Tue, 23 Feb 2010 23:13:23 -0500
6928059: G1: command line parameter renaming
Summary: Rename G1 parameters to make them more consistent.
Reviewed-by: jmasa, johnc
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 _conc_mark_initiated(false),
182 _should_initiate_conc_mark(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 {
204 // Set up the region size and associated fields. Given that the
205 // policy is created before the heap, we have to set this up here,
206 // so it's done as soon as possible.
207 HeapRegion::setup_heap_region_size(Arguments::min_heap_size());
208 HeapRegionRemSet::setup_remset_size();
210 _recent_prev_end_times_for_all_gcs_sec->add(os::elapsedTime());
211 _prev_collection_pause_end_ms = os::elapsedTime() * 1000.0;
213 _par_last_ext_root_scan_times_ms = new double[_parallel_gc_threads];
214 _par_last_mark_stack_scan_times_ms = new double[_parallel_gc_threads];
215 _par_last_scan_only_times_ms = new double[_parallel_gc_threads];
216 _par_last_scan_only_regions_scanned = new double[_parallel_gc_threads];
218 _par_last_update_rs_start_times_ms = new double[_parallel_gc_threads];
219 _par_last_update_rs_times_ms = new double[_parallel_gc_threads];
220 _par_last_update_rs_processed_buffers = new double[_parallel_gc_threads];
222 _par_last_scan_rs_start_times_ms = new double[_parallel_gc_threads];
223 _par_last_scan_rs_times_ms = new double[_parallel_gc_threads];
224 _par_last_scan_new_refs_times_ms = new double[_parallel_gc_threads];
226 _par_last_obj_copy_times_ms = new double[_parallel_gc_threads];
228 _par_last_termination_times_ms = new double[_parallel_gc_threads];
230 // start conservatively
231 _expensive_region_limit_ms = 0.5 * (double) MaxGCPauseMillis;
233 // <NEW PREDICTION>
235 int index;
236 if (ParallelGCThreads == 0)
237 index = 0;
238 else if (ParallelGCThreads > 8)
239 index = 7;
240 else
241 index = ParallelGCThreads - 1;
243 _pending_card_diff_seq->add(0.0);
244 _rs_length_diff_seq->add(rs_length_diff_defaults[index]);
245 _cost_per_card_ms_seq->add(cost_per_card_ms_defaults[index]);
246 _cost_per_scan_only_region_ms_seq->add(
247 cost_per_scan_only_region_ms_defaults[index]);
248 _fully_young_cards_per_entry_ratio_seq->add(
249 fully_young_cards_per_entry_ratio_defaults[index]);
250 _cost_per_entry_ms_seq->add(cost_per_entry_ms_defaults[index]);
251 _cost_per_byte_ms_seq->add(cost_per_byte_ms_defaults[index]);
252 _constant_other_time_ms_seq->add(constant_other_time_ms_defaults[index]);
253 _young_other_cost_per_region_ms_seq->add(
254 young_other_cost_per_region_ms_defaults[index]);
255 _non_young_other_cost_per_region_ms_seq->add(
256 non_young_other_cost_per_region_ms_defaults[index]);
258 // </NEW PREDICTION>
260 double time_slice = (double) GCPauseIntervalMillis / 1000.0;
261 double max_gc_time = (double) MaxGCPauseMillis / 1000.0;
262 guarantee(max_gc_time < time_slice,
263 "Max GC time should not be greater than the time slice");
264 _mmu_tracker = new G1MMUTrackerQueue(time_slice, max_gc_time);
265 _sigma = (double) G1ConfidencePercent / 100.0;
267 // start conservatively (around 50ms is about right)
268 _concurrent_mark_init_times_ms->add(0.05);
269 _concurrent_mark_remark_times_ms->add(0.05);
270 _concurrent_mark_cleanup_times_ms->add(0.20);
271 _tenuring_threshold = MaxTenuringThreshold;
273 // if G1FixedSurvivorSpaceSize is 0 which means the size is not
274 // fixed, then _max_survivor_regions will be calculated at
275 // calculate_young_list_target_config during initialization
276 _max_survivor_regions = G1FixedSurvivorSpaceSize / HeapRegion::GrainBytes;
278 initialize_all();
279 }
281 // Increment "i", mod "len"
282 static void inc_mod(int& i, int len) {
283 i++; if (i == len) i = 0;
284 }
286 void G1CollectorPolicy::initialize_flags() {
287 set_min_alignment(HeapRegion::GrainBytes);
288 set_max_alignment(GenRemSet::max_alignment_constraint(rem_set_name()));
289 if (SurvivorRatio < 1) {
290 vm_exit_during_initialization("Invalid survivor ratio specified");
291 }
292 CollectorPolicy::initialize_flags();
293 }
295 void G1CollectorPolicy::init() {
296 // Set aside an initial future to_space.
297 _g1 = G1CollectedHeap::heap();
298 size_t regions = Universe::heap()->capacity() / HeapRegion::GrainBytes;
300 assert(Heap_lock->owned_by_self(), "Locking discipline.");
302 if (G1SteadyStateUsed < 50) {
303 vm_exit_during_initialization("G1SteadyStateUsed must be at least 50%.");
304 }
306 initialize_gc_policy_counters();
308 if (G1Gen) {
309 _in_young_gc_mode = true;
311 if (G1YoungGenSize == 0) {
312 set_adaptive_young_list_length(true);
313 _young_list_fixed_length = 0;
314 } else {
315 set_adaptive_young_list_length(false);
316 _young_list_fixed_length = (G1YoungGenSize / HeapRegion::GrainBytes);
317 }
318 _free_regions_at_end_of_collection = _g1->free_regions();
319 _scan_only_regions_at_end_of_collection = 0;
320 calculate_young_list_min_length();
321 guarantee( _young_list_min_length == 0, "invariant, not enough info" );
322 calculate_young_list_target_config();
323 } else {
324 _young_list_fixed_length = 0;
325 _in_young_gc_mode = false;
326 }
327 }
329 // Create the jstat counters for the policy.
330 void G1CollectorPolicy::initialize_gc_policy_counters()
331 {
332 _gc_policy_counters = new GCPolicyCounters("GarbageFirst", 1, 2 + G1Gen);
333 }
335 void G1CollectorPolicy::calculate_young_list_min_length() {
336 _young_list_min_length = 0;
338 if (!adaptive_young_list_length())
339 return;
341 if (_alloc_rate_ms_seq->num() > 3) {
342 double now_sec = os::elapsedTime();
343 double when_ms = _mmu_tracker->when_max_gc_sec(now_sec) * 1000.0;
344 double alloc_rate_ms = predict_alloc_rate_ms();
345 int min_regions = (int) ceil(alloc_rate_ms * when_ms);
346 int current_region_num = (int) _g1->young_list_length();
347 _young_list_min_length = min_regions + current_region_num;
348 }
349 }
351 void G1CollectorPolicy::calculate_young_list_target_config() {
352 if (adaptive_young_list_length()) {
353 size_t rs_lengths = (size_t) get_new_prediction(_rs_lengths_seq);
354 calculate_young_list_target_config(rs_lengths);
355 } else {
356 if (full_young_gcs())
357 _young_list_target_length = _young_list_fixed_length;
358 else
359 _young_list_target_length = _young_list_fixed_length / 2;
360 _young_list_target_length = MAX2(_young_list_target_length, (size_t)1);
361 size_t so_length = calculate_optimal_so_length(_young_list_target_length);
362 guarantee( so_length < _young_list_target_length, "invariant" );
363 _young_list_so_prefix_length = so_length;
364 }
365 calculate_survivors_policy();
366 }
368 // This method calculate the optimal scan-only set for a fixed young
369 // gen size. I couldn't work out how to reuse the more elaborate one,
370 // i.e. calculate_young_list_target_config(rs_length), as the loops are
371 // fundamentally different (the other one finds a config for different
372 // S-O lengths, whereas here we need to do the opposite).
373 size_t G1CollectorPolicy::calculate_optimal_so_length(
374 size_t young_list_length) {
375 if (!G1UseScanOnlyPrefix)
376 return 0;
378 if (_all_pause_times_ms->num() < 3) {
379 // we won't use a scan-only set at the beginning to allow the rest
380 // of the predictors to warm up
381 return 0;
382 }
384 if (_cost_per_scan_only_region_ms_seq->num() < 3) {
385 // then, we'll only set the S-O set to 1 for a little bit of time,
386 // to get enough information on the scanning cost
387 return 1;
388 }
390 size_t pending_cards = (size_t) get_new_prediction(_pending_cards_seq);
391 size_t rs_lengths = (size_t) get_new_prediction(_rs_lengths_seq);
392 size_t adj_rs_lengths = rs_lengths + predict_rs_length_diff();
393 size_t scanned_cards;
394 if (full_young_gcs())
395 scanned_cards = predict_young_card_num(adj_rs_lengths);
396 else
397 scanned_cards = predict_non_young_card_num(adj_rs_lengths);
398 double base_time_ms = predict_base_elapsed_time_ms(pending_cards,
399 scanned_cards);
401 size_t so_length = 0;
402 double max_gc_eff = 0.0;
403 for (size_t i = 0; i < young_list_length; ++i) {
404 double gc_eff = 0.0;
405 double pause_time_ms = 0.0;
406 predict_gc_eff(young_list_length, i, base_time_ms,
407 &gc_eff, &pause_time_ms);
408 if (gc_eff > max_gc_eff) {
409 max_gc_eff = gc_eff;
410 so_length = i;
411 }
412 }
414 // set it to 95% of the optimal to make sure we sample the "area"
415 // around the optimal length to get up-to-date survival rate data
416 return so_length * 950 / 1000;
417 }
419 // This is a really cool piece of code! It finds the best
420 // target configuration (young length / scan-only prefix length) so
421 // that GC efficiency is maximized and that we also meet a pause
422 // time. It's a triple nested loop. These loops are explained below
423 // from the inside-out :-)
424 //
425 // (a) The innermost loop will try to find the optimal young length
426 // for a fixed S-O length. It uses a binary search to speed up the
427 // process. We assume that, for a fixed S-O length, as we add more
428 // young regions to the CSet, the GC efficiency will only go up (I'll
429 // skip the proof). So, using a binary search to optimize this process
430 // makes perfect sense.
431 //
432 // (b) The middle loop will fix the S-O length before calling the
433 // innermost one. It will vary it between two parameters, increasing
434 // it by a given increment.
435 //
436 // (c) The outermost loop will call the middle loop three times.
437 // (1) The first time it will explore all possible S-O length values
438 // from 0 to as large as it can get, using a coarse increment (to
439 // quickly "home in" to where the optimal seems to be).
440 // (2) The second time it will explore the values around the optimal
441 // that was found by the first iteration using a fine increment.
442 // (3) Once the optimal config has been determined by the second
443 // iteration, we'll redo the calculation, but setting the S-O length
444 // to 95% of the optimal to make sure we sample the "area"
445 // around the optimal length to get up-to-date survival rate data
446 //
447 // Termination conditions for the iterations are several: the pause
448 // time is over the limit, we do not have enough to-space, etc.
450 void G1CollectorPolicy::calculate_young_list_target_config(size_t rs_lengths) {
451 guarantee( adaptive_young_list_length(), "pre-condition" );
453 double start_time_sec = os::elapsedTime();
454 size_t min_reserve_perc = MAX2((size_t)2, (size_t)G1ReservePercent);
455 min_reserve_perc = MIN2((size_t) 50, min_reserve_perc);
456 size_t reserve_regions =
457 (size_t) ((double) min_reserve_perc * (double) _g1->n_regions() / 100.0);
459 if (full_young_gcs() && _free_regions_at_end_of_collection > 0) {
460 // we are in fully-young mode and there are free regions in the heap
462 double survivor_regions_evac_time =
463 predict_survivor_regions_evac_time();
465 size_t min_so_length = 0;
466 size_t max_so_length = 0;
468 if (G1UseScanOnlyPrefix) {
469 if (_all_pause_times_ms->num() < 3) {
470 // we won't use a scan-only set at the beginning to allow the rest
471 // of the predictors to warm up
472 min_so_length = 0;
473 max_so_length = 0;
474 } else if (_cost_per_scan_only_region_ms_seq->num() < 3) {
475 // then, we'll only set the S-O set to 1 for a little bit of time,
476 // to get enough information on the scanning cost
477 min_so_length = 1;
478 max_so_length = 1;
479 } else if (_in_marking_window || _last_full_young_gc) {
480 // no S-O prefix during a marking phase either, as at the end
481 // of the marking phase we'll have to use a very small young
482 // length target to fill up the rest of the CSet with
483 // non-young regions and, if we have lots of scan-only regions
484 // left-over, we will not be able to add any more non-young
485 // regions.
486 min_so_length = 0;
487 max_so_length = 0;
488 } else {
489 // this is the common case; we'll never reach the maximum, we
490 // one of the end conditions will fire well before that
491 // (hopefully!)
492 min_so_length = 0;
493 max_so_length = _free_regions_at_end_of_collection - 1;
494 }
495 } else {
496 // no S-O prefix, as the switch is not set, but we still need to
497 // do one iteration to calculate the best young target that
498 // meets the pause time; this way we reuse the same code instead
499 // of replicating it
500 min_so_length = 0;
501 max_so_length = 0;
502 }
504 double target_pause_time_ms = _mmu_tracker->max_gc_time() * 1000.0;
505 size_t pending_cards = (size_t) get_new_prediction(_pending_cards_seq);
506 size_t adj_rs_lengths = rs_lengths + predict_rs_length_diff();
507 size_t scanned_cards;
508 if (full_young_gcs())
509 scanned_cards = predict_young_card_num(adj_rs_lengths);
510 else
511 scanned_cards = predict_non_young_card_num(adj_rs_lengths);
512 // calculate this once, so that we don't have to recalculate it in
513 // the innermost loop
514 double base_time_ms = predict_base_elapsed_time_ms(pending_cards, scanned_cards)
515 + survivor_regions_evac_time;
516 // the result
517 size_t final_young_length = 0;
518 size_t final_so_length = 0;
519 double final_gc_eff = 0.0;
520 // we'll also keep track of how many times we go into the inner loop
521 // this is for profiling reasons
522 size_t calculations = 0;
524 // this determines which of the three iterations the outer loop is in
525 typedef enum {
526 pass_type_coarse,
527 pass_type_fine,
528 pass_type_final
529 } pass_type_t;
531 // range of the outer loop's iteration
532 size_t from_so_length = min_so_length;
533 size_t to_so_length = max_so_length;
534 guarantee( from_so_length <= to_so_length, "invariant" );
536 // this will keep the S-O length that's found by the second
537 // iteration of the outer loop; we'll keep it just in case the third
538 // iteration fails to find something
539 size_t fine_so_length = 0;
541 // the increment step for the coarse (first) iteration
542 size_t so_coarse_increments = 5;
544 // the common case, we'll start with the coarse iteration
545 pass_type_t pass = pass_type_coarse;
546 size_t so_length_incr = so_coarse_increments;
548 if (from_so_length == to_so_length) {
549 // not point in doing the coarse iteration, we'll go directly into
550 // the fine one (we essentially trying to find the optimal young
551 // length for a fixed S-O length).
