Tue, 23 Feb 2010 23:14:34 -0500
6928073: G1: use existing command line parameters for marking cycle initiation
Summary: replace the combination of the G1SteadyStateUsed / G1SteadyStateUsedDelta parameteres to decide the marking initiation threshold and instead use InitiatingHeapOccupancyPercent.
Reviewed-by: ysr, 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 initialize_gc_policy_counters();
304 if (G1Gen) {
305 _in_young_gc_mode = true;
307 if (G1YoungGenSize == 0) {
308 set_adaptive_young_list_length(true);
309 _young_list_fixed_length = 0;
310 } else {
311 set_adaptive_young_list_length(false);
312 _young_list_fixed_length = (G1YoungGenSize / HeapRegion::GrainBytes);
313 }
314 _free_regions_at_end_of_collection = _g1->free_regions();
315 _scan_only_regions_at_end_of_collection = 0;
316 calculate_young_list_min_length();
317 guarantee( _young_list_min_length == 0, "invariant, not enough info" );
318 calculate_young_list_target_config();
319 } else {
320 _young_list_fixed_length = 0;
321 _in_young_gc_mode = false;
322 }
323 }
325 // Create the jstat counters for the policy.
326 void G1CollectorPolicy::initialize_gc_policy_counters()
327 {
328 _gc_policy_counters = new GCPolicyCounters("GarbageFirst", 1, 2 + G1Gen);
329 }
331 void G1CollectorPolicy::calculate_young_list_min_length() {
332 _young_list_min_length = 0;
334 if (!adaptive_young_list_length())
335 return;
337 if (_alloc_rate_ms_seq->num() > 3) {
338 double now_sec = os::elapsedTime();
339 double when_ms = _mmu_tracker->when_max_gc_sec(now_sec) * 1000.0;
340 double alloc_rate_ms = predict_alloc_rate_ms();
341 int min_regions = (int) ceil(alloc_rate_ms * when_ms);
342 int current_region_num = (int) _g1->young_list_length();
343 _young_list_min_length = min_regions + current_region_num;
344 }
345 }
347 void G1CollectorPolicy::calculate_young_list_target_config() {
348 if (adaptive_young_list_length()) {
349 size_t rs_lengths = (size_t) get_new_prediction(_rs_lengths_seq);
350 calculate_young_list_target_config(rs_lengths);
351 } else {
352 if (full_young_gcs())
353 _young_list_target_length = _young_list_fixed_length;
354 else
355 _young_list_target_length = _young_list_fixed_length / 2;
356 _young_list_target_length = MAX2(_young_list_target_length, (size_t)1);
357 size_t so_length = calculate_optimal_so_length(_young_list_target_length);
358 guarantee( so_length < _young_list_target_length, "invariant" );
359 _young_list_so_prefix_length = so_length;
360 }
361 calculate_survivors_policy();
362 }
364 // This method calculate the optimal scan-only set for a fixed young
365 // gen size. I couldn't work out how to reuse the more elaborate one,
366 // i.e. calculate_young_list_target_config(rs_length), as the loops are
367 // fundamentally different (the other one finds a config for different
368 // S-O lengths, whereas here we need to do the opposite).
369 size_t G1CollectorPolicy::calculate_optimal_so_length(
370 size_t young_list_length) {
371 if (!G1UseScanOnlyPrefix)
372 return 0;
374 if (_all_pause_times_ms->num() < 3) {
375 // we won't use a scan-only set at the beginning to allow the rest
376 // of the predictors to warm up
377 return 0;
378 }
380 if (_cost_per_scan_only_region_ms_seq->num() < 3) {
381 // then, we'll only set the S-O set to 1 for a little bit of time,
382 // to get enough information on the scanning cost
383 return 1;
384 }
386 size_t pending_cards = (size_t) get_new_prediction(_pending_cards_seq);
387 size_t rs_lengths = (size_t) get_new_prediction(_rs_lengths_seq);
388 size_t adj_rs_lengths = rs_lengths + predict_rs_length_diff();
389 size_t scanned_cards;
390 if (full_young_gcs())
391 scanned_cards = predict_young_card_num(adj_rs_lengths);
392 else
393 scanned_cards = predict_non_young_card_num(adj_rs_lengths);
394 double base_time_ms = predict_base_elapsed_time_ms(pending_cards,
395 scanned_cards);
397 size_t so_length = 0;
398 double max_gc_eff = 0.0;
399 for (size_t i = 0; i < young_list_length; ++i) {
400 double gc_eff = 0.0;
401 double pause_time_ms = 0.0;
402 predict_gc_eff(young_list_length, i, base_time_ms,
403 &gc_eff, &pause_time_ms);
404 if (gc_eff > max_gc_eff) {
405 max_gc_eff = gc_eff;
406 so_length = i;
407 }
408 }
410 // set it to 95% of the optimal to make sure we sample the "area"
411 // around the optimal length to get up-to-date survival rate data
412 return so_length * 950 / 1000;
413 }
415 // This is a really cool piece of code! It finds the best
416 // target configuration (young length / scan-only prefix length) so
417 // that GC efficiency is maximized and that we also meet a pause
418 // time. It's a triple nested loop. These loops are explained below
419 // from the inside-out :-)
420 //
421 // (a) The innermost loop will try to find the optimal young length
422 // for a fixed S-O length. It uses a binary search to speed up the
423 // process. We assume that, for a fixed S-O length, as we add more
424 // young regions to the CSet, the GC efficiency will only go up (I'll
425 // skip the proof). So, using a binary search to optimize this process
426 // makes perfect sense.
427 //
428 // (b) The middle loop will fix the S-O length before calling the
429 // innermost one. It will vary it between two parameters, increasing
430 // it by a given increment.
431 //
432 // (c) The outermost loop will call the middle loop three times.
433 // (1) The first time it will explore all possible S-O length values
434 // from 0 to as large as it can get, using a coarse increment (to
435 // quickly "home in" to where the optimal seems to be).
436 // (2) The second time it will explore the values around the optimal
437 // that was found by the first iteration using a fine increment.
438 // (3) Once the optimal config has been determined by the second
439 // iteration, we'll redo the calculation, but setting the S-O length
440 // to 95% of the optimal to make sure we sample the "area"
441 // around the optimal length to get up-to-date survival rate data
442 //
443 // Termination conditions for the iterations are several: the pause
444 // time is over the limit, we do not have enough to-space, etc.
446 void G1CollectorPolicy::calculate_young_list_target_config(size_t rs_lengths) {
447 guarantee( adaptive_young_list_length(), "pre-condition" );
449 double start_time_sec = os::elapsedTime();
450 size_t min_reserve_perc = MAX2((size_t)2, (size_t)G1ReservePercent);
451 min_reserve_perc = MIN2((size_t) 50, min_reserve_perc);
452 size_t reserve_regions =
453 (size_t) ((double) min_reserve_perc * (double) _g1->n_regions() / 100.0);
455 if (full_young_gcs() && _free_regions_at_end_of_collection > 0) {
456 // we are in fully-young mode and there are free regions in the heap
458 double survivor_regions_evac_time =
459 predict_survivor_regions_evac_time();
461 size_t min_so_length = 0;
462 size_t max_so_length = 0;
464 if (G1UseScanOnlyPrefix) {
465 if (_all_pause_times_ms->num() < 3) {
466 // we won't use a scan-only set at the beginning to allow the rest
467 // of the predictors to warm up
468 min_so_length = 0;
469 max_so_length = 0;
470 } else if (_cost_per_scan_only_region_ms_seq->num() < 3) {
471 // then, we'll only set the S-O set to 1 for a little bit of time,
472 // to get enough information on the scanning cost
473 min_so_length = 1;
474 max_so_length = 1;
475 } else if (_in_marking_window || _last_full_young_gc) {
476 // no S-O prefix during a marking phase either, as at the end
477 // of the marking phase we'll have to use a very small young
478 // length target to fill up the rest of the CSet with
479 // non-young regions and, if we have lots of scan-only regions
480 // left-over, we will not be able to add any more non-young
481 // regions.
