Tue, 30 Mar 2010 15:43:03 -0400
6937142: G1: improvements to debugging output (S-M)
Summary: Various fixes to the G1 debugging output.
Reviewed-by: johnc, iveresov
1 /*
2 * Copyright 2001-2009 Sun Microsystems, Inc. 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 Sun Microsystems, Inc., 4150 Network Circle, Santa Clara,
20 * CA 95054 USA or visit www.sun.com if you need additional information or
21 * have any questions.
22 *
23 */
25 #include "incls/_precompiled.incl"
26 #include "incls/_g1CollectorPolicy.cpp.incl"
28 #define PREDICTIONS_VERBOSE 0
30 // <NEW PREDICTION>
32 // Different defaults for different number of GC threads
33 // They were chosen by running GCOld and SPECjbb on debris with different
34 // numbers of GC threads and choosing them based on the results
36 // all the same
37 static double rs_length_diff_defaults[] = {
38 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0
39 };
41 static double cost_per_card_ms_defaults[] = {
42 0.01, 0.005, 0.005, 0.003, 0.003, 0.002, 0.002, 0.0015
43 };
45 static double cost_per_scan_only_region_ms_defaults[] = {
46 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0
47 };
49 // all the same
50 static double fully_young_cards_per_entry_ratio_defaults[] = {
51 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0
52 };
54 static double cost_per_entry_ms_defaults[] = {
55 0.015, 0.01, 0.01, 0.008, 0.008, 0.0055, 0.0055, 0.005
56 };
58 static double cost_per_byte_ms_defaults[] = {
59 0.00006, 0.00003, 0.00003, 0.000015, 0.000015, 0.00001, 0.00001, 0.000009
60 };
62 // these should be pretty consistent
63 static double constant_other_time_ms_defaults[] = {
64 5.0, 5.0, 5.0, 5.0, 5.0, 5.0, 5.0, 5.0
65 };
68 static double young_other_cost_per_region_ms_defaults[] = {
69 0.3, 0.2, 0.2, 0.15, 0.15, 0.12, 0.12, 0.1
70 };
72 static double non_young_other_cost_per_region_ms_defaults[] = {
73 1.0, 0.7, 0.7, 0.5, 0.5, 0.42, 0.42, 0.30
74 };
76 // </NEW PREDICTION>
78 G1CollectorPolicy::G1CollectorPolicy() :
79 _parallel_gc_threads((ParallelGCThreads > 0) ? ParallelGCThreads : 1),
80 _n_pauses(0),
81 _recent_CH_strong_roots_times_ms(new TruncatedSeq(NumPrevPausesForHeuristics)),
82 _recent_G1_strong_roots_times_ms(new TruncatedSeq(NumPrevPausesForHeuristics)),
83 _recent_evac_times_ms(new TruncatedSeq(NumPrevPausesForHeuristics)),
84 _recent_pause_times_ms(new TruncatedSeq(NumPrevPausesForHeuristics)),
85 _recent_rs_sizes(new TruncatedSeq(NumPrevPausesForHeuristics)),
86 _recent_gc_times_ms(new TruncatedSeq(NumPrevPausesForHeuristics)),
87 _all_pause_times_ms(new NumberSeq()),
88 _stop_world_start(0.0),
89 _all_stop_world_times_ms(new NumberSeq()),
90 _all_yield_times_ms(new NumberSeq()),
92 _all_mod_union_times_ms(new NumberSeq()),
94 _summary(new Summary()),
95 _abandoned_summary(new AbandonedSummary()),
97 #ifndef PRODUCT
98 _cur_clear_ct_time_ms(0.0),
99 _min_clear_cc_time_ms(-1.0),
100 _max_clear_cc_time_ms(-1.0),
101 _cur_clear_cc_time_ms(0.0),
102 _cum_clear_cc_time_ms(0.0),
103 _num_cc_clears(0L),
104 #endif
106 _region_num_young(0),
107 _region_num_tenured(0),
108 _prev_region_num_young(0),
109 _prev_region_num_tenured(0),
111 _aux_num(10),
112 _all_aux_times_ms(new NumberSeq[_aux_num]),
113 _cur_aux_start_times_ms(new double[_aux_num]),
114 _cur_aux_times_ms(new double[_aux_num]),
115 _cur_aux_times_set(new bool[_aux_num]),
117 _concurrent_mark_init_times_ms(new TruncatedSeq(NumPrevPausesForHeuristics)),
118 _concurrent_mark_remark_times_ms(new TruncatedSeq(NumPrevPausesForHeuristics)),
119 _concurrent_mark_cleanup_times_ms(new TruncatedSeq(NumPrevPausesForHeuristics)),
121 // <NEW PREDICTION>
123 _alloc_rate_ms_seq(new TruncatedSeq(TruncatedSeqLength)),
124 _prev_collection_pause_end_ms(0.0),
125 _pending_card_diff_seq(new TruncatedSeq(TruncatedSeqLength)),
126 _rs_length_diff_seq(new TruncatedSeq(TruncatedSeqLength)),
127 _cost_per_card_ms_seq(new TruncatedSeq(TruncatedSeqLength)),
128 _cost_per_scan_only_region_ms_seq(new TruncatedSeq(TruncatedSeqLength)),
129 _fully_young_cards_per_entry_ratio_seq(new TruncatedSeq(TruncatedSeqLength)),
130 _partially_young_cards_per_entry_ratio_seq(
131 new TruncatedSeq(TruncatedSeqLength)),
132 _cost_per_entry_ms_seq(new TruncatedSeq(TruncatedSeqLength)),
133 _partially_young_cost_per_entry_ms_seq(new TruncatedSeq(TruncatedSeqLength)),
134 _cost_per_byte_ms_seq(new TruncatedSeq(TruncatedSeqLength)),
135 _cost_per_byte_ms_during_cm_seq(new TruncatedSeq(TruncatedSeqLength)),
136 _cost_per_scan_only_region_ms_during_cm_seq(new TruncatedSeq(TruncatedSeqLength)),
137 _constant_other_time_ms_seq(new TruncatedSeq(TruncatedSeqLength)),
138 _young_other_cost_per_region_ms_seq(new TruncatedSeq(TruncatedSeqLength)),
139 _non_young_other_cost_per_region_ms_seq(
140 new TruncatedSeq(TruncatedSeqLength)),
142 _pending_cards_seq(new TruncatedSeq(TruncatedSeqLength)),
143 _scanned_cards_seq(new TruncatedSeq(TruncatedSeqLength)),
144 _rs_lengths_seq(new TruncatedSeq(TruncatedSeqLength)),
146 _pause_time_target_ms((double) MaxGCPauseMillis),
148 // </NEW PREDICTION>
150 _in_young_gc_mode(false),
151 _full_young_gcs(true),
152 _full_young_pause_num(0),
153 _partial_young_pause_num(0),
155 _during_marking(false),
156 _in_marking_window(false),
157 _in_marking_window_im(false),
159 _known_garbage_ratio(0.0),
160 _known_garbage_bytes(0),
162 _young_gc_eff_seq(new TruncatedSeq(TruncatedSeqLength)),
163 _target_pause_time_ms(-1.0),
165 _recent_prev_end_times_for_all_gcs_sec(new TruncatedSeq(NumPrevPausesForHeuristics)),
167 _recent_CS_bytes_used_before(new TruncatedSeq(NumPrevPausesForHeuristics)),
168 _recent_CS_bytes_surviving(new TruncatedSeq(NumPrevPausesForHeuristics)),
170 _recent_avg_pause_time_ratio(0.0),
171 _num_markings(0),
172 _n_marks(0),
173 _n_pauses_at_mark_end(0),
175 _all_full_gc_times_ms(new NumberSeq()),
177 // G1PausesBtwnConcMark defaults to -1
178 // so the hack is to do the cast QQQ FIXME
179 _pauses_btwn_concurrent_mark((size_t)G1PausesBtwnConcMark),
180 _n_marks_since_last_pause(0),
181 _initiate_conc_mark_if_possible(false),
182 _during_initial_mark_pause(false),
183 _should_revert_to_full_young_gcs(false),
184 _last_full_young_gc(false),
186 _prev_collection_pause_used_at_end_bytes(0),
188 _collection_set(NULL),
189 #ifdef _MSC_VER // the use of 'this' below gets a warning, make it go away
190 #pragma warning( disable:4355 ) // 'this' : used in base member initializer list
191 #endif // _MSC_VER
193 _short_lived_surv_rate_group(new SurvRateGroup(this, "Short Lived",
194 G1YoungSurvRateNumRegionsSummary)),
195 _survivor_surv_rate_group(new SurvRateGroup(this, "Survivor",
196 G1YoungSurvRateNumRegionsSummary)),
197 // add here any more surv rate groups
198 _recorded_survivor_regions(0),
199 _recorded_survivor_head(NULL),
200 _recorded_survivor_tail(NULL),
201 _survivors_age_table(true),
203 _gc_overhead_perc(0.0)
205 {
206 // Set up the region size and associated fields. Given that the
207 // policy is created before the heap, we have to set this up here,
208 // so it's done as soon as possible.
209 HeapRegion::setup_heap_region_size(Arguments::min_heap_size());
210 HeapRegionRemSet::setup_remset_size();
212 _recent_prev_end_times_for_all_gcs_sec->add(os::elapsedTime());
213 _prev_collection_pause_end_ms = os::elapsedTime() * 1000.0;
215 _par_last_ext_root_scan_times_ms = new double[_parallel_gc_threads];
216 _par_last_mark_stack_scan_times_ms = new double[_parallel_gc_threads];
217 _par_last_scan_only_times_ms = new double[_parallel_gc_threads];
218 _par_last_scan_only_regions_scanned = new double[_parallel_gc_threads];
220 _par_last_update_rs_start_times_ms = new double[_parallel_gc_threads];
221 _par_last_update_rs_times_ms = new double[_parallel_gc_threads];
222 _par_last_update_rs_processed_buffers = new double[_parallel_gc_threads];
224 _par_last_scan_rs_start_times_ms = new double[_parallel_gc_threads];
225 _par_last_scan_rs_times_ms = new double[_parallel_gc_threads];
226 _par_last_scan_new_refs_times_ms = new double[_parallel_gc_threads];
228 _par_last_obj_copy_times_ms = new double[_parallel_gc_threads];
230 _par_last_termination_times_ms = new double[_parallel_gc_threads];
232 // start conservatively
233 _expensive_region_limit_ms = 0.5 * (double) MaxGCPauseMillis;
235 // <NEW PREDICTION>
237 int index;
238 if (ParallelGCThreads == 0)
239 index = 0;
240 else if (ParallelGCThreads > 8)
241 index = 7;
242 else
243 index = ParallelGCThreads - 1;
245 _pending_card_diff_seq->add(0.0);
246 _rs_length_diff_seq->add(rs_length_diff_defaults[index]);
247 _cost_per_card_ms_seq->add(cost_per_card_ms_defaults[index]);
248 _cost_per_scan_only_region_ms_seq->add(
249 cost_per_scan_only_region_ms_defaults[index]);
250 _fully_young_cards_per_entry_ratio_seq->add(
251 fully_young_cards_per_entry_ratio_defaults[index]);
252 _cost_per_entry_ms_seq->add(cost_per_entry_ms_defaults[index]);
253 _cost_per_byte_ms_seq->add(cost_per_byte_ms_defaults[index]);
254 _constant_other_time_ms_seq->add(constant_other_time_ms_defaults[index]);
255 _young_other_cost_per_region_ms_seq->add(
256 young_other_cost_per_region_ms_defaults[index]);
257 _non_young_other_cost_per_region_ms_seq->add(
258 non_young_other_cost_per_region_ms_defaults[index]);
260 // </NEW PREDICTION>
262 double time_slice = (double) GCPauseIntervalMillis / 1000.0;
263 double max_gc_time = (double) MaxGCPauseMillis / 1000.0;
264 guarantee(max_gc_time < time_slice,
265 "Max GC time should not be greater than the time slice");
266 _mmu_tracker = new G1MMUTrackerQueue(time_slice, max_gc_time);
267 _sigma = (double) G1ConfidencePercent / 100.0;
269 // start conservatively (around 50ms is about right)
270 _concurrent_mark_init_times_ms->add(0.05);
271 _concurrent_mark_remark_times_ms->add(0.05);
272 _concurrent_mark_cleanup_times_ms->add(0.20);
273 _tenuring_threshold = MaxTenuringThreshold;
275 // if G1FixedSurvivorSpaceSize is 0 which means the size is not
276 // fixed, then _max_survivor_regions will be calculated at
277 // calculate_young_list_target_config during initialization
278 _max_survivor_regions = G1FixedSurvivorSpaceSize / HeapRegion::GrainBytes;
280 assert(GCTimeRatio > 0,
281 "we should have set it to a default value set_g1_gc_flags() "
282 "if a user set it to 0");
283 _gc_overhead_perc = 100.0 * (1.0 / (1.0 + GCTimeRatio));
285 initialize_all();
286 }
288 // Increment "i", mod "len"
289 static void inc_mod(int& i, int len) {
290 i++; if (i == len) i = 0;
291 }
293 void G1CollectorPolicy::initialize_flags() {
294 set_min_alignment(HeapRegion::GrainBytes);
295 set_max_alignment(GenRemSet::max_alignment_constraint(rem_set_name()));
296 if (SurvivorRatio < 1) {
297 vm_exit_during_initialization("Invalid survivor ratio specified");
298 }
299 CollectorPolicy::initialize_flags();
300 }
302 // The easiest way to deal with the parsing of the NewSize /
303 // MaxNewSize / etc. parameteres is to re-use the code in the
304 // TwoGenerationCollectorPolicy class. This is similar to what
305 // ParallelScavenge does with its GenerationSizer class (see
306 // ParallelScavengeHeap::initialize()). We might change this in the
307 // future, but it's a good start.
308 class G1YoungGenSizer : public TwoGenerationCollectorPolicy {
309 size_t size_to_region_num(size_t byte_size) {
310 return MAX2((size_t) 1, byte_size / HeapRegion::GrainBytes);
311 }
313 public:
314 G1YoungGenSizer() {
315 initialize_flags();
316 initialize_size_info();
317 }
319 size_t min_young_region_num() {
320 return size_to_region_num(_min_gen0_size);
321 }
322 size_t initial_young_region_num() {
323 return size_to_region_num(_initial_gen0_size);
324 }
325 size_t max_young_region_num() {
326 return size_to_region_num(_max_gen0_size);
327 }
328 };
330 void G1CollectorPolicy::init() {
331 // Set aside an initial future to_space.
332 _g1 = G1CollectedHeap::heap();
334 assert(Heap_lock->owned_by_self(), "Locking discipline.");
336 initialize_gc_policy_counters();
338 if (G1Gen) {
339 _in_young_gc_mode = true;
341 G1YoungGenSizer sizer;
342 size_t initial_region_num = sizer.initial_young_region_num();
344 if (UseAdaptiveSizePolicy) {
345 set_adaptive_young_list_length(true);
346 _young_list_fixed_length = 0;
347 } else {
348 set_adaptive_young_list_length(false);
349 _young_list_fixed_length = initial_region_num;
350 }
351 _free_regions_at_end_of_collection = _g1->free_regions();
352 _scan_only_regions_at_end_of_collection = 0;
353 calculate_young_list_min_length();
354 guarantee( _young_list_min_length == 0, "invariant, not enough info" );
355 calculate_young_list_target_config();
356 } else {
357 _young_list_fixed_length = 0;
358 _in_young_gc_mode = false;
359 }
360 }
362 // Create the jstat counters for the policy.
363 void G1CollectorPolicy::initialize_gc_policy_counters()
364 {
365 _gc_policy_counters = new GCPolicyCounters("GarbageFirst", 1, 2 + G1Gen);
366 }
368 void G1CollectorPolicy::calculate_young_list_min_length() {
369 _young_list_min_length = 0;
371 if (!adaptive_young_list_length())
372 return;
374 if (_alloc_rate_ms_seq->num() > 3) {
375 double now_sec = os::elapsedTime();
376 double when_ms = _mmu_tracker->when_max_gc_sec(now_sec) * 1000.0;
377 double alloc_rate_ms = predict_alloc_rate_ms();
378 int min_regions = (int) ceil(alloc_rate_ms * when_ms);
379 int current_region_num = (int) _g1->young_list_length();
380 _young_list_min_length = min_regions + current_region_num;
381 }
382 }
384 void G1CollectorPolicy::calculate_young_list_target_config() {
385 if (adaptive_young_list_length()) {
386 size_t rs_lengths = (size_t) get_new_prediction(_rs_lengths_seq);
387 calculate_young_list_target_config(rs_lengths);
388 } else {
389 if (full_young_gcs())
390 _young_list_target_length = _young_list_fixed_length;
391 else
392 _young_list_target_length = _young_list_fixed_length / 2;
393 _young_list_target_length = MAX2(_young_list_target_length, (size_t)1);
394 size_t so_length = calculate_optimal_so_length(_young_list_target_length);
395 guarantee( so_length < _young_list_target_length, "invariant" );
396 _young_list_so_prefix_length = so_length;
397 }
398 calculate_survivors_policy();
399 }
401 // This method calculate the optimal scan-only set for a fixed young
402 // gen size. I couldn't work out how to reuse the more elaborate one,
403 // i.e. calculate_young_list_target_config(rs_length), as the loops are
404 // fundamentally different (the other one finds a config for different
405 // S-O lengths, whereas here we need to do the opposite).