552 so_length_incr = 1;
553 pass = pass_type_final;
554 } else if (to_so_length - from_so_length < 3 * so_coarse_increments) {
555 // again, the range is too short so no point in foind the coarse
556 // iteration either
557 so_length_incr = 1;
558 pass = pass_type_fine;
559 }
561 bool done = false;
562 // this is the outermost loop
563 while (!done) {
564 #ifdef TRACE_CALC_YOUNG_CONFIG
565 // leave this in for debugging, just in case
566 gclog_or_tty->print_cr("searching between " SIZE_FORMAT " and " SIZE_FORMAT
567 ", incr " SIZE_FORMAT ", pass %s",
568 from_so_length, to_so_length, so_length_incr,
569 (pass == pass_type_coarse) ? "coarse" :
570 (pass == pass_type_fine) ? "fine" : "final");
571 #endif // TRACE_CALC_YOUNG_CONFIG
573 size_t so_length = from_so_length;
574 size_t init_free_regions =
575 MAX2((size_t)0,
576 _free_regions_at_end_of_collection +
577 _scan_only_regions_at_end_of_collection - reserve_regions);
579 // this determines whether a configuration was found
580 bool gc_eff_set = false;
581 // this is the middle loop
582 while (so_length <= to_so_length) {
583 // base time, which excludes region-related time; again we
584 // calculate it once to avoid recalculating it in the
585 // innermost loop
586 double base_time_with_so_ms =
587 base_time_ms + predict_scan_only_time_ms(so_length);
588 // it's already over the pause target, go around
589 if (base_time_with_so_ms > target_pause_time_ms)
590 break;
592 size_t starting_young_length = so_length+1;
594 // we make sure that the short young length that makes sense
595 // (one more than the S-O length) is feasible
596 size_t min_young_length = starting_young_length;
597 double min_gc_eff;
598 bool min_ok;
599 ++calculations;
600 min_ok = predict_gc_eff(min_young_length, so_length,
601 base_time_with_so_ms,
602 init_free_regions, target_pause_time_ms,
603 &min_gc_eff);
605 if (min_ok) {
606 // the shortest young length is indeed feasible; we'll know
607 // set up the max young length and we'll do a binary search
608 // between min_young_length and max_young_length
609 size_t max_young_length = _free_regions_at_end_of_collection - 1;
610 double max_gc_eff = 0.0;
611 bool max_ok = false;
613 // the innermost loop! (finally!)
614 while (max_young_length > min_young_length) {
615 // we'll make sure that min_young_length is always at a
616 // feasible config
617 guarantee( min_ok, "invariant" );
619 ++calculations;
620 max_ok = predict_gc_eff(max_young_length, so_length,
621 base_time_with_so_ms,
622 init_free_regions, target_pause_time_ms,
623 &max_gc_eff);
625 size_t diff = (max_young_length - min_young_length) / 2;
626 if (max_ok) {
627 min_young_length = max_young_length;
628 min_gc_eff = max_gc_eff;
629 min_ok = true;
630 }
631 max_young_length = min_young_length + diff;
632 }
634 // the innermost loop found a config
635 guarantee( min_ok, "invariant" );
636 if (min_gc_eff > final_gc_eff) {
637 // it's the best config so far, so we'll keep it
638 final_gc_eff = min_gc_eff;
639 final_young_length = min_young_length;
640 final_so_length = so_length;
641 gc_eff_set = true;
642 }
643 }
645 // incremental the fixed S-O length and go around
646 so_length += so_length_incr;
647 }
649 // this is the end of the outermost loop and we need to decide
650 // what to do during the next iteration
651 if (pass == pass_type_coarse) {
652 // we just did the coarse pass (first iteration)
654 if (!gc_eff_set)
655 // we didn't find a feasible config so we'll just bail out; of
656 // course, it might be the case that we missed it; but I'd say
657 // it's a bit unlikely
658 done = true;
659 else {
660 // We did find a feasible config with optimal GC eff during
661 // the first pass. So the second pass we'll only consider the
662 // S-O lengths around that config with a fine increment.
664 guarantee( so_length_incr == so_coarse_increments, "invariant" );
665 guarantee( final_so_length >= min_so_length, "invariant" );
667 #ifdef TRACE_CALC_YOUNG_CONFIG
668 // leave this in for debugging, just in case
669 gclog_or_tty->print_cr(" coarse pass: SO length " SIZE_FORMAT,
670 final_so_length);
671 #endif // TRACE_CALC_YOUNG_CONFIG
673 from_so_length =
674 (final_so_length - min_so_length > so_coarse_increments) ?
675 final_so_length - so_coarse_increments + 1 : min_so_length;
676 to_so_length =
677 (max_so_length - final_so_length > so_coarse_increments) ?
678 final_so_length + so_coarse_increments - 1 : max_so_length;
680 pass = pass_type_fine;
681 so_length_incr = 1;
682 }
683 } else if (pass == pass_type_fine) {
684 // we just finished the second pass
686 if (!gc_eff_set) {
687 // we didn't find a feasible config (yes, it's possible;
688 // notice that, sometimes, we go directly into the fine
689 // iteration and skip the coarse one) so we bail out
690 done = true;
691 } else {
692 // We did find a feasible config with optimal GC eff
693 guarantee( so_length_incr == 1, "invariant" );
695 if (final_so_length == 0) {
696 // The config is of an empty S-O set, so we'll just bail out
697 done = true;
698 } else {
699 // we'll go around once more, setting the S-O length to 95%
700 // of the optimal
701 size_t new_so_length = 950 * final_so_length / 1000;
703 #ifdef TRACE_CALC_YOUNG_CONFIG
704 // leave this in for debugging, just in case
705 gclog_or_tty->print_cr(" fine pass: SO length " SIZE_FORMAT
706 ", setting it to " SIZE_FORMAT,
707 final_so_length, new_so_length);
708 #endif // TRACE_CALC_YOUNG_CONFIG
710 from_so_length = new_so_length;
711 to_so_length = new_so_length;
712 fine_so_length = final_so_length;
714 pass = pass_type_final;
715 }
716 }
717 } else if (pass == pass_type_final) {
718 // we just finished the final (third) pass
720 if (!gc_eff_set)
721 // we didn't find a feasible config, so we'll just use the one
722 // we found during the second pass, which we saved
723 final_so_length = fine_so_length;
725 // and we're done!
726 done = true;
727 } else {
728 guarantee( false, "should never reach here" );
729 }
731 // we now go around the outermost loop
732 }
734 // we should have at least one region in the target young length
735 _young_list_target_length =
736 MAX2((size_t) 1, final_young_length + _recorded_survivor_regions);
737 if (final_so_length >= final_young_length)
738 // and we need to ensure that the S-O length is not greater than
739 // the target young length (this is being a bit careful)
740 final_so_length = 0;
741 _young_list_so_prefix_length = final_so_length;
742 guarantee( !_in_marking_window || !_last_full_young_gc ||
743 _young_list_so_prefix_length == 0, "invariant" );
745 // let's keep an eye of how long we spend on this calculation
746 // right now, I assume that we'll print it when we need it; we
747 // should really adde it to the breakdown of a pause
748 double end_time_sec = os::elapsedTime();
749 double elapsed_time_ms = (end_time_sec - start_time_sec) * 1000.0;
751 #ifdef TRACE_CALC_YOUNG_CONFIG
752 // leave this in for debugging, just in case
753 gclog_or_tty->print_cr("target = %1.1lf ms, young = " SIZE_FORMAT
754 ", SO = " SIZE_FORMAT ", "
755 "elapsed %1.2lf ms, calcs: " SIZE_FORMAT " (%s%s) "
756 SIZE_FORMAT SIZE_FORMAT,
757 target_pause_time_ms,
758 _young_list_target_length - _young_list_so_prefix_length,
759 _young_list_so_prefix_length,
760 elapsed_time_ms,
761 calculations,
762 full_young_gcs() ? "full" : "partial",
763 should_initiate_conc_mark() ? " i-m" : "",
764 _in_marking_window,
765 _in_marking_window_im);
766 #endif // TRACE_CALC_YOUNG_CONFIG
768 if (_young_list_target_length < _young_list_min_length) {
769 // bummer; this means that, if we do a pause when the optimal
770 // config dictates, we'll violate the pause spacing target (the
771 // min length was calculate based on the application's current
772 // alloc rate);
774 // so, we have to bite the bullet, and allocate the minimum
775 // number. We'll violate our target, but we just can't meet it.
777 size_t so_length = 0;
778 // a note further up explains why we do not want an S-O length
779 // during marking
780 if (!_in_marking_window && !_last_full_young_gc)
781 // but we can still try to see whether we can find an optimal
782 // S-O length
783 so_length = calculate_optimal_so_length(_young_list_min_length);
785 #ifdef TRACE_CALC_YOUNG_CONFIG
786 // leave this in for debugging, just in case
787 gclog_or_tty->print_cr("adjusted target length from "
788 SIZE_FORMAT " to " SIZE_FORMAT
789 ", SO " SIZE_FORMAT,
790 _young_list_target_length, _young_list_min_length,
791 so_length);
792 #endif // TRACE_CALC_YOUNG_CONFIG
794 _young_list_target_length =
795 MAX2(_young_list_min_length, (size_t)1);
796 _young_list_so_prefix_length = so_length;
797 }
798 } else {
799 // we are in a partially-young mode or we've run out of regions (due
800 // to evacuation failure)
802 #ifdef TRACE_CALC_YOUNG_CONFIG
803 // leave this in for debugging, just in case
804 gclog_or_tty->print_cr("(partial) setting target to " SIZE_FORMAT
805 ", SO " SIZE_FORMAT,
806 _young_list_min_length, 0);
807 #endif // TRACE_CALC_YOUNG_CONFIG
809 // we'll do the pause as soon as possible and with no S-O prefix
810 // (see above for the reasons behind the latter)
811 _young_list_target_length =
812 MAX2(_young_list_min_length, (size_t) 1);
813 _young_list_so_prefix_length = 0;
814 }
816 _rs_lengths_prediction = rs_lengths;
817 }
819 // This is used by: calculate_optimal_so_length(length). It returns
820 // the GC eff and predicted pause time for a particular config
821 void
822 G1CollectorPolicy::predict_gc_eff(size_t young_length,
823 size_t so_length,
824 double base_time_ms,
825 double* ret_gc_eff,
826 double* ret_pause_time_ms) {
827 double so_time_ms = predict_scan_only_time_ms(so_length);
828 double accum_surv_rate_adj = 0.0;
829 if (so_length > 0)
830 accum_surv_rate_adj = accum_yg_surv_rate_pred((int)(so_length - 1));
831 double accum_surv_rate =
832 accum_yg_surv_rate_pred((int)(young_length - 1)) - accum_surv_rate_adj;
833 size_t bytes_to_copy =
834 (size_t) (accum_surv_rate * (double) HeapRegion::GrainBytes);
835 double copy_time_ms = predict_object_copy_time_ms(bytes_to_copy);
836 double young_other_time_ms =
837 predict_young_other_time_ms(young_length - so_length);
838 double pause_time_ms =
839 base_time_ms + so_time_ms + copy_time_ms + young_other_time_ms;
840 size_t reclaimed_bytes =
841 (young_length - so_length) * HeapRegion::GrainBytes - bytes_to_copy;
842 double gc_eff = (double) reclaimed_bytes / pause_time_ms;
844 *ret_gc_eff = gc_eff;
845 *ret_pause_time_ms = pause_time_ms;
846 }
848 // This is used by: calculate_young_list_target_config(rs_length). It
849 // returns the GC eff of a particular config. It returns false if that
850 // config violates any of the end conditions of the search in the
851 // calling method, or true upon success. The end conditions were put
852 // here since it's called twice and it was best not to replicate them
853 // in the caller. Also, passing the parameteres avoids having to
854 // recalculate them in the innermost loop.
855 bool
856 G1CollectorPolicy::predict_gc_eff(size_t young_length,
857 size_t so_length,
858 double base_time_with_so_ms,
859 size_t init_free_regions,
860 double target_pause_time_ms,
861 double* ret_gc_eff) {
862 *ret_gc_eff = 0.0;
864 if (young_length >= init_free_regions)
865 // end condition 1: not enough space for the young regions
866 return false;
868 double accum_surv_rate_adj = 0.0;
869 if (so_length > 0)
870 accum_surv_rate_adj = accum_yg_surv_rate_pred((int)(so_length - 1));
871 double accum_surv_rate =
872 accum_yg_surv_rate_pred((int)(young_length - 1)) - accum_surv_rate_adj;
873 size_t bytes_to_copy =
874 (size_t) (accum_surv_rate * (double) HeapRegion::GrainBytes);
875 double copy_time_ms = predict_object_copy_time_ms(bytes_to_copy);
876 double young_other_time_ms =
877 predict_young_other_time_ms(young_length - so_length);
878 double pause_time_ms =
879 base_time_with_so_ms + copy_time_ms + young_other_time_ms;
881 if (pause_time_ms > target_pause_time_ms)
882 // end condition 2: over the target pause time
883 return false;
885 size_t reclaimed_bytes =
886 (young_length - so_length) * HeapRegion::GrainBytes - bytes_to_copy;
887 size_t free_bytes =
888 (init_free_regions - young_length) * HeapRegion::GrainBytes;
890 if ((2.0 + sigma()) * (double) bytes_to_copy > (double) free_bytes)
891 // end condition 3: out of to-space (conservatively)
892 return false;
894 // success!
895 double gc_eff = (double) reclaimed_bytes / pause_time_ms;
896 *ret_gc_eff = gc_eff;
898 return true;
899 }
901 double G1CollectorPolicy::predict_survivor_regions_evac_time() {
902 double survivor_regions_evac_time = 0.0;
903 for (HeapRegion * r = _recorded_survivor_head;
904 r != NULL && r != _recorded_survivor_tail->get_next_young_region();
905 r = r->get_next_young_region()) {
906 survivor_regions_evac_time += predict_region_elapsed_time_ms(r, true);
907 }
908 return survivor_regions_evac_time;
909 }
911 void G1CollectorPolicy::check_prediction_validity() {
912 guarantee( adaptive_young_list_length(), "should not call this otherwise" );
914 size_t rs_lengths = _g1->young_list_sampled_rs_lengths();
915 if (rs_lengths > _rs_lengths_prediction) {
916 // add 10% to avoid having to recalculate often
917 size_t rs_lengths_prediction = rs_lengths * 1100 / 1000;
918 calculate_young_list_target_config(rs_lengths_prediction);
919 }
920 }
922 HeapWord* G1CollectorPolicy::mem_allocate_work(size_t size,
923 bool is_tlab,
924 bool* gc_overhead_limit_was_exceeded) {
925 guarantee(false, "Not using this policy feature yet.");
926 return NULL;
927 }
929 // This method controls how a collector handles one or more
930 // of its generations being fully allocated.