482 min_so_length = 0;
483 max_so_length = 0;
484 } else {
485 // this is the common case; we'll never reach the maximum, we
486 // one of the end conditions will fire well before that
487 // (hopefully!)
488 min_so_length = 0;
489 max_so_length = _free_regions_at_end_of_collection - 1;
490 }
491 } else {
492 // no S-O prefix, as the switch is not set, but we still need to
493 // do one iteration to calculate the best young target that
494 // meets the pause time; this way we reuse the same code instead
495 // of replicating it
496 min_so_length = 0;
497 max_so_length = 0;
498 }
500 double target_pause_time_ms = _mmu_tracker->max_gc_time() * 1000.0;
501 size_t pending_cards = (size_t) get_new_prediction(_pending_cards_seq);
502 size_t adj_rs_lengths = rs_lengths + predict_rs_length_diff();
503 size_t scanned_cards;
504 if (full_young_gcs())
505 scanned_cards = predict_young_card_num(adj_rs_lengths);
506 else
507 scanned_cards = predict_non_young_card_num(adj_rs_lengths);
508 // calculate this once, so that we don't have to recalculate it in
509 // the innermost loop
510 double base_time_ms = predict_base_elapsed_time_ms(pending_cards, scanned_cards)
511 + survivor_regions_evac_time;
512 // the result
513 size_t final_young_length = 0;
514 size_t final_so_length = 0;
515 double final_gc_eff = 0.0;
516 // we'll also keep track of how many times we go into the inner loop
517 // this is for profiling reasons
518 size_t calculations = 0;
520 // this determines which of the three iterations the outer loop is in
521 typedef enum {
522 pass_type_coarse,
523 pass_type_fine,
524 pass_type_final
525 } pass_type_t;
527 // range of the outer loop's iteration
528 size_t from_so_length = min_so_length;
529 size_t to_so_length = max_so_length;
530 guarantee( from_so_length <= to_so_length, "invariant" );
532 // this will keep the S-O length that's found by the second
533 // iteration of the outer loop; we'll keep it just in case the third
534 // iteration fails to find something
535 size_t fine_so_length = 0;
537 // the increment step for the coarse (first) iteration
538 size_t so_coarse_increments = 5;
540 // the common case, we'll start with the coarse iteration
541 pass_type_t pass = pass_type_coarse;
542 size_t so_length_incr = so_coarse_increments;
544 if (from_so_length == to_so_length) {
545 // not point in doing the coarse iteration, we'll go directly into
546 // the fine one (we essentially trying to find the optimal young
547 // length for a fixed S-O length).
548 so_length_incr = 1;
549 pass = pass_type_final;
550 } else if (to_so_length - from_so_length < 3 * so_coarse_increments) {
551 // again, the range is too short so no point in foind the coarse
552 // iteration either
553 so_length_incr = 1;
554 pass = pass_type_fine;
555 }
557 bool done = false;
558 // this is the outermost loop
559 while (!done) {
560 #ifdef TRACE_CALC_YOUNG_CONFIG
561 // leave this in for debugging, just in case
562 gclog_or_tty->print_cr("searching between " SIZE_FORMAT " and " SIZE_FORMAT
563 ", incr " SIZE_FORMAT ", pass %s",
564 from_so_length, to_so_length, so_length_incr,
565 (pass == pass_type_coarse) ? "coarse" :
566 (pass == pass_type_fine) ? "fine" : "final");
567 #endif // TRACE_CALC_YOUNG_CONFIG
569 size_t so_length = from_so_length;
570 size_t init_free_regions =
571 MAX2((size_t)0,
572 _free_regions_at_end_of_collection +
573 _scan_only_regions_at_end_of_collection - reserve_regions);
575 // this determines whether a configuration was found
576 bool gc_eff_set = false;
577 // this is the middle loop
578 while (so_length <= to_so_length) {
579 // base time, which excludes region-related time; again we
580 // calculate it once to avoid recalculating it in the
581 // innermost loop
582 double base_time_with_so_ms =
583 base_time_ms + predict_scan_only_time_ms(so_length);
584 // it's already over the pause target, go around
585 if (base_time_with_so_ms > target_pause_time_ms)
586 break;
588 size_t starting_young_length = so_length+1;
590 // we make sure that the short young length that makes sense
591 // (one more than the S-O length) is feasible
592 size_t min_young_length = starting_young_length;
593 double min_gc_eff;
594 bool min_ok;
595 ++calculations;
596 min_ok = predict_gc_eff(min_young_length, so_length,
597 base_time_with_so_ms,
598 init_free_regions, target_pause_time_ms,
599 &min_gc_eff);
601 if (min_ok) {
602 // the shortest young length is indeed feasible; we'll know
603 // set up the max young length and we'll do a binary search
604 // between min_young_length and max_young_length
605 size_t max_young_length = _free_regions_at_end_of_collection - 1;
606 double max_gc_eff = 0.0;
607 bool max_ok = false;
609 // the innermost loop! (finally!)
610 while (max_young_length > min_young_length) {
611 // we'll make sure that min_young_length is always at a
612 // feasible config
613 guarantee( min_ok, "invariant" );
615 ++calculations;
616 max_ok = predict_gc_eff(max_young_length, so_length,
617 base_time_with_so_ms,
618 init_free_regions, target_pause_time_ms,
619 &max_gc_eff);
621 size_t diff = (max_young_length - min_young_length) / 2;
622 if (max_ok) {
623 min_young_length = max_young_length;
624 min_gc_eff = max_gc_eff;
625 min_ok = true;
626 }
627 max_young_length = min_young_length + diff;
628 }
630 // the innermost loop found a config
631 guarantee( min_ok, "invariant" );
632 if (min_gc_eff > final_gc_eff) {
633 // it's the best config so far, so we'll keep it
634 final_gc_eff = min_gc_eff;
635 final_young_length = min_young_length;
636 final_so_length = so_length;
637 gc_eff_set = true;
638 }
639 }
641 // incremental the fixed S-O length and go around
642 so_length += so_length_incr;
643 }
645 // this is the end of the outermost loop and we need to decide
646 // what to do during the next iteration
647 if (pass == pass_type_coarse) {
648 // we just did the coarse pass (first iteration)
650 if (!gc_eff_set)
651 // we didn't find a feasible config so we'll just bail out; of
652 // course, it might be the case that we missed it; but I'd say
653 // it's a bit unlikely
654 done = true;
655 else {
656 // We did find a feasible config with optimal GC eff during
657 // the first pass. So the second pass we'll only consider the
658 // S-O lengths around that config with a fine increment.