406 size_t G1CollectorPolicy::calculate_optimal_so_length(
407 size_t young_list_length) {
408 if (!G1UseScanOnlyPrefix)
409 return 0;
411 if (_all_pause_times_ms->num() < 3) {
412 // we won't use a scan-only set at the beginning to allow the rest
413 // of the predictors to warm up
414 return 0;
415 }
417 if (_cost_per_scan_only_region_ms_seq->num() < 3) {
418 // then, we'll only set the S-O set to 1 for a little bit of time,
419 // to get enough information on the scanning cost
420 return 1;
421 }
423 size_t pending_cards = (size_t) get_new_prediction(_pending_cards_seq);
424 size_t rs_lengths = (size_t) get_new_prediction(_rs_lengths_seq);
425 size_t adj_rs_lengths = rs_lengths + predict_rs_length_diff();
426 size_t scanned_cards;
427 if (full_young_gcs())
428 scanned_cards = predict_young_card_num(adj_rs_lengths);
429 else
430 scanned_cards = predict_non_young_card_num(adj_rs_lengths);
431 double base_time_ms = predict_base_elapsed_time_ms(pending_cards,
432 scanned_cards);
434 size_t so_length = 0;
435 double max_gc_eff = 0.0;
436 for (size_t i = 0; i < young_list_length; ++i) {
437 double gc_eff = 0.0;
438 double pause_time_ms = 0.0;
439 predict_gc_eff(young_list_length, i, base_time_ms,
440 &gc_eff, &pause_time_ms);
441 if (gc_eff > max_gc_eff) {
442 max_gc_eff = gc_eff;
443 so_length = i;
444 }
445 }
447 // set it to 95% of the optimal to make sure we sample the "area"
448 // around the optimal length to get up-to-date survival rate data
449 return so_length * 950 / 1000;
450 }
452 // This is a really cool piece of code! It finds the best
453 // target configuration (young length / scan-only prefix length) so
454 // that GC efficiency is maximized and that we also meet a pause
455 // time. It's a triple nested loop. These loops are explained below
456 // from the inside-out :-)
457 //
458 // (a) The innermost loop will try to find the optimal young length
459 // for a fixed S-O length. It uses a binary search to speed up the
460 // process. We assume that, for a fixed S-O length, as we add more
461 // young regions to the CSet, the GC efficiency will only go up (I'll
462 // skip the proof). So, using a binary search to optimize this process
463 // makes perfect sense.
464 //
465 // (b) The middle loop will fix the S-O length before calling the
466 // innermost one. It will vary it between two parameters, increasing
467 // it by a given increment.
468 //
469 // (c) The outermost loop will call the middle loop three times.
470 // (1) The first time it will explore all possible S-O length values
471 // from 0 to as large as it can get, using a coarse increment (to
472 // quickly "home in" to where the optimal seems to be).
473 // (2) The second time it will explore the values around the optimal
474 // that was found by the first iteration using a fine increment.
475 // (3) Once the optimal config has been determined by the second
476 // iteration, we'll redo the calculation, but setting the S-O length
477 // to 95% of the optimal to make sure we sample the "area"
478 // around the optimal length to get up-to-date survival rate data
479 //
480 // Termination conditions for the iterations are several: the pause
481 // time is over the limit, we do not have enough to-space, etc.
483 void G1CollectorPolicy::calculate_young_list_target_config(size_t rs_lengths) {
484 guarantee( adaptive_young_list_length(), "pre-condition" );
486 double start_time_sec = os::elapsedTime();
487 size_t min_reserve_perc = MAX2((size_t)2, (size_t)G1ReservePercent);
488 min_reserve_perc = MIN2((size_t) 50, min_reserve_perc);
489 size_t reserve_regions =
490 (size_t) ((double) min_reserve_perc * (double) _g1->n_regions() / 100.0);
492 if (full_young_gcs() && _free_regions_at_end_of_collection > 0) {
493 // we are in fully-young mode and there are free regions in the heap
495 double survivor_regions_evac_time =
496 predict_survivor_regions_evac_time();
498 size_t min_so_length = 0;
499 size_t max_so_length = 0;
501 if (G1UseScanOnlyPrefix) {
502 if (_all_pause_times_ms->num() < 3) {
503 // we won't use a scan-only set at the beginning to allow the rest
504 // of the predictors to warm up
505 min_so_length = 0;
506 max_so_length = 0;
507 } else if (_cost_per_scan_only_region_ms_seq->num() < 3) {
508 // then, we'll only set the S-O set to 1 for a little bit of time,
509 // to get enough information on the scanning cost
510 min_so_length = 1;
511 max_so_length = 1;
512 } else if (_in_marking_window || _last_full_young_gc) {
513 // no S-O prefix during a marking phase either, as at the end
514 // of the marking phase we'll have to use a very small young
515 // length target to fill up the rest of the CSet with
516 // non-young regions and, if we have lots of scan-only regions
517 // left-over, we will not be able to add any more non-young
518 // regions.
519 min_so_length = 0;
520 max_so_length = 0;
521 } else {
522 // this is the common case; we'll never reach the maximum, we
523 // one of the end conditions will fire well before that
524 // (hopefully!)
525 min_so_length = 0;
526 max_so_length = _free_regions_at_end_of_collection - 1;
527 }
528 } else {
529 // no S-O prefix, as the switch is not set, but we still need to
530 // do one iteration to calculate the best young target that
531 // meets the pause time; this way we reuse the same code instead
532 // of replicating it
533 min_so_length = 0;
534 max_so_length = 0;
535 }
537 double target_pause_time_ms = _mmu_tracker->max_gc_time() * 1000.0;
538 size_t pending_cards = (size_t) get_new_prediction(_pending_cards_seq);
539 size_t adj_rs_lengths = rs_lengths + predict_rs_length_diff();
540 size_t scanned_cards;
541 if (full_young_gcs())
542 scanned_cards = predict_young_card_num(adj_rs_lengths);
543 else
544 scanned_cards = predict_non_young_card_num(adj_rs_lengths);
545 // calculate this once, so that we don't have to recalculate it in
546 // the innermost loop
547 double base_time_ms = predict_base_elapsed_time_ms(pending_cards, scanned_cards)
548 + survivor_regions_evac_time;
549 // the result
550 size_t final_young_length = 0;
551 size_t final_so_length = 0;
552 double final_gc_eff = 0.0;
553 // we'll also keep track of how many times we go into the inner loop
554 // this is for profiling reasons
555 size_t calculations = 0;
557 // this determines which of the three iterations the outer loop is in
558 typedef enum {
559 pass_type_coarse,
560 pass_type_fine,
561 pass_type_final
562 } pass_type_t;
564 // range of the outer loop's iteration
565 size_t from_so_length = min_so_length;
566 size_t to_so_length = max_so_length;
567 guarantee( from_so_length <= to_so_length, "invariant" );
569 // this will keep the S-O length that's found by the second
570 // iteration of the outer loop; we'll keep it just in case the third
571 // iteration fails to find something
572 size_t fine_so_length = 0;
574 // the increment step for the coarse (first) iteration
575 size_t so_coarse_increments = 5;
577 // the common case, we'll start with the coarse iteration
578 pass_type_t pass = pass_type_coarse;
579 size_t so_length_incr = so_coarse_increments;
581 if (from_so_length == to_so_length) {
582 // not point in doing the coarse iteration, we'll go directly into
583 // the fine one (we essentially trying to find the optimal young
584 // length for a fixed S-O length).
585 so_length_incr = 1;
586 pass = pass_type_final;
587 } else if (to_so_length - from_so_length < 3 * so_coarse_increments) {
588 // again, the range is too short so no point in foind the coarse
589 // iteration either
590 so_length_incr = 1;
591 pass = pass_type_fine;
592 }
594 bool done = false;
595 // this is the outermost loop
596 while (!done) {
597 #ifdef TRACE_CALC_YOUNG_CONFIG
598 // leave this in for debugging, just in case
599 gclog_or_tty->print_cr("searching between " SIZE_FORMAT " and " SIZE_FORMAT
600 ", incr " SIZE_FORMAT ", pass %s",
601 from_so_length, to_so_length, so_length_incr,
602 (pass == pass_type_coarse) ? "coarse" :
603 (pass == pass_type_fine) ? "fine" : "final");
604 #endif // TRACE_CALC_YOUNG_CONFIG
606 size_t so_length = from_so_length;
607 size_t init_free_regions =
608 MAX2((size_t)0,
609 _free_regions_at_end_of_collection +
610 _scan_only_regions_at_end_of_collection - reserve_regions);
612 // this determines whether a configuration was found
613 bool gc_eff_set = false;
614 // this is the middle loop
615 while (so_length <= to_so_length) {
616 // base time, which excludes region-related time; again we
617 // calculate it once to avoid recalculating it in the
618 // innermost loop
619 double base_time_with_so_ms =
620 base_time_ms + predict_scan_only_time_ms(so_length);
621 // it's already over the pause target, go around
622 if (base_time_with_so_ms > target_pause_time_ms)
623 break;
625 size_t starting_young_length = so_length+1;
627 // we make sure that the short young length that makes sense
628 // (one more than the S-O length) is feasible
629 size_t min_young_length = starting_young_length;
630 double min_gc_eff;
631 bool min_ok;
632 ++calculations;
633 min_ok = predict_gc_eff(min_young_length, so_length,
634 base_time_with_so_ms,
635 init_free_regions, target_pause_time_ms,
636 &min_gc_eff);
638 if (min_ok) {
639 // the shortest young length is indeed feasible; we'll know
640 // set up the max young length and we'll do a binary search
641 // between min_young_length and max_young_length
642 size_t max_young_length = _free_regions_at_end_of_collection - 1;
643 double max_gc_eff = 0.0;
644 bool max_ok = false;
646 // the innermost loop! (finally!)
647 while (max_young_length > min_young_length) {
648 // we'll make sure that min_young_length is always at a
649 // feasible config
650 guarantee( min_ok, "invariant" );
652 ++calculations;
653 max_ok = predict_gc_eff(max_young_length, so_length,
654 base_time_with_so_ms,
655 init_free_regions, target_pause_time_ms,
656 &max_gc_eff);
658 size_t diff = (max_young_length - min_young_length) / 2;
659 if (max_ok) {
660 min_young_length = max_young_length;
661 min_gc_eff = max_gc_eff;
662 min_ok = true;
663 }
664 max_young_length = min_young_length + diff;
665 }
667 // the innermost loop found a config
668 guarantee( min_ok, "invariant" );
669 if (min_gc_eff > final_gc_eff) {
670 // it's the best config so far, so we'll keep it
671 final_gc_eff = min_gc_eff;
672 final_young_length = min_young_length;
673 final_so_length = so_length;
674 gc_eff_set = true;
675 }
676 }
678 // incremental the fixed S-O length and go around
679 so_length += so_length_incr;
680 }
682 // this is the end of the outermost loop and we need to decide
683 // what to do during the next iteration
684 if (pass == pass_type_coarse) {
685 // we just did the coarse pass (first iteration)
687 if (!gc_eff_set)
688 // we didn't find a feasible config so we'll just bail out; of
689 // course, it might be the case that we missed it; but I'd say
690 // it's a bit unlikely
691 done = true;
692 else {
693 // We did find a feasible config with optimal GC eff during
694 // the first pass. So the second pass we'll only consider the
695 // S-O lengths around that config with a fine increment.
697 guarantee( so_length_incr == so_coarse_increments, "invariant" );
698 guarantee( final_so_length >= min_so_length, "invariant" );
700 #ifdef TRACE_CALC_YOUNG_CONFIG
701 // leave this in for debugging, just in case
702 gclog_or_tty->print_cr(" coarse pass: SO length " SIZE_FORMAT,
703 final_so_length);
704 #endif // TRACE_CALC_YOUNG_CONFIG
706 from_so_length =
707 (final_so_length - min_so_length > so_coarse_increments) ?
708 final_so_length - so_coarse_increments + 1 : min_so_length;
709 to_so_length =
710 (max_so_length - final_so_length > so_coarse_increments) ?
711 final_so_length + so_coarse_increments - 1 : max_so_length;
713 pass = pass_type_fine;
714 so_length_incr = 1;
715 }
716 } else if (pass == pass_type_fine) {
717 // we just finished the second pass
719 if (!gc_eff_set) {
720 // we didn't find a feasible config (yes, it's possible;
721 // notice that, sometimes, we go directly into the fine
722 // iteration and skip the coarse one) so we bail out
723 done = true;
724 } else {
725 // We did find a feasible config with optimal GC eff
726 guarantee( so_length_incr == 1, "invariant" );
728 if (final_so_length == 0) {
729 // The config is of an empty S-O set, so we'll just bail out
730 done = true;
731 } else {
732 // we'll go around once more, setting the S-O length to 95%
733 // of the optimal
734 size_t new_so_length = 950 * final_so_length / 1000;
736 #ifdef TRACE_CALC_YOUNG_CONFIG
737 // leave this in for debugging, just in case
738 gclog_or_tty->print_cr(" fine pass: SO length " SIZE_FORMAT
739 ", setting it to " SIZE_FORMAT,
740 final_so_length, new_so_length);
741 #endif // TRACE_CALC_YOUNG_CONFIG
743 from_so_length = new_so_length;
744 to_so_length = new_so_length;
745 fine_so_length = final_so_length;
747 pass = pass_type_final;
748 }
749 }
750 } else if (pass == pass_type_final) {
751 // we just finished the final (third) pass
753 if (!gc_eff_set)
754 // we didn't find a feasible config, so we'll just use the one
755 // we found during the second pass, which we saved
756 final_so_length = fine_so_length;
758 // and we're done!
759 done = true;
760 } else {
761 guarantee( false, "should never reach here" );
762 }
764 // we now go around the outermost loop
765 }
767 // we should have at least one region in the target young length
768 _young_list_target_length =
769 MAX2((size_t) 1, final_young_length + _recorded_survivor_regions);
770 if (final_so_length >= final_young_length)
771 // and we need to ensure that the S-O length is not greater than
772 // the target young length (this is being a bit careful)
773 final_so_length = 0;
774 _young_list_so_prefix_length = final_so_length;
775 guarantee( !_in_marking_window || !_last_full_young_gc ||
776 _young_list_so_prefix_length == 0, "invariant" );
778 // let's keep an eye of how long we spend on this calculation
779 // right now, I assume that we'll print it when we need it; we
780 // should really adde it to the breakdown of a pause
781 double end_time_sec = os::elapsedTime();
782 double elapsed_time_ms = (end_time_sec - start_time_sec) * 1000.0;
784 #ifdef TRACE_CALC_YOUNG_CONFIG
785 // leave this in for debugging, just in case
786 gclog_or_tty->print_cr("target = %1.1lf ms, young = " SIZE_FORMAT
787 ", SO = " SIZE_FORMAT ", "
788 "elapsed %1.2lf ms, calcs: " SIZE_FORMAT " (%s%s) "
789 SIZE_FORMAT SIZE_FORMAT,
790 target_pause_time_ms,
791 _young_list_target_length - _young_list_so_prefix_length,
792 _young_list_so_prefix_length,
793 elapsed_time_ms,
794 calculations,
795 full_young_gcs() ? "full" : "partial",
796 during_initial_mark_pause() ? " i-m" : "",
797 _in_marking_window,
798 _in_marking_window_im);
799 #endif // TRACE_CALC_YOUNG_CONFIG
801 if (_young_list_target_length < _young_list_min_length) {
802 // bummer; this means that, if we do a pause when the optimal
803 // config dictates, we'll violate the pause spacing target (the
804 // min length was calculate based on the application's current
805 // alloc rate);
807 // so, we have to bite the bullet, and allocate the minimum
808 // number. We'll violate our target, but we just can't meet it.