931 HeapWord* G1CollectorPolicy::satisfy_failed_allocation(size_t size,
932 bool is_tlab) {
933 guarantee(false, "Not using this policy feature yet.");
934 return NULL;
935 }
938 #ifndef PRODUCT
939 bool G1CollectorPolicy::verify_young_ages() {
940 HeapRegion* head = _g1->young_list_first_region();
941 return
942 verify_young_ages(head, _short_lived_surv_rate_group);
943 // also call verify_young_ages on any additional surv rate groups
944 }
946 bool
947 G1CollectorPolicy::verify_young_ages(HeapRegion* head,
948 SurvRateGroup *surv_rate_group) {
949 guarantee( surv_rate_group != NULL, "pre-condition" );
951 const char* name = surv_rate_group->name();
952 bool ret = true;
953 int prev_age = -1;
955 for (HeapRegion* curr = head;
956 curr != NULL;
957 curr = curr->get_next_young_region()) {
958 SurvRateGroup* group = curr->surv_rate_group();
959 if (group == NULL && !curr->is_survivor()) {
960 gclog_or_tty->print_cr("## %s: encountered NULL surv_rate_group", name);
961 ret = false;
962 }
964 if (surv_rate_group == group) {
965 int age = curr->age_in_surv_rate_group();
967 if (age < 0) {
968 gclog_or_tty->print_cr("## %s: encountered negative age", name);
969 ret = false;
970 }
972 if (age <= prev_age) {
973 gclog_or_tty->print_cr("## %s: region ages are not strictly increasing "
974 "(%d, %d)", name, age, prev_age);
975 ret = false;
976 }
977 prev_age = age;
978 }
979 }
981 return ret;
982 }
983 #endif // PRODUCT
985 void G1CollectorPolicy::record_full_collection_start() {
986 _cur_collection_start_sec = os::elapsedTime();
987 // Release the future to-space so that it is available for compaction into.
988 _g1->set_full_collection();
989 }
991 void G1CollectorPolicy::record_full_collection_end() {
992 // Consider this like a collection pause for the purposes of allocation
993 // since last pause.
994 double end_sec = os::elapsedTime();
995 double full_gc_time_sec = end_sec - _cur_collection_start_sec;
996 double full_gc_time_ms = full_gc_time_sec * 1000.0;
998 _all_full_gc_times_ms->add(full_gc_time_ms);
1000 update_recent_gc_times(end_sec, full_gc_time_ms);
1002 _g1->clear_full_collection();
1004 // "Nuke" the heuristics that control the fully/partially young GC
1005 // transitions and make sure we start with fully young GCs after the
1006 // Full GC.
1007 set_full_young_gcs(true);
1008 _last_full_young_gc = false;
1009 _should_revert_to_full_young_gcs = false;
1010 _should_initiate_conc_mark = false;
1011 _known_garbage_bytes = 0;
1012 _known_garbage_ratio = 0.0;
1013 _in_marking_window = false;
1014 _in_marking_window_im = false;
1016 _short_lived_surv_rate_group->record_scan_only_prefix(0);
1017 _short_lived_surv_rate_group->start_adding_regions();
1018 // also call this on any additional surv rate groups
1020 record_survivor_regions(0, NULL, NULL);
1022 _prev_region_num_young = _region_num_young;
1023 _prev_region_num_tenured = _region_num_tenured;
1025 _free_regions_at_end_of_collection = _g1->free_regions();
1026 _scan_only_regions_at_end_of_collection = 0;
1027 // Reset survivors SurvRateGroup.
1028 _survivor_surv_rate_group->reset();
1029 calculate_young_list_min_length();
1030 calculate_young_list_target_config();
1031 }
1033 void G1CollectorPolicy::record_before_bytes(size_t bytes) {
1034 _bytes_in_to_space_before_gc += bytes;
1035 }
1037 void G1CollectorPolicy::record_after_bytes(size_t bytes) {
1038 _bytes_in_to_space_after_gc += bytes;
1039 }
1041 void G1CollectorPolicy::record_stop_world_start() {
1042 _stop_world_start = os::elapsedTime();
1043 }
1045 void G1CollectorPolicy::record_collection_pause_start(double start_time_sec,
1046 size_t start_used) {
1047 if (PrintGCDetails) {
1048 gclog_or_tty->stamp(PrintGCTimeStamps);
1049 gclog_or_tty->print("[GC pause");
1050 if (in_young_gc_mode())
1051 gclog_or_tty->print(" (%s)", full_young_gcs() ? "young" : "partial");
1052 }
1054 assert(_g1->used_regions() == _g1->recalculate_used_regions(),
1055 "sanity");
1056 assert(_g1->used() == _g1->recalculate_used(), "sanity");
1058 double s_w_t_ms = (start_time_sec - _stop_world_start) * 1000.0;
1059 _all_stop_world_times_ms->add(s_w_t_ms);
1060 _stop_world_start = 0.0;
1062 _cur_collection_start_sec = start_time_sec;
1063 _cur_collection_pause_used_at_start_bytes = start_used;
1064 _cur_collection_pause_used_regions_at_start = _g1->used_regions();
1065 _pending_cards = _g1->pending_card_num();
1066 _max_pending_cards = _g1->max_pending_card_num();
1068 _bytes_in_to_space_before_gc = 0;
1069 _bytes_in_to_space_after_gc = 0;
1070 _bytes_in_collection_set_before_gc = 0;
1072 #ifdef DEBUG
1073 // initialise these to something well known so that we can spot
1074 // if they are not set properly
1076 for (int i = 0; i < _parallel_gc_threads; ++i) {
1077 _par_last_ext_root_scan_times_ms[i] = -666.0;
1078 _par_last_mark_stack_scan_times_ms[i] = -666.0;
1079 _par_last_scan_only_times_ms[i] = -666.0;
1080 _par_last_scan_only_regions_scanned[i] = -666.0;
1081 _par_last_update_rs_start_times_ms[i] = -666.0;
1082 _par_last_update_rs_times_ms[i] = -666.0;
1083 _par_last_update_rs_processed_buffers[i] = -666.0;
1084 _par_last_scan_rs_start_times_ms[i] = -666.0;
1085 _par_last_scan_rs_times_ms[i] = -666.0;
1086 _par_last_scan_new_refs_times_ms[i] = -666.0;
1087 _par_last_obj_copy_times_ms[i] = -666.0;
1088 _par_last_termination_times_ms[i] = -666.0;
1089 }
1090 #endif
1092 for (int i = 0; i < _aux_num; ++i) {
1093 _cur_aux_times_ms[i] = 0.0;
1094 _cur_aux_times_set[i] = false;
1095 }
1097 _satb_drain_time_set = false;
1098 _last_satb_drain_processed_buffers = -1;
1100 if (in_young_gc_mode())
1101 _last_young_gc_full = false;
1104 // do that for any other surv rate groups
1105 _short_lived_surv_rate_group->stop_adding_regions();
1106 size_t short_lived_so_length = _young_list_so_prefix_length;
1107 _short_lived_surv_rate_group->record_scan_only_prefix(short_lived_so_length);
1108 tag_scan_only(short_lived_so_length);
1109 _survivors_age_table.clear();
1111 assert( verify_young_ages(), "region age verification" );
1112 }
1114 void G1CollectorPolicy::tag_scan_only(size_t short_lived_scan_only_length) {
1115 // done in a way that it can be extended for other surv rate groups too...
1117 HeapRegion* head = _g1->young_list_first_region();
1118 bool finished_short_lived = (short_lived_scan_only_length == 0);
1120 if (finished_short_lived)
1121 return;
1123 for (HeapRegion* curr = head;
1124 curr != NULL;
1125 curr = curr->get_next_young_region()) {
1126 SurvRateGroup* surv_rate_group = curr->surv_rate_group();
1127 int age = curr->age_in_surv_rate_group();
1129 if (surv_rate_group == _short_lived_surv_rate_group) {
1130 if ((size_t)age < short_lived_scan_only_length)
1131 curr->set_scan_only();
1132 else
1133 finished_short_lived = true;
1134 }
1137 if (finished_short_lived)
1138 return;
1139 }
1141 guarantee( false, "we should never reach here" );
1142 }
1144 void G1CollectorPolicy::record_mark_closure_time(double mark_closure_time_ms) {
1145 _mark_closure_time_ms = mark_closure_time_ms;
1146 }
1148 void G1CollectorPolicy::record_concurrent_mark_init_start() {
1149 _mark_init_start_sec = os::elapsedTime();
1150 guarantee(!in_young_gc_mode(), "should not do be here in young GC mode");
1151 }
1153 void G1CollectorPolicy::record_concurrent_mark_init_end_pre(double
1154 mark_init_elapsed_time_ms) {
1155 _during_marking = true;
1156 _should_initiate_conc_mark = false;
1157 _cur_mark_stop_world_time_ms = mark_init_elapsed_time_ms;
1158 }
1160 void G1CollectorPolicy::record_concurrent_mark_init_end() {
1161 double end_time_sec = os::elapsedTime();
1162 double elapsed_time_ms = (end_time_sec - _mark_init_start_sec) * 1000.0;
1163 _concurrent_mark_init_times_ms->add(elapsed_time_ms);
1164 record_concurrent_mark_init_end_pre(elapsed_time_ms);
1166 _mmu_tracker->add_pause(_mark_init_start_sec, end_time_sec, true);
1167 }
1169 void G1CollectorPolicy::record_concurrent_mark_remark_start() {
1170 _mark_remark_start_sec = os::elapsedTime();
1171 _during_marking = false;
1172 }
1174 void G1CollectorPolicy::record_concurrent_mark_remark_end() {
1175 double end_time_sec = os::elapsedTime();
1176 double elapsed_time_ms = (end_time_sec - _mark_remark_start_sec)*1000.0;
1177 _concurrent_mark_remark_times_ms->add(elapsed_time_ms);
1178 _cur_mark_stop_world_time_ms += elapsed_time_ms;
1179 _prev_collection_pause_end_ms += elapsed_time_ms;
1181 _mmu_tracker->add_pause(_mark_remark_start_sec, end_time_sec, true);
1182 }
1184 void G1CollectorPolicy::record_concurrent_mark_cleanup_start() {
1185 _mark_cleanup_start_sec = os::elapsedTime();
1186 }
1188 void
1189 G1CollectorPolicy::record_concurrent_mark_cleanup_end(size_t freed_bytes,
1190 size_t max_live_bytes) {
1191 record_concurrent_mark_cleanup_end_work1(freed_bytes, max_live_bytes);
1192 record_concurrent_mark_cleanup_end_work2();
1193 }
1195 void
1196 G1CollectorPolicy::
1197 record_concurrent_mark_cleanup_end_work1(size_t freed_bytes,
1198 size_t max_live_bytes) {
1199 if (_n_marks < 2) _n_marks++;
1200 if (G1PolicyVerbose > 0)
1201 gclog_or_tty->print_cr("At end of marking, max_live is " SIZE_FORMAT " MB "
1202 " (of " SIZE_FORMAT " MB heap).",
1203 max_live_bytes/M, _g1->capacity()/M);
1204 }
1206 // The important thing about this is that it includes "os::elapsedTime".
1207 void G1CollectorPolicy::record_concurrent_mark_cleanup_end_work2() {
1208 double end_time_sec = os::elapsedTime();
1209 double elapsed_time_ms = (end_time_sec - _mark_cleanup_start_sec)*1000.0;
1210 _concurrent_mark_cleanup_times_ms->add(elapsed_time_ms);
1211 _cur_mark_stop_world_time_ms += elapsed_time_ms;
1212 _prev_collection_pause_end_ms += elapsed_time_ms;
1214 _mmu_tracker->add_pause(_mark_cleanup_start_sec, end_time_sec, true);
1216 _num_markings++;
1218 // We did a marking, so reset the "since_last_mark" variables.
1219 double considerConcMarkCost = 1.0;
1220 // If there are available processors, concurrent activity is free...
