when adding a region
1051 // (should not be called directly).
1052 void add_region_to_incremental_cset_common(HeapRegion* hr);
1054 public:
1055 // Add hr to the LHS of the incremental collection set.
1056 void add_region_to_incremental_cset_lhs(HeapRegion* hr);
1058 // Add hr to the RHS of the incremental collection set.
1059 void add_region_to_incremental_cset_rhs(HeapRegion* hr);
1061 #ifndef PRODUCT
1062 void print_collection_set(HeapRegion* list_head, outputStream* st);
1063 #endif // !PRODUCT
1065 bool initiate_conc_mark_if_possible() { return _initiate_conc_mark_if_possible; }
1066 void set_initiate_conc_mark_if_possible() { _initiate_conc_mark_if_possible = true; }
1067 void clear_initiate_conc_mark_if_possible() { _initiate_conc_mark_if_possible = false; }
1069 bool during_initial_mark_pause() { return _during_initial_mark_pause; }
1070 void set_during_initial_mark_pause() { _during_initial_mark_pause = true; }
1071 void clear_during_initial_mark_pause(){ _during_initial_mark_pause = false; }
1073 // This sets the initiate_conc_mark_if_possible() flag to start a
1074 // new cycle, as long as we are not already in one. It's best if it
1075 // is called during a safepoint when the test whether a cycle is in
1076 // progress or not is stable.
1077 bool force_initial_mark_if_outside_cycle(GCCause::Cause gc_cause);
1079 // This is called at the very beginning of an evacuation pause (it
1080 // has to be the first thing that the pause does). If
1081 // initiate_conc_mark_if_possible() is true, and the concurrent
1082 // marking thread has completed its work during the previous cycle,
1083 // it will set during_initial_mark_pause() to so that the pause does
1084 // the initial-mark work and start a marking cycle.
1085 void decide_on_conc_mark_initiation();
1087 // If an expansion would be appropriate, because recent GC overhead had
1088 // exceeded the desired limit, return an amount to expand by.
1089 size_t expansion_amount();
1091 #ifndef PRODUCT
1092 // Check any appropriate marked bytes info, asserting false if
1093 // something's wrong, else returning "true".
1094 bool assertMarkedBytesDataOK();
1095 #endif
1097 // Print tracing information.
1098 void print_tracing_info() const;
1100 // Print stats on young survival ratio
1101 void print_yg_surv_rate_info() const;
1103 void finished_recalculating_age_indexes(bool is_survivors) {
1104 if (is_survivors) {
1105 _survivor_surv_rate_group->finished_recalculating_age_indexes();
1106 } else {
1107 _short_lived_surv_rate_group->finished_recalculating_age_indexes();
1108 }
1109 // do that for any other surv rate groups
1110 }
1112 bool is_young_list_full() {
1113 size_t young_list_length = _g1->young_list()->length();
1114 size_t young_list_target_length = _young_list_target_length;
1115 return young_list_length >= young_list_target_length;
1116 }
1118 bool can_expand_young_list() {
1119 size_t young_list_length = _g1->young_list()->length();
1120 size_t young_list_max_length = _young_list_max_length;
1121 return young_list_length < young_list_max_length;
1122 }
1124 size_t young_list_max_length() {
1125 return _young_list_max_length;
1126 }
1128 void update_region_num(bool young);
1130 bool full_young_gcs() {
1131 return _full_young_gcs;
1132 }
1133 void set_full_young_gcs(bool full_young_gcs) {
1134 _full_young_gcs = full_young_gcs;
1135 }
1137 bool adaptive_young_list_length() {
1138 return _adaptive_young_list_length;
1139 }
1140 void set_adaptive_young_list_length(bool adaptive_young_list_length) {
1141 _adaptive_young_list_length = adaptive_young_list_length;
1142 }
1144 inline double get_gc_eff_factor() {
1145 double ratio = _known_garbage_ratio;
1147 double square = ratio * ratio;
1148 // square = square * square;
1149 double ret = square * 9.0 + 1.0;
1150 #if 0
1151 gclog_or_tty->print_cr("ratio = %1.2lf, ret = %1.2lf", ratio, ret);
1152 #endif // 0
1153 guarantee(0.0 <= ret && ret < 10.0, "invariant!");
1154 return ret;
1155 }
1157 private:
1158 //
1159 // Survivor regions policy.
1160 //
1162 // Current tenuring threshold, set to 0 if the collector reaches the
1163 // maximum amount of suvivors regions.
1164 int _tenuring_threshold;
1166 // The limit on the number of regions allocated for survivors.
1167 size_t _max_survivor_regions;
1169 // For reporting purposes.
1170 size_t _eden_bytes_before_gc;
1171 size_t _survivor_bytes_before_gc;
1172 size_t _capacity_before_gc;
1174 // The amount of survor regions after a collection.
1175 size_t _recorded_survivor_regions;
1176 // List of survivor regions.
1177 HeapRegion* _recorded_survivor_head;
1178 HeapRegion* _recorded_survivor_tail;
1180 ageTable _survivors_age_table;
1182 public:
1184 inline GCAllocPurpose
1185 evacuation_destination(HeapRegion* src_region, int age, size_t word_sz) {
1186 if (age < _tenuring_threshold && src_region->is_young()) {
1187 return GCAllocForSurvived;
1188 } else {
1189 return GCAllocForTenured;
1190 }
1191 }
1193 inline bool track_object_age(GCAllocPurpose purpose) {
1194 return purpose == GCAllocForSurvived;
1195 }
1197 static const size_t REGIONS_UNLIMITED = ~(size_t)0;
1199 size_t max_regions(int purpose);
1201 // The limit on regions for a particular purpose is reached.
1202 void note_alloc_region_limit_reached(int purpose) {
1203 if (purpose == GCAllocForSurvived) {
1204 _tenuring_threshold = 0;
1205 }
1206 }
1208 void note_start_adding_survivor_regions() {
1209 _survivor_surv_rate_group->start_adding_regions();
1210 }
1212 void note_stop_adding_survivor_regions() {
1213 _survivor_surv_rate_group->stop_adding_regions();
1214 }
1216 void record_survivor_regions(size_t regions,
1217 HeapRegion* head,
1218 HeapRegion* tail) {
1219 _recorded_survivor_regions = regions;
1220 _recorded_survivor_head = head;
1221 _recorded_survivor_tail = tail;
1222 }
1224 size_t recorded_survivor_regions() {
1225 return _recorded_survivor_regions;
1226 }
1228 void record_thread_age_table(ageTable* age_table)
1229 {
1230 _survivors_age_table.merge_par(age_table);
1231 }
1233 void update_max_gc_locker_expansion();
1235 // Calculates survivor space parameters.
1236 void update_survivors_policy();
1238 };
1240 // This should move to some place more general...
1242 // If we have "n" measurements, and we've kept track of their "sum" and the
1243 // "sum_of_squares" of the measurements, this returns the variance of the
1244 // sequence.
1245 inline double variance(int n, double sum_of_squares, double sum) {
1246 double n_d = (double)n;
1247 double avg = sum/n_d;
1248 return (sum_of_squares - 2.0 * avg * sum + n_d * avg * avg) / n_d;
1249 }
1251 #endif // SHARE_VM_GC_IMPLEMENTATION_G1_G1COLLECTORPOLICY_HPP