810 size_t so_length = 0;
811 // a note further up explains why we do not want an S-O length
812 // during marking
813 if (!_in_marking_window && !_last_full_young_gc)
814 // but we can still try to see whether we can find an optimal
815 // S-O length
816 so_length = calculate_optimal_so_length(_young_list_min_length);
818 #ifdef TRACE_CALC_YOUNG_CONFIG
819 // leave this in for debugging, just in case
820 gclog_or_tty->print_cr("adjusted target length from "
821 SIZE_FORMAT " to " SIZE_FORMAT
822 ", SO " SIZE_FORMAT,
823 _young_list_target_length, _young_list_min_length,
824 so_length);
825 #endif // TRACE_CALC_YOUNG_CONFIG
827 _young_list_target_length =
828 MAX2(_young_list_min_length, (size_t)1);
829 _young_list_so_prefix_length = so_length;
830 }
831 } else {
832 // we are in a partially-young mode or we've run out of regions (due
833 // to evacuation failure)
835 #ifdef TRACE_CALC_YOUNG_CONFIG
836 // leave this in for debugging, just in case
837 gclog_or_tty->print_cr("(partial) setting target to " SIZE_FORMAT
838 ", SO " SIZE_FORMAT,
839 _young_list_min_length, 0);
840 #endif // TRACE_CALC_YOUNG_CONFIG
842 // we'll do the pause as soon as possible and with no S-O prefix
843 // (see above for the reasons behind the latter)
844 _young_list_target_length =
845 MAX2(_young_list_min_length, (size_t) 1);
846 _young_list_so_prefix_length = 0;
847 }
849 _rs_lengths_prediction = rs_lengths;
850 }
852 // This is used by: calculate_optimal_so_length(length). It returns
853 // the GC eff and predicted pause time for a particular config
854 void
855 G1CollectorPolicy::predict_gc_eff(size_t young_length,
856 size_t so_length,
857 double base_time_ms,
858 double* ret_gc_eff,
859 double* ret_pause_time_ms) {
860 double so_time_ms = predict_scan_only_time_ms(so_length);
861 double accum_surv_rate_adj = 0.0;
862 if (so_length > 0)
863 accum_surv_rate_adj = accum_yg_surv_rate_pred((int)(so_length - 1));
864 double accum_surv_rate =
865 accum_yg_surv_rate_pred((int)(young_length - 1)) - accum_surv_rate_adj;
866 size_t bytes_to_copy =
867 (size_t) (accum_surv_rate * (double) HeapRegion::GrainBytes);
868 double copy_time_ms = predict_object_copy_time_ms(bytes_to_copy);
869 double young_other_time_ms =
870 predict_young_other_time_ms(young_length - so_length);
871 double pause_time_ms =
872 base_time_ms + so_time_ms + copy_time_ms + young_other_time_ms;
873 size_t reclaimed_bytes =
874 (young_length - so_length) * HeapRegion::GrainBytes - bytes_to_copy;
875 double gc_eff = (double) reclaimed_bytes / pause_time_ms;
877 *ret_gc_eff = gc_eff;
878 *ret_pause_time_ms = pause_time_ms;
879 }
881 // This is used by: calculate_young_list_target_config(rs_length). It
882 // returns the GC eff of a particular config. It returns false if that
883 // config violates any of the end conditions of the search in the
884 // calling method, or true upon success. The end conditions were put
885 // here since it's called twice and it was best not to replicate them
886 // in the caller. Also, passing the parameteres avoids having to
887 // recalculate them in the innermost loop.
888 bool
889 G1CollectorPolicy::predict_gc_eff(size_t young_length,
890 size_t so_length,
891 double base_time_with_so_ms,
892 size_t init_free_regions,
893 double target_pause_time_ms,
894 double* ret_gc_eff) {
895 *ret_gc_eff = 0.0;
897 if (young_length >= init_free_regions)
898 // end condition 1: not enough space for the young regions
899 return false;
901 double accum_surv_rate_adj = 0.0;
902 if (so_length > 0)
903 accum_surv_rate_adj = accum_yg_surv_rate_pred((int)(so_length - 1));
904 double accum_surv_rate =
905 accum_yg_surv_rate_pred((int)(young_length - 1)) - accum_surv_rate_adj;
906 size_t bytes_to_copy =
907 (size_t) (accum_surv_rate * (double) HeapRegion::GrainBytes);
908 double copy_time_ms = predict_object_copy_time_ms(bytes_to_copy);
909 double young_other_time_ms =
910 predict_young_other_time_ms(young_length - so_length);
911 double pause_time_ms =
912 base_time_with_so_ms + copy_time_ms + young_other_time_ms;
914 if (pause_time_ms > target_pause_time_ms)
915 // end condition 2: over the target pause time
916 return false;
918 size_t reclaimed_bytes =
919 (young_length - so_length) * HeapRegion::GrainBytes - bytes_to_copy;
920 size_t free_bytes =
921 (init_free_regions - young_length) * HeapRegion::GrainBytes;
923 if ((2.0 + sigma()) * (double) bytes_to_copy > (double) free_bytes)
924 // end condition 3: out of to-space (conservatively)
925 return false;
927 // success!
928 double gc_eff = (double) reclaimed_bytes / pause_time_ms;
929 *ret_gc_eff = gc_eff;
931 return true;
932 }
934 double G1CollectorPolicy::predict_survivor_regions_evac_time() {
935 double survivor_regions_evac_time = 0.0;
936 for (HeapRegion * r = _recorded_survivor_head;
937 r != NULL && r != _recorded_survivor_tail->get_next_young_region();
938 r = r->get_next_young_region()) {
939 survivor_regions_evac_time += predict_region_elapsed_time_ms(r, true);
940 }
941 return survivor_regions_evac_time;
942 }
944 void G1CollectorPolicy::check_prediction_validity() {
945 guarantee( adaptive_young_list_length(), "should not call this otherwise" );
947 size_t rs_lengths = _g1->young_list_sampled_rs_lengths();
948 if (rs_lengths > _rs_lengths_prediction) {
949 // add 10% to avoid having to recalculate often
950 size_t rs_lengths_prediction = rs_lengths * 1100 / 1000;
951 calculate_young_list_target_config(rs_lengths_prediction);
952 }
953 }
955 HeapWord* G1CollectorPolicy::mem_allocate_work(size_t size,
956 bool is_tlab,
957 bool* gc_overhead_limit_was_exceeded) {
958 guarantee(false, "Not using this policy feature yet.");
959 return NULL;
960 }
962 // This method controls how a collector handles one or more
963 // of its generations being fully allocated.
964 HeapWord* G1CollectorPolicy::satisfy_failed_allocation(size_t size,
965 bool is_tlab) {
966 guarantee(false, "Not using this policy feature yet.");
967 return NULL;
968 }
971 #ifndef PRODUCT
972 bool G1CollectorPolicy::verify_young_ages() {
973 HeapRegion* head = _g1->young_list_first_region();
974 return
975 verify_young_ages(head, _short_lived_surv_rate_group);
976 // also call verify_young_ages on any additional surv rate groups
977 }
979 bool
980 G1CollectorPolicy::verify_young_ages(HeapRegion* head,
981 SurvRateGroup *surv_rate_group) {
982 guarantee( surv_rate_group != NULL, "pre-condition" );
984 const char* name = surv_rate_group->name();
985 bool ret = true;
986 int prev_age = -1;
988 for (HeapRegion* curr = head;
989 curr != NULL;
990 curr = curr->get_next_young_region()) {
991 SurvRateGroup* group = curr->surv_rate_group();
992 if (group == NULL && !curr->is_survivor()) {
993 gclog_or_tty->print_cr("## %s: encountered NULL surv_rate_group", name);
994 ret = false;
995 }
997 if (surv_rate_group == group) {
998 int age = curr->age_in_surv_rate_group();
1000 if (age < 0) {
1001 gclog_or_tty->print_cr("## %s: encountered negative age", name);
1002 ret = false;
1003 }
1005 if (age <= prev_age) {
1006 gclog_or_tty->print_cr("## %s: region ages are not strictly increasing "
1007 "(%d, %d)", name, age, prev_age);
1008 ret = false;
1009 }
1010 prev_age = age;
1011 }
1012 }
1014 return ret;
1015 }
1016 #endif // PRODUCT
1018 void G1CollectorPolicy::record_full_collection_start() {
1019 _cur_collection_start_sec = os::elapsedTime();
1020 // Release the future to-space so that it is available for compaction into.
1021 _g1->set_full_collection();
1022 }
1024 void G1CollectorPolicy::record_full_collection_end() {
1025 // Consider this like a collection pause for the purposes of allocation
1026 // since last pause.
1027 double end_sec = os::elapsedTime();
1028 double full_gc_time_sec = end_sec - _cur_collection_start_sec;
1029 double full_gc_time_ms = full_gc_time_sec * 1000.0;
1031 _all_full_gc_times_ms->add(full_gc_time_ms);
1033 update_recent_gc_times(end_sec, full_gc_time_ms);
1035 _g1->clear_full_collection();
1037 // "Nuke" the heuristics that control the fully/partially young GC
1038 // transitions and make sure we start with fully young GCs after the
1039 // Full GC.
1040 set_full_young_gcs(true);
1041 _last_full_young_gc = false;
1042 _should_revert_to_full_young_gcs = false;
1043 clear_initiate_conc_mark_if_possible();
1044 clear_during_initial_mark_pause();
1045 _known_garbage_bytes = 0;
1046 _known_garbage_ratio = 0.0;
1047 _in_marking_window = false;
1048 _in_marking_window_im = false;
1050 _short_lived_surv_rate_group->record_scan_only_prefix(0);
1051 _short_lived_surv_rate_group->start_adding_regions();
1052 // also call this on any additional surv rate groups
1054 record_survivor_regions(0, NULL, NULL);
1056 _prev_region_num_young = _region_num_young;
1057 _prev_region_num_tenured = _region_num_tenured;
1059 _free_regions_at_end_of_collection = _g1->free_regions();
1060 _scan_only_regions_at_end_of_collection = 0;
1061 // Reset survivors SurvRateGroup.
1062 _survivor_surv_rate_group->reset();
1063 calculate_young_list_min_length();
1064 calculate_young_list_target_config();
1065 }
1067 void G1CollectorPolicy::record_before_bytes(size_t bytes) {
1068 _bytes_in_to_space_before_gc += bytes;
1069 }
1071 void G1CollectorPolicy::record_after_bytes(size_t bytes) {
1072 _bytes_in_to_space_after_gc += bytes;
1073 }
1075 void G1CollectorPolicy::record_stop_world_start() {
1076 _stop_world_start = os::elapsedTime();
1077 }
1079 void G1CollectorPolicy::record_collection_pause_start(double start_time_sec,
1080 size_t start_used) {
1081 if (PrintGCDetails) {
1082 gclog_or_tty->stamp(PrintGCTimeStamps);
1083 gclog_or_tty->print("[GC pause");
1084 if (in_young_gc_mode())
1085 gclog_or_tty->print(" (%s)", full_young_gcs() ? "young" : "partial");
1086 }
1088 assert(_g1->used_regions() == _g1->recalculate_used_regions(),
1089 "sanity");
1090 assert(_g1->used() == _g1->recalculate_used(), "sanity");
1092 double s_w_t_ms = (start_time_sec - _stop_world_start) * 1000.0;
1093 _all_stop_world_times_ms->add(s_w_t_ms);
1094 _stop_world_start = 0.0;
1096 _cur_collection_start_sec = start_time_sec;
1097 _cur_collection_pause_used_at_start_bytes = start_used;
1098 _cur_collection_pause_used_regions_at_start = _g1->used_regions();
1099 _pending_cards = _g1->pending_card_num();
1100 _max_pending_cards = _g1->max_pending_card_num();
1102 _bytes_in_to_space_before_gc = 0;
1103 _bytes_in_to_space_after_gc = 0;
1104 _bytes_in_collection_set_before_gc = 0;
1106 #ifdef DEBUG
1107 // initialise these to something well known so that we can spot
1108 // if they are not set properly
1110 for (int i = 0; i < _parallel_gc_threads; ++i) {
1111 _par_last_ext_root_scan_times_ms[i] = -666.0;
1112 _par_last_mark_stack_scan_times_ms[i] = -666.0;
1113 _par_last_scan_only_times_ms[i] = -666.0;
1114 _par_last_scan_only_regions_scanned[i] = -666.0;
1115 _par_last_update_rs_start_times_ms[i] = -666.0;
1116 _par_last_update_rs_times_ms[i] = -666.0;
1117 _par_last_update_rs_processed_buffers[i] = -666.0;
1118 _par_last_scan_rs_start_times_ms[i] = -666.0;
1119 _par_last_scan_rs_times_ms[i] = -666.0;
1120 _par_last_scan_new_refs_times_ms[i] = -666.0;
1121 _par_last_obj_copy_times_ms[i] = -666.0;
1122 _par_last_termination_times_ms[i] = -666.0;
1123 }
1124 #endif
1126 for (int i = 0; i < _aux_num; ++i) {
1127 _cur_aux_times_ms[i] = 0.0;
1128 _cur_aux_times_set[i] = false;
1129 }
1131 _satb_drain_time_set = false;
1132 _last_satb_drain_processed_buffers = -1;
1134 if (in_young_gc_mode())
1135 _last_young_gc_full = false;
1138 // do that for any other surv rate groups
1139 _short_lived_surv_rate_group->stop_adding_regions();
1140 size_t short_lived_so_length = _young_list_so_prefix_length;
1141 _short_lived_surv_rate_group->record_scan_only_prefix(short_lived_so_length);
1142 tag_scan_only(short_lived_so_length);
1143 _survivors_age_table.clear();
1145 assert( verify_young_ages(), "region age verification" );
1146 }
1148 void G1CollectorPolicy::tag_scan_only(size_t short_lived_scan_only_length) {
1149 // done in a way that it can be extended for other surv rate groups too...
1151 HeapRegion* head = _g1->young_list_first_region();
1152 bool finished_short_lived = (short_lived_scan_only_length == 0);
1154 if (finished_short_lived)
1155 return;
1157 for (HeapRegion* curr = head;
1158 curr != NULL;
1159 curr = curr->get_next_young_region()) {
1160 SurvRateGroup* surv_rate_group = curr->surv_rate_group();
1161 int age = curr->age_in_surv_rate_group();
1163 if (surv_rate_group == _short_lived_surv_rate_group) {
1164 if ((size_t)age < short_lived_scan_only_length)
1165 curr->set_scan_only();
1166 else
1167 finished_short_lived = true;
1168 }
1171 if (finished_short_lived)
1172 return;
1173 }
1175 guarantee( false, "we should never reach here" );
1176 }
1178 void G1CollectorPolicy::record_mark_closure_time(double mark_closure_time_ms) {
1179 _mark_closure_time_ms = mark_closure_time_ms;
1180 }
1182 void G1CollectorPolicy::record_concurrent_mark_init_start() {
1183 _mark_init_start_sec = os::elapsedTime();
1184 guarantee(!in_young_gc_mode(), "should not do be here in young GC mode");
1185 }
1187 void G1CollectorPolicy::record_concurrent_mark_init_end_pre(double
1188 mark_init_elapsed_time_ms) {
1189 _during_marking = true;
1190 assert(!initiate_conc_mark_if_possible(), "we should have cleared it by now");
1191 clear_during_initial_mark_pause();
1192 _cur_mark_stop_world_time_ms = mark_init_elapsed_time_ms;
1193 }
1195 void G1CollectorPolicy::record_concurrent_mark_init_end() {
1196 double end_time_sec = os::elapsedTime();
1197 double elapsed_time_ms = (end_time_sec - _mark_init_start_sec) * 1000.0;
1198 _concurrent_mark_init_times_ms->add(elapsed_time_ms);
1199 record_concurrent_mark_init_end_pre(elapsed_time_ms);
1201 _mmu_tracker->add_pause(_mark_init_start_sec, end_time_sec, true);
1202 }
1204 void G1CollectorPolicy::record_concurrent_mark_remark_start() {
1205 _mark_remark_start_sec = os::elapsedTime();
1206 _during_marking = false;
1207 }
1209 void G1CollectorPolicy::record_concurrent_mark_remark_end() {
1210 double end_time_sec = os::elapsedTime();
1211 double elapsed_time_ms = (end_time_sec - _mark_remark_start_sec)*1000.0;
1212 _concurrent_mark_remark_times_ms->add(elapsed_time_ms);
1213 _cur_mark_stop_world_time_ms += elapsed_time_ms;
1214 _prev_collection_pause_end_ms += elapsed_time_ms;
1216 _mmu_tracker->add_pause(_mark_remark_start_sec, end_time_sec, true);
1217 }
1219 void G1CollectorPolicy::record_concurrent_mark_cleanup_start() {
1220 _mark_cleanup_start_sec = os::elapsedTime();
1221 }
1223 void
1224 G1CollectorPolicy::record_concurrent_mark_cleanup_end(size_t freed_bytes,
1225 size_t max_live_bytes) {
1226 record_concurrent_mark_cleanup_end_work1(freed_bytes, max_live_bytes);
1227 record_concurrent_mark_cleanup_end_work2();
1228 }
1230 void
1231 G1CollectorPolicy::
1232 record_concurrent_mark_cleanup_end_work1(size_t freed_bytes,
1233 size_t max_live_bytes) {
1234 if (_n_marks < 2) _n_marks++;
1235 if (G1PolicyVerbose > 0)
1236 gclog_or_tty->print_cr("At end of marking, max_live is " SIZE_FORMAT " MB "
1237 " (of " SIZE_FORMAT " MB heap).",
1238 max_live_bytes/M, _g1->capacity()/M);
1239 }
1241 // The important thing about this is that it includes "os::elapsedTime".