1221 if (Threads::number_of_non_daemon_threads() * 2 <
1222 os::active_processor_count()) {
1223 considerConcMarkCost = 0.0;
1224 }
1225 _n_pauses_at_mark_end = _n_pauses;
1226 _n_marks_since_last_pause++;
1227 _conc_mark_initiated = false;
1228 }
1230 void
1231 G1CollectorPolicy::record_concurrent_mark_cleanup_completed() {
1232 if (in_young_gc_mode()) {
1233 _should_revert_to_full_young_gcs = false;
1234 _last_full_young_gc = true;
1235 _in_marking_window = false;
1236 if (adaptive_young_list_length())
1237 calculate_young_list_target_config();
1238 }
1239 }
1241 void G1CollectorPolicy::record_concurrent_pause() {
1242 if (_stop_world_start > 0.0) {
1243 double yield_ms = (os::elapsedTime() - _stop_world_start) * 1000.0;
1244 _all_yield_times_ms->add(yield_ms);
1245 }
1246 }
1248 void G1CollectorPolicy::record_concurrent_pause_end() {
1249 }
1251 void G1CollectorPolicy::record_collection_pause_end_CH_strong_roots() {
1252 _cur_CH_strong_roots_end_sec = os::elapsedTime();
1253 _cur_CH_strong_roots_dur_ms =
1254 (_cur_CH_strong_roots_end_sec - _cur_collection_start_sec) * 1000.0;
1255 }
1257 void G1CollectorPolicy::record_collection_pause_end_G1_strong_roots() {
1258 _cur_G1_strong_roots_end_sec = os::elapsedTime();
1259 _cur_G1_strong_roots_dur_ms =
1260 (_cur_G1_strong_roots_end_sec - _cur_CH_strong_roots_end_sec) * 1000.0;
1261 }
1263 template<class T>
1264 T sum_of(T* sum_arr, int start, int n, int N) {
1265 T sum = (T)0;
1266 for (int i = 0; i < n; i++) {
1267 int j = (start + i) % N;
1268 sum += sum_arr[j];
1269 }
1270 return sum;
1271 }
1273 void G1CollectorPolicy::print_par_stats (int level,
1274 const char* str,
1275 double* data,
1276 bool summary) {
1277 double min = data[0], max = data[0];
1278 double total = 0.0;
1279 int j;
1280 for (j = 0; j < level; ++j)
1281 gclog_or_tty->print(" ");
1282 gclog_or_tty->print("[%s (ms):", str);
1283 for (uint i = 0; i < ParallelGCThreads; ++i) {
1284 double val = data[i];
1285 if (val < min)
1286 min = val;
1287 if (val > max)
1288 max = val;
1289 total += val;
1290 gclog_or_tty->print(" %3.1lf", val);
1291 }
1292 if (summary) {
1293 gclog_or_tty->print_cr("");
1294 double avg = total / (double) ParallelGCThreads;
1295 gclog_or_tty->print(" ");
1296 for (j = 0; j < level; ++j)
1297 gclog_or_tty->print(" ");
1298 gclog_or_tty->print("Avg: %5.1lf, Min: %5.1lf, Max: %5.1lf",
1299 avg, min, max);
1300 }
1301 gclog_or_tty->print_cr("]");
1302 }
1304 void G1CollectorPolicy::print_par_buffers (int level,
1305 const char* str,
1306 double* data,
1307 bool summary) {
1308 double min = data[0], max = data[0];
1309 double total = 0.0;
1310 int j;
1311 for (j = 0; j < level; ++j)
1312 gclog_or_tty->print(" ");
1313 gclog_or_tty->print("[%s :", str);
1314 for (uint i = 0; i < ParallelGCThreads; ++i) {
1315 double val = data[i];
1316 if (val < min)
1317 min = val;
1318 if (val > max)
1319 max = val;
1320 total += val;
1321 gclog_or_tty->print(" %d", (int) val);
1322 }
1323 if (summary) {
1324 gclog_or_tty->print_cr("");
1325 double avg = total / (double) ParallelGCThreads;
1326 gclog_or_tty->print(" ");
1327 for (j = 0; j < level; ++j)
1328 gclog_or_tty->print(" ");
1329 gclog_or_tty->print("Sum: %d, Avg: %d, Min: %d, Max: %d",
1330 (int)total, (int)avg, (int)min, (int)max);
1331 }
1332 gclog_or_tty->print_cr("]");
1333 }
1335 void G1CollectorPolicy::print_stats (int level,
1336 const char* str,
1337 double value) {
1338 for (int j = 0; j < level; ++j)
1339 gclog_or_tty->print(" ");
1340 gclog_or_tty->print_cr("[%s: %5.1lf ms]", str, value);
1341 }
1343 void G1CollectorPolicy::print_stats (int level,
1344 const char* str,
1345 int value) {
1346 for (int j = 0; j < level; ++j)
1347 gclog_or_tty->print(" ");
1348 gclog_or_tty->print_cr("[%s: %d]", str, value);
1349 }
1351 double G1CollectorPolicy::avg_value (double* data) {
1352 if (ParallelGCThreads > 0) {
1353 double ret = 0.0;
1354 for (uint i = 0; i < ParallelGCThreads; ++i)
1355 ret += data[i];
1356 return ret / (double) ParallelGCThreads;
1357 } else {
1358 return data[0];
1359 }
1360 }
1362 double G1CollectorPolicy::max_value (double* data) {
1363 if (ParallelGCThreads > 0) {
1364 double ret = data[0];
1365 for (uint i = 1; i < ParallelGCThreads; ++i)
1366 if (data[i] > ret)
1367 ret = data[i];
1368 return ret;
1369 } else {
1370 return data[0];
1371 }
1372 }
1374 double G1CollectorPolicy::sum_of_values (double* data) {
1375 if (ParallelGCThreads > 0) {
1376 double sum = 0.0;
1377 for (uint i = 0; i < ParallelGCThreads; i++)
1378 sum += data[i];
1379 return sum;
1380 } else {
1381 return data[0];
1382 }
1383 }
1385 double G1CollectorPolicy::max_sum (double* data1,
1386 double* data2) {
1387 double ret = data1[0] + data2[0];
1389 if (ParallelGCThreads > 0) {
1390 for (uint i = 1; i < ParallelGCThreads; ++i) {
1391 double data = data1[i] + data2[i];
1392 if (data > ret)
1393 ret = data;
1394 }
1395 }
1396 return ret;
1397 }
1399 // Anything below that is considered to be zero
1400 #define MIN_TIMER_GRANULARITY 0.0000001
1402 void G1CollectorPolicy::record_collection_pause_end(bool abandoned) {
1403 double end_time_sec = os::elapsedTime();
1404 double elapsed_ms = _last_pause_time_ms;
1405 bool parallel = ParallelGCThreads > 0;
1406 double evac_ms = (end_time_sec - _cur_G1_strong_roots_end_sec) * 1000.0;
1407 size_t rs_size =
1408 _cur_collection_pause_used_regions_at_start - collection_set_size();
1409 size_t cur_used_bytes = _g1->used();
1410 assert(cur_used_bytes == _g1->recalculate_used(), "It should!");
1411 bool last_pause_included_initial_mark = false;
1412 bool update_stats = !abandoned && !_g1->evacuation_failed();
1414 #ifndef PRODUCT
1415 if (G1YoungSurvRateVerbose) {
1416 gclog_or_tty->print_cr("");
1417 _short_lived_surv_rate_group->print();
1418 // do that for any other surv rate groups too
1419 }
1420 #endif // PRODUCT
1422 if (in_young_gc_mode()) {
1423 last_pause_included_initial_mark = _should_initiate_conc_mark;
1424 if (last_pause_included_initial_mark)
1425 record_concurrent_mark_init_end_pre(0.0);
1427 size_t min_used_targ =
1428 (_g1->capacity() / 100) * (G1SteadyStateUsed - G1SteadyStateUsedDelta);
1430 if (cur_used_bytes > min_used_targ) {
1431 if (cur_used_bytes <= _prev_collection_pause_used_at_end_bytes) {
1432 } else if (!_g1->mark_in_progress() && !_last_full_young_gc) {
1433 _should_initiate_conc_mark = true;
1434 }
1435 }
1437 _prev_collection_pause_used_at_end_bytes = cur_used_bytes;
1438 }
1440 _mmu_tracker->add_pause(end_time_sec - elapsed_ms/1000.0,
1441 end_time_sec, false);
1443 guarantee(_cur_collection_pause_used_regions_at_start >=
1444 collection_set_size(),
1445 "Negative RS size?");
1447 // This assert is exempted when we're doing parallel collection pauses,
1448 // because the fragmentation caused by the parallel GC allocation buffers
1449 // can lead to more memory being used during collection than was used
1450 // before. Best leave this out until the fragmentation problem is fixed.
1451 // Pauses in which evacuation failed can also lead to negative
1452 // collections, since no space is reclaimed from a region containing an
1453 // object whose evacuation failed.
1454 // Further, we're now always doing parallel collection. But I'm still
1455 // leaving this here as a placeholder for a more precise assertion later.
1456 // (DLD, 10/05.)
1457 assert((true || parallel) // Always using GC LABs now.
1458 || _g1->evacuation_failed()
1459 || _cur_collection_pause_used_at_start_bytes >= cur_used_bytes,
1460 "Negative collection");
1462 size_t freed_bytes =
1463 _cur_collection_pause_used_at_start_bytes - cur_used_bytes;
1464 size_t surviving_bytes = _collection_set_bytes_used_before - freed_bytes;
1465 double survival_fraction =
1466 (double)surviving_bytes/
1467 (double)_collection_set_bytes_used_before;
1469 _n_pauses++;
1471 if (update_stats) {
1472 _recent_CH_strong_roots_times_ms->add(_cur_CH_strong_roots_dur_ms);
1473 _recent_G1_strong_roots_times_ms->add(_cur_G1_strong_roots_dur_ms);
1474 _recent_evac_times_ms->add(evac_ms);
1475 _recent_pause_times_ms->add(elapsed_ms);
1477 _recent_rs_sizes->add(rs_size);
1479 // We exempt parallel collection from this check because Alloc Buffer
1480 // fragmentation can produce negative collections. Same with evac
1481 // failure.
1482 // Further, we're now always doing parallel collection. But I'm still
1483 // leaving this here as a placeholder for a more precise assertion later.
1484 // (DLD, 10/05.
1485 assert((true || parallel)
1486 || _g1->evacuation_failed()
1487 || surviving_bytes <= _collection_set_bytes_used_before,
1488 "Or else negative collection!");
1489 _recent_CS_bytes_used_before->add(_collection_set_bytes_used_before);
1490 _recent_CS_bytes_surviving->add(surviving_bytes);
1492 // this is where we update the allocation rate of the application
1493 double app_time_ms =
1494 (_cur_collection_start_sec * 1000.0 - _prev_collection_pause_end_ms);
1495 if (app_time_ms < MIN_TIMER_GRANULARITY) {
1496 // This usually happens due to the timer not having the required
1497 // granularity. Some Linuxes are the usual culprits.
1498 // We'll just set it to something (arbitrarily) small.
1499 app_time_ms = 1.0;
1500 }
1501 size_t regions_allocated =
1502 (_region_num_young - _prev_region_num_young) +
1503 (_region_num_tenured - _prev_region_num_tenured);
1504 double alloc_rate_ms = (double) regions_allocated / app_time_ms;
1505 _alloc_rate_ms_seq->add(alloc_rate_ms);
1506 _prev_region_num_young = _region_num_young;
1507 _prev_region_num_tenured = _region_num_tenured;
1509 double interval_ms =
1510 (end_time_sec - _recent_prev_end_times_for_all_gcs_sec->oldest()) * 1000.0;
1511 update_recent_gc_times(end_time_sec, elapsed_ms);
1512 _recent_avg_pause_time_ratio = _recent_gc_times_ms->sum()/interval_ms;
1513 if (recent_avg_pause_time_ratio() < 0.0 ||
1514 (recent_avg_pause_time_ratio() - 1.0 > 0.0)) {
1515 #ifndef PRODUCT
1516 // Dump info to allow post-facto debugging
1517 gclog_or_tty->print_cr("recent_avg_pause_time_ratio() out of bounds");
1518 gclog_or_tty->print_cr("-------------------------------------------");
1519 gclog_or_tty->print_cr("Recent GC Times (ms):");
1520 _recent_gc_times_ms->dump();
1521 gclog_or_tty->print_cr("(End Time=%3.3f) Recent GC End Times (s):", end_time_sec);
1522 _recent_prev_end_times_for_all_gcs_sec->dump();
1523 gclog_or_tty->print_cr("GC = %3.3f, Interval = %3.3f, Ratio = %3.3f",
1524 _recent_gc_times_ms->sum(), interval_ms, recent_avg_pause_time_ratio());
1525 // In debug mode, terminate the JVM if the user wants to debug at this point.
1526 assert(!G1FailOnFPError, "Debugging data for CR 6898948 has been dumped above");
1527 #endif // !PRODUCT
1528 // Clip ratio between 0.0 and 1.0, and continue. This will be fixed in
1529 // CR 6902692 by redoing the manner in which the ratio is incrementally computed.
1530 if (_recent_avg_pause_time_ratio < 0.0) {
1531 _recent_avg_pause_time_ratio = 0.0;
1532 } else {
1533 assert(_recent_avg_pause_time_ratio - 1.0 > 0.0, "Ctl-point invariant");
1534 _recent_avg_pause_time_ratio = 1.0;
1535 }
1536 }
1537 }
1539 if (G1PolicyVerbose > 1) {
1540 gclog_or_tty->print_cr(" Recording collection pause(%d)", _n_pauses);
1541 }
1543 PauseSummary* summary;
1544 if (abandoned) {
1545 summary = _abandoned_summary;
1546 } else {
1547 summary = _summary;
1548 }
1550 double ext_root_scan_time = avg_value(_par_last_ext_root_scan_times_ms);
1551 double mark_stack_scan_time = avg_value(_par_last_mark_stack_scan_times_ms);
1552 double scan_only_time = avg_value(_par_last_scan_only_times_ms);
1553 double scan_only_regions_scanned =
1554 sum_of_values(_par_last_scan_only_regions_scanned);
1555 double update_rs_time = avg_value(_par_last_update_rs_times_ms);
1556 double update_rs_processed_buffers =
1557 sum_of_values(_par_last_update_rs_processed_buffers);
1558 double scan_rs_time = avg_value(_par_last_scan_rs_times_ms);
1559 double obj_copy_time = avg_value(_par_last_obj_copy_times_ms);
1560 double termination_time = avg_value(_par_last_termination_times_ms);
1562 double parallel_other_time = _cur_collection_par_time_ms -
1563 (update_rs_time + ext_root_scan_time + mark_stack_scan_time +
1564 scan_only_time + scan_rs_time + obj_copy_time + termination_time);
1565 if (update_stats) {
1566 MainBodySummary* body_summary = summary->main_body_summary();
1567 guarantee(body_summary != NULL, "should not be null!");
1569 if (_satb_drain_time_set)
1570 body_summary->record_satb_drain_time_ms(_cur_satb_drain_time_ms);
1571 else
1572 body_summary->record_satb_drain_time_ms(0.0);
1573 body_summary->record_ext_root_scan_time_ms(ext_root_scan_time);
1574 body_summary->record_mark_stack_scan_time_ms(mark_stack_scan_time);
1575 body_summary->record_scan_only_time_ms(scan_only_time);
1576 body_summary->record_update_rs_time_ms(update_rs_time);
1577 body_summary->record_scan_rs_time_ms(scan_rs_time);
1578 body_summary->record_obj_copy_time_ms(obj_copy_time);
1579 if (parallel) {
1580 body_summary->record_parallel_time_ms(_cur_collection_par_time_ms);
1581 body_summary->record_clear_ct_time_ms(_cur_clear_ct_time_ms);
1582 body_summary->record_termination_time_ms(termination_time);
1583 body_summary->record_parallel_other_time_ms(parallel_other_time);
1584 }
1585 body_summary->record_mark_closure_time_ms(_mark_closure_time_ms);
1586 }
1588 if (G1PolicyVerbose > 1) {
1589 gclog_or_tty->print_cr(" ET: %10.6f ms (avg: %10.6f ms)\n"
1590 " CH Strong: %10.6f ms (avg: %10.6f ms)\n"
1591 " G1 Strong: %10.6f ms (avg: %10.6f ms)\n"
1592 " Evac: %10.6f ms (avg: %10.6f ms)\n"
1593 " ET-RS: %10.6f ms (avg: %10.6f ms)\n"
1594 " |RS|: " SIZE_FORMAT,
1595 elapsed_ms, recent_avg_time_for_pauses_ms(),
1596 _cur_CH_strong_roots_dur_ms, recent_avg_time_for_CH_strong_ms(),
1597 _cur_G1_strong_roots_dur_ms, recent_avg_time_for_G1_strong_ms(),
1598 evac_ms, recent_avg_time_for_evac_ms(),
1599 scan_rs_time,
1600 recent_avg_time_for_pauses_ms() -
1601 recent_avg_time_for_G1_strong_ms(),
1602 rs_size);
1604 gclog_or_tty->print_cr(" Used at start: " SIZE_FORMAT"K"
1605 " At end " SIZE_FORMAT "K\n"
1606 " garbage : " SIZE_FORMAT "K"
1607 " of " SIZE_FORMAT "K\n"
1608 " survival : %6.2f%% (%6.2f%% avg)",
1609 _cur_collection_pause_used_at_start_bytes/K,
1610 _g1->used()/K, freed_bytes/K,
1611 _collection_set_bytes_used_before/K,
1612 survival_fraction*100.0,
1613 recent_avg_survival_fraction()*100.0);
1614 gclog_or_tty->print_cr(" Recent %% gc pause time: %6.2f",
1615 recent_avg_pause_time_ratio() * 100.0);
1616 }
1618 double other_time_ms = elapsed_ms;
1620 if (!abandoned) {
1621 if (_satb_drain_time_set)
1622 other_time_ms -= _cur_satb_drain_time_ms;
1624 if (parallel)
1625 other_time_ms -= _cur_collection_par_time_ms + _cur_clear_ct_time_ms;
1626 else
1627 other_time_ms -=
1628 update_rs_time +
1629 ext_root_scan_time + mark_stack_scan_time + scan_only_time +
1630 scan_rs_time + obj_copy_time;
1631 }
1633 if (PrintGCDetails) {
1634 gclog_or_tty->print_cr("%s%s, %1.8lf secs]",
1635 abandoned ? " (abandoned)" : "",
1636 (last_pause_included_initial_mark) ? " (initial-mark)" : "",
1637 elapsed_ms / 1000.0);
1639 if (!abandoned) {
1640 if (_satb_drain_time_set) {
1641 print_stats(1, "SATB Drain Time", _cur_satb_drain_time_ms);
1642 }
1643 if (_last_satb_drain_processed_buffers >= 0) {
1644 print_stats(2, "Processed Buffers", _last_satb_drain_processed_buffers);
1645 }
1646 if (parallel) {
1647 print_stats(1, "Parallel Time", _cur_collection_par_time_ms);
1648 print_par_stats(2, "Update RS (Start)", _par_last_update_rs_start_times_ms, false);
1649 print_par_stats(2, "Update RS", _par_last_update_rs_times_ms);
1650 print_par_buffers(3, "Processed Buffers",
1651 _par_last_update_rs_processed_buffers, true);
1652 print_par_stats(2, "Ext Root Scanning", _par_last_ext_root_scan_times_ms);
1653 print_par_stats(2, "Mark Stack Scanning", _par_last_mark_stack_scan_times_ms);
1654 print_par_stats(2, "Scan-Only Scanning", _par_last_scan_only_times_ms);
1655 print_par_buffers(3, "Scan-Only Regions",
1656 _par_last_scan_only_regions_scanned, true);
1657 print_par_stats(2, "Scan RS", _par_last_scan_rs_times_ms);
1658 print_par_stats(2, "Object Copy", _par_last_obj_copy_times_ms);
1659 print_par_stats(2, "Termination", _par_last_termination_times_ms);
1660 print_stats(2, "Other", parallel_other_time);
1661 print_stats(1, "Clear CT", _cur_clear_ct_time_ms);
1662 } else {
1663 print_stats(1, "Update RS", update_rs_time);
1664 print_stats(2, "Processed Buffers",
1665 (int)update_rs_processed_buffers);
1666 print_stats(1, "Ext Root Scanning", ext_root_scan_time);
1667 print_stats(1, "Mark Stack Scanning", mark_stack_scan_time);
1668 print_stats(1, "Scan-Only Scanning", scan_only_time);
1669 print_stats(1, "Scan RS", scan_rs_time);
1670 print_stats(1, "Object Copying", obj_copy_time);
1671 }
1672 }
1673 #ifndef PRODUCT
1674 print_stats(1, "Cur Clear CC", _cur_clear_cc_time_ms);
1675 print_stats(1, "Cum Clear CC", _cum_clear_cc_time_ms);
1676 print_stats(1, "Min Clear CC", _min_clear_cc_time_ms);
1677 print_stats(1, "Max Clear CC", _max_clear_cc_time_ms);
1678 if (_num_cc_clears > 0) {
1679 print_stats(1, "Avg Clear CC", _cum_clear_cc_time_ms / ((double)_num_cc_clears));
1680 }
1681 #endif
1682 print_stats(1, "Other", other_time_ms);
1683 for (int i = 0; i < _aux_num; ++i) {
1684 if (_cur_aux_times_set[i]) {
1685 char buffer[96];
1686 sprintf(buffer, "Aux%d", i);
1687 print_stats(1, buffer, _cur_aux_times_ms[i]);
1688 }
1689 }
1690 }
1691 if (PrintGCDetails)
1692 gclog_or_tty->print(" [");
1693 if (PrintGC || PrintGCDetails)
1694 _g1->print_size_transition(gclog_or_tty,
1695 _cur_collection_pause_used_at_start_bytes,
1696 _g1->used(), _g1->capacity());
1697 if (PrintGCDetails)
1698 gclog_or_tty->print_cr("]");
1700 _all_pause_times_ms->add(elapsed_ms);
1701 if (update_stats) {
1702 summary->record_total_time_ms(elapsed_ms);
1703 summary->record_other_time_ms(other_time_ms);
1704 }
1705 for (int i = 0; i < _aux_num; ++i)
1706 if (_cur_aux_times_set[i])
1707 _all_aux_times_ms[i].add(_cur_aux_times_ms[i]);
1709 // Reset marks-between-pauses counter.