660 guarantee( so_length_incr == so_coarse_increments, "invariant" );
661 guarantee( final_so_length >= min_so_length, "invariant" );
663 #ifdef TRACE_CALC_YOUNG_CONFIG
664 // leave this in for debugging, just in case
665 gclog_or_tty->print_cr(" coarse pass: SO length " SIZE_FORMAT,
666 final_so_length);
667 #endif // TRACE_CALC_YOUNG_CONFIG
669 from_so_length =
670 (final_so_length - min_so_length > so_coarse_increments) ?
671 final_so_length - so_coarse_increments + 1 : min_so_length;
672 to_so_length =
673 (max_so_length - final_so_length > so_coarse_increments) ?
674 final_so_length + so_coarse_increments - 1 : max_so_length;
676 pass = pass_type_fine;
677 so_length_incr = 1;
678 }
679 } else if (pass == pass_type_fine) {
680 // we just finished the second pass
682 if (!gc_eff_set) {
683 // we didn't find a feasible config (yes, it's possible;
684 // notice that, sometimes, we go directly into the fine
685 // iteration and skip the coarse one) so we bail out
686 done = true;
687 } else {
688 // We did find a feasible config with optimal GC eff
689 guarantee( so_length_incr == 1, "invariant" );
691 if (final_so_length == 0) {
692 // The config is of an empty S-O set, so we'll just bail out
693 done = true;
694 } else {
695 // we'll go around once more, setting the S-O length to 95%
696 // of the optimal
697 size_t new_so_length = 950 * final_so_length / 1000;
699 #ifdef TRACE_CALC_YOUNG_CONFIG
700 // leave this in for debugging, just in case
701 gclog_or_tty->print_cr(" fine pass: SO length " SIZE_FORMAT
702 ", setting it to " SIZE_FORMAT,
703 final_so_length, new_so_length);
704 #endif // TRACE_CALC_YOUNG_CONFIG
706 from_so_length = new_so_length;
707 to_so_length = new_so_length;
708 fine_so_length = final_so_length;
710 pass = pass_type_final;
711 }
712 }
713 } else if (pass == pass_type_final) {
714 // we just finished the final (third) pass
716 if (!gc_eff_set)
717 // we didn't find a feasible config, so we'll just use the one
718 // we found during the second pass, which we saved
719 final_so_length = fine_so_length;
721 // and we're done!
722 done = true;
723 } else {
724 guarantee( false, "should never reach here" );
725 }
727 // we now go around the outermost loop
728 }
730 // we should have at least one region in the target young length
731 _young_list_target_length =
732 MAX2((size_t) 1, final_young_length + _recorded_survivor_regions);
733 if (final_so_length >= final_young_length)
734 // and we need to ensure that the S-O length is not greater than
735 // the target young length (this is being a bit careful)
736 final_so_length = 0;
737 _young_list_so_prefix_length = final_so_length;
738 guarantee( !_in_marking_window || !_last_full_young_gc ||
739 _young_list_so_prefix_length == 0, "invariant" );
741 // let's keep an eye of how long we spend on this calculation
742 // right now, I assume that we'll print it when we need it; we
743 // should really adde it to the breakdown of a pause
744 double end_time_sec = os::elapsedTime();
745 double elapsed_time_ms = (end_time_sec - start_time_sec) * 1000.0;
747 #ifdef TRACE_CALC_YOUNG_CONFIG
748 // leave this in for debugging, just in case
749 gclog_or_tty->print_cr("target = %1.1lf ms, young = " SIZE_FORMAT
750 ", SO = " SIZE_FORMAT ", "
751 "elapsed %1.2lf ms, calcs: " SIZE_FORMAT " (%s%s) "
752 SIZE_FORMAT SIZE_FORMAT,
753 target_pause_time_ms,
754 _young_list_target_length - _young_list_so_prefix_length,
755 _young_list_so_prefix_length,
756 elapsed_time_ms,
757 calculations,
758 full_young_gcs() ? "full" : "partial",
759 should_initiate_conc_mark() ? " i-m" : "",
760 _in_marking_window,
761 _in_marking_window_im);
762 #endif // TRACE_CALC_YOUNG_CONFIG
764 if (_young_list_target_length < _young_list_min_length) {
765 // bummer; this means that, if we do a pause when the optimal
766 // config dictates, we'll violate the pause spacing target (the
767 // min length was calculate based on the application's current
768 // alloc rate);
770 // so, we have to bite the bullet, and allocate the minimum
771 // number. We'll violate our target, but we just can't meet it.
773 size_t so_length = 0;
774 // a note further up explains why we do not want an S-O length
775 // during marking
776 if (!_in_marking_window && !_last_full_young_gc)
777 // but we can still try to see whether we can find an optimal
778 // S-O length
779 so_length = calculate_optimal_so_length(_young_list_min_length);
781 #ifdef TRACE_CALC_YOUNG_CONFIG
782 // leave this in for debugging, just in case
783 gclog_or_tty->print_cr("adjusted target length from "
784 SIZE_FORMAT " to " SIZE_FORMAT
785 ", SO " SIZE_FORMAT,
786 _young_list_target_length, _young_list_min_length,
787 so_length);
788 #endif // TRACE_CALC_YOUNG_CONFIG
790 _young_list_target_length =
791 MAX2(_young_list_min_length, (size_t)1);
792 _young_list_so_prefix_length = so_length;
793 }
794 } else {
795 // we are in a partially-young mode or we've run out of regions (due
796 // to evacuation failure)
798 #ifdef TRACE_CALC_YOUNG_CONFIG
799 // leave this in for debugging, just in case
800 gclog_or_tty->print_cr("(partial) setting target to " SIZE_FORMAT
801 ", SO " SIZE_FORMAT,
802 _young_list_min_length, 0);
803 #endif // TRACE_CALC_YOUNG_CONFIG
805 // we'll do the pause as soon as possible and with no S-O prefix
806 // (see above for the reasons behind the latter)
807 _young_list_target_length =
808 MAX2(_young_list_min_length, (size_t) 1);
809 _young_list_so_prefix_length = 0;
810 }
812 _rs_lengths_prediction = rs_lengths;
813 }
815 // This is used by: calculate_optimal_so_length(length). It returns
816 // the GC eff and predicted pause time for a particular config
817 void
818 G1CollectorPolicy::predict_gc_eff(size_t young_length,
819 size_t so_length,
820 double base_time_ms,
821 double* ret_gc_eff,
822 double* ret_pause_time_ms) {
823 double so_time_ms = predict_scan_only_time_ms(so_length);
824 double accum_surv_rate_adj = 0.0;
825 if (so_length > 0)
826 accum_surv_rate_adj = accum_yg_surv_rate_pred((int)(so_length - 1));
827 double accum_surv_rate =
828 accum_yg_surv_rate_pred((int)(young_length - 1)) - accum_surv_rate_adj;
829 size_t bytes_to_copy =
830 (size_t) (accum_surv_rate * (double) HeapRegion::GrainBytes);
831 double copy_time_ms = predict_object_copy_time_ms(bytes_to_copy);
832 double young_other_time_ms =
833 predict_young_other_time_ms(young_length - so_length);
834 double pause_time_ms =
835 base_time_ms + so_time_ms + copy_time_ms + young_other_time_ms;
836 size_t reclaimed_bytes =
837 (young_length - so_length) * HeapRegion::GrainBytes - bytes_to_copy;
838 double gc_eff = (double) reclaimed_bytes / pause_time_ms;
840 *ret_gc_eff = gc_eff;
841 *ret_pause_time_ms = pause_time_ms;
842 }
844 // This is used by: calculate_young_list_target_config(rs_length). It
845 // returns the GC eff of a particular config. It returns false if that
846 // config violates any of the end conditions of the search in the
847 // calling method, or true upon success. The end conditions were put
848 // here since it's called twice and it was best not to replicate them
849 // in the caller. Also, passing the parameteres avoids having to
850 // recalculate them in the innermost loop.