1242 void G1CollectorPolicy::record_concurrent_mark_cleanup_end_work2() {
1243 double end_time_sec = os::elapsedTime();
1244 double elapsed_time_ms = (end_time_sec - _mark_cleanup_start_sec)*1000.0;
1245 _concurrent_mark_cleanup_times_ms->add(elapsed_time_ms);
1246 _cur_mark_stop_world_time_ms += elapsed_time_ms;
1247 _prev_collection_pause_end_ms += elapsed_time_ms;
1249 _mmu_tracker->add_pause(_mark_cleanup_start_sec, end_time_sec, true);
1251 _num_markings++;
1253 // We did a marking, so reset the "since_last_mark" variables.
1254 double considerConcMarkCost = 1.0;
1255 // If there are available processors, concurrent activity is free...
1256 if (Threads::number_of_non_daemon_threads() * 2 <
1257 os::active_processor_count()) {
1258 considerConcMarkCost = 0.0;
1259 }
1260 _n_pauses_at_mark_end = _n_pauses;
1261 _n_marks_since_last_pause++;
1262 }
1264 void
1265 G1CollectorPolicy::record_concurrent_mark_cleanup_completed() {
1266 if (in_young_gc_mode()) {
1267 _should_revert_to_full_young_gcs = false;
1268 _last_full_young_gc = true;
1269 _in_marking_window = false;
1270 if (adaptive_young_list_length())
1271 calculate_young_list_target_config();
1272 }
1273 }
1275 void G1CollectorPolicy::record_concurrent_pause() {
1276 if (_stop_world_start > 0.0) {
1277 double yield_ms = (os::elapsedTime() - _stop_world_start) * 1000.0;
1278 _all_yield_times_ms->add(yield_ms);
1279 }
1280 }
1282 void G1CollectorPolicy::record_concurrent_pause_end() {
1283 }
1285 void G1CollectorPolicy::record_collection_pause_end_CH_strong_roots() {
1286 _cur_CH_strong_roots_end_sec = os::elapsedTime();
1287 _cur_CH_strong_roots_dur_ms =
1288 (_cur_CH_strong_roots_end_sec - _cur_collection_start_sec) * 1000.0;
1289 }
1291 void G1CollectorPolicy::record_collection_pause_end_G1_strong_roots() {
1292 _cur_G1_strong_roots_end_sec = os::elapsedTime();
1293 _cur_G1_strong_roots_dur_ms =
1294 (_cur_G1_strong_roots_end_sec - _cur_CH_strong_roots_end_sec) * 1000.0;
1295 }
1297 template<class T>
1298 T sum_of(T* sum_arr, int start, int n, int N) {
1299 T sum = (T)0;
1300 for (int i = 0; i < n; i++) {
1301 int j = (start + i) % N;
1302 sum += sum_arr[j];
1303 }
1304 return sum;
1305 }
1307 void G1CollectorPolicy::print_par_stats (int level,
1308 const char* str,
1309 double* data,
1310 bool summary) {
1311 double min = data[0], max = data[0];
1312 double total = 0.0;
1313 int j;
1314 for (j = 0; j < level; ++j)
1315 gclog_or_tty->print(" ");
1316 gclog_or_tty->print("[%s (ms):", str);
1317 for (uint i = 0; i < ParallelGCThreads; ++i) {
1318 double val = data[i];
1319 if (val < min)
1320 min = val;
1321 if (val > max)
1322 max = val;
1323 total += val;
1324 gclog_or_tty->print(" %3.1lf", val);
1325 }
1326 if (summary) {
1327 gclog_or_tty->print_cr("");
1328 double avg = total / (double) ParallelGCThreads;
1329 gclog_or_tty->print(" ");
1330 for (j = 0; j < level; ++j)
1331 gclog_or_tty->print(" ");
1332 gclog_or_tty->print("Avg: %5.1lf, Min: %5.1lf, Max: %5.1lf",
1333 avg, min, max);
1334 }
1335 gclog_or_tty->print_cr("]");
1336 }
1338 void G1CollectorPolicy::print_par_buffers (int level,
1339 const char* str,
1340 double* data,
1341 bool summary) {
1342 double min = data[0], max = data[0];
1343 double total = 0.0;
1344 int j;
1345 for (j = 0; j < level; ++j)
1346 gclog_or_tty->print(" ");
1347 gclog_or_tty->print("[%s :", str);
1348 for (uint i = 0; i < ParallelGCThreads; ++i) {
1349 double val = data[i];
1350 if (val < min)
1351 min = val;
1352 if (val > max)
1353 max = val;
1354 total += val;
1355 gclog_or_tty->print(" %d", (int) val);
1356 }
1357 if (summary) {
1358 gclog_or_tty->print_cr("");
1359 double avg = total / (double) ParallelGCThreads;
1360 gclog_or_tty->print(" ");
1361 for (j = 0; j < level; ++j)
1362 gclog_or_tty->print(" ");
1363 gclog_or_tty->print("Sum: %d, Avg: %d, Min: %d, Max: %d",
1364 (int)total, (int)avg, (int)min, (int)max);
1365 }
1366 gclog_or_tty->print_cr("]");
1367 }
1369 void G1CollectorPolicy::print_stats (int level,
1370 const char* str,
1371 double value) {
1372 for (int j = 0; j < level; ++j)
1373 gclog_or_tty->print(" ");
1374 gclog_or_tty->print_cr("[%s: %5.1lf ms]", str, value);
1375 }
1377 void G1CollectorPolicy::print_stats (int level,
1378 const char* str,
1379 int value) {
1380 for (int j = 0; j < level; ++j)
1381 gclog_or_tty->print(" ");
1382 gclog_or_tty->print_cr("[%s: %d]", str, value);
1383 }
1385 double G1CollectorPolicy::avg_value (double* data) {
1386 if (ParallelGCThreads > 0) {
1387 double ret = 0.0;
1388 for (uint i = 0; i < ParallelGCThreads; ++i)
1389 ret += data[i];
1390 return ret / (double) ParallelGCThreads;
1391 } else {
1392 return data[0];
1393 }
1394 }
1396 double G1CollectorPolicy::max_value (double* data) {
1397 if (ParallelGCThreads > 0) {
1398 double ret = data[0];
1399 for (uint i = 1; i < ParallelGCThreads; ++i)
1400 if (data[i] > ret)
1401 ret = data[i];
1402 return ret;
1403 } else {
1404 return data[0];
1405 }
1406 }
1408 double G1CollectorPolicy::sum_of_values (double* data) {
1409 if (ParallelGCThreads > 0) {
1410 double sum = 0.0;
1411 for (uint i = 0; i < ParallelGCThreads; i++)
1412 sum += data[i];
1413 return sum;
1414 } else {
1415 return data[0];
1416 }
1417 }
1419 double G1CollectorPolicy::max_sum (double* data1,
1420 double* data2) {
1421 double ret = data1[0] + data2[0];
1423 if (ParallelGCThreads > 0) {
1424 for (uint i = 1; i < ParallelGCThreads; ++i) {
1425 double data = data1[i] + data2[i];
1426 if (data > ret)
1427 ret = data;
1428 }
1429 }
1430 return ret;
1431 }
1433 // Anything below that is considered to be zero
1434 #define MIN_TIMER_GRANULARITY 0.0000001
1436 void G1CollectorPolicy::record_collection_pause_end(bool abandoned) {
1437 double end_time_sec = os::elapsedTime();
1438 double elapsed_ms = _last_pause_time_ms;
1439 bool parallel = ParallelGCThreads > 0;
1440 double evac_ms = (end_time_sec - _cur_G1_strong_roots_end_sec) * 1000.0;
1441 size_t rs_size =
1442 _cur_collection_pause_used_regions_at_start - collection_set_size();
1443 size_t cur_used_bytes = _g1->used();
1444 assert(cur_used_bytes == _g1->recalculate_used(), "It should!");
1445 bool last_pause_included_initial_mark = false;
1446 bool update_stats = !abandoned && !_g1->evacuation_failed();
1448 #ifndef PRODUCT
1449 if (G1YoungSurvRateVerbose) {
1450 gclog_or_tty->print_cr("");
1451 _short_lived_surv_rate_group->print();
1452 // do that for any other surv rate groups too
1453 }
1454 #endif // PRODUCT
1456 if (in_young_gc_mode()) {
1457 last_pause_included_initial_mark = during_initial_mark_pause();
1458 if (last_pause_included_initial_mark)
1459 record_concurrent_mark_init_end_pre(0.0);
1461 size_t min_used_targ =
1462 (_g1->capacity() / 100) * InitiatingHeapOccupancyPercent;
1465 if (!_g1->mark_in_progress() && !_last_full_young_gc) {
1466 assert(!last_pause_included_initial_mark, "invariant");
1467 if (cur_used_bytes > min_used_targ &&
1468 cur_used_bytes > _prev_collection_pause_used_at_end_bytes) {
1469 assert(!during_initial_mark_pause(), "we should not see this here");
1471 // Note: this might have already been set, if during the last
1472 // pause we decided to start a cycle but at the beginning of
1473 // this pause we decided to postpone it. That's OK.
1474 set_initiate_conc_mark_if_possible();
1475 }
1476 }
1478 _prev_collection_pause_used_at_end_bytes = cur_used_bytes;
1479 }
1481 _mmu_tracker->add_pause(end_time_sec - elapsed_ms/1000.0,
1482 end_time_sec, false);
1484 guarantee(_cur_collection_pause_used_regions_at_start >=
1485 collection_set_size(),
1486 "Negative RS size?");
1488 // This assert is exempted when we're doing parallel collection pauses,
1489 // because the fragmentation caused by the parallel GC allocation buffers
1490 // can lead to more memory being used during collection than was used
1491 // before. Best leave this out until the fragmentation problem is fixed.
1492 // Pauses in which evacuation failed can also lead to negative
1493 // collections, since no space is reclaimed from a region containing an
1494 // object whose evacuation failed.
1495 // Further, we're now always doing parallel collection. But I'm still
1496 // leaving this here as a placeholder for a more precise assertion later.
1497 // (DLD, 10/05.)
1498 assert((true || parallel) // Always using GC LABs now.
1499 || _g1->evacuation_failed()
1500 || _cur_collection_pause_used_at_start_bytes >= cur_used_bytes,
1501 "Negative collection");
1503 size_t freed_bytes =
1504 _cur_collection_pause_used_at_start_bytes - cur_used_bytes;
1505 size_t surviving_bytes = _collection_set_bytes_used_before - freed_bytes;
1506 double survival_fraction =
1507 (double)surviving_bytes/
1508 (double)_collection_set_bytes_used_before;
1510 _n_pauses++;
1512 if (update_stats) {
1513 _recent_CH_strong_roots_times_ms->add(_cur_CH_strong_roots_dur_ms);
1514 _recent_G1_strong_roots_times_ms->add(_cur_G1_strong_roots_dur_ms);
1515 _recent_evac_times_ms->add(evac_ms);
1516 _recent_pause_times_ms->add(elapsed_ms);
1518 _recent_rs_sizes->add(rs_size);
1520 // We exempt parallel collection from this check because Alloc Buffer
1521 // fragmentation can produce negative collections. Same with evac
1522 // failure.
1523 // Further, we're now always doing parallel collection. But I'm still
1524 // leaving this here as a placeholder for a more precise assertion later.
1525 // (DLD, 10/05.
1526 assert((true || parallel)
1527 || _g1->evacuation_failed()
1528 || surviving_bytes <= _collection_set_bytes_used_before,
1529 "Or else negative collection!");
1530 _recent_CS_bytes_used_before->add(_collection_set_bytes_used_before);
1531 _recent_CS_bytes_surviving->add(surviving_bytes);
1533 // this is where we update the allocation rate of the application
1534 double app_time_ms =
1535 (_cur_collection_start_sec * 1000.0 - _prev_collection_pause_end_ms);
1536 if (app_time_ms < MIN_TIMER_GRANULARITY) {
1537 // This usually happens due to the timer not having the required
1538 // granularity. Some Linuxes are the usual culprits.
1539 // We'll just set it to something (arbitrarily) small.
1540 app_time_ms = 1.0;
1541 }
1542 size_t regions_allocated =
1543 (_region_num_young - _prev_region_num_young) +
1544 (_region_num_tenured - _prev_region_num_tenured);
1545 double alloc_rate_ms = (double) regions_allocated / app_time_ms;
1546 _alloc_rate_ms_seq->add(alloc_rate_ms);
1547 _prev_region_num_young = _region_num_young;
1548 _prev_region_num_tenured = _region_num_tenured;
1550 double interval_ms =
1551 (end_time_sec - _recent_prev_end_times_for_all_gcs_sec->oldest()) * 1000.0;
1552 update_recent_gc_times(end_time_sec, elapsed_ms);
1553 _recent_avg_pause_time_ratio = _recent_gc_times_ms->sum()/interval_ms;
1554 if (recent_avg_pause_time_ratio() < 0.0 ||
1555 (recent_avg_pause_time_ratio() - 1.0 > 0.0)) {
1556 #ifndef PRODUCT
1557 // Dump info to allow post-facto debugging
1558 gclog_or_tty->print_cr("recent_avg_pause_time_ratio() out of bounds");
1559 gclog_or_tty->print_cr("-------------------------------------------");
1560 gclog_or_tty->print_cr("Recent GC Times (ms):");
1561 _recent_gc_times_ms->dump();
1562 gclog_or_tty->print_cr("(End Time=%3.3f) Recent GC End Times (s):", end_time_sec);
1563 _recent_prev_end_times_for_all_gcs_sec->dump();
1564 gclog_or_tty->print_cr("GC = %3.3f, Interval = %3.3f, Ratio = %3.3f",
1565 _recent_gc_times_ms->sum(), interval_ms, recent_avg_pause_time_ratio());
1566 // In debug mode, terminate the JVM if the user wants to debug at this point.
1567 assert(!G1FailOnFPError, "Debugging data for CR 6898948 has been dumped above");
1568 #endif // !PRODUCT
1569 // Clip ratio between 0.0 and 1.0, and continue. This will be fixed in
1570 // CR 6902692 by redoing the manner in which the ratio is incrementally computed.