1710 _n_marks_since_last_pause = 0;
1712 // Update the efficiency-since-mark vars.
1713 double proc_ms = elapsed_ms * (double) _parallel_gc_threads;
1714 if (elapsed_ms < MIN_TIMER_GRANULARITY) {
1715 // This usually happens due to the timer not having the required
1716 // granularity. Some Linuxes are the usual culprits.
1717 // We'll just set it to something (arbitrarily) small.
1718 proc_ms = 1.0;
1719 }
1720 double cur_efficiency = (double) freed_bytes / proc_ms;
1722 bool new_in_marking_window = _in_marking_window;
1723 bool new_in_marking_window_im = false;
1724 if (_should_initiate_conc_mark) {
1725 new_in_marking_window = true;
1726 new_in_marking_window_im = true;
1727 }
1729 if (in_young_gc_mode()) {
1730 if (_last_full_young_gc) {
1731 set_full_young_gcs(false);
1732 _last_full_young_gc = false;
1733 }
1735 if ( !_last_young_gc_full ) {
1736 if ( _should_revert_to_full_young_gcs ||
1737 _known_garbage_ratio < 0.05 ||
1738 (adaptive_young_list_length() &&
1739 (get_gc_eff_factor() * cur_efficiency < predict_young_gc_eff())) ) {
1740 set_full_young_gcs(true);
1741 }
1742 }
1743 _should_revert_to_full_young_gcs = false;
1745 if (_last_young_gc_full && !_during_marking)
1746 _young_gc_eff_seq->add(cur_efficiency);
1747 }
1749 _short_lived_surv_rate_group->start_adding_regions();
1750 // do that for any other surv rate groupsx
1752 // <NEW PREDICTION>
1754 if (update_stats) {
1755 double pause_time_ms = elapsed_ms;
1757 size_t diff = 0;
1758 if (_max_pending_cards >= _pending_cards)
1759 diff = _max_pending_cards - _pending_cards;
1760 _pending_card_diff_seq->add((double) diff);
1762 double cost_per_card_ms = 0.0;
1763 if (_pending_cards > 0) {
1764 cost_per_card_ms = update_rs_time / (double) _pending_cards;
1765 _cost_per_card_ms_seq->add(cost_per_card_ms);
1766 }
1768 double cost_per_scan_only_region_ms = 0.0;
1769 if (scan_only_regions_scanned > 0.0) {
1770 cost_per_scan_only_region_ms =
1771 scan_only_time / scan_only_regions_scanned;
1772 if (_in_marking_window_im)
1773 _cost_per_scan_only_region_ms_during_cm_seq->add(cost_per_scan_only_region_ms);
1774 else
1775 _cost_per_scan_only_region_ms_seq->add(cost_per_scan_only_region_ms);
1776 }
1778 size_t cards_scanned = _g1->cards_scanned();
1780 double cost_per_entry_ms = 0.0;
1781 if (cards_scanned > 10) {
1782 cost_per_entry_ms = scan_rs_time / (double) cards_scanned;
1783 if (_last_young_gc_full)
1784 _cost_per_entry_ms_seq->add(cost_per_entry_ms);
1785 else
1786 _partially_young_cost_per_entry_ms_seq->add(cost_per_entry_ms);
1787 }
1789 if (_max_rs_lengths > 0) {
1790 double cards_per_entry_ratio =
1791 (double) cards_scanned / (double) _max_rs_lengths;
1792 if (_last_young_gc_full)
1793 _fully_young_cards_per_entry_ratio_seq->add(cards_per_entry_ratio);
1794 else
1795 _partially_young_cards_per_entry_ratio_seq->add(cards_per_entry_ratio);
1796 }
1798 size_t rs_length_diff = _max_rs_lengths - _recorded_rs_lengths;
1799 if (rs_length_diff >= 0)
1800 _rs_length_diff_seq->add((double) rs_length_diff);
1802 size_t copied_bytes = surviving_bytes;
1803 double cost_per_byte_ms = 0.0;
1804 if (copied_bytes > 0) {
1805 cost_per_byte_ms = obj_copy_time / (double) copied_bytes;
1806 if (_in_marking_window)
1807 _cost_per_byte_ms_during_cm_seq->add(cost_per_byte_ms);
1808 else
1809 _cost_per_byte_ms_seq->add(cost_per_byte_ms);
1810 }
1812 double all_other_time_ms = pause_time_ms -
1813 (update_rs_time + scan_only_time + scan_rs_time + obj_copy_time +
1814 _mark_closure_time_ms + termination_time);
1816 double young_other_time_ms = 0.0;
1817 if (_recorded_young_regions > 0) {
1818 young_other_time_ms =
1819 _recorded_young_cset_choice_time_ms +
1820 _recorded_young_free_cset_time_ms;
1821 _young_other_cost_per_region_ms_seq->add(young_other_time_ms /
1822 (double) _recorded_young_regions);
1823 }
1824 double non_young_other_time_ms = 0.0;
1825 if (_recorded_non_young_regions > 0) {
1826 non_young_other_time_ms =
1827 _recorded_non_young_cset_choice_time_ms +
1828 _recorded_non_young_free_cset_time_ms;
1830 _non_young_other_cost_per_region_ms_seq->add(non_young_other_time_ms /
1831 (double) _recorded_non_young_regions);
1832 }
1834 double constant_other_time_ms = all_other_time_ms -
1835 (young_other_time_ms + non_young_other_time_ms);
1836 _constant_other_time_ms_seq->add(constant_other_time_ms);
1838 double survival_ratio = 0.0;
1839 if (_bytes_in_collection_set_before_gc > 0) {
1840 survival_ratio = (double) bytes_in_to_space_during_gc() /
1841 (double) _bytes_in_collection_set_before_gc;
1842 }
1844 _pending_cards_seq->add((double) _pending_cards);
1845 _scanned_cards_seq->add((double) cards_scanned);
1846 _rs_lengths_seq->add((double) _max_rs_lengths);
1848 double expensive_region_limit_ms =
1849 (double) MaxGCPauseMillis - predict_constant_other_time_ms();
1850 if (expensive_region_limit_ms < 0.0) {
1851 // this means that the other time was predicted to be longer than
1852 // than the max pause time
1853 expensive_region_limit_ms = (double) MaxGCPauseMillis;
1854 }
1855 _expensive_region_limit_ms = expensive_region_limit_ms;
1857 if (PREDICTIONS_VERBOSE) {
1858 gclog_or_tty->print_cr("");
1859 gclog_or_tty->print_cr("PREDICTIONS %1.4lf %d "
1860 "REGIONS %d %d %d %d "
1861 "PENDING_CARDS %d %d "
1862 "CARDS_SCANNED %d %d "
1863 "RS_LENGTHS %d %d "
1864 "SCAN_ONLY_SCAN %1.6lf %1.6lf "
1865 "RS_UPDATE %1.6lf %1.6lf RS_SCAN %1.6lf %1.6lf "
1866 "SURVIVAL_RATIO %1.6lf %1.6lf "
1867 "OBJECT_COPY %1.6lf %1.6lf OTHER_CONSTANT %1.6lf %1.6lf "
1868 "OTHER_YOUNG %1.6lf %1.6lf "
1869 "OTHER_NON_YOUNG %1.6lf %1.6lf "
1870 "VTIME_DIFF %1.6lf TERMINATION %1.6lf "
1871 "ELAPSED %1.6lf %1.6lf ",
1872 _cur_collection_start_sec,
1873 (!_last_young_gc_full) ? 2 :
1874 (last_pause_included_initial_mark) ? 1 : 0,
1875 _recorded_region_num,
1876 _recorded_young_regions,
1877 _recorded_scan_only_regions,
1878 _recorded_non_young_regions,
1879 _predicted_pending_cards, _pending_cards,
1880 _predicted_cards_scanned, cards_scanned,
1881 _predicted_rs_lengths, _max_rs_lengths,
1882 _predicted_scan_only_scan_time_ms, scan_only_time,
1883 _predicted_rs_update_time_ms, update_rs_time,
1884 _predicted_rs_scan_time_ms, scan_rs_time,
1885 _predicted_survival_ratio, survival_ratio,
1886 _predicted_object_copy_time_ms, obj_copy_time,
1887 _predicted_constant_other_time_ms, constant_other_time_ms,
1888 _predicted_young_other_time_ms, young_other_time_ms,
1889 _predicted_non_young_other_time_ms,
1890 non_young_other_time_ms,
1891 _vtime_diff_ms, termination_time,
1892 _predicted_pause_time_ms, elapsed_ms);
1893 }
1895 if (G1PolicyVerbose > 0) {
1896 gclog_or_tty->print_cr("Pause Time, predicted: %1.4lfms (predicted %s), actual: %1.4lfms",
1897 _predicted_pause_time_ms,
1898 (_within_target) ? "within" : "outside",
1899 elapsed_ms);
1900 }
1902 }
1904 _in_marking_window = new_in_marking_window;
1905 _in_marking_window_im = new_in_marking_window_im;
1906 _free_regions_at_end_of_collection = _g1->free_regions();
1907 _scan_only_regions_at_end_of_collection = _g1->young_list_length();
1908 calculate_young_list_min_length();
1909 calculate_young_list_target_config();
1911 // Note that _mmu_tracker->max_gc_time() returns the time in seconds.
1912 double update_rs_time_goal_ms = _mmu_tracker->max_gc_time() * MILLIUNITS * G1RSetUpdatingPauseTimePercent / 100.0;
1913 adjust_concurrent_refinement(update_rs_time, update_rs_processed_buffers, update_rs_time_goal_ms);
1915 // </NEW PREDICTION>
1917 _target_pause_time_ms = -1.0;
1918 }
1920 // <NEW PREDICTION>
1922 void G1CollectorPolicy::adjust_concurrent_refinement(double update_rs_time,
1923 double update_rs_processed_buffers,
1924 double goal_ms) {
1925 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set();
1926 ConcurrentG1Refine *cg1r = G1CollectedHeap::heap()->concurrent_g1_refine();
1928 if (G1UseAdaptiveConcRefinement) {
1929 const int k_gy = 3, k_gr = 6;
1930 const double inc_k = 1.1, dec_k = 0.9;
1932 int g = cg1r->green_zone();
1933 if (update_rs_time > goal_ms) {
1934 g = (int)(g * dec_k); // Can become 0, that's OK. That would mean a mutator-only processing.