851 bool
852 G1CollectorPolicy::predict_gc_eff(size_t young_length,
853 size_t so_length,
854 double base_time_with_so_ms,
855 size_t init_free_regions,
856 double target_pause_time_ms,
857 double* ret_gc_eff) {
858 *ret_gc_eff = 0.0;
860 if (young_length >= init_free_regions)
861 // end condition 1: not enough space for the young regions
862 return false;
864 double accum_surv_rate_adj = 0.0;
865 if (so_length > 0)
866 accum_surv_rate_adj = accum_yg_surv_rate_pred((int)(so_length - 1));
867 double accum_surv_rate =
868 accum_yg_surv_rate_pred((int)(young_length - 1)) - accum_surv_rate_adj;
869 size_t bytes_to_copy =
870 (size_t) (accum_surv_rate * (double) HeapRegion::GrainBytes);
871 double copy_time_ms = predict_object_copy_time_ms(bytes_to_copy);
872 double young_other_time_ms =
873 predict_young_other_time_ms(young_length - so_length);
874 double pause_time_ms =
875 base_time_with_so_ms + copy_time_ms + young_other_time_ms;
877 if (pause_time_ms > target_pause_time_ms)
878 // end condition 2: over the target pause time
879 return false;
881 size_t reclaimed_bytes =
882 (young_length - so_length) * HeapRegion::GrainBytes - bytes_to_copy;
883 size_t free_bytes =
884 (init_free_regions - young_length) * HeapRegion::GrainBytes;
886 if ((2.0 + sigma()) * (double) bytes_to_copy > (double) free_bytes)
887 // end condition 3: out of to-space (conservatively)
888 return false;
890 // success!
891 double gc_eff = (double) reclaimed_bytes / pause_time_ms;
892 *ret_gc_eff = gc_eff;
894 return true;
895 }
897 double G1CollectorPolicy::predict_survivor_regions_evac_time() {
898 double survivor_regions_evac_time = 0.0;
899 for (HeapRegion * r = _recorded_survivor_head;
900 r != NULL && r != _recorded_survivor_tail->get_next_young_region();
901 r = r->get_next_young_region()) {
902 survivor_regions_evac_time += predict_region_elapsed_time_ms(r, true);
903 }
904 return survivor_regions_evac_time;
905 }
907 void G1CollectorPolicy::check_prediction_validity() {
908 guarantee( adaptive_young_list_length(), "should not call this otherwise" );
910 size_t rs_lengths = _g1->young_list_sampled_rs_lengths();
911 if (rs_lengths > _rs_lengths_prediction) {
912 // add 10% to avoid having to recalculate often
913 size_t rs_lengths_prediction = rs_lengths * 1100 / 1000;
914 calculate_young_list_target_config(rs_lengths_prediction);
915 }
916 }
918 HeapWord* G1CollectorPolicy::mem_allocate_work(size_t size,
919 bool is_tlab,
920 bool* gc_overhead_limit_was_exceeded) {
921 guarantee(false, "Not using this policy feature yet.");
922 return NULL;
923 }
925 // This method controls how a collector handles one or more
926 // of its generations being fully allocated.
927 HeapWord* G1CollectorPolicy::satisfy_failed_allocation(size_t size,
928 bool is_tlab) {
929 guarantee(false, "Not using this policy feature yet.");
930 return NULL;
931 }
934 #ifndef PRODUCT
935 bool G1CollectorPolicy::verify_young_ages() {
936 HeapRegion* head = _g1->young_list_first_region();
937 return
938 verify_young_ages(head, _short_lived_surv_rate_group);
939 // also call verify_young_ages on any additional surv rate groups
940 }
942 bool
943 G1CollectorPolicy::verify_young_ages(HeapRegion* head,
944 SurvRateGroup *surv_rate_group) {
945 guarantee( surv_rate_group != NULL, "pre-condition" );
947 const char* name = surv_rate_group->name();
948 bool ret = true;
949 int prev_age = -1;
951 for (HeapRegion* curr = head;
952 curr != NULL;
953 curr = curr->get_next_young_region()) {
954 SurvRateGroup* group = curr->surv_rate_group();
955 if (group == NULL && !curr->is_survivor()) {
956 gclog_or_tty->print_cr("## %s: encountered NULL surv_rate_group", name);
957 ret = false;
958 }
960 if (surv_rate_group == group) {
961 int age = curr->age_in_surv_rate_group();
963 if (age < 0) {
964 gclog_or_tty->print_cr("## %s: encountered negative age", name);
965 ret = false;
966 }
968 if (age <= prev_age) {
969 gclog_or_tty->print_cr("## %s: region ages are not strictly increasing "
970 "(%d, %d)", name, age, prev_age);
971 ret = false;
972 }
973 prev_age = age;
974 }
975 }
977 return ret;
978 }
979 #endif // PRODUCT
981 void G1CollectorPolicy::record_full_collection_start() {
982 _cur_collection_start_sec = os::elapsedTime();
983 // Release the future to-space so that it is available for compaction into.
984 _g1->set_full_collection();
985 }
987 void G1CollectorPolicy::record_full_collection_end() {
988 // Consider this like a collection pause for the purposes of allocation
989 // since last pause.
990 double end_sec = os::elapsedTime();
991 double full_gc_time_sec = end_sec - _cur_collection_start_sec;
992 double full_gc_time_ms = full_gc_time_sec * 1000.0;
994 _all_full_gc_times_ms->add(full_gc_time_ms);
996 update_recent_gc_times(end_sec, full_gc_time_ms);
998 _g1->clear_full_collection();
1000 // "Nuke" the heuristics that control the fully/partially young GC
1001 // transitions and make sure we start with fully young GCs after the
1002 // Full GC.
1003 set_full_young_gcs(true);
1004 _last_full_young_gc = false;
1005 _should_revert_to_full_young_gcs = false;
1006 _should_initiate_conc_mark = false;
1007 _known_garbage_bytes = 0;
1008 _known_garbage_ratio = 0.0;
1009 _in_marking_window = false;
1010 _in_marking_window_im = false;
1012 _short_lived_surv_rate_group->record_scan_only_prefix(0);
1013 _short_lived_surv_rate_group->start_adding_regions();
1014 // also call this on any additional surv rate groups
1016 record_survivor_regions(0, NULL, NULL);
1018 _prev_region_num_young = _region_num_young;
1019 _prev_region_num_tenured = _region_num_tenured;
1021 _free_regions_at_end_of_collection = _g1->free_regions();
1022 _scan_only_regions_at_end_of_collection = 0;
1023 // Reset survivors SurvRateGroup.