1571 if (_recent_avg_pause_time_ratio < 0.0) {
1572 _recent_avg_pause_time_ratio = 0.0;
1573 } else {
1574 assert(_recent_avg_pause_time_ratio - 1.0 > 0.0, "Ctl-point invariant");
1575 _recent_avg_pause_time_ratio = 1.0;
1576 }
1577 }
1578 }
1580 if (G1PolicyVerbose > 1) {
1581 gclog_or_tty->print_cr(" Recording collection pause(%d)", _n_pauses);
1582 }
1584 PauseSummary* summary;
1585 if (abandoned) {
1586 summary = _abandoned_summary;
1587 } else {
1588 summary = _summary;
1589 }
1591 double ext_root_scan_time = avg_value(_par_last_ext_root_scan_times_ms);
1592 double mark_stack_scan_time = avg_value(_par_last_mark_stack_scan_times_ms);
1593 double scan_only_time = avg_value(_par_last_scan_only_times_ms);
1594 double scan_only_regions_scanned =
1595 sum_of_values(_par_last_scan_only_regions_scanned);
1596 double update_rs_time = avg_value(_par_last_update_rs_times_ms);
1597 double update_rs_processed_buffers =
1598 sum_of_values(_par_last_update_rs_processed_buffers);
1599 double scan_rs_time = avg_value(_par_last_scan_rs_times_ms);
1600 double obj_copy_time = avg_value(_par_last_obj_copy_times_ms);
1601 double termination_time = avg_value(_par_last_termination_times_ms);
1603 double parallel_other_time = _cur_collection_par_time_ms -
1604 (update_rs_time + ext_root_scan_time + mark_stack_scan_time +
1605 scan_only_time + scan_rs_time + obj_copy_time + termination_time);
1606 if (update_stats) {
1607 MainBodySummary* body_summary = summary->main_body_summary();
1608 guarantee(body_summary != NULL, "should not be null!");
1610 if (_satb_drain_time_set)
1611 body_summary->record_satb_drain_time_ms(_cur_satb_drain_time_ms);
1612 else
1613 body_summary->record_satb_drain_time_ms(0.0);
1614 body_summary->record_ext_root_scan_time_ms(ext_root_scan_time);
1615 body_summary->record_mark_stack_scan_time_ms(mark_stack_scan_time);
1616 body_summary->record_scan_only_time_ms(scan_only_time);
1617 body_summary->record_update_rs_time_ms(update_rs_time);
1618 body_summary->record_scan_rs_time_ms(scan_rs_time);
1619 body_summary->record_obj_copy_time_ms(obj_copy_time);
1620 if (parallel) {
1621 body_summary->record_parallel_time_ms(_cur_collection_par_time_ms);
1622 body_summary->record_clear_ct_time_ms(_cur_clear_ct_time_ms);
1623 body_summary->record_termination_time_ms(termination_time);
1624 body_summary->record_parallel_other_time_ms(parallel_other_time);
1625 }
1626 body_summary->record_mark_closure_time_ms(_mark_closure_time_ms);
1627 }
1629 if (G1PolicyVerbose > 1) {
1630 gclog_or_tty->print_cr(" ET: %10.6f ms (avg: %10.6f ms)\n"
1631 " CH Strong: %10.6f ms (avg: %10.6f ms)\n"
1632 " G1 Strong: %10.6f ms (avg: %10.6f ms)\n"
1633 " Evac: %10.6f ms (avg: %10.6f ms)\n"
1634 " ET-RS: %10.6f ms (avg: %10.6f ms)\n"
1635 " |RS|: " SIZE_FORMAT,
1636 elapsed_ms, recent_avg_time_for_pauses_ms(),
1637 _cur_CH_strong_roots_dur_ms, recent_avg_time_for_CH_strong_ms(),
1638 _cur_G1_strong_roots_dur_ms, recent_avg_time_for_G1_strong_ms(),
1639 evac_ms, recent_avg_time_for_evac_ms(),
1640 scan_rs_time,
1641 recent_avg_time_for_pauses_ms() -
1642 recent_avg_time_for_G1_strong_ms(),
1643 rs_size);
1645 gclog_or_tty->print_cr(" Used at start: " SIZE_FORMAT"K"
1646 " At end " SIZE_FORMAT "K\n"
1647 " garbage : " SIZE_FORMAT "K"
1648 " of " SIZE_FORMAT "K\n"
1649 " survival : %6.2f%% (%6.2f%% avg)",
1650 _cur_collection_pause_used_at_start_bytes/K,
1651 _g1->used()/K, freed_bytes/K,
1652 _collection_set_bytes_used_before/K,
1653 survival_fraction*100.0,
1654 recent_avg_survival_fraction()*100.0);
1655 gclog_or_tty->print_cr(" Recent %% gc pause time: %6.2f",
1656 recent_avg_pause_time_ratio() * 100.0);
1657 }
1659 double other_time_ms = elapsed_ms;
1661 if (!abandoned) {
1662 if (_satb_drain_time_set)
1663 other_time_ms -= _cur_satb_drain_time_ms;
1665 if (parallel)
1666 other_time_ms -= _cur_collection_par_time_ms + _cur_clear_ct_time_ms;
1667 else
1668 other_time_ms -=
1669 update_rs_time +
1670 ext_root_scan_time + mark_stack_scan_time + scan_only_time +
1671 scan_rs_time + obj_copy_time;
1672 }
1674 if (PrintGCDetails) {
1675 gclog_or_tty->print_cr("%s%s, %1.8lf secs]",
1676 abandoned ? " (abandoned)" : "",
1677 (last_pause_included_initial_mark) ? " (initial-mark)" : "",
1678 elapsed_ms / 1000.0);
1680 if (!abandoned) {
1681 if (_satb_drain_time_set) {
1682 print_stats(1, "SATB Drain Time", _cur_satb_drain_time_ms);
1683 }
1684 if (_last_satb_drain_processed_buffers >= 0) {
1685 print_stats(2, "Processed Buffers", _last_satb_drain_processed_buffers);
1686 }
1687 if (parallel) {
1688 print_stats(1, "Parallel Time", _cur_collection_par_time_ms);
1689 print_par_stats(2, "Update RS (Start)", _par_last_update_rs_start_times_ms, false);
1690 print_par_stats(2, "Update RS", _par_last_update_rs_times_ms);
1691 print_par_buffers(3, "Processed Buffers",
1692 _par_last_update_rs_processed_buffers, true);
1693 print_par_stats(2, "Ext Root Scanning", _par_last_ext_root_scan_times_ms);
1694 print_par_stats(2, "Mark Stack Scanning", _par_last_mark_stack_scan_times_ms);
1695 print_par_stats(2, "Scan-Only Scanning", _par_last_scan_only_times_ms);
1696 print_par_buffers(3, "Scan-Only Regions",
1697 _par_last_scan_only_regions_scanned, true);
1698 print_par_stats(2, "Scan RS", _par_last_scan_rs_times_ms);
1699 print_par_stats(2, "Object Copy", _par_last_obj_copy_times_ms);
1700 print_par_stats(2, "Termination", _par_last_termination_times_ms);
1701 print_stats(2, "Other", parallel_other_time);
1702 print_stats(1, "Clear CT", _cur_clear_ct_time_ms);
1703 } else {
1704 print_stats(1, "Update RS", update_rs_time);
1705 print_stats(2, "Processed Buffers",
1706 (int)update_rs_processed_buffers);
1707 print_stats(1, "Ext Root Scanning", ext_root_scan_time);
1708 print_stats(1, "Mark Stack Scanning", mark_stack_scan_time);
1709 print_stats(1, "Scan-Only Scanning", scan_only_time);
1710 print_stats(1, "Scan RS", scan_rs_time);
1711 print_stats(1, "Object Copying", obj_copy_time);
1712 }
1713 }
1714 #ifndef PRODUCT
1715 print_stats(1, "Cur Clear CC", _cur_clear_cc_time_ms);
1716 print_stats(1, "Cum Clear CC", _cum_clear_cc_time_ms);
1717 print_stats(1, "Min Clear CC", _min_clear_cc_time_ms);
1718 print_stats(1, "Max Clear CC", _max_clear_cc_time_ms);
1719 if (_num_cc_clears > 0) {
1720 print_stats(1, "Avg Clear CC", _cum_clear_cc_time_ms / ((double)_num_cc_clears));
1721 }
1722 #endif
1723 print_stats(1, "Other", other_time_ms);
1724 for (int i = 0; i < _aux_num; ++i) {
1725 if (_cur_aux_times_set[i]) {
1726 char buffer[96];
1727 sprintf(buffer, "Aux%d", i);
1728 print_stats(1, buffer, _cur_aux_times_ms[i]);
1729 }
1730 }
1731 }
1732 if (PrintGCDetails)
1733 gclog_or_tty->print(" [");
1734 if (PrintGC || PrintGCDetails)
1735 _g1->print_size_transition(gclog_or_tty,
1736 _cur_collection_pause_used_at_start_bytes,
1737 _g1->used(), _g1->capacity());
1738 if (PrintGCDetails)
1739 gclog_or_tty->print_cr("]");
1741 _all_pause_times_ms->add(elapsed_ms);
1742 if (update_stats) {
1743 summary->record_total_time_ms(elapsed_ms);
1744 summary->record_other_time_ms(other_time_ms);
1745 }
1746 for (int i = 0; i < _aux_num; ++i)
1747 if (_cur_aux_times_set[i])
1748 _all_aux_times_ms[i].add(_cur_aux_times_ms[i]);
1750 // Reset marks-between-pauses counter.
1751 _n_marks_since_last_pause = 0;
1753 // Update the efficiency-since-mark vars.
1754 double proc_ms = elapsed_ms * (double) _parallel_gc_threads;
1755 if (elapsed_ms < MIN_TIMER_GRANULARITY) {
1756 // This usually happens due to the timer not having the required
1757 // granularity. Some Linuxes are the usual culprits.
1758 // We'll just set it to something (arbitrarily) small.
1759 proc_ms = 1.0;
1760 }
1761 double cur_efficiency = (double) freed_bytes / proc_ms;
1763 bool new_in_marking_window = _in_marking_window;
1764 bool new_in_marking_window_im = false;
1765 if (during_initial_mark_pause()) {
1766 new_in_marking_window = true;
1767 new_in_marking_window_im = true;
1768 }
1770 if (in_young_gc_mode()) {
1771 if (_last_full_young_gc) {
1772 set_full_young_gcs(false);
1773 _last_full_young_gc = false;
1774 }
1776 if ( !_last_young_gc_full ) {
1777 if ( _should_revert_to_full_young_gcs ||
1778 _known_garbage_ratio < 0.05 ||
1779 (adaptive_young_list_length() &&
1780 (get_gc_eff_factor() * cur_efficiency < predict_young_gc_eff())) ) {
1781 set_full_young_gcs(true);
1782 }
1783 }
1784 _should_revert_to_full_young_gcs = false;
1786 if (_last_young_gc_full && !_during_marking)
1787 _young_gc_eff_seq->add(cur_efficiency);
1788 }
1790 _short_lived_surv_rate_group->start_adding_regions();
1791 // do that for any other surv rate groupsx
1793 // <NEW PREDICTION>
1795 if (update_stats) {
1796 double pause_time_ms = elapsed_ms;
1798 size_t diff = 0;
1799 if (_max_pending_cards >= _pending_cards)
1800 diff = _max_pending_cards - _pending_cards;
1801 _pending_card_diff_seq->add((double) diff);
1803 double cost_per_card_ms = 0.0;
1804 if (_pending_cards > 0) {
1805 cost_per_card_ms = update_rs_time / (double) _pending_cards;
1806 _cost_per_card_ms_seq->add(cost_per_card_ms);
1807 }
1809 double cost_per_scan_only_region_ms = 0.0;
1810 if (scan_only_regions_scanned > 0.0) {
1811 cost_per_scan_only_region_ms =
1812 scan_only_time / scan_only_regions_scanned;
1813 if (_in_marking_window_im)
1814 _cost_per_scan_only_region_ms_during_cm_seq->add(cost_per_scan_only_region_ms);
1815 else
1816 _cost_per_scan_only_region_ms_seq->add(cost_per_scan_only_region_ms);
1817 }
1819 size_t cards_scanned = _g1->cards_scanned();
1821 double cost_per_entry_ms = 0.0;
1822 if (cards_scanned > 10) {
1823 cost_per_entry_ms = scan_rs_time / (double) cards_scanned;
1824 if (_last_young_gc_full)
1825 _cost_per_entry_ms_seq->add(cost_per_entry_ms);
1826 else
1827 _partially_young_cost_per_entry_ms_seq->add(cost_per_entry_ms);
1828 }
1830 if (_max_rs_lengths > 0) {
1831 double cards_per_entry_ratio =
1832 (double) cards_scanned / (double) _max_rs_lengths;
1833 if (_last_young_gc_full)
1834 _fully_young_cards_per_entry_ratio_seq->add(cards_per_entry_ratio);
1835 else
1836 _partially_young_cards_per_entry_ratio_seq->add(cards_per_entry_ratio);
1837 }
1839 size_t rs_length_diff = _max_rs_lengths - _recorded_rs_lengths;
1840 if (rs_length_diff >= 0)
1841 _rs_length_diff_seq->add((double) rs_length_diff);
1843 size_t copied_bytes = surviving_bytes;
1844 double cost_per_byte_ms = 0.0;
1845 if (copied_bytes > 0) {
1846 cost_per_byte_ms = obj_copy_time / (double) copied_bytes;
1847 if (_in_marking_window)
1848 _cost_per_byte_ms_during_cm_seq->add(cost_per_byte_ms);
1849 else
1850 _cost_per_byte_ms_seq->add(cost_per_byte_ms);
1851 }
1853 double all_other_time_ms = pause_time_ms -
1854 (update_rs_time + scan_only_time + scan_rs_time + obj_copy_time +
1855 _mark_closure_time_ms + termination_time);
1857 double young_other_time_ms = 0.0;
1858 if (_recorded_young_regions > 0) {
1859 young_other_time_ms =
1860 _recorded_young_cset_choice_time_ms +
1861 _recorded_young_free_cset_time_ms;
1862 _young_other_cost_per_region_ms_seq->add(young_other_time_ms /
1863 (double) _recorded_young_regions);
1864 }
1865 double non_young_other_time_ms = 0.0;
1866 if (_recorded_non_young_regions > 0) {
1867 non_young_other_time_ms =
1868 _recorded_non_young_cset_choice_time_ms +
1869 _recorded_non_young_free_cset_time_ms;
1871 _non_young_other_cost_per_region_ms_seq->add(non_young_other_time_ms /
1872 (double) _recorded_non_young_regions);
1873 }
1875 double constant_other_time_ms = all_other_time_ms -
1876 (young_other_time_ms + non_young_other_time_ms);
1877 _constant_other_time_ms_seq->add(constant_other_time_ms);
1879 double survival_ratio = 0.0;
1880 if (_bytes_in_collection_set_before_gc > 0) {
1881 survival_ratio = (double) bytes_in_to_space_during_gc() /
1882 (double) _bytes_in_collection_set_before_gc;
1883 }
1885 _pending_cards_seq->add((double) _pending_cards);
1886 _scanned_cards_seq->add((double) cards_scanned);
1887 _rs_lengths_seq->add((double) _max_rs_lengths);
1889 double expensive_region_limit_ms =
1890 (double) MaxGCPauseMillis - predict_constant_other_time_ms();
1891 if (expensive_region_limit_ms < 0.0) {
1892 // this means that the other time was predicted to be longer than
1893 // than the max pause time
1894 expensive_region_limit_ms = (double) MaxGCPauseMillis;
1895 }
1896 _expensive_region_limit_ms = expensive_region_limit_ms;
1898 if (PREDICTIONS_VERBOSE) {
1899 gclog_or_tty->print_cr("");
1900 gclog_or_tty->print_cr("PREDICTIONS %1.4lf %d "
1901 "REGIONS %d %d %d %d "
1902 "PENDING_CARDS %d %d "
1903 "CARDS_SCANNED %d %d "
1904 "RS_LENGTHS %d %d "
1905 "SCAN_ONLY_SCAN %1.6lf %1.6lf "
1906 "RS_UPDATE %1.6lf %1.6lf RS_SCAN %1.6lf %1.6lf "
1907 "SURVIVAL_RATIO %1.6lf %1.6lf "
1908 "OBJECT_COPY %1.6lf %1.6lf OTHER_CONSTANT %1.6lf %1.6lf "
1909 "OTHER_YOUNG %1.6lf %1.6lf "
1910 "OTHER_NON_YOUNG %1.6lf %1.6lf "
1911 "VTIME_DIFF %1.6lf TERMINATION %1.6lf "
1912 "ELAPSED %1.6lf %1.6lf ",
1913 _cur_collection_start_sec,
1914 (!_last_young_gc_full) ? 2 :
1915 (last_pause_included_initial_mark) ? 1 : 0,
1916 _recorded_region_num,
1917 _recorded_young_regions,
1918 _recorded_scan_only_regions,
1919 _recorded_non_young_regions,
1920 _predicted_pending_cards, _pending_cards,
1921 _predicted_cards_scanned, cards_scanned,
1922 _predicted_rs_lengths, _max_rs_lengths,
1923 _predicted_scan_only_scan_time_ms, scan_only_time,
1924 _predicted_rs_update_time_ms, update_rs_time,
1925 _predicted_rs_scan_time_ms, scan_rs_time,
1926 _predicted_survival_ratio, survival_ratio,
1927 _predicted_object_copy_time_ms, obj_copy_time,
1928 _predicted_constant_other_time_ms, constant_other_time_ms,
1929 _predicted_young_other_time_ms, young_other_time_ms,
1930 _predicted_non_young_other_time_ms,
1931 non_young_other_time_ms,
1932 _vtime_diff_ms, termination_time,
1933 _predicted_pause_time_ms, elapsed_ms);
1934 }
1936 if (G1PolicyVerbose > 0) {
1937 gclog_or_tty->print_cr("Pause Time, predicted: %1.4lfms (predicted %s), actual: %1.4lfms",
1938 _predicted_pause_time_ms,
1939 (_within_target) ? "within" : "outside",
1940 elapsed_ms);
1941 }
1943 }
1945 _in_marking_window = new_in_marking_window;
1946 _in_marking_window_im = new_in_marking_window_im;
1947 _free_regions_at_end_of_collection = _g1->free_regions();
1948 _scan_only_regions_at_end_of_collection = _g1->young_list_length();
1949 calculate_young_list_min_length();
1950 calculate_young_list_target_config();
1952 // Note that _mmu_tracker->max_gc_time() returns the time in seconds.