1935 } else {
1936 if (update_rs_time < goal_ms && update_rs_processed_buffers > g) {
1937 g = (int)MAX2(g * inc_k, g + 1.0);
1938 }
1939 }
1940 // Change the refinement threads params
1941 cg1r->set_green_zone(g);
1942 cg1r->set_yellow_zone(g * k_gy);
1943 cg1r->set_red_zone(g * k_gr);
1944 cg1r->reinitialize_threads();
1946 int processing_threshold_delta = MAX2((int)(cg1r->green_zone() * sigma()), 1);
1947 int processing_threshold = MIN2(cg1r->green_zone() + processing_threshold_delta,
1948 cg1r->yellow_zone());
1949 // Change the barrier params
1950 dcqs.set_process_completed_threshold(processing_threshold);
1951 dcqs.set_max_completed_queue(cg1r->red_zone());
1952 }
1954 int curr_queue_size = dcqs.completed_buffers_num();
1955 if (curr_queue_size >= cg1r->yellow_zone()) {
1956 dcqs.set_completed_queue_padding(curr_queue_size);
1957 } else {
1958 dcqs.set_completed_queue_padding(0);
1959 }
1960 dcqs.notify_if_necessary();
1961 }
1963 double
1964 G1CollectorPolicy::
1965 predict_young_collection_elapsed_time_ms(size_t adjustment) {
1966 guarantee( adjustment == 0 || adjustment == 1, "invariant" );
1968 G1CollectedHeap* g1h = G1CollectedHeap::heap();
1969 size_t young_num = g1h->young_list_length();
1970 if (young_num == 0)
1971 return 0.0;
1973 young_num += adjustment;
1974 size_t pending_cards = predict_pending_cards();
1975 size_t rs_lengths = g1h->young_list_sampled_rs_lengths() +
1976 predict_rs_length_diff();
1977 size_t card_num;
1978 if (full_young_gcs())
1979 card_num = predict_young_card_num(rs_lengths);
1980 else
1981 card_num = predict_non_young_card_num(rs_lengths);
1982 size_t young_byte_size = young_num * HeapRegion::GrainBytes;
1983 double accum_yg_surv_rate =
1984 _short_lived_surv_rate_group->accum_surv_rate(adjustment);
1986 size_t bytes_to_copy =
1987 (size_t) (accum_yg_surv_rate * (double) HeapRegion::GrainBytes);
1989 return
1990 predict_rs_update_time_ms(pending_cards) +
1991 predict_rs_scan_time_ms(card_num) +
1992 predict_object_copy_time_ms(bytes_to_copy) +
1993 predict_young_other_time_ms(young_num) +
1994 predict_constant_other_time_ms();
1995 }
1997 double
1998 G1CollectorPolicy::predict_base_elapsed_time_ms(size_t pending_cards) {
1999 size_t rs_length = predict_rs_length_diff();
2000 size_t card_num;
2001 if (full_young_gcs())
2002 card_num = predict_young_card_num(rs_length);
2003 else
2004 card_num = predict_non_young_card_num(rs_length);
2005 return predict_base_elapsed_time_ms(pending_cards, card_num);
2006 }
2008 double
2009 G1CollectorPolicy::predict_base_elapsed_time_ms(size_t pending_cards,
2010 size_t scanned_cards) {
2011 return
2012 predict_rs_update_time_ms(pending_cards) +
2013 predict_rs_scan_time_ms(scanned_cards) +
2014 predict_constant_other_time_ms();
2015 }
2017 double
2018 G1CollectorPolicy::predict_region_elapsed_time_ms(HeapRegion* hr,
2019 bool young) {
2020 size_t rs_length = hr->rem_set()->occupied();
2021 size_t card_num;
2022 if (full_young_gcs())
2023 card_num = predict_young_card_num(rs_length);
2024 else
2025 card_num = predict_non_young_card_num(rs_length);
2026 size_t bytes_to_copy = predict_bytes_to_copy(hr);
2028 double region_elapsed_time_ms =
2029 predict_rs_scan_time_ms(card_num) +
2030 predict_object_copy_time_ms(bytes_to_copy);
2032 if (young)
2033 region_elapsed_time_ms += predict_young_other_time_ms(1);
2034 else
2035 region_elapsed_time_ms += predict_non_young_other_time_ms(1);
2037 return region_elapsed_time_ms;
2038 }
2040 size_t
2041 G1CollectorPolicy::predict_bytes_to_copy(HeapRegion* hr) {
2042 size_t bytes_to_copy;
2043 if (hr->is_marked())
2044 bytes_to_copy = hr->max_live_bytes();
2045 else {
2046 guarantee( hr->is_young() && hr->age_in_surv_rate_group() != -1,
2047 "invariant" );
2048 int age = hr->age_in_surv_rate_group();
2049 double yg_surv_rate = predict_yg_surv_rate(age, hr->surv_rate_group());
2050 bytes_to_copy = (size_t) ((double) hr->used() * yg_surv_rate);
2051 }
2053 return bytes_to_copy;
2054 }
2056 void
2057 G1CollectorPolicy::start_recording_regions() {
2058 _recorded_rs_lengths = 0;
2059 _recorded_scan_only_regions = 0;
2060 _recorded_young_regions = 0;
2061 _recorded_non_young_regions = 0;
2063 #if PREDICTIONS_VERBOSE
2064 _predicted_rs_lengths = 0;
2065 _predicted_cards_scanned = 0;
2067 _recorded_marked_bytes = 0;
2068 _recorded_young_bytes = 0;
2069 _predicted_bytes_to_copy = 0;
2070 #endif // PREDICTIONS_VERBOSE
2071 }
2073 void
2074 G1CollectorPolicy::record_cset_region(HeapRegion* hr, bool young) {
2075 if (young) {
2076 ++_recorded_young_regions;
2077 } else {
2078 ++_recorded_non_young_regions;
2079 }
2080 #if PREDICTIONS_VERBOSE
2081 if (young) {
2082 _recorded_young_bytes += hr->used();
2083 } else {
2084 _recorded_marked_bytes += hr->max_live_bytes();
2085 }
2086 _predicted_bytes_to_copy += predict_bytes_to_copy(hr);
2087 #endif // PREDICTIONS_VERBOSE
2089 size_t rs_length = hr->rem_set()->occupied();
2090 _recorded_rs_lengths += rs_length;
2091 }
2093 void
2094 G1CollectorPolicy::record_scan_only_regions(size_t scan_only_length) {
2095 _recorded_scan_only_regions = scan_only_length;
2096 }
2098 void
2099 G1CollectorPolicy::end_recording_regions() {
2100 #if PREDICTIONS_VERBOSE
2101 _predicted_pending_cards = predict_pending_cards();
2102 _predicted_rs_lengths = _recorded_rs_lengths + predict_rs_length_diff();
2103 if (full_young_gcs())
2104 _predicted_cards_scanned += predict_young_card_num(_predicted_rs_lengths);
2105 else
2106 _predicted_cards_scanned +=
2107 predict_non_young_card_num(_predicted_rs_lengths);
2108 _recorded_region_num = _recorded_young_regions + _recorded_non_young_regions;
2110 _predicted_scan_only_scan_time_ms =
2111 predict_scan_only_time_ms(_recorded_scan_only_regions);
2112 _predicted_rs_update_time_ms =
2113 predict_rs_update_time_ms(_g1->pending_card_num());
2114 _predicted_rs_scan_time_ms =
2115 predict_rs_scan_time_ms(_predicted_cards_scanned);
2116 _predicted_object_copy_time_ms =
2117 predict_object_copy_time_ms(_predicted_bytes_to_copy);
2118 _predicted_constant_other_time_ms =
2119 predict_constant_other_time_ms();
2120 _predicted_young_other_time_ms =
2121 predict_young_other_time_ms(_recorded_young_regions);
2122 _predicted_non_young_other_time_ms =
2123 predict_non_young_other_time_ms(_recorded_non_young_regions);
2125 _predicted_pause_time_ms =
2126 _predicted_scan_only_scan_time_ms +
2127 _predicted_rs_update_time_ms +
2128 _predicted_rs_scan_time_ms +
2129 _predicted_object_copy_time_ms +
2130 _predicted_constant_other_time_ms +
2131 _predicted_young_other_time_ms +
2132 _predicted_non_young_other_time_ms;
2133 #endif // PREDICTIONS_VERBOSE
2134 }
2136 void G1CollectorPolicy::check_if_region_is_too_expensive(double
2137 predicted_time_ms) {
2138 // I don't think we need to do this when in young GC mode since
2139 // marking will be initiated next time we hit the soft limit anyway...
2140 if (predicted_time_ms > _expensive_region_limit_ms) {
2141 if (!in_young_gc_mode()) {
2142 set_full_young_gcs(true);
2143 _should_initiate_conc_mark = true;
2144 } else
2145 // no point in doing another partial one
2146 _should_revert_to_full_young_gcs = true;
2147 }
2148 }
2150 // </NEW PREDICTION>
2153 void G1CollectorPolicy::update_recent_gc_times(double end_time_sec,
2154 double elapsed_ms) {
2155 _recent_gc_times_ms->add(elapsed_ms);
2156 _recent_prev_end_times_for_all_gcs_sec->add(end_time_sec);
2157 _prev_collection_pause_end_ms = end_time_sec * 1000.0;
2158 }
2160 double G1CollectorPolicy::recent_avg_time_for_pauses_ms() {
2161 if (_recent_pause_times_ms->num() == 0) return (double) MaxGCPauseMillis;
2162 else return _recent_pause_times_ms->avg();
2163 }
2165 double G1CollectorPolicy::recent_avg_time_for_CH_strong_ms() {
2166 if (_recent_CH_strong_roots_times_ms->num() == 0)
2167 return (double)MaxGCPauseMillis/3.0;
2168 else return _recent_CH_strong_roots_times_ms->avg();
2169 }
2171 double G1CollectorPolicy::recent_avg_time_for_G1_strong_ms() {
2172 if (_recent_G1_strong_roots_times_ms->num() == 0)
2173 return (double)MaxGCPauseMillis/3.0;
2174 else return _recent_G1_strong_roots_times_ms->avg();
2175 }
2177 double G1CollectorPolicy::recent_avg_time_for_evac_ms() {
2178 if (_recent_evac_times_ms->num() == 0) return (double)MaxGCPauseMillis/3.0;
2179 else return _recent_evac_times_ms->avg();
2180 }
2182 int G1CollectorPolicy::number_of_recent_gcs() {
2183 assert(_recent_CH_strong_roots_times_ms->num() ==
2184 _recent_G1_strong_roots_times_ms->num(), "Sequence out of sync");
2185 assert(_recent_G1_strong_roots_times_ms->num() ==
2186 _recent_evac_times_ms->num(), "Sequence out of sync");
2187 assert(_recent_evac_times_ms->num() ==
2188 _recent_pause_times_ms->num(), "Sequence out of sync");
2189 assert(_recent_pause_times_ms->num() ==
2190 _recent_CS_bytes_used_before->num(), "Sequence out of sync");
2191 assert(_recent_CS_bytes_used_before->num() ==
2192 _recent_CS_bytes_surviving->num(), "Sequence out of sync");
2193 return _recent_pause_times_ms->num();
2194 }
2196 double G1CollectorPolicy::recent_avg_survival_fraction() {
2197 return recent_avg_survival_fraction_work(_recent_CS_bytes_surviving,
2198 _recent_CS_bytes_used_before);
2199 }
2201 double G1CollectorPolicy::last_survival_fraction() {
2202 return last_survival_fraction_work(_recent_CS_bytes_surviving,
2203 _recent_CS_bytes_used_before);
2204 }
2206 double
2207 G1CollectorPolicy::recent_avg_survival_fraction_work(TruncatedSeq* surviving,
2208 TruncatedSeq* before) {
2209 assert(surviving->num() == before->num(), "Sequence out of sync");
2210 if (before->sum() > 0.0) {
2211 double recent_survival_rate = surviving->sum() / before->sum();
2212 // We exempt parallel collection from this check because Alloc Buffer
2213 // fragmentation can produce negative collections.
2214 // Further, we're now always doing parallel collection. But I'm still
2215 // leaving this here as a placeholder for a more precise assertion later.
2216 // (DLD, 10/05.)
2217 assert((true || ParallelGCThreads > 0) ||
2218 _g1->evacuation_failed() ||
2219 recent_survival_rate <= 1.0, "Or bad frac");
2220 return recent_survival_rate;
2221 } else {
2222 return 1.0; // Be conservative.
2223 }
2224 }
2226 double
2227 G1CollectorPolicy::last_survival_fraction_work(TruncatedSeq* surviving,
2228 TruncatedSeq* before) {
2229 assert(surviving->num() == before->num(), "Sequence out of sync");
2230 if (surviving->num() > 0 && before->last() > 0.0) {
2231 double last_survival_rate = surviving->last() / before->last();
2232 // We exempt parallel collection from this check because Alloc Buffer
2233 // fragmentation can produce negative collections.
2234 // Further, we're now always doing parallel collection. But I'm still
2235 // leaving this here as a placeholder for a more precise assertion later.
2236 // (DLD, 10/05.)
2237 assert((true || ParallelGCThreads > 0) ||
2238 last_survival_rate <= 1.0, "Or bad frac");
2239 return last_survival_rate;
2240 } else {
2241 return 1.0;
2242 }
2243 }
2245 static const int survival_min_obs = 5;
2246 static double survival_min_obs_limits[] = { 0.9, 0.7, 0.5, 0.3, 0.1 };
2247 static const double min_survival_rate = 0.1;
2249 double
2250 G1CollectorPolicy::conservative_avg_survival_fraction_work(double avg,
2251 double latest) {
2252 double res = avg;
2253 if (number_of_recent_gcs() < survival_min_obs) {
2254 res = MAX2(res, survival_min_obs_limits[number_of_recent_gcs()]);
2255 }
2256 res = MAX2(res, latest);
2257 res = MAX2(res, min_survival_rate);
2258 // In the parallel case, LAB fragmentation can produce "negative
2259 // collections"; so can evac failure. Cap at 1.0
2260 res = MIN2(res, 1.0);
2261 return res;
2262 }
2264 size_t G1CollectorPolicy::expansion_amount() {
2265 if ((int)(recent_avg_pause_time_ratio() * 100.0) > G1GCPercent) {
2266 // We will double the existing space, or take
2267 // G1ExpandByPercentOfAvailable % of the available expansion
2268 // space, whichever is smaller, bounded below by a minimum
2269 // expansion (unless that's all that's left.)