1024 _survivor_surv_rate_group->reset();
1025 calculate_young_list_min_length();
1026 calculate_young_list_target_config();
1027 }
1029 void G1CollectorPolicy::record_before_bytes(size_t bytes) {
1030 _bytes_in_to_space_before_gc += bytes;
1031 }
1033 void G1CollectorPolicy::record_after_bytes(size_t bytes) {
1034 _bytes_in_to_space_after_gc += bytes;
1035 }
1037 void G1CollectorPolicy::record_stop_world_start() {
1038 _stop_world_start = os::elapsedTime();
1039 }
1041 void G1CollectorPolicy::record_collection_pause_start(double start_time_sec,
1042 size_t start_used) {
1043 if (PrintGCDetails) {
1044 gclog_or_tty->stamp(PrintGCTimeStamps);
1045 gclog_or_tty->print("[GC pause");
1046 if (in_young_gc_mode())
1047 gclog_or_tty->print(" (%s)", full_young_gcs() ? "young" : "partial");
1048 }
1050 assert(_g1->used_regions() == _g1->recalculate_used_regions(),
1051 "sanity");
1052 assert(_g1->used() == _g1->recalculate_used(), "sanity");
1054 double s_w_t_ms = (start_time_sec - _stop_world_start) * 1000.0;
1055 _all_stop_world_times_ms->add(s_w_t_ms);
1056 _stop_world_start = 0.0;
1058 _cur_collection_start_sec = start_time_sec;
1059 _cur_collection_pause_used_at_start_bytes = start_used;
1060 _cur_collection_pause_used_regions_at_start = _g1->used_regions();
1061 _pending_cards = _g1->pending_card_num();
1062 _max_pending_cards = _g1->max_pending_card_num();
1064 _bytes_in_to_space_before_gc = 0;
1065 _bytes_in_to_space_after_gc = 0;
1066 _bytes_in_collection_set_before_gc = 0;
1068 #ifdef DEBUG
1069 // initialise these to something well known so that we can spot
1070 // if they are not set properly
1072 for (int i = 0; i < _parallel_gc_threads; ++i) {
1073 _par_last_ext_root_scan_times_ms[i] = -666.0;
1074 _par_last_mark_stack_scan_times_ms[i] = -666.0;
1075 _par_last_scan_only_times_ms[i] = -666.0;
1076 _par_last_scan_only_regions_scanned[i] = -666.0;
1077 _par_last_update_rs_start_times_ms[i] = -666.0;
1078 _par_last_update_rs_times_ms[i] = -666.0;
1079 _par_last_update_rs_processed_buffers[i] = -666.0;
1080 _par_last_scan_rs_start_times_ms[i] = -666.0;
1081 _par_last_scan_rs_times_ms[i] = -666.0;
1082 _par_last_scan_new_refs_times_ms[i] = -666.0;
1083 _par_last_obj_copy_times_ms[i] = -666.0;
1084 _par_last_termination_times_ms[i] = -666.0;
1085 }
1086 #endif
1088 for (int i = 0; i < _aux_num; ++i) {
1089 _cur_aux_times_ms[i] = 0.0;
1090 _cur_aux_times_set[i] = false;
1091 }
1093 _satb_drain_time_set = false;
1094 _last_satb_drain_processed_buffers = -1;
1096 if (in_young_gc_mode())
1097 _last_young_gc_full = false;
1100 // do that for any other surv rate groups
1101 _short_lived_surv_rate_group->stop_adding_regions();
1102 size_t short_lived_so_length = _young_list_so_prefix_length;
1103 _short_lived_surv_rate_group->record_scan_only_prefix(short_lived_so_length);
1104 tag_scan_only(short_lived_so_length);
1105 _survivors_age_table.clear();
1107 assert( verify_young_ages(), "region age verification" );
1108 }
1110 void G1CollectorPolicy::tag_scan_only(size_t short_lived_scan_only_length) {
1111 // done in a way that it can be extended for other surv rate groups too...
1113 HeapRegion* head = _g1->young_list_first_region();
1114 bool finished_short_lived = (short_lived_scan_only_length == 0);
1116 if (finished_short_lived)
1117 return;
1119 for (HeapRegion* curr = head;
1120 curr != NULL;
1121 curr = curr->get_next_young_region()) {
1122 SurvRateGroup* surv_rate_group = curr->surv_rate_group();
1123 int age = curr->age_in_surv_rate_group();
1125 if (surv_rate_group == _short_lived_surv_rate_group) {
1126 if ((size_t)age < short_lived_scan_only_length)
1127 curr->set_scan_only();
1128 else
1129 finished_short_lived = true;
1130 }
1133 if (finished_short_lived)
1134 return;
1135 }
1137 guarantee( false, "we should never reach here" );
1138 }
1140 void G1CollectorPolicy::record_mark_closure_time(double mark_closure_time_ms) {
1141 _mark_closure_time_ms = mark_closure_time_ms;
1142 }
1144 void G1CollectorPolicy::record_concurrent_mark_init_start() {
1145 _mark_init_start_sec = os::elapsedTime();
1146 guarantee(!in_young_gc_mode(), "should not do be here in young GC mode");
1147 }
1149 void G1CollectorPolicy::record_concurrent_mark_init_end_pre(double
1150 mark_init_elapsed_time_ms) {
1151 _during_marking = true;
1152 _should_initiate_conc_mark = false;
1153 _cur_mark_stop_world_time_ms = mark_init_elapsed_time_ms;
1154 }
1156 void G1CollectorPolicy::record_concurrent_mark_init_end() {
1157 double end_time_sec = os::elapsedTime();
1158 double elapsed_time_ms = (end_time_sec - _mark_init_start_sec) * 1000.0;
1159 _concurrent_mark_init_times_ms->add(elapsed_time_ms);
1160 record_concurrent_mark_init_end_pre(elapsed_time_ms);
1162 _mmu_tracker->add_pause(_mark_init_start_sec, end_time_sec, true);
1163 }
1165 void G1CollectorPolicy::record_concurrent_mark_remark_start() {
1166 _mark_remark_start_sec = os::elapsedTime();
1167 _during_marking = false;
1168 }
1170 void G1CollectorPolicy::record_concurrent_mark_remark_end() {
1171 double end_time_sec = os::elapsedTime();
1172 double elapsed_time_ms = (end_time_sec - _mark_remark_start_sec)*1000.0;
1173 _concurrent_mark_remark_times_ms->add(elapsed_time_ms);
1174 _cur_mark_stop_world_time_ms += elapsed_time_ms;
1175 _prev_collection_pause_end_ms += elapsed_time_ms;
1177 _mmu_tracker->add_pause(_mark_remark_start_sec, end_time_sec, true);
1178 }
1180 void G1CollectorPolicy::record_concurrent_mark_cleanup_start() {
1181 _mark_cleanup_start_sec = os::elapsedTime();
1182 }
1184 void
1185 G1CollectorPolicy::record_concurrent_mark_cleanup_end(size_t freed_bytes,
1186 size_t max_live_bytes) {
1187 record_concurrent_mark_cleanup_end_work1(freed_bytes, max_live_bytes);
1188 record_concurrent_mark_cleanup_end_work2();
1189 }
1191 void
1192 G1CollectorPolicy::
1193 record_concurrent_mark_cleanup_end_work1(size_t freed_bytes,
1194 size_t max_live_bytes) {
1195 if (_n_marks < 2) _n_marks++;
1196 if (G1PolicyVerbose > 0)
1197 gclog_or_tty->print_cr("At end of marking, max_live is " SIZE_FORMAT " MB "
1198 " (of " SIZE_FORMAT " MB heap).",
1199 max_live_bytes/M, _g1->capacity()/M);
1200 }
1202 // The important thing about this is that it includes "os::elapsedTime".