1953 double update_rs_time_goal_ms = _mmu_tracker->max_gc_time() * MILLIUNITS * G1RSetUpdatingPauseTimePercent / 100.0;
1954 adjust_concurrent_refinement(update_rs_time, update_rs_processed_buffers, update_rs_time_goal_ms);
1956 // </NEW PREDICTION>
1958 _target_pause_time_ms = -1.0;
1959 }
1961 // <NEW PREDICTION>
1963 void G1CollectorPolicy::adjust_concurrent_refinement(double update_rs_time,
1964 double update_rs_processed_buffers,
1965 double goal_ms) {
1966 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set();
1967 ConcurrentG1Refine *cg1r = G1CollectedHeap::heap()->concurrent_g1_refine();
1969 if (G1UseAdaptiveConcRefinement) {
1970 const int k_gy = 3, k_gr = 6;
1971 const double inc_k = 1.1, dec_k = 0.9;
1973 int g = cg1r->green_zone();
1974 if (update_rs_time > goal_ms) {
1975 g = (int)(g * dec_k); // Can become 0, that's OK. That would mean a mutator-only processing.
1976 } else {
1977 if (update_rs_time < goal_ms && update_rs_processed_buffers > g) {
1978 g = (int)MAX2(g * inc_k, g + 1.0);
1979 }
1980 }
1981 // Change the refinement threads params
1982 cg1r->set_green_zone(g);
1983 cg1r->set_yellow_zone(g * k_gy);
1984 cg1r->set_red_zone(g * k_gr);
1985 cg1r->reinitialize_threads();
1987 int processing_threshold_delta = MAX2((int)(cg1r->green_zone() * sigma()), 1);
1988 int processing_threshold = MIN2(cg1r->green_zone() + processing_threshold_delta,
1989 cg1r->yellow_zone());
1990 // Change the barrier params
1991 dcqs.set_process_completed_threshold(processing_threshold);
1992 dcqs.set_max_completed_queue(cg1r->red_zone());
1993 }
1995 int curr_queue_size = dcqs.completed_buffers_num();
1996 if (curr_queue_size >= cg1r->yellow_zone()) {
1997 dcqs.set_completed_queue_padding(curr_queue_size);
1998 } else {
1999 dcqs.set_completed_queue_padding(0);
2000 }
2001 dcqs.notify_if_necessary();
2002 }
2004 double
2005 G1CollectorPolicy::
2006 predict_young_collection_elapsed_time_ms(size_t adjustment) {
2007 guarantee( adjustment == 0 || adjustment == 1, "invariant" );
2009 G1CollectedHeap* g1h = G1CollectedHeap::heap();
2010 size_t young_num = g1h->young_list_length();
2011 if (young_num == 0)
2012 return 0.0;
2014 young_num += adjustment;
2015 size_t pending_cards = predict_pending_cards();
2016 size_t rs_lengths = g1h->young_list_sampled_rs_lengths() +
2017 predict_rs_length_diff();
2018 size_t card_num;
2019 if (full_young_gcs())
2020 card_num = predict_young_card_num(rs_lengths);
2021 else
2022 card_num = predict_non_young_card_num(rs_lengths);
2023 size_t young_byte_size = young_num * HeapRegion::GrainBytes;
2024 double accum_yg_surv_rate =
2025 _short_lived_surv_rate_group->accum_surv_rate(adjustment);
2027 size_t bytes_to_copy =
2028 (size_t) (accum_yg_surv_rate * (double) HeapRegion::GrainBytes);
2030 return
2031 predict_rs_update_time_ms(pending_cards) +
2032 predict_rs_scan_time_ms(card_num) +
2033 predict_object_copy_time_ms(bytes_to_copy) +
2034 predict_young_other_time_ms(young_num) +
2035 predict_constant_other_time_ms();
2036 }
2038 double
2039 G1CollectorPolicy::predict_base_elapsed_time_ms(size_t pending_cards) {
2040 size_t rs_length = predict_rs_length_diff();
2041 size_t card_num;
2042 if (full_young_gcs())
2043 card_num = predict_young_card_num(rs_length);
2044 else
2045 card_num = predict_non_young_card_num(rs_length);
2046 return predict_base_elapsed_time_ms(pending_cards, card_num);
2047 }
2049 double
2050 G1CollectorPolicy::predict_base_elapsed_time_ms(size_t pending_cards,
2051 size_t scanned_cards) {
2052 return
2053 predict_rs_update_time_ms(pending_cards) +
2054 predict_rs_scan_time_ms(scanned_cards) +
2055 predict_constant_other_time_ms();
2056 }
2058 double
2059 G1CollectorPolicy::predict_region_elapsed_time_ms(HeapRegion* hr,
2060 bool young) {
2061 size_t rs_length = hr->rem_set()->occupied();
2062 size_t card_num;
2063 if (full_young_gcs())
2064 card_num = predict_young_card_num(rs_length);
2065 else
2066 card_num = predict_non_young_card_num(rs_length);
2067 size_t bytes_to_copy = predict_bytes_to_copy(hr);
2069 double region_elapsed_time_ms =
2070 predict_rs_scan_time_ms(card_num) +
2071 predict_object_copy_time_ms(bytes_to_copy);
2073 if (young)
2074 region_elapsed_time_ms += predict_young_other_time_ms(1);
2075 else
2076 region_elapsed_time_ms += predict_non_young_other_time_ms(1);
2078 return region_elapsed_time_ms;
2079 }
2081 size_t
2082 G1CollectorPolicy::predict_bytes_to_copy(HeapRegion* hr) {
2083 size_t bytes_to_copy;
2084 if (hr->is_marked())
2085 bytes_to_copy = hr->max_live_bytes();
2086 else {
2087 guarantee( hr->is_young() && hr->age_in_surv_rate_group() != -1,
2088 "invariant" );
2089 int age = hr->age_in_surv_rate_group();
2090 double yg_surv_rate = predict_yg_surv_rate(age, hr->surv_rate_group());
2091 bytes_to_copy = (size_t) ((double) hr->used() * yg_surv_rate);
2092 }
2094 return bytes_to_copy;
2095 }
2097 void
2098 G1CollectorPolicy::start_recording_regions() {
2099 _recorded_rs_lengths = 0;
2100 _recorded_scan_only_regions = 0;
2101 _recorded_young_regions = 0;
2102 _recorded_non_young_regions = 0;
2104 #if PREDICTIONS_VERBOSE
2105 _predicted_rs_lengths = 0;
2106 _predicted_cards_scanned = 0;
2108 _recorded_marked_bytes = 0;
2109 _recorded_young_bytes = 0;
2110 _predicted_bytes_to_copy = 0;
2111 #endif // PREDICTIONS_VERBOSE
2112 }
2114 void
2115 G1CollectorPolicy::record_cset_region(HeapRegion* hr, bool young) {
2116 if (young) {
2117 ++_recorded_young_regions;
2118 } else {
2119 ++_recorded_non_young_regions;
2120 }
2121 #if PREDICTIONS_VERBOSE
2122 if (young) {
2123 _recorded_young_bytes += hr->used();
2124 } else {
2125 _recorded_marked_bytes += hr->max_live_bytes();
2126 }
2127 _predicted_bytes_to_copy += predict_bytes_to_copy(hr);
2128 #endif // PREDICTIONS_VERBOSE
2130 size_t rs_length = hr->rem_set()->occupied();
2131 _recorded_rs_lengths += rs_length;
2132 }
2134 void
2135 G1CollectorPolicy::record_scan_only_regions(size_t scan_only_length) {
2136 _recorded_scan_only_regions = scan_only_length;
2137 }
2139 void
2140 G1CollectorPolicy::end_recording_regions() {
2141 #if PREDICTIONS_VERBOSE
2142 _predicted_pending_cards = predict_pending_cards();
2143 _predicted_rs_lengths = _recorded_rs_lengths + predict_rs_length_diff();
2144 if (full_young_gcs())
2145 _predicted_cards_scanned += predict_young_card_num(_predicted_rs_lengths);
2146 else
2147 _predicted_cards_scanned +=
2148 predict_non_young_card_num(_predicted_rs_lengths);
2149 _recorded_region_num = _recorded_young_regions + _recorded_non_young_regions;
2151 _predicted_scan_only_scan_time_ms =
2152 predict_scan_only_time_ms(_recorded_scan_only_regions);
2153 _predicted_rs_update_time_ms =
2154 predict_rs_update_time_ms(_g1->pending_card_num());
2155 _predicted_rs_scan_time_ms =
2156 predict_rs_scan_time_ms(_predicted_cards_scanned);
2157 _predicted_object_copy_time_ms =
2158 predict_object_copy_time_ms(_predicted_bytes_to_copy);
2159 _predicted_constant_other_time_ms =
2160 predict_constant_other_time_ms();
2161 _predicted_young_other_time_ms =
2162 predict_young_other_time_ms(_recorded_young_regions);
2163 _predicted_non_young_other_time_ms =
2164 predict_non_young_other_time_ms(_recorded_non_young_regions);
2166 _predicted_pause_time_ms =
2167 _predicted_scan_only_scan_time_ms +
2168 _predicted_rs_update_time_ms +
2169 _predicted_rs_scan_time_ms +
2170 _predicted_object_copy_time_ms +
2171 _predicted_constant_other_time_ms +
2172 _predicted_young_other_time_ms +
2173 _predicted_non_young_other_time_ms;
2174 #endif // PREDICTIONS_VERBOSE
2175 }
2177 void G1CollectorPolicy::check_if_region_is_too_expensive(double
2178 predicted_time_ms) {
2179 // I don't think we need to do this when in young GC mode since
2180 // marking will be initiated next time we hit the soft limit anyway...
2181 if (predicted_time_ms > _expensive_region_limit_ms) {
2182 if (!in_young_gc_mode()) {
2183 set_full_young_gcs(true);
2184 // We might want to do something different here. However,
2185 // right now we don't support the non-generational G1 mode
2186 // (and in fact we are planning to remove the associated code,
2187 // see CR 6814390). So, let's leave it as is and this will be
2188 // removed some time in the future
2189 ShouldNotReachHere();
2190 set_during_initial_mark_pause();
2191 } else
2192 // no point in doing another partial one
2193 _should_revert_to_full_young_gcs = true;
2194 }
2195 }
2197 // </NEW PREDICTION>
2200 void G1CollectorPolicy::update_recent_gc_times(double end_time_sec,
2201 double elapsed_ms) {
2202 _recent_gc_times_ms->add(elapsed_ms);
2203 _recent_prev_end_times_for_all_gcs_sec->add(end_time_sec);
2204 _prev_collection_pause_end_ms = end_time_sec * 1000.0;
2205 }
2207 double G1CollectorPolicy::recent_avg_time_for_pauses_ms() {
2208 if (_recent_pause_times_ms->num() == 0) return (double) MaxGCPauseMillis;
2209 else return _recent_pause_times_ms->avg();
2210 }
2212 double G1CollectorPolicy::recent_avg_time_for_CH_strong_ms() {
2213 if (_recent_CH_strong_roots_times_ms->num() == 0)
2214 return (double)MaxGCPauseMillis/3.0;
2215 else return _recent_CH_strong_roots_times_ms->avg();
2216 }
2218 double G1CollectorPolicy::recent_avg_time_for_G1_strong_ms() {
2219 if (_recent_G1_strong_roots_times_ms->num() == 0)
2220 return (double)MaxGCPauseMillis/3.0;
2221 else return _recent_G1_strong_roots_times_ms->avg();
2222 }
2224 double G1CollectorPolicy::recent_avg_time_for_evac_ms() {
2225 if (_recent_evac_times_ms->num() == 0) return (double)MaxGCPauseMillis/3.0;
2226 else return _recent_evac_times_ms->avg();
2227 }
2229 int G1CollectorPolicy::number_of_recent_gcs() {
2230 assert(_recent_CH_strong_roots_times_ms->num() ==
2231 _recent_G1_strong_roots_times_ms->num(), "Sequence out of sync");
2232 assert(_recent_G1_strong_roots_times_ms->num() ==
2233 _recent_evac_times_ms->num(), "Sequence out of sync");
2234 assert(_recent_evac_times_ms->num() ==
2235 _recent_pause_times_ms->num(), "Sequence out of sync");
2236 assert(_recent_pause_times_ms->num() ==
2237 _recent_CS_bytes_used_before->num(), "Sequence out of sync");
2238 assert(_recent_CS_bytes_used_before->num() ==
2239 _recent_CS_bytes_surviving->num(), "Sequence out of sync");
2240 return _recent_pause_times_ms->num();
2241 }
2243 double G1CollectorPolicy::recent_avg_survival_fraction() {
2244 return recent_avg_survival_fraction_work(_recent_CS_bytes_surviving,
2245 _recent_CS_bytes_used_before);
2246 }
2248 double G1CollectorPolicy::last_survival_fraction() {
2249 return last_survival_fraction_work(_recent_CS_bytes_surviving,
2250 _recent_CS_bytes_used_before);
2251 }
2253 double
2254 G1CollectorPolicy::recent_avg_survival_fraction_work(TruncatedSeq* surviving,
2255 TruncatedSeq* before) {
2256 assert(surviving->num() == before->num(), "Sequence out of sync");
2257 if (before->sum() > 0.0) {
2258 double recent_survival_rate = surviving->sum() / before->sum();
2259 // We exempt parallel collection from this check because Alloc Buffer
2260 // fragmentation can produce negative collections.
2261 // Further, we're now always doing parallel collection. But I'm still
2262 // leaving this here as a placeholder for a more precise assertion later.
2263 // (DLD, 10/05.)
2264 assert((true || ParallelGCThreads > 0) ||
2265 _g1->evacuation_failed() ||
2266 recent_survival_rate <= 1.0, "Or bad frac");
2267 return recent_survival_rate;
2268 } else {
2269 return 1.0; // Be conservative.
2270 }
2271 }
2273 double
2274 G1CollectorPolicy::last_survival_fraction_work(TruncatedSeq* surviving,
2275 TruncatedSeq* before) {
2276 assert(surviving->num() == before->num(), "Sequence out of sync");
2277 if (surviving->num() > 0 && before->last() > 0.0) {
2278 double last_survival_rate = surviving->last() / before->last();
2279 // We exempt parallel collection from this check because Alloc Buffer
2280 // fragmentation can produce negative collections.
2281 // Further, we're now always doing parallel collection. But I'm still
2282 // leaving this here as a placeholder for a more precise assertion later.
2283 // (DLD, 10/05.)
2284 assert((true || ParallelGCThreads > 0) ||
2285 last_survival_rate <= 1.0, "Or bad frac");
2286 return last_survival_rate;
2287 } else {
2288 return 1.0;
2289 }
2290 }
2292 static const int survival_min_obs = 5;
2293 static double survival_min_obs_limits[] = { 0.9, 0.7, 0.5, 0.3, 0.1 };
2294 static const double min_survival_rate = 0.1;
2296 double
2297 G1CollectorPolicy::conservative_avg_survival_fraction_work(double avg,
2298 double latest) {
2299 double res = avg;
2300 if (number_of_recent_gcs() < survival_min_obs) {
2301 res = MAX2(res, survival_min_obs_limits[number_of_recent_gcs()]);
2302 }
2303 res = MAX2(res, latest);
2304 res = MAX2(res, min_survival_rate);
2305 // In the parallel case, LAB fragmentation can produce "negative
2306 // collections"; so can evac failure. Cap at 1.0
2307 res = MIN2(res, 1.0);
2308 return res;
2309 }
2311 size_t G1CollectorPolicy::expansion_amount() {
2312 if ((recent_avg_pause_time_ratio() * 100.0) > _gc_overhead_perc) {
2313 // We will double the existing space, or take
2314 // G1ExpandByPercentOfAvailable % of the available expansion
2315 // space, whichever is smaller, bounded below by a minimum
2316 // expansion (unless that's all that's left.)