2270 const size_t min_expand_bytes = 1*M;
2271 size_t reserved_bytes = _g1->g1_reserved_obj_bytes();
2272 size_t committed_bytes = _g1->capacity();
2273 size_t uncommitted_bytes = reserved_bytes - committed_bytes;
2274 size_t expand_bytes;
2275 size_t expand_bytes_via_pct =
2276 uncommitted_bytes * G1ExpandByPercentOfAvailable / 100;
2277 expand_bytes = MIN2(expand_bytes_via_pct, committed_bytes);
2278 expand_bytes = MAX2(expand_bytes, min_expand_bytes);
2279 expand_bytes = MIN2(expand_bytes, uncommitted_bytes);
2280 if (G1PolicyVerbose > 1) {
2281 gclog_or_tty->print("Decided to expand: ratio = %5.2f, "
2282 "committed = %d%s, uncommited = %d%s, via pct = %d%s.\n"
2283 " Answer = %d.\n",
2284 recent_avg_pause_time_ratio(),
2285 byte_size_in_proper_unit(committed_bytes),
2286 proper_unit_for_byte_size(committed_bytes),
2287 byte_size_in_proper_unit(uncommitted_bytes),
2288 proper_unit_for_byte_size(uncommitted_bytes),
2289 byte_size_in_proper_unit(expand_bytes_via_pct),
2290 proper_unit_for_byte_size(expand_bytes_via_pct),
2291 byte_size_in_proper_unit(expand_bytes),
2292 proper_unit_for_byte_size(expand_bytes));
2293 }
2294 return expand_bytes;
2295 } else {
2296 return 0;
2297 }
2298 }
2300 void G1CollectorPolicy::note_start_of_mark_thread() {
2301 _mark_thread_startup_sec = os::elapsedTime();
2302 }
2304 class CountCSClosure: public HeapRegionClosure {
2305 G1CollectorPolicy* _g1_policy;
2306 public:
2307 CountCSClosure(G1CollectorPolicy* g1_policy) :
2308 _g1_policy(g1_policy) {}
2309 bool doHeapRegion(HeapRegion* r) {
2310 _g1_policy->_bytes_in_collection_set_before_gc += r->used();
2311 return false;
2312 }
2313 };
2315 void G1CollectorPolicy::count_CS_bytes_used() {
2316 CountCSClosure cs_closure(this);
2317 _g1->collection_set_iterate(&cs_closure);
2318 }
2320 static void print_indent(int level) {
2321 for (int j = 0; j < level+1; ++j)
2322 gclog_or_tty->print(" ");
2323 }
2325 void G1CollectorPolicy::print_summary (int level,
2326 const char* str,
2327 NumberSeq* seq) const {
2328 double sum = seq->sum();
2329 print_indent(level);
2330 gclog_or_tty->print_cr("%-24s = %8.2lf s (avg = %8.2lf ms)",
2331 str, sum / 1000.0, seq->avg());
2332 }
2334 void G1CollectorPolicy::print_summary_sd (int level,
2335 const char* str,
2336 NumberSeq* seq) const {
2337 print_summary(level, str, seq);
2338 print_indent(level + 5);
2339 gclog_or_tty->print_cr("(num = %5d, std dev = %8.2lf ms, max = %8.2lf ms)",
2340 seq->num(), seq->sd(), seq->maximum());
2341 }
2343 void G1CollectorPolicy::check_other_times(int level,
2344 NumberSeq* other_times_ms,
2345 NumberSeq* calc_other_times_ms) const {
2346 bool should_print = false;
2348 double max_sum = MAX2(fabs(other_times_ms->sum()),
2349 fabs(calc_other_times_ms->sum()));
2350 double min_sum = MIN2(fabs(other_times_ms->sum()),
2351 fabs(calc_other_times_ms->sum()));
2352 double sum_ratio = max_sum / min_sum;
2353 if (sum_ratio > 1.1) {
2354 should_print = true;
2355 print_indent(level + 1);
2356 gclog_or_tty->print_cr("## CALCULATED OTHER SUM DOESN'T MATCH RECORDED ###");
2357 }
2359 double max_avg = MAX2(fabs(other_times_ms->avg()),
2360 fabs(calc_other_times_ms->avg()));
2361 double min_avg = MIN2(fabs(other_times_ms->avg()),
2362 fabs(calc_other_times_ms->avg()));
2363 double avg_ratio = max_avg / min_avg;
2364 if (avg_ratio > 1.1) {
2365 should_print = true;
2366 print_indent(level + 1);
2367 gclog_or_tty->print_cr("## CALCULATED OTHER AVG DOESN'T MATCH RECORDED ###");
2368 }
2370 if (other_times_ms->sum() < -0.01) {
2371 print_indent(level + 1);
2372 gclog_or_tty->print_cr("## RECORDED OTHER SUM IS NEGATIVE ###");
2373 }
2375 if (other_times_ms->avg() < -0.01) {
2376 print_indent(level + 1);
2377 gclog_or_tty->print_cr("## RECORDED OTHER AVG IS NEGATIVE ###");
2378 }
2380 if (calc_other_times_ms->sum() < -0.01) {
2381 should_print = true;
2382 print_indent(level + 1);
2383 gclog_or_tty->print_cr("## CALCULATED OTHER SUM IS NEGATIVE ###");
2384 }
2386 if (calc_other_times_ms->avg() < -0.01) {
2387 should_print = true;
2388 print_indent(level + 1);
2389 gclog_or_tty->print_cr("## CALCULATED OTHER AVG IS NEGATIVE ###");
2390 }
2392 if (should_print)
2393 print_summary(level, "Other(Calc)", calc_other_times_ms);
2394 }
2396 void G1CollectorPolicy::print_summary(PauseSummary* summary) const {
2397 bool parallel = ParallelGCThreads > 0;
2398 MainBodySummary* body_summary = summary->main_body_summary();
2399 if (summary->get_total_seq()->num() > 0) {
2400 print_summary_sd(0, "Evacuation Pauses", summary->get_total_seq());
2401 if (body_summary != NULL) {
2402 print_summary(1, "SATB Drain", body_summary->get_satb_drain_seq());
2403 if (parallel) {
2404 print_summary(1, "Parallel Time", body_summary->get_parallel_seq());
2405 print_summary(2, "Update RS", body_summary->get_update_rs_seq());
2406 print_summary(2, "Ext Root Scanning",
2407 body_summary->get_ext_root_scan_seq());
2408 print_summary(2, "Mark Stack Scanning",
2409 body_summary->get_mark_stack_scan_seq());
2410 print_summary(2, "Scan-Only Scanning",
2411 body_summary->get_scan_only_seq());
2412 print_summary(2, "Scan RS", body_summary->get_scan_rs_seq());
2413 print_summary(2, "Object Copy", body_summary->get_obj_copy_seq());
2414 print_summary(2, "Termination", body_summary->get_termination_seq());
2415 print_summary(2, "Other", body_summary->get_parallel_other_seq());
2416 {
2417 NumberSeq* other_parts[] = {
2418 body_summary->get_update_rs_seq(),
2419 body_summary->get_ext_root_scan_seq(),
2420 body_summary->get_mark_stack_scan_seq(),
2421 body_summary->get_scan_only_seq(),
2422 body_summary->get_scan_rs_seq(),
2423 body_summary->get_obj_copy_seq(),
2424 body_summary->get_termination_seq()
2425 };
2426 NumberSeq calc_other_times_ms(body_summary->get_parallel_seq(),
2427 7, other_parts);
2428 check_other_times(2, body_summary->get_parallel_other_seq(),
2429 &calc_other_times_ms);
2430 }
2431 print_summary(1, "Mark Closure", body_summary->get_mark_closure_seq());
2432 print_summary(1, "Clear CT", body_summary->get_clear_ct_seq());
2433 } else {
2434 print_summary(1, "Update RS", body_summary->get_update_rs_seq());
2435 print_summary(1, "Ext Root Scanning",
2436 body_summary->get_ext_root_scan_seq());
2437 print_summary(1, "Mark Stack Scanning",
2438 body_summary->get_mark_stack_scan_seq());
2439 print_summary(1, "Scan-Only Scanning",
2440 body_summary->get_scan_only_seq());
2441 print_summary(1, "Scan RS", body_summary->get_scan_rs_seq());
2442 print_summary(1, "Object Copy", body_summary->get_obj_copy_seq());
2443 }
2444 }
2445 print_summary(1, "Other", summary->get_other_seq());
2446 {
2447 NumberSeq calc_other_times_ms;
2448 if (body_summary != NULL) {
2449 // not abandoned
2450 if (parallel) {
2451 // parallel
2452 NumberSeq* other_parts[] = {
2453 body_summary->get_satb_drain_seq(),
2454 body_summary->get_parallel_seq(),
2455 body_summary->get_clear_ct_seq()
2456 };
2457 calc_other_times_ms = NumberSeq(summary->get_total_seq(),
2458 3, other_parts);
2459 } else {
2460 // serial
2461 NumberSeq* other_parts[] = {
2462 body_summary->get_satb_drain_seq(),
2463 body_summary->get_update_rs_seq(),
2464 body_summary->get_ext_root_scan_seq(),
2465 body_summary->get_mark_stack_scan_seq(),
2466 body_summary->get_scan_only_seq(),
2467 body_summary->get_scan_rs_seq(),
2468 body_summary->get_obj_copy_seq()
2469 };
2470 calc_other_times_ms = NumberSeq(summary->get_total_seq(),
2471 7, other_parts);
2472 }
2473 } else {
2474 // abandoned
2475 calc_other_times_ms = NumberSeq();
2476 }
2477 check_other_times(1, summary->get_other_seq(), &calc_other_times_ms);
2478 }
2479 } else {
2480 print_indent(0);
2481 gclog_or_tty->print_cr("none");
2482 }
2483 gclog_or_tty->print_cr("");
2484 }
2486 void
2487 G1CollectorPolicy::print_abandoned_summary(PauseSummary* summary) const {
2488 bool printed = false;
2489 if (summary->get_total_seq()->num() > 0) {
2490 printed = true;
2491 print_summary(summary);
2492 }
2493 if (!printed) {
2494 print_indent(0);
2495 gclog_or_tty->print_cr("none");
2496 gclog_or_tty->print_cr("");
2497 }
2498 }
2500 void G1CollectorPolicy::print_tracing_info() const {
2501 if (TraceGen0Time) {
2502 gclog_or_tty->print_cr("ALL PAUSES");
2503 print_summary_sd(0, "Total", _all_pause_times_ms);
2504 gclog_or_tty->print_cr("");
2505 gclog_or_tty->print_cr("");
2506 gclog_or_tty->print_cr(" Full Young GC Pauses: %8d", _full_young_pause_num);
2507 gclog_or_tty->print_cr(" Partial Young GC Pauses: %8d", _partial_young_pause_num);
2508 gclog_or_tty->print_cr("");
2510 gclog_or_tty->print_cr("EVACUATION PAUSES");
2511 print_summary(_summary);
2513 gclog_or_tty->print_cr("ABANDONED PAUSES");
2514 print_abandoned_summary(_abandoned_summary);
2516 gclog_or_tty->print_cr("MISC");
2517 print_summary_sd(0, "Stop World", _all_stop_world_times_ms);
2518 print_summary_sd(0, "Yields", _all_yield_times_ms);
2519 for (int i = 0; i < _aux_num; ++i) {
2520 if (_all_aux_times_ms[i].num() > 0) {
2521 char buffer[96];
2522 sprintf(buffer, "Aux%d", i);
2523 print_summary_sd(0, buffer, &_all_aux_times_ms[i]);
2524 }
2525 }
2527 size_t all_region_num = _region_num_young + _region_num_tenured;
2528 gclog_or_tty->print_cr(" New Regions %8d, Young %8d (%6.2lf%%), "
2529 "Tenured %8d (%6.2lf%%)",
2530 all_region_num,
2531 _region_num_young,
2532 (double) _region_num_young / (double) all_region_num * 100.0,
2533 _region_num_tenured,
2534 (double) _region_num_tenured / (double) all_region_num * 100.0);
2535 }
2536 if (TraceGen1Time) {
2537 if (_all_full_gc_times_ms->num() > 0) {
2538 gclog_or_tty->print("\n%4d full_gcs: total time = %8.2f s",
2539 _all_full_gc_times_ms->num(),
2540 _all_full_gc_times_ms->sum() / 1000.0);
2541 gclog_or_tty->print_cr(" (avg = %8.2fms).", _all_full_gc_times_ms->avg());
2542 gclog_or_tty->print_cr(" [std. dev = %8.2f ms, max = %8.2f ms]",
2543 _all_full_gc_times_ms->sd(),
2544 _all_full_gc_times_ms->maximum());
2545 }
2546 }
2547 }
2549 void G1CollectorPolicy::print_yg_surv_rate_info() const {
2550 #ifndef PRODUCT
2551 _short_lived_surv_rate_group->print_surv_rate_summary();
2552 // add this call for any other surv rate groups
2553 #endif // PRODUCT
2554 }
2556 bool
2557 G1CollectorPolicy::should_add_next_region_to_young_list() {
2558 assert(in_young_gc_mode(), "should be in young GC mode");
2559 bool ret;
2560 size_t young_list_length = _g1->young_list_length();
2561 size_t young_list_max_length = _young_list_target_length;
2562 if (G1FixedEdenSize) {
2563 young_list_max_length -= _max_survivor_regions;
2564 }
2565 if (young_list_length < young_list_max_length) {
2566 ret = true;
2567 ++_region_num_young;
2568 } else {
2569 ret = false;
2570 ++_region_num_tenured;
2571 }
2573 return ret;
2574 }
2576 #ifndef PRODUCT
2577 // for debugging, bit of a hack...
2578 static char*
2579 region_num_to_mbs(int length) {
2580 static char buffer[64];
2581 double bytes = (double) (length * HeapRegion::GrainBytes);
2582 double mbs = bytes / (double) (1024 * 1024);
2583 sprintf(buffer, "%7.2lfMB", mbs);
2584 return buffer;
2585 }
2586 #endif // PRODUCT
2588 size_t G1CollectorPolicy::max_regions(int purpose) {
2589 switch (purpose) {
2590 case GCAllocForSurvived:
2591 return _max_survivor_regions;
2592 case GCAllocForTenured:
2593 return REGIONS_UNLIMITED;
2594 default:
2595 ShouldNotReachHere();
2596 return REGIONS_UNLIMITED;
2597 };
2598 }
2600 // Calculates survivor space parameters.
2601 void G1CollectorPolicy::calculate_survivors_policy()
2602 {
2603 if (G1FixedSurvivorSpaceSize == 0) {
2604 _max_survivor_regions = _young_list_target_length / SurvivorRatio;
2605 } else {
2606 _max_survivor_regions = G1FixedSurvivorSpaceSize / HeapRegion::GrainBytes;
2607 }
2609 if (G1FixedTenuringThreshold) {
2610 _tenuring_threshold = MaxTenuringThreshold;
2611 } else {
2612 _tenuring_threshold = _survivors_age_table.compute_tenuring_threshold(
2613 HeapRegion::GrainWords * _max_survivor_regions);
2614 }
2615 }
2617 bool
2618 G1CollectorPolicy_BestRegionsFirst::should_do_collection_pause(size_t
2619 word_size) {
2620 assert(_g1->regions_accounted_for(), "Region leakage!");
2621 // Initiate a pause when we reach the steady-state "used" target.
2622 size_t used_hard = (_g1->capacity() / 100) * G1SteadyStateUsed;
2623 size_t used_soft =
2624 MAX2((_g1->capacity() / 100) * (G1SteadyStateUsed - G1SteadyStateUsedDelta),
2625 used_hard/2);
2626 size_t used = _g1->used();
2628 double max_pause_time_ms = _mmu_tracker->max_gc_time() * 1000.0;
2630 size_t young_list_length = _g1->young_list_length();
2631 size_t young_list_max_length = _young_list_target_length;
2632 if (G1FixedEdenSize) {
2633 young_list_max_length -= _max_survivor_regions;
2634 }
2635 bool reached_target_length = young_list_length >= young_list_max_length;
2637 if (in_young_gc_mode()) {
2638 if (reached_target_length) {
2639 assert( young_list_length > 0 && _g1->young_list_length() > 0,
2640 "invariant" );
2641 _target_pause_time_ms = max_pause_time_ms;
2642 return true;
2643 }
2644 } else {
2645 guarantee( false, "should not reach here" );
2646 }
2648 return false;
2649 }
2651 #ifndef PRODUCT
2652 class HRSortIndexIsOKClosure: public HeapRegionClosure {
2653 CollectionSetChooser* _chooser;
2654 public:
2655 HRSortIndexIsOKClosure(CollectionSetChooser* chooser) :
2656 _chooser(chooser) {}
2658 bool doHeapRegion(HeapRegion* r) {
2659 if (!r->continuesHumongous()) {
2660 assert(_chooser->regionProperlyOrdered(r), "Ought to be.");
2661 }
2662 return false;
2663 }
2664 };
2666 bool G1CollectorPolicy_BestRegionsFirst::assertMarkedBytesDataOK() {
2667 HRSortIndexIsOKClosure cl(_collectionSetChooser);
2668 _g1->heap_region_iterate(&cl);
2669 return true;
2670 }
2671 #endif
2673 void
2674 G1CollectorPolicy_BestRegionsFirst::
2675 record_collection_pause_start(double start_time_sec, size_t start_used) {
2676 G1CollectorPolicy::record_collection_pause_start(start_time_sec, start_used);
2677 }
2679 class NextNonCSElemFinder: public HeapRegionClosure {
2680 HeapRegion* _res;
2681 public:
2682 NextNonCSElemFinder(): _res(NULL) {}
2683 bool doHeapRegion(HeapRegion* r) {
2684 if (!r->in_collection_set()) {
2685 _res = r;
2686 return true;
2687 } else {
2688 return false;
2689 }
2690 }
2691 HeapRegion* res() { return _res; }
2692 };
2694 class KnownGarbageClosure: public HeapRegionClosure {
2695 CollectionSetChooser* _hrSorted;
2697 public:
2698 KnownGarbageClosure(CollectionSetChooser* hrSorted) :
2699 _hrSorted(hrSorted)
2700 {}
2702 bool doHeapRegion(HeapRegion* r) {
2703 // We only include humongous regions in collection
2704 // sets when concurrent mark shows that their contained object is
2705 // unreachable.