1203 void G1CollectorPolicy::record_concurrent_mark_cleanup_end_work2() {
1204 double end_time_sec = os::elapsedTime();
1205 double elapsed_time_ms = (end_time_sec - _mark_cleanup_start_sec)*1000.0;
1206 _concurrent_mark_cleanup_times_ms->add(elapsed_time_ms);
1207 _cur_mark_stop_world_time_ms += elapsed_time_ms;
1208 _prev_collection_pause_end_ms += elapsed_time_ms;
1210 _mmu_tracker->add_pause(_mark_cleanup_start_sec, end_time_sec, true);
1212 _num_markings++;
1214 // We did a marking, so reset the "since_last_mark" variables.
1215 double considerConcMarkCost = 1.0;
1216 // If there are available processors, concurrent activity is free...
1217 if (Threads::number_of_non_daemon_threads() * 2 <
1218 os::active_processor_count()) {
1219 considerConcMarkCost = 0.0;
1220 }
1221 _n_pauses_at_mark_end = _n_pauses;
1222 _n_marks_since_last_pause++;
1223 _conc_mark_initiated = false;
1224 }
1226 void
1227 G1CollectorPolicy::record_concurrent_mark_cleanup_completed() {
1228 if (in_young_gc_mode()) {
1229 _should_revert_to_full_young_gcs = false;
1230 _last_full_young_gc = true;
1231 _in_marking_window = false;
1232 if (adaptive_young_list_length())
1233 calculate_young_list_target_config();
1234 }
1235 }
1237 void G1CollectorPolicy::record_concurrent_pause() {
1238 if (_stop_world_start > 0.0) {
1239 double yield_ms = (os::elapsedTime() - _stop_world_start) * 1000.0;
1240 _all_yield_times_ms->add(yield_ms);
1241 }
1242 }
1244 void G1CollectorPolicy::record_concurrent_pause_end() {
1245 }
1247 void G1CollectorPolicy::record_collection_pause_end_CH_strong_roots() {
1248 _cur_CH_strong_roots_end_sec = os::elapsedTime();
1249 _cur_CH_strong_roots_dur_ms =
1250 (_cur_CH_strong_roots_end_sec - _cur_collection_start_sec) * 1000.0;
1251 }
1253 void G1CollectorPolicy::record_collection_pause_end_G1_strong_roots() {
1254 _cur_G1_strong_roots_end_sec = os::elapsedTime();
1255 _cur_G1_strong_roots_dur_ms =
1256 (_cur_G1_strong_roots_end_sec - _cur_CH_strong_roots_end_sec) * 1000.0;
1257 }
1259 template<class T>
1260 T sum_of(T* sum_arr, int start, int n, int N) {
1261 T sum = (T)0;
1262 for (int i = 0; i < n; i++) {
1263 int j = (start + i) % N;
1264 sum += sum_arr[j];
1265 }
1266 return sum;
1267 }
1269 void G1CollectorPolicy::print_par_stats (int level,
1270 const char* str,
1271 double* data,
1272 bool summary) {
1273 double min = data[0], max = data[0];
1274 double total = 0.0;
1275 int j;
1276 for (j = 0; j < level; ++j)
1277 gclog_or_tty->print(" ");
1278 gclog_or_tty->print("[%s (ms):", str);
1279 for (uint i = 0; i < ParallelGCThreads; ++i) {
1280 double val = data[i];
1281 if (val < min)
1282 min = val;
1283 if (val > max)
1284 max = val;
1285 total += val;
1286 gclog_or_tty->print(" %3.1lf", val);
1287 }
1288 if (summary) {
1289 gclog_or_tty->print_cr("");
1290 double avg = total / (double) ParallelGCThreads;
1291 gclog_or_tty->print(" ");
1292 for (j = 0; j < level; ++j)
1293 gclog_or_tty->print(" ");
1294 gclog_or_tty->print("Avg: %5.1lf, Min: %5.1lf, Max: %5.1lf",
1295 avg, min, max);
1296 }
1297 gclog_or_tty->print_cr("]");
1298 }
1300 void G1CollectorPolicy::print_par_buffers (int level,
1301 const char* str,
1302 double* data,
1303 bool summary) {
1304 double min = data[0], max = data[0];
1305 double total = 0.0;
1306 int j;
1307 for (j = 0; j < level; ++j)
1308 gclog_or_tty->print(" ");
1309 gclog_or_tty->print("[%s :", str);
1310 for (uint i = 0; i < ParallelGCThreads; ++i) {
1311 double val = data[i];
1312 if (val < min)
1313 min = val;
1314 if (val > max)
1315 max = val;
1316 total += val;
1317 gclog_or_tty->print(" %d", (int) val);
1318 }
1319 if (summary) {
1320 gclog_or_tty->print_cr("");
1321 double avg = total / (double) ParallelGCThreads;
1322 gclog_or_tty->print(" ");
1323 for (j = 0; j < level; ++j)
1324 gclog_or_tty->print(" ");
1325 gclog_or_tty->print("Sum: %d, Avg: %d, Min: %d, Max: %d",
1326 (int)total, (int)avg, (int)min, (int)max);
1327 }
1328 gclog_or_tty->print_cr("]");
1329 }
1331 void G1CollectorPolicy::print_stats (int level,
1332 const char* str,
1333 double value) {
1334 for (int j = 0; j < level; ++j)
1335 gclog_or_tty->print(" ");
1336 gclog_or_tty->print_cr("[%s: %5.1lf ms]", str, value);
1337 }
1339 void G1CollectorPolicy::print_stats (int level,
1340 const char* str,
1341 int value) {
1342 for (int j = 0; j < level; ++j)
1343 gclog_or_tty->print(" ");
1344 gclog_or_tty->print_cr("[%s: %d]", str, value);
1345 }
1347 double G1CollectorPolicy::avg_value (double* data) {
1348 if (ParallelGCThreads > 0) {
1349 double ret = 0.0;
1350 for (uint i = 0; i < ParallelGCThreads; ++i)
1351 ret += data[i];
1352 return ret / (double) ParallelGCThreads;
1353 } else {
1354 return data[0];
1355 }
1356 }
1358 double G1CollectorPolicy::max_value (double* data) {
1359 if (ParallelGCThreads > 0) {
1360 double ret = data[0];
1361 for (uint i = 1; i < ParallelGCThreads; ++i)
1362 if (data[i] > ret)
1363 ret = data[i];
1364 return ret;
1365 } else {
1366 return data[0];
1367 }
1368 }
1370 double G1CollectorPolicy::sum_of_values (double* data) {
1371 if (ParallelGCThreads > 0) {
1372 double sum = 0.0;
1373 for (uint i = 0; i < ParallelGCThreads; i++)
1374 sum += data[i];
1375 return sum;
1376 } else {
1377 return data[0];
1378 }
1379 }
1381 double G1CollectorPolicy::max_sum (double* data1,
1382 double* data2) {
1383 double ret = data1[0] + data2[0];
1385 if (ParallelGCThreads > 0) {
1386 for (uint i = 1; i < ParallelGCThreads; ++i) {
1387 double data = data1[i] + data2[i];
1388 if (data > ret)
1389 ret = data;
1390 }
1391 }
1392 return ret;
1393 }
1395 // Anything below that is considered to be zero
1396 #define MIN_TIMER_GRANULARITY 0.0000001
1398 void G1CollectorPolicy::record_collection_pause_end(bool abandoned) {
1399 double end_time_sec = os::elapsedTime();
1400 double elapsed_ms = _last_pause_time_ms;
1401 bool parallel = ParallelGCThreads > 0;
1402 double evac_ms = (end_time_sec - _cur_G1_strong_roots_end_sec) * 1000.0;
1403 size_t rs_size =
1404 _cur_collection_pause_used_regions_at_start - collection_set_size();
1405 size_t cur_used_bytes = _g1->used();
1406 assert(cur_used_bytes == _g1->recalculate_used(), "It should!");
1407 bool last_pause_included_initial_mark = false;
1408 bool update_stats = !abandoned && !_g1->evacuation_failed();
1410 #ifndef PRODUCT
1411 if (G1YoungSurvRateVerbose) {
1412 gclog_or_tty->print_cr("");
1413 _short_lived_surv_rate_group->print();
1414 // do that for any other surv rate groups too
1415 }
1416 #endif // PRODUCT
1418 if (in_young_gc_mode()) {
1419 last_pause_included_initial_mark = _should_initiate_conc_mark;
1420 if (last_pause_included_initial_mark)
1421 record_concurrent_mark_init_end_pre(0.0);
1423 size_t min_used_targ =
1424 (_g1->capacity() / 100) * InitiatingHeapOccupancyPercent;
1426 if (cur_used_bytes > min_used_targ) {
1427 if (cur_used_bytes <= _prev_collection_pause_used_at_end_bytes) {
1428 } else if (!_g1->mark_in_progress() && !_last_full_young_gc) {
1429 _should_initiate_conc_mark = true;
1430 }
1431 }
1433 _prev_collection_pause_used_at_end_bytes = cur_used_bytes;
1434 }
1436 _mmu_tracker->add_pause(end_time_sec - elapsed_ms/1000.0,
1437 end_time_sec, false);
1439 guarantee(_cur_collection_pause_used_regions_at_start >=
1440 collection_set_size(),
1441 "Negative RS size?");
1443 // This assert is exempted when we're doing parallel collection pauses,
1444 // because the fragmentation caused by the parallel GC allocation buffers
1445 // can lead to more memory being used during collection than was used
1446 // before. Best leave this out until the fragmentation problem is fixed.