2317 const size_t min_expand_bytes = 1*M;
2318 size_t reserved_bytes = _g1->g1_reserved_obj_bytes();
2319 size_t committed_bytes = _g1->capacity();
2320 size_t uncommitted_bytes = reserved_bytes - committed_bytes;
2321 size_t expand_bytes;
2322 size_t expand_bytes_via_pct =
2323 uncommitted_bytes * G1ExpandByPercentOfAvailable / 100;
2324 expand_bytes = MIN2(expand_bytes_via_pct, committed_bytes);
2325 expand_bytes = MAX2(expand_bytes, min_expand_bytes);
2326 expand_bytes = MIN2(expand_bytes, uncommitted_bytes);
2327 if (G1PolicyVerbose > 1) {
2328 gclog_or_tty->print("Decided to expand: ratio = %5.2f, "
2329 "committed = %d%s, uncommited = %d%s, via pct = %d%s.\n"
2330 " Answer = %d.\n",
2331 recent_avg_pause_time_ratio(),
2332 byte_size_in_proper_unit(committed_bytes),
2333 proper_unit_for_byte_size(committed_bytes),
2334 byte_size_in_proper_unit(uncommitted_bytes),
2335 proper_unit_for_byte_size(uncommitted_bytes),
2336 byte_size_in_proper_unit(expand_bytes_via_pct),
2337 proper_unit_for_byte_size(expand_bytes_via_pct),
2338 byte_size_in_proper_unit(expand_bytes),
2339 proper_unit_for_byte_size(expand_bytes));
2340 }
2341 return expand_bytes;
2342 } else {
2343 return 0;
2344 }
2345 }
2347 void G1CollectorPolicy::note_start_of_mark_thread() {
2348 _mark_thread_startup_sec = os::elapsedTime();
2349 }
2351 class CountCSClosure: public HeapRegionClosure {
2352 G1CollectorPolicy* _g1_policy;
2353 public:
2354 CountCSClosure(G1CollectorPolicy* g1_policy) :
2355 _g1_policy(g1_policy) {}
2356 bool doHeapRegion(HeapRegion* r) {
2357 _g1_policy->_bytes_in_collection_set_before_gc += r->used();
2358 return false;
2359 }
2360 };
2362 void G1CollectorPolicy::count_CS_bytes_used() {
2363 CountCSClosure cs_closure(this);
2364 _g1->collection_set_iterate(&cs_closure);
2365 }
2367 static void print_indent(int level) {
2368 for (int j = 0; j < level+1; ++j)
2369 gclog_or_tty->print(" ");
2370 }
2372 void G1CollectorPolicy::print_summary (int level,
2373 const char* str,
2374 NumberSeq* seq) const {
2375 double sum = seq->sum();
2376 print_indent(level);
2377 gclog_or_tty->print_cr("%-24s = %8.2lf s (avg = %8.2lf ms)",
2378 str, sum / 1000.0, seq->avg());
2379 }
2381 void G1CollectorPolicy::print_summary_sd (int level,
2382 const char* str,
2383 NumberSeq* seq) const {
2384 print_summary(level, str, seq);
2385 print_indent(level + 5);
2386 gclog_or_tty->print_cr("(num = %5d, std dev = %8.2lf ms, max = %8.2lf ms)",
2387 seq->num(), seq->sd(), seq->maximum());
2388 }
2390 void G1CollectorPolicy::check_other_times(int level,
2391 NumberSeq* other_times_ms,
2392 NumberSeq* calc_other_times_ms) const {
2393 bool should_print = false;
2395 double max_sum = MAX2(fabs(other_times_ms->sum()),
2396 fabs(calc_other_times_ms->sum()));
2397 double min_sum = MIN2(fabs(other_times_ms->sum()),
2398 fabs(calc_other_times_ms->sum()));
2399 double sum_ratio = max_sum / min_sum;
2400 if (sum_ratio > 1.1) {
2401 should_print = true;
2402 print_indent(level + 1);
2403 gclog_or_tty->print_cr("## CALCULATED OTHER SUM DOESN'T MATCH RECORDED ###");
2404 }
2406 double max_avg = MAX2(fabs(other_times_ms->avg()),
2407 fabs(calc_other_times_ms->avg()));
2408 double min_avg = MIN2(fabs(other_times_ms->avg()),
2409 fabs(calc_other_times_ms->avg()));
2410 double avg_ratio = max_avg / min_avg;
2411 if (avg_ratio > 1.1) {
2412 should_print = true;
2413 print_indent(level + 1);
2414 gclog_or_tty->print_cr("## CALCULATED OTHER AVG DOESN'T MATCH RECORDED ###");
2415 }
2417 if (other_times_ms->sum() < -0.01) {
2418 print_indent(level + 1);
2419 gclog_or_tty->print_cr("## RECORDED OTHER SUM IS NEGATIVE ###");
2420 }
2422 if (other_times_ms->avg() < -0.01) {
2423 print_indent(level + 1);
2424 gclog_or_tty->print_cr("## RECORDED OTHER AVG IS NEGATIVE ###");
2425 }
2427 if (calc_other_times_ms->sum() < -0.01) {
2428 should_print = true;
2429 print_indent(level + 1);
2430 gclog_or_tty->print_cr("## CALCULATED OTHER SUM IS NEGATIVE ###");
2431 }
2433 if (calc_other_times_ms->avg() < -0.01) {
2434 should_print = true;
2435 print_indent(level + 1);
2436 gclog_or_tty->print_cr("## CALCULATED OTHER AVG IS NEGATIVE ###");
2437 }
2439 if (should_print)
2440 print_summary(level, "Other(Calc)", calc_other_times_ms);
2441 }
2443 void G1CollectorPolicy::print_summary(PauseSummary* summary) const {
2444 bool parallel = ParallelGCThreads > 0;
2445 MainBodySummary* body_summary = summary->main_body_summary();
2446 if (summary->get_total_seq()->num() > 0) {
2447 print_summary_sd(0, "Evacuation Pauses", summary->get_total_seq());
2448 if (body_summary != NULL) {
2449 print_summary(1, "SATB Drain", body_summary->get_satb_drain_seq());
2450 if (parallel) {
2451 print_summary(1, "Parallel Time", body_summary->get_parallel_seq());
2452 print_summary(2, "Update RS", body_summary->get_update_rs_seq());
2453 print_summary(2, "Ext Root Scanning",
2454 body_summary->get_ext_root_scan_seq());
2455 print_summary(2, "Mark Stack Scanning",
2456 body_summary->get_mark_stack_scan_seq());
2457 print_summary(2, "Scan-Only Scanning",
2458 body_summary->get_scan_only_seq());
2459 print_summary(2, "Scan RS", body_summary->get_scan_rs_seq());
2460 print_summary(2, "Object Copy", body_summary->get_obj_copy_seq());
2461 print_summary(2, "Termination", body_summary->get_termination_seq());
2462 print_summary(2, "Other", body_summary->get_parallel_other_seq());
2463 {
2464 NumberSeq* other_parts[] = {
2465 body_summary->get_update_rs_seq(),
2466 body_summary->get_ext_root_scan_seq(),
2467 body_summary->get_mark_stack_scan_seq(),
2468 body_summary->get_scan_only_seq(),
2469 body_summary->get_scan_rs_seq(),
2470 body_summary->get_obj_copy_seq(),
2471 body_summary->get_termination_seq()
2472 };
2473 NumberSeq calc_other_times_ms(body_summary->get_parallel_seq(),
2474 7, other_parts);
2475 check_other_times(2, body_summary->get_parallel_other_seq(),
2476 &calc_other_times_ms);
2477 }
2478 print_summary(1, "Mark Closure", body_summary->get_mark_closure_seq());
2479 print_summary(1, "Clear CT", body_summary->get_clear_ct_seq());
2480 } else {
2481 print_summary(1, "Update RS", body_summary->get_update_rs_seq());
2482 print_summary(1, "Ext Root Scanning",
2483 body_summary->get_ext_root_scan_seq());
2484 print_summary(1, "Mark Stack Scanning",
2485 body_summary->get_mark_stack_scan_seq());
2486 print_summary(1, "Scan-Only Scanning",
2487 body_summary->get_scan_only_seq());
2488 print_summary(1, "Scan RS", body_summary->get_scan_rs_seq());
2489 print_summary(1, "Object Copy", body_summary->get_obj_copy_seq());
2490 }
2491 }
2492 print_summary(1, "Other", summary->get_other_seq());
2493 {
2494 NumberSeq calc_other_times_ms;
2495 if (body_summary != NULL) {
2496 // not abandoned
2497 if (parallel) {
2498 // parallel
2499 NumberSeq* other_parts[] = {
2500 body_summary->get_satb_drain_seq(),
2501 body_summary->get_parallel_seq(),
2502 body_summary->get_clear_ct_seq()
2503 };
2504 calc_other_times_ms = NumberSeq(summary->get_total_seq(),
2505 3, other_parts);
2506 } else {
2507 // serial
2508 NumberSeq* other_parts[] = {
2509 body_summary->get_satb_drain_seq(),
2510 body_summary->get_update_rs_seq(),
2511 body_summary->get_ext_root_scan_seq(),
2512 body_summary->get_mark_stack_scan_seq(),
2513 body_summary->get_scan_only_seq(),
2514 body_summary->get_scan_rs_seq(),
2515 body_summary->get_obj_copy_seq()
2516 };
2517 calc_other_times_ms = NumberSeq(summary->get_total_seq(),
2518 7, other_parts);
2519 }
2520 } else {
2521 // abandoned
2522 calc_other_times_ms = NumberSeq();
2523 }
2524 check_other_times(1, summary->get_other_seq(), &calc_other_times_ms);
2525 }
2526 } else {
2527 print_indent(0);
2528 gclog_or_tty->print_cr("none");
2529 }
2530 gclog_or_tty->print_cr("");
2531 }
2533 void
2534 G1CollectorPolicy::print_abandoned_summary(PauseSummary* summary) const {
2535 bool printed = false;
2536 if (summary->get_total_seq()->num() > 0) {
2537 printed = true;
2538 print_summary(summary);
2539 }
2540 if (!printed) {
2541 print_indent(0);
2542 gclog_or_tty->print_cr("none");
2543 gclog_or_tty->print_cr("");
2544 }
2545 }
2547 void G1CollectorPolicy::print_tracing_info() const {
2548 if (TraceGen0Time) {
2549 gclog_or_tty->print_cr("ALL PAUSES");
2550 print_summary_sd(0, "Total", _all_pause_times_ms);
2551 gclog_or_tty->print_cr("");
2552 gclog_or_tty->print_cr("");
2553 gclog_or_tty->print_cr(" Full Young GC Pauses: %8d", _full_young_pause_num);
2554 gclog_or_tty->print_cr(" Partial Young GC Pauses: %8d", _partial_young_pause_num);
2555 gclog_or_tty->print_cr("");
2557 gclog_or_tty->print_cr("EVACUATION PAUSES");
2558 print_summary(_summary);
2560 gclog_or_tty->print_cr("ABANDONED PAUSES");
2561 print_abandoned_summary(_abandoned_summary);
2563 gclog_or_tty->print_cr("MISC");
2564 print_summary_sd(0, "Stop World", _all_stop_world_times_ms);
2565 print_summary_sd(0, "Yields", _all_yield_times_ms);
2566 for (int i = 0; i < _aux_num; ++i) {
2567 if (_all_aux_times_ms[i].num() > 0) {
2568 char buffer[96];
2569 sprintf(buffer, "Aux%d", i);
2570 print_summary_sd(0, buffer, &_all_aux_times_ms[i]);
2571 }
2572 }
2574 size_t all_region_num = _region_num_young + _region_num_tenured;
2575 gclog_or_tty->print_cr(" New Regions %8d, Young %8d (%6.2lf%%), "
2576 "Tenured %8d (%6.2lf%%)",
2577 all_region_num,
2578 _region_num_young,
2579 (double) _region_num_young / (double) all_region_num * 100.0,
2580 _region_num_tenured,
2581 (double) _region_num_tenured / (double) all_region_num * 100.0);
2582 }
2583 if (TraceGen1Time) {
2584 if (_all_full_gc_times_ms->num() > 0) {
2585 gclog_or_tty->print("\n%4d full_gcs: total time = %8.2f s",
2586 _all_full_gc_times_ms->num(),
2587 _all_full_gc_times_ms->sum() / 1000.0);
2588 gclog_or_tty->print_cr(" (avg = %8.2fms).", _all_full_gc_times_ms->avg());
2589 gclog_or_tty->print_cr(" [std. dev = %8.2f ms, max = %8.2f ms]",
2590 _all_full_gc_times_ms->sd(),
2591 _all_full_gc_times_ms->maximum());
2592 }
2593 }
2594 }
2596 void G1CollectorPolicy::print_yg_surv_rate_info() const {
2597 #ifndef PRODUCT
2598 _short_lived_surv_rate_group->print_surv_rate_summary();
2599 // add this call for any other surv rate groups
2600 #endif // PRODUCT
2601 }
2603 bool
2604 G1CollectorPolicy::should_add_next_region_to_young_list() {
2605 assert(in_young_gc_mode(), "should be in young GC mode");
2606 bool ret;
2607 size_t young_list_length = _g1->young_list_length();
2608 size_t young_list_max_length = _young_list_target_length;
2609 if (G1FixedEdenSize) {
2610 young_list_max_length -= _max_survivor_regions;
2611 }
2612 if (young_list_length < young_list_max_length) {
2613 ret = true;
2614 ++_region_num_young;
2615 } else {
2616 ret = false;
2617 ++_region_num_tenured;
2618 }
2620 return ret;
2621 }
2623 #ifndef PRODUCT
2624 // for debugging, bit of a hack...
2625 static char*
2626 region_num_to_mbs(int length) {
2627 static char buffer[64];
2628 double bytes = (double) (length * HeapRegion::GrainBytes);
2629 double mbs = bytes / (double) (1024 * 1024);
2630 sprintf(buffer, "%7.2lfMB", mbs);
2631 return buffer;
2632 }
2633 #endif // PRODUCT
2635 size_t G1CollectorPolicy::max_regions(int purpose) {
2636 switch (purpose) {
2637 case GCAllocForSurvived:
2638 return _max_survivor_regions;
2639 case GCAllocForTenured:
2640 return REGIONS_UNLIMITED;
2641 default:
2642 ShouldNotReachHere();
2643 return REGIONS_UNLIMITED;
2644 };
2645 }
2647 // Calculates survivor space parameters.
2648 void G1CollectorPolicy::calculate_survivors_policy()
2649 {
2650 if (G1FixedSurvivorSpaceSize == 0) {
2651 _max_survivor_regions = _young_list_target_length / SurvivorRatio;
2652 } else {
2653 _max_survivor_regions = G1FixedSurvivorSpaceSize / HeapRegion::GrainBytes;
2654 }
2656 if (G1FixedTenuringThreshold) {
2657 _tenuring_threshold = MaxTenuringThreshold;
2658 } else {
2659 _tenuring_threshold = _survivors_age_table.compute_tenuring_threshold(
2660 HeapRegion::GrainWords * _max_survivor_regions);
2661 }
2662 }
2664 bool
2665 G1CollectorPolicy_BestRegionsFirst::should_do_collection_pause(size_t
2666 word_size) {
2667 assert(_g1->regions_accounted_for(), "Region leakage!");
2668 double max_pause_time_ms = _mmu_tracker->max_gc_time() * 1000.0;
2670 size_t young_list_length = _g1->young_list_length();
2671 size_t young_list_max_length = _young_list_target_length;
2672 if (G1FixedEdenSize) {
2673 young_list_max_length -= _max_survivor_regions;
2674 }
2675 bool reached_target_length = young_list_length >= young_list_max_length;
2677 if (in_young_gc_mode()) {
2678 if (reached_target_length) {
2679 assert( young_list_length > 0 && _g1->young_list_length() > 0,
2680 "invariant" );
2681 _target_pause_time_ms = max_pause_time_ms;
2682 return true;
2683 }
2684 } else {
2685 guarantee( false, "should not reach here" );
2686 }
2688 return false;
2689 }
2691 #ifndef PRODUCT
2692 class HRSortIndexIsOKClosure: public HeapRegionClosure {
2693 CollectionSetChooser* _chooser;
2694 public:
2695 HRSortIndexIsOKClosure(CollectionSetChooser* chooser) :
2696 _chooser(chooser) {}
2698 bool doHeapRegion(HeapRegion* r) {
2699 if (!r->continuesHumongous()) {
2700 assert(_chooser->regionProperlyOrdered(r), "Ought to be.");
2701 }
2702 return false;
2703 }
2704 };
2706 bool G1CollectorPolicy_BestRegionsFirst::assertMarkedBytesDataOK() {
2707 HRSortIndexIsOKClosure cl(_collectionSetChooser);
2708 _g1->heap_region_iterate(&cl);
2709 return true;
2710 }
2711 #endif
2713 void
2714 G1CollectorPolicy::decide_on_conc_mark_initiation() {
2715 // We are about to decide on whether this pause will be an
2716 // initial-mark pause.
2718 // First, during_initial_mark_pause() should not be already set. We
2719 // will set it here if we have to. However, it should be cleared by
2720 // the end of the pause (it's only set for the duration of an
2721 // initial-mark pause).
2722 assert(!during_initial_mark_pause(), "pre-condition");
2724 if (initiate_conc_mark_if_possible()) {
2725 // We had noticed on a previous pause that the heap occupancy has
2726 // gone over the initiating threshold and we should start a
2727 // concurrent marking cycle. So we might initiate one.