2707 // Do we have any marking information for this region?
2708 if (r->is_marked()) {
2709 // We don't include humongous regions in collection
2710 // sets because we collect them immediately at the end of a marking
2711 // cycle. We also don't include young regions because we *must*
2712 // include them in the next collection pause.
2713 if (!r->isHumongous() && !r->is_young()) {
2714 _hrSorted->addMarkedHeapRegion(r);
2715 }
2716 }
2717 return false;
2718 }
2719 };
2721 class ParKnownGarbageHRClosure: public HeapRegionClosure {
2722 CollectionSetChooser* _hrSorted;
2723 jint _marked_regions_added;
2724 jint _chunk_size;
2725 jint _cur_chunk_idx;
2726 jint _cur_chunk_end; // Cur chunk [_cur_chunk_idx, _cur_chunk_end)
2727 int _worker;
2728 int _invokes;
2730 void get_new_chunk() {
2731 _cur_chunk_idx = _hrSorted->getParMarkedHeapRegionChunk(_chunk_size);
2732 _cur_chunk_end = _cur_chunk_idx + _chunk_size;
2733 }
2734 void add_region(HeapRegion* r) {
2735 if (_cur_chunk_idx == _cur_chunk_end) {
2736 get_new_chunk();
2737 }
2738 assert(_cur_chunk_idx < _cur_chunk_end, "postcondition");
2739 _hrSorted->setMarkedHeapRegion(_cur_chunk_idx, r);
2740 _marked_regions_added++;
2741 _cur_chunk_idx++;
2742 }
2744 public:
2745 ParKnownGarbageHRClosure(CollectionSetChooser* hrSorted,
2746 jint chunk_size,
2747 int worker) :
2748 _hrSorted(hrSorted), _chunk_size(chunk_size), _worker(worker),
2749 _marked_regions_added(0), _cur_chunk_idx(0), _cur_chunk_end(0),
2750 _invokes(0)
2751 {}
2753 bool doHeapRegion(HeapRegion* r) {
2754 // We only include humongous regions in collection
2755 // sets when concurrent mark shows that their contained object is
2756 // unreachable.
2757 _invokes++;
2759 // Do we have any marking information for this region?
2760 if (r->is_marked()) {
2761 // We don't include humongous regions in collection
2762 // sets because we collect them immediately at the end of a marking
2763 // cycle.
2764 // We also do not include young regions in collection sets
2765 if (!r->isHumongous() && !r->is_young()) {
2766 add_region(r);
2767 }
2768 }
2769 return false;
2770 }
2771 jint marked_regions_added() { return _marked_regions_added; }
2772 int invokes() { return _invokes; }
2773 };
2775 class ParKnownGarbageTask: public AbstractGangTask {
2776 CollectionSetChooser* _hrSorted;
2777 jint _chunk_size;
2778 G1CollectedHeap* _g1;
2779 public:
2780 ParKnownGarbageTask(CollectionSetChooser* hrSorted, jint chunk_size) :
2781 AbstractGangTask("ParKnownGarbageTask"),
2782 _hrSorted(hrSorted), _chunk_size(chunk_size),
2783 _g1(G1CollectedHeap::heap())
2784 {}
2786 void work(int i) {
2787 ParKnownGarbageHRClosure parKnownGarbageCl(_hrSorted, _chunk_size, i);
2788 // Back to zero for the claim value.
2789 _g1->heap_region_par_iterate_chunked(&parKnownGarbageCl, i,
2790 HeapRegion::InitialClaimValue);
2791 jint regions_added = parKnownGarbageCl.marked_regions_added();
2792 _hrSorted->incNumMarkedHeapRegions(regions_added);
2793 if (G1PrintParCleanupStats) {
2794 gclog_or_tty->print(" Thread %d called %d times, added %d regions to list.\n",
2795 i, parKnownGarbageCl.invokes(), regions_added);
2796 }
2797 }
2798 };
2800 void
2801 G1CollectorPolicy_BestRegionsFirst::
2802 record_concurrent_mark_cleanup_end(size_t freed_bytes,
2803 size_t max_live_bytes) {
2804 double start;
2805 if (G1PrintParCleanupStats) start = os::elapsedTime();
2806 record_concurrent_mark_cleanup_end_work1(freed_bytes, max_live_bytes);
2808 _collectionSetChooser->clearMarkedHeapRegions();
2809 double clear_marked_end;
2810 if (G1PrintParCleanupStats) {
2811 clear_marked_end = os::elapsedTime();
2812 gclog_or_tty->print_cr(" clear marked regions + work1: %8.3f ms.",
2813 (clear_marked_end - start)*1000.0);
2814 }
2815 if (ParallelGCThreads > 0) {
2816 const size_t OverpartitionFactor = 4;
2817 const size_t MinChunkSize = 8;
2818 const size_t ChunkSize =
2819 MAX2(_g1->n_regions() / (ParallelGCThreads * OverpartitionFactor),
2820 MinChunkSize);
2821 _collectionSetChooser->prepareForAddMarkedHeapRegionsPar(_g1->n_regions(),
2822 ChunkSize);
2823 ParKnownGarbageTask parKnownGarbageTask(_collectionSetChooser,
2824 (int) ChunkSize);
2825 _g1->workers()->run_task(&parKnownGarbageTask);
2827 assert(_g1->check_heap_region_claim_values(HeapRegion::InitialClaimValue),
2828 "sanity check");
2829 } else {
2830 KnownGarbageClosure knownGarbagecl(_collectionSetChooser);
2831 _g1->heap_region_iterate(&knownGarbagecl);
2832 }
2833 double known_garbage_end;
2834 if (G1PrintParCleanupStats) {
2835 known_garbage_end = os::elapsedTime();
2836 gclog_or_tty->print_cr(" compute known garbage: %8.3f ms.",
2837 (known_garbage_end - clear_marked_end)*1000.0);
2838 }
2839 _collectionSetChooser->sortMarkedHeapRegions();
2840 double sort_end;
2841 if (G1PrintParCleanupStats) {
2842 sort_end = os::elapsedTime();
2843 gclog_or_tty->print_cr(" sorting: %8.3f ms.",
2844 (sort_end - known_garbage_end)*1000.0);
2845 }
2847 record_concurrent_mark_cleanup_end_work2();
2848 double work2_end;
2849 if (G1PrintParCleanupStats) {
2850 work2_end = os::elapsedTime();
2851 gclog_or_tty->print_cr(" work2: %8.3f ms.",
2852 (work2_end - sort_end)*1000.0);
2853 }
2854 }
2856 // Add the heap region to the collection set and return the conservative
2857 // estimate of the number of live bytes.
2858 void G1CollectorPolicy::
2859 add_to_collection_set(HeapRegion* hr) {
2860 if (G1PrintHeapRegions) {
2861 gclog_or_tty->print_cr("added region to cset %d:["PTR_FORMAT", "PTR_FORMAT"], "
2862 "top "PTR_FORMAT", young %s",
2863 hr->hrs_index(), hr->bottom(), hr->end(),
2864 hr->top(), (hr->is_young()) ? "YES" : "NO");
2865 }
2867 if (_g1->mark_in_progress())
2868 _g1->concurrent_mark()->registerCSetRegion(hr);
2870 assert(!hr->in_collection_set(),
2871 "should not already be in the CSet");
2872 hr->set_in_collection_set(true);
2873 hr->set_next_in_collection_set(_collection_set);
2874 _collection_set = hr;
2875 _collection_set_size++;
2876 _collection_set_bytes_used_before += hr->used();
2877 _g1->register_region_with_in_cset_fast_test(hr);
2878 }
2880 void
2881 G1CollectorPolicy_BestRegionsFirst::
2882 choose_collection_set() {
2883 double non_young_start_time_sec;
2884 start_recording_regions();
2886 guarantee(_target_pause_time_ms > -1.0
2887 NOT_PRODUCT(|| Universe::heap()->gc_cause() == GCCause::_scavenge_alot),
2888 "_target_pause_time_ms should have been set!");
2889 #ifndef PRODUCT
2890 if (_target_pause_time_ms <= -1.0) {
2891 assert(ScavengeALot && Universe::heap()->gc_cause() == GCCause::_scavenge_alot, "Error");
2892 _target_pause_time_ms = _mmu_tracker->max_gc_time() * 1000.0;
2893 }
2894 #endif
2895 assert(_collection_set == NULL, "Precondition");
2897 double base_time_ms = predict_base_elapsed_time_ms(_pending_cards);
2898 double predicted_pause_time_ms = base_time_ms;
2900 double target_time_ms = _target_pause_time_ms;
2901 double time_remaining_ms = target_time_ms - base_time_ms;
2903 // the 10% and 50% values are arbitrary...
2904 if (time_remaining_ms < 0.10*target_time_ms) {
2905 time_remaining_ms = 0.50 * target_time_ms;
2906 _within_target = false;
2907 } else {
2908 _within_target = true;
2909 }
2911 // We figure out the number of bytes available for future to-space.
2912 // For new regions without marking information, we must assume the
2913 // worst-case of complete survival. If we have marking information for a
2914 // region, we can bound the amount of live data. We can add a number of
2915 // such regions, as long as the sum of the live data bounds does not
2916 // exceed the available evacuation space.
2917 size_t max_live_bytes = _g1->free_regions() * HeapRegion::GrainBytes;
2919 size_t expansion_bytes =
2920 _g1->expansion_regions() * HeapRegion::GrainBytes;
2922 _collection_set_bytes_used_before = 0;
2923 _collection_set_size = 0;
2925 // Adjust for expansion and slop.
2926 max_live_bytes = max_live_bytes + expansion_bytes;
2928 assert(_g1->regions_accounted_for(), "Region leakage!");
2930 HeapRegion* hr;
2931 if (in_young_gc_mode()) {
2932 double young_start_time_sec = os::elapsedTime();
2934 if (G1PolicyVerbose > 0) {
2935 gclog_or_tty->print_cr("Adding %d young regions to the CSet",
2936 _g1->young_list_length());
2937 }
2938 _young_cset_length = 0;
2939 _last_young_gc_full = full_young_gcs() ? true : false;
2940 if (_last_young_gc_full)
2941 ++_full_young_pause_num;
2942 else
2943 ++_partial_young_pause_num;
2944 hr = _g1->pop_region_from_young_list();
2945 while (hr != NULL) {
2947 assert( hr->young_index_in_cset() == -1, "invariant" );
2948 assert( hr->age_in_surv_rate_group() != -1, "invariant" );
2949 hr->set_young_index_in_cset((int) _young_cset_length);
2951 ++_young_cset_length;
2952 double predicted_time_ms = predict_region_elapsed_time_ms(hr, true);
2953 time_remaining_ms -= predicted_time_ms;
2954 predicted_pause_time_ms += predicted_time_ms;
2955 assert(!hr->in_collection_set(), "invariant");
2956 add_to_collection_set(hr);
2957 record_cset_region(hr, true);
2958 max_live_bytes -= MIN2(hr->max_live_bytes(), max_live_bytes);
2959 if (G1PolicyVerbose > 0) {
2960 gclog_or_tty->print_cr(" Added [" PTR_FORMAT ", " PTR_FORMAT") to CS.",
2961 hr->bottom(), hr->end());
2962 gclog_or_tty->print_cr(" (" SIZE_FORMAT " KB left in heap.)",
2963 max_live_bytes/K);
2964 }
2965 hr = _g1->pop_region_from_young_list();
2966 }
2968 record_scan_only_regions(_g1->young_list_scan_only_length());
2970 double young_end_time_sec = os::elapsedTime();
2971 _recorded_young_cset_choice_time_ms =
2972 (young_end_time_sec - young_start_time_sec) * 1000.0;
2974 non_young_start_time_sec = os::elapsedTime();
2976 if (_young_cset_length > 0 && _last_young_gc_full) {
2977 // don't bother adding more regions...
2978 goto choose_collection_set_end;
2979 }
2980 }
2982 if (!in_young_gc_mode() || !full_young_gcs()) {
2983 bool should_continue = true;
2984 NumberSeq seq;
2985 double avg_prediction = 100000000000000000.0; // something very large
2986 do {
2987 hr = _collectionSetChooser->getNextMarkedRegion(time_remaining_ms,
2988 avg_prediction);
2989 if (hr != NULL) {
2990 double predicted_time_ms = predict_region_elapsed_time_ms(hr, false);
2991 time_remaining_ms -= predicted_time_ms;
2992 predicted_pause_time_ms += predicted_time_ms;
2993 add_to_collection_set(hr);
2994 record_cset_region(hr, false);
2995 max_live_bytes -= MIN2(hr->max_live_bytes(), max_live_bytes);
2996 if (G1PolicyVerbose > 0) {
2997 gclog_or_tty->print_cr(" (" SIZE_FORMAT " KB left in heap.)",
2998 max_live_bytes/K);
2999 }
3000 seq.add(predicted_time_ms);
3001 avg_prediction = seq.avg() + seq.sd();
3002 }
3003 should_continue =
3004 ( hr != NULL) &&
3005 ( (adaptive_young_list_length()) ? time_remaining_ms > 0.0
3006 : _collection_set_size < _young_list_fixed_length );
3007 } while (should_continue);
3009 if (!adaptive_young_list_length() &&
3010 _collection_set_size < _young_list_fixed_length)
3011 _should_revert_to_full_young_gcs = true;
3012 }
3014 choose_collection_set_end:
3015 count_CS_bytes_used();
3017 end_recording_regions();
3019 double non_young_end_time_sec = os::elapsedTime();
3020 _recorded_non_young_cset_choice_time_ms =
3021 (non_young_end_time_sec - non_young_start_time_sec) * 1000.0;
3022 }
3024 void G1CollectorPolicy_BestRegionsFirst::record_full_collection_end() {
3025 G1CollectorPolicy::record_full_collection_end();
3026 _collectionSetChooser->updateAfterFullCollection();
3027 }
3029 void G1CollectorPolicy_BestRegionsFirst::
3030 expand_if_possible(size_t numRegions) {
3031 size_t expansion_bytes = numRegions * HeapRegion::GrainBytes;
3032 _g1->expand(expansion_bytes);
3033 }
3035 void G1CollectorPolicy_BestRegionsFirst::
3036 record_collection_pause_end(bool abandoned) {
3037 G1CollectorPolicy::record_collection_pause_end(abandoned);
3038 assert(assertMarkedBytesDataOK(), "Marked regions not OK at pause end.");
3039 }