1447 // Pauses in which evacuation failed can also lead to negative
1448 // collections, since no space is reclaimed from a region containing an
1449 // object whose evacuation failed.
1450 // Further, we're now always doing parallel collection. But I'm still
1451 // leaving this here as a placeholder for a more precise assertion later.
1452 // (DLD, 10/05.)
1453 assert((true || parallel) // Always using GC LABs now.
1454 || _g1->evacuation_failed()
1455 || _cur_collection_pause_used_at_start_bytes >= cur_used_bytes,
1456 "Negative collection");
1458 size_t freed_bytes =
1459 _cur_collection_pause_used_at_start_bytes - cur_used_bytes;
1460 size_t surviving_bytes = _collection_set_bytes_used_before - freed_bytes;
1461 double survival_fraction =
1462 (double)surviving_bytes/
1463 (double)_collection_set_bytes_used_before;
1465 _n_pauses++;
1467 if (update_stats) {
1468 _recent_CH_strong_roots_times_ms->add(_cur_CH_strong_roots_dur_ms);
1469 _recent_G1_strong_roots_times_ms->add(_cur_G1_strong_roots_dur_ms);
1470 _recent_evac_times_ms->add(evac_ms);
1471 _recent_pause_times_ms->add(elapsed_ms);
1473 _recent_rs_sizes->add(rs_size);
1475 // We exempt parallel collection from this check because Alloc Buffer
1476 // fragmentation can produce negative collections. Same with evac
1477 // failure.
1478 // Further, we're now always doing parallel collection. But I'm still
1479 // leaving this here as a placeholder for a more precise assertion later.
1480 // (DLD, 10/05.
1481 assert((true || parallel)
1482 || _g1->evacuation_failed()
1483 || surviving_bytes <= _collection_set_bytes_used_before,
1484 "Or else negative collection!");
1485 _recent_CS_bytes_used_before->add(_collection_set_bytes_used_before);
1486 _recent_CS_bytes_surviving->add(surviving_bytes);
1488 // this is where we update the allocation rate of the application
1489 double app_time_ms =
1490 (_cur_collection_start_sec * 1000.0 - _prev_collection_pause_end_ms);
1491 if (app_time_ms < MIN_TIMER_GRANULARITY) {
1492 // This usually happens due to the timer not having the required
1493 // granularity. Some Linuxes are the usual culprits.
1494 // We'll just set it to something (arbitrarily) small.
1495 app_time_ms = 1.0;
1496 }
1497 size_t regions_allocated =
1498 (_region_num_young - _prev_region_num_young) +
1499 (_region_num_tenured - _prev_region_num_tenured);
1500 double alloc_rate_ms = (double) regions_allocated / app_time_ms;
1501 _alloc_rate_ms_seq->add(alloc_rate_ms);
1502 _prev_region_num_young = _region_num_young;
1503 _prev_region_num_tenured = _region_num_tenured;
1505 double interval_ms =
1506 (end_time_sec - _recent_prev_end_times_for_all_gcs_sec->oldest()) * 1000.0;
1507 update_recent_gc_times(end_time_sec, elapsed_ms);
1508 _recent_avg_pause_time_ratio = _recent_gc_times_ms->sum()/interval_ms;
1509 if (recent_avg_pause_time_ratio() < 0.0 ||
1510 (recent_avg_pause_time_ratio() - 1.0 > 0.0)) {
1511 #ifndef PRODUCT
1512 // Dump info to allow post-facto debugging
1513 gclog_or_tty->print_cr("recent_avg_pause_time_ratio() out of bounds");
1514 gclog_or_tty->print_cr("-------------------------------------------");
1515 gclog_or_tty->print_cr("Recent GC Times (ms):");
1516 _recent_gc_times_ms->dump();
1517 gclog_or_tty->print_cr("(End Time=%3.3f) Recent GC End Times (s):", end_time_sec);
1518 _recent_prev_end_times_for_all_gcs_sec->dump();
1519 gclog_or_tty->print_cr("GC = %3.3f, Interval = %3.3f, Ratio = %3.3f",
1520 _recent_gc_times_ms->sum(), interval_ms, recent_avg_pause_time_ratio());
1521 // In debug mode, terminate the JVM if the user wants to debug at this point.
1522 assert(!G1FailOnFPError, "Debugging data for CR 6898948 has been dumped above");
1523 #endif // !PRODUCT
1524 // Clip ratio between 0.0 and 1.0, and continue. This will be fixed in
1525 // CR 6902692 by redoing the manner in which the ratio is incrementally computed.