2729 bool during_cycle = _g1->concurrent_mark()->cmThread()->during_cycle();
2730 if (!during_cycle) {
2731 // The concurrent marking thread is not "during a cycle", i.e.,
2732 // it has completed the last one. So we can go ahead and
2733 // initiate a new cycle.
2735 set_during_initial_mark_pause();
2737 // And we can now clear initiate_conc_mark_if_possible() as
2738 // we've already acted on it.
2739 clear_initiate_conc_mark_if_possible();
2740 } else {
2741 // The concurrent marking thread is still finishing up the
2742 // previous cycle. If we start one right now the two cycles
2743 // overlap. In particular, the concurrent marking thread might
2744 // be in the process of clearing the next marking bitmap (which
2745 // we will use for the next cycle if we start one). Starting a
2746 // cycle now will be bad given that parts of the marking
2747 // information might get cleared by the marking thread. And we
2748 // cannot wait for the marking thread to finish the cycle as it
2749 // periodically yields while clearing the next marking bitmap
2750 // and, if it's in a yield point, it's waiting for us to
2751 // finish. So, at this point we will not start a cycle and we'll
2752 // let the concurrent marking thread complete the last one.
2753 }
2754 }
2755 }
2757 void
2758 G1CollectorPolicy_BestRegionsFirst::
2759 record_collection_pause_start(double start_time_sec, size_t start_used) {
2760 G1CollectorPolicy::record_collection_pause_start(start_time_sec, start_used);
2761 }
2763 class NextNonCSElemFinder: public HeapRegionClosure {
2764 HeapRegion* _res;
2765 public:
2766 NextNonCSElemFinder(): _res(NULL) {}
2767 bool doHeapRegion(HeapRegion* r) {
2768 if (!r->in_collection_set()) {
2769 _res = r;
2770 return true;
2771 } else {
2772 return false;
2773 }
2774 }
2775 HeapRegion* res() { return _res; }
2776 };
2778 class KnownGarbageClosure: public HeapRegionClosure {
2779 CollectionSetChooser* _hrSorted;
2781 public:
2782 KnownGarbageClosure(CollectionSetChooser* hrSorted) :
2783 _hrSorted(hrSorted)
2784 {}
2786 bool doHeapRegion(HeapRegion* r) {
2787 // We only include humongous regions in collection
2788 // sets when concurrent mark shows that their contained object is
2789 // unreachable.
2791 // Do we have any marking information for this region?
2792 if (r->is_marked()) {
2793 // We don't include humongous regions in collection
2794 // sets because we collect them immediately at the end of a marking
2795 // cycle. We also don't include young regions because we *must*
2796 // include them in the next collection pause.
2797 if (!r->isHumongous() && !r->is_young()) {
2798 _hrSorted->addMarkedHeapRegion(r);
2799 }
2800 }
2801 return false;
2802 }
2803 };
2805 class ParKnownGarbageHRClosure: public HeapRegionClosure {
2806 CollectionSetChooser* _hrSorted;
2807 jint _marked_regions_added;
2808 jint _chunk_size;
2809 jint _cur_chunk_idx;
2810 jint _cur_chunk_end; // Cur chunk [_cur_chunk_idx, _cur_chunk_end)
2811 int _worker;
2812 int _invokes;
2814 void get_new_chunk() {
2815 _cur_chunk_idx = _hrSorted->getParMarkedHeapRegionChunk(_chunk_size);
2816 _cur_chunk_end = _cur_chunk_idx + _chunk_size;
2817 }
2818 void add_region(HeapRegion* r) {
2819 if (_cur_chunk_idx == _cur_chunk_end) {
2820 get_new_chunk();
2821 }
2822 assert(_cur_chunk_idx < _cur_chunk_end, "postcondition");
2823 _hrSorted->setMarkedHeapRegion(_cur_chunk_idx, r);
2824 _marked_regions_added++;
2825 _cur_chunk_idx++;
2826 }
2828 public:
2829 ParKnownGarbageHRClosure(CollectionSetChooser* hrSorted,
2830 jint chunk_size,
2831 int worker) :
2832 _hrSorted(hrSorted), _chunk_size(chunk_size), _worker(worker),
2833 _marked_regions_added(0), _cur_chunk_idx(0), _cur_chunk_end(0),
2834 _invokes(0)
2835 {}
2837 bool doHeapRegion(HeapRegion* r) {
2838 // We only include humongous regions in collection
2839 // sets when concurrent mark shows that their contained object is
2840 // unreachable.
2841 _invokes++;
2843 // Do we have any marking information for this region?
2844 if (r->is_marked()) {
2845 // We don't include humongous regions in collection
2846 // sets because we collect them immediately at the end of a marking
2847 // cycle.
2848 // We also do not include young regions in collection sets
2849 if (!r->isHumongous() && !r->is_young()) {
2850 add_region(r);
2851 }
2852 }
2853 return false;
2854 }
2855 jint marked_regions_added() { return _marked_regions_added; }
2856 int invokes() { return _invokes; }
2857 };
2859 class ParKnownGarbageTask: public AbstractGangTask {
2860 CollectionSetChooser* _hrSorted;
2861 jint _chunk_size;
2862 G1CollectedHeap* _g1;
2863 public:
2864 ParKnownGarbageTask(CollectionSetChooser* hrSorted, jint chunk_size) :
2865 AbstractGangTask("ParKnownGarbageTask"),
2866 _hrSorted(hrSorted), _chunk_size(chunk_size),
2867 _g1(G1CollectedHeap::heap())
2868 {}
2870 void work(int i) {
2871 ParKnownGarbageHRClosure parKnownGarbageCl(_hrSorted, _chunk_size, i);
2872 // Back to zero for the claim value.
2873 _g1->heap_region_par_iterate_chunked(&parKnownGarbageCl, i,
2874 HeapRegion::InitialClaimValue);
2875 jint regions_added = parKnownGarbageCl.marked_regions_added();
2876 _hrSorted->incNumMarkedHeapRegions(regions_added);
2877 if (G1PrintParCleanupStats) {
2878 gclog_or_tty->print(" Thread %d called %d times, added %d regions to list.\n",
2879 i, parKnownGarbageCl.invokes(), regions_added);
2880 }
2881 }
2882 };
2884 void
2885 G1CollectorPolicy_BestRegionsFirst::
2886 record_concurrent_mark_cleanup_end(size_t freed_bytes,
2887 size_t max_live_bytes) {
2888 double start;
2889 if (G1PrintParCleanupStats) start = os::elapsedTime();
2890 record_concurrent_mark_cleanup_end_work1(freed_bytes, max_live_bytes);
2892 _collectionSetChooser->clearMarkedHeapRegions();
2893 double clear_marked_end;
2894 if (G1PrintParCleanupStats) {
2895 clear_marked_end = os::elapsedTime();
2896 gclog_or_tty->print_cr(" clear marked regions + work1: %8.3f ms.",
2897 (clear_marked_end - start)*1000.0);
2898 }
2899 if (ParallelGCThreads > 0) {
2900 const size_t OverpartitionFactor = 4;
2901 const size_t MinChunkSize = 8;
2902 const size_t ChunkSize =
2903 MAX2(_g1->n_regions() / (ParallelGCThreads * OverpartitionFactor),
2904 MinChunkSize);
2905 _collectionSetChooser->prepareForAddMarkedHeapRegionsPar(_g1->n_regions(),
2906 ChunkSize);
2907 ParKnownGarbageTask parKnownGarbageTask(_collectionSetChooser,
2908 (int) ChunkSize);
2909 _g1->workers()->run_task(&parKnownGarbageTask);
2911 assert(_g1->check_heap_region_claim_values(HeapRegion::InitialClaimValue),
2912 "sanity check");
2913 } else {
2914 KnownGarbageClosure knownGarbagecl(_collectionSetChooser);
2915 _g1->heap_region_iterate(&knownGarbagecl);
2916 }
2917 double known_garbage_end;
2918 if (G1PrintParCleanupStats) {
2919 known_garbage_end = os::elapsedTime();
2920 gclog_or_tty->print_cr(" compute known garbage: %8.3f ms.",
2921 (known_garbage_end - clear_marked_end)*1000.0);
2922 }
2923 _collectionSetChooser->sortMarkedHeapRegions();
2924 double sort_end;
2925 if (G1PrintParCleanupStats) {
2926 sort_end = os::elapsedTime();
2927 gclog_or_tty->print_cr(" sorting: %8.3f ms.",
2928 (sort_end - known_garbage_end)*1000.0);
2929 }
2931 record_concurrent_mark_cleanup_end_work2();
2932 double work2_end;
2933 if (G1PrintParCleanupStats) {
2934 work2_end = os::elapsedTime();
2935 gclog_or_tty->print_cr(" work2: %8.3f ms.",
2936 (work2_end - sort_end)*1000.0);
2937 }
2938 }
2940 // Add the heap region to the collection set and return the conservative
2941 // estimate of the number of live bytes.
2942 void G1CollectorPolicy::
2943 add_to_collection_set(HeapRegion* hr) {
2944 if (G1PrintHeapRegions) {
2945 gclog_or_tty->print_cr("added region to cset "
2946 "%d:["PTR_FORMAT", "PTR_FORMAT"], "
2947 "top "PTR_FORMAT", %s",
2948 hr->hrs_index(), hr->bottom(), hr->end(),
2949 hr->top(), hr->is_young() ? "YOUNG" : "NOT_YOUNG");
2950 }
2952 if (_g1->mark_in_progress())
2953 _g1->concurrent_mark()->registerCSetRegion(hr);
2955 assert(!hr->in_collection_set(),
2956 "should not already be in the CSet");
2957 hr->set_in_collection_set(true);
2958 hr->set_next_in_collection_set(_collection_set);
2959 _collection_set = hr;
2960 _collection_set_size++;
2961 _collection_set_bytes_used_before += hr->used();
2962 _g1->register_region_with_in_cset_fast_test(hr);
2963 }
2965 void
2966 G1CollectorPolicy_BestRegionsFirst::
2967 choose_collection_set() {
2968 double non_young_start_time_sec;
2969 start_recording_regions();
2971 guarantee(_target_pause_time_ms > -1.0
2972 NOT_PRODUCT(|| Universe::heap()->gc_cause() == GCCause::_scavenge_alot),
2973 "_target_pause_time_ms should have been set!");
2974 #ifndef PRODUCT
2975 if (_target_pause_time_ms <= -1.0) {
2976 assert(ScavengeALot && Universe::heap()->gc_cause() == GCCause::_scavenge_alot, "Error");
2977 _target_pause_time_ms = _mmu_tracker->max_gc_time() * 1000.0;
2978 }
2979 #endif
2980 assert(_collection_set == NULL, "Precondition");
2982 double base_time_ms = predict_base_elapsed_time_ms(_pending_cards);
2983 double predicted_pause_time_ms = base_time_ms;
2985 double target_time_ms = _target_pause_time_ms;
2986 double time_remaining_ms = target_time_ms - base_time_ms;
2988 // the 10% and 50% values are arbitrary...
2989 if (time_remaining_ms < 0.10*target_time_ms) {
2990 time_remaining_ms = 0.50 * target_time_ms;
2991 _within_target = false;
2992 } else {
2993 _within_target = true;
2994 }
2996 // We figure out the number of bytes available for future to-space.
2997 // For new regions without marking information, we must assume the
2998 // worst-case of complete survival. If we have marking information for a
2999 // region, we can bound the amount of live data. We can add a number of
3000 // such regions, as long as the sum of the live data bounds does not
3001 // exceed the available evacuation space.
3002 size_t max_live_bytes = _g1->free_regions() * HeapRegion::GrainBytes;
3004 size_t expansion_bytes =
3005 _g1->expansion_regions() * HeapRegion::GrainBytes;
3007 _collection_set_bytes_used_before = 0;
3008 _collection_set_size = 0;
3010 // Adjust for expansion and slop.
3011 max_live_bytes = max_live_bytes + expansion_bytes;
3013 assert(_g1->regions_accounted_for(), "Region leakage!");
3015 HeapRegion* hr;
3016 if (in_young_gc_mode()) {
3017 double young_start_time_sec = os::elapsedTime();
3019 if (G1PolicyVerbose > 0) {
3020 gclog_or_tty->print_cr("Adding %d young regions to the CSet",
3021 _g1->young_list_length());
3022 }
3023 _young_cset_length = 0;
3024 _last_young_gc_full = full_young_gcs() ? true : false;
3025 if (_last_young_gc_full)
3026 ++_full_young_pause_num;
3027 else
3028 ++_partial_young_pause_num;
3029 hr = _g1->pop_region_from_young_list();
3030 while (hr != NULL) {
3032 assert( hr->young_index_in_cset() == -1, "invariant" );
3033 assert( hr->age_in_surv_rate_group() != -1, "invariant" );
3034 hr->set_young_index_in_cset((int) _young_cset_length);
3036 ++_young_cset_length;
3037 double predicted_time_ms = predict_region_elapsed_time_ms(hr, true);
3038 time_remaining_ms -= predicted_time_ms;
3039 predicted_pause_time_ms += predicted_time_ms;
3040 assert(!hr->in_collection_set(), "invariant");
3041 add_to_collection_set(hr);
3042 record_cset_region(hr, true);
3043 max_live_bytes -= MIN2(hr->max_live_bytes(), max_live_bytes);
3044 if (G1PolicyVerbose > 0) {
3045 gclog_or_tty->print_cr(" Added [" PTR_FORMAT ", " PTR_FORMAT") to CS.",
3046 hr->bottom(), hr->end());
3047 gclog_or_tty->print_cr(" (" SIZE_FORMAT " KB left in heap.)",
3048 max_live_bytes/K);
3049 }
3050 hr = _g1->pop_region_from_young_list();
3051 }
3053 record_scan_only_regions(_g1->young_list_scan_only_length());
3055 double young_end_time_sec = os::elapsedTime();
3056 _recorded_young_cset_choice_time_ms =
3057 (young_end_time_sec - young_start_time_sec) * 1000.0;
3059 non_young_start_time_sec = os::elapsedTime();
3061 if (_young_cset_length > 0 && _last_young_gc_full) {
3062 // don't bother adding more regions...
3063 goto choose_collection_set_end;
3064 }
3065 }
3067 if (!in_young_gc_mode() || !full_young_gcs()) {
3068 bool should_continue = true;
3069 NumberSeq seq;
3070 double avg_prediction = 100000000000000000.0; // something very large
3071 do {
3072 hr = _collectionSetChooser->getNextMarkedRegion(time_remaining_ms,
3073 avg_prediction);
3074 if (hr != NULL) {
3075 double predicted_time_ms = predict_region_elapsed_time_ms(hr, false);
3076 time_remaining_ms -= predicted_time_ms;
3077 predicted_pause_time_ms += predicted_time_ms;
3078 add_to_collection_set(hr);
3079 record_cset_region(hr, false);
3080 max_live_bytes -= MIN2(hr->max_live_bytes(), max_live_bytes);
3081 if (G1PolicyVerbose > 0) {
3082 gclog_or_tty->print_cr(" (" SIZE_FORMAT " KB left in heap.)",
3083 max_live_bytes/K);
3084 }
3085 seq.add(predicted_time_ms);
3086 avg_prediction = seq.avg() + seq.sd();
3087 }
3088 should_continue =
3089 ( hr != NULL) &&
3090 ( (adaptive_young_list_length()) ? time_remaining_ms > 0.0
3091 : _collection_set_size < _young_list_fixed_length );
3092 } while (should_continue);
3094 if (!adaptive_young_list_length() &&
3095 _collection_set_size < _young_list_fixed_length)
3096 _should_revert_to_full_young_gcs = true;
3097 }
3099 choose_collection_set_end:
3100 count_CS_bytes_used();
3102 end_recording_regions();
3104 double non_young_end_time_sec = os::elapsedTime();
3105 _recorded_non_young_cset_choice_time_ms =
3106 (non_young_end_time_sec - non_young_start_time_sec) * 1000.0;
3107 }
3109 void G1CollectorPolicy_BestRegionsFirst::record_full_collection_end() {
3110 G1CollectorPolicy::record_full_collection_end();
3111 _collectionSetChooser->updateAfterFullCollection();
3112 }
3114 void G1CollectorPolicy_BestRegionsFirst::
3115 expand_if_possible(size_t numRegions) {
3116 size_t expansion_bytes = numRegions * HeapRegion::GrainBytes;
3117 _g1->expand(expansion_bytes);
3118 }
3120 void G1CollectorPolicy_BestRegionsFirst::
3121 record_collection_pause_end(bool abandoned) {
3122 G1CollectorPolicy::record_collection_pause_end(abandoned);
3123 assert(assertMarkedBytesDataOK(), "Marked regions not OK at pause end.");
3124 }