1526 if (_recent_avg_pause_time_ratio < 0.0) {
1527 _recent_avg_pause_time_ratio = 0.0;
1528 } else {
1529 assert(_recent_avg_pause_time_ratio - 1.0 > 0.0, "Ctl-point invariant");
1530 _recent_avg_pause_time_ratio = 1.0;
1531 }
1532 }
1533 }
1535 if (G1PolicyVerbose > 1) {
1536 gclog_or_tty->print_cr(" Recording collection pause(%d)", _n_pauses);
1537 }
1539 PauseSummary* summary;
1540 if (abandoned) {
1541 summary = _abandoned_summary;
1542 } else {
1543 summary = _summary;
1544 }
1546 double ext_root_scan_time = avg_value(_par_last_ext_root_scan_times_ms);
1547 double mark_stack_scan_time = avg_value(_par_last_mark_stack_scan_times_ms);
1548 double scan_only_time = avg_value(_par_last_scan_only_times_ms);
1549 double scan_only_regions_scanned =
1550 sum_of_values(_par_last_scan_only_regions_scanned);
1551 double update_rs_time = avg_value(_par_last_update_rs_times_ms);
1552 double update_rs_processed_buffers =
1553 sum_of_values(_par_last_update_rs_processed_buffers);
1554 double scan_rs_time = avg_value(_par_last_scan_rs_times_ms);
1555 double obj_copy_time = avg_value(_par_last_obj_copy_times_ms);
1556 double termination_time = avg_value(_par_last_termination_times_ms);
1558 double parallel_other_time = _cur_collection_par_time_ms -
1559 (update_rs_time + ext_root_scan_time + mark_stack_scan_time +
1560 scan_only_time + scan_rs_time + obj_copy_time + termination_time);
1561 if (update_stats) {
1562 MainBodySummary* body_summary = summary->main_body_summary();
1563 guarantee(body_summary != NULL, "should not be null!");
1565 if (_satb_drain_time_set)
1566 body_summary->record_satb_drain_time_ms(_cur_satb_drain_time_ms);
1567 else
1568 body_summary->record_satb_drain_time_ms(0.0);
1569 body_summary->record_ext_root_scan_time_ms(ext_root_scan_time);
1570 body_summary->record_mark_stack_scan_time_ms(mark_stack_scan_time);
1571 body_summary->record_scan_only_time_ms(scan_only_time);
1572 body_summary->record_update_rs_time_ms(update_rs_time);
1573 body_summary->record_scan_rs_time_ms(scan_rs_time);
1574 body_summary->record_obj_copy_time_ms(obj_copy_time);
1575 if (parallel) {
1576 body_summary->record_parallel_time_ms(_cur_collection_par_time_ms);
1577 body_summary->record_clear_ct_time_ms(_cur_clear_ct_time_ms);
1578 body_summary->record_termination_time_ms(termination_time);
1579 body_summary->record_parallel_other_time_ms(parallel_other_time);
1580 }
1581 body_summary->record_mark_closure_time_ms(_mark_closure_time_ms);
1582 }
1584 if (G1PolicyVerbose > 1) {
1585 gclog_or_tty->print_cr(" ET: %10.6f ms (avg: %10.6f ms)\n"
1586 " CH Strong: %10.6f ms (avg: %10.6f ms)\n"
1587 " G1 Strong: %10.6f ms (avg: %10.6f ms)\n"
1588 " Evac: %10.6f ms (avg: %10.6f ms)\n"
1589 " ET-RS: %10.6f ms (avg: %10.6f ms)\n"
1590 " |RS|: " SIZE_FORMAT,
1591 elapsed_ms, recent_avg_time_for_pauses_ms(),
1592 _cur_CH_strong_roots_dur_ms, recent_avg_time_for_CH_strong_ms(),
1593 _cur_G1_strong_roots_dur_ms, recent_avg_time_for_G1_strong_ms(),
1594 evac_ms, recent_avg_time_for_evac_ms(),
1595 scan_rs_time,
1596 recent_avg_time_for_pauses_ms() -
1597 recent_avg_time_for_G1_strong_ms(),
1598 rs_size);
1600 gclog_or_tty->print_cr(" Used at start: " SIZE_FORMAT"K"
1601 " At end " SIZE_FORMAT "K\n"
1602 " garbage : " SIZE_FORMAT "K"
1603 " of " SIZE_FORMAT "K\n"
1604 " survival : %6.2f%% (%6.2f%% avg)",
1605 _cur_collection_pause_used_at_start_bytes/K,
1606 _g1->used()/K, freed_bytes/K,
1607 _collection_set_bytes_used_before/K,
1608 survival_fraction*100.0,
1609 recent_avg_survival_fraction()*100.0);
1610 gclog_or_tty->print_cr(" Recent %% gc pause time: %6.2f",
1611 recent_avg_pause_time_ratio() * 100.0);
1612 }
1614 double other_time_ms = elapsed_ms;
1616 if (!abandoned) {
1617 if (_satb_drain_time_set)
1618 other_time_ms -= _cur_satb_drain_time_ms;
1620 if (parallel)
1621 other_time_ms -= _cur_collection_par_time_ms + _cur_clear_ct_time_ms;
1622 else
1623 other_time_ms -=
1624 update_rs_time +
1625 ext_root_scan_time + mark_stack_scan_time + scan_only_time +
1626 scan_rs_time + obj_copy_time;
1627 }
1629 if (PrintGCDetails) {
1630 gclog_or_tty->print_cr("%s%s, %1.8lf secs]",
1631 abandoned ? " (abandoned)" : "",
1632 (last_pause_included_initial_mark) ? " (initial-mark)" : "",
1633 elapsed_ms / 1000.0);
1635 if (!abandoned) {
1636 if (_satb_drain_time_set) {
1637 print_stats(1, "SATB Drain Time", _cur_satb_drain_time_ms);
1638 }
1639 if (_last_satb_drain_processed_buffers >= 0) {
1640 print_stats(2, "Processed Buffers", _last_satb_drain_processed_buffers);
1641 }
1642 if (parallel) {
1643 print_stats(1, "Parallel Time", _cur_collection_par_time_ms);
1644 print_par_stats(2, "Update RS (Start)", _par_last_update_rs_start_times_ms, false);
1645 print_par_stats(2, "Update RS", _par_last_update_rs_times_ms);
1646 print_par_buffers(3, "Processed Buffers",
1647 _par_last_update_rs_processed_buffers, true);
1648 print_par_stats(2, "Ext Root Scanning", _par_last_ext_root_scan_times_ms);
1649 print_par_stats(2, "Mark Stack Scanning", _par_last_mark_stack_scan_times_ms);
1650 print_par_stats(2, "Scan-Only Scanning", _par_last_scan_only_times_ms);
1651 print_par_buffers(3, "Scan-Only Regions",
1652 _par_last_scan_only_regions_scanned, true);
1653 print_par_stats(2, "Scan RS", _par_last_scan_rs_times_ms);
1654 print_par_stats(2, "Object Copy", _par_last_obj_copy_times_ms);
1655 print_par_stats(2, "Termination", _par_last_termination_times_ms);
1656 print_stats(2, "Other", parallel_other_time);
1657 print_stats(1, "Clear CT", _cur_clear_ct_time_ms);
1658 } else {
1659 print_stats(1, "Update RS", update_rs_time);
1660 print_stats(2, "Processed Buffers",
1661 (int)update_rs_processed_buffers);
1662 print_stats(1, "Ext Root Scanning", ext_root_scan_time);
1663 print_stats(1, "Mark Stack Scanning", mark_stack_scan_time);
1664 print_stats(1, "Scan-Only Scanning", scan_only_time);
1665 print_stats(1, "Scan RS", scan_rs_time);
1666 print_stats(1, "Object Copying", obj_copy_time);
1667 }
1668 }
1669 #ifndef PRODUCT
1670 print_stats(1, "Cur Clear CC", _cur_clear_cc_time_ms);
1671 print_stats(1, "Cum Clear CC", _cum_clear_cc_time_ms);
1672 print_stats(1, "Min Clear CC", _min_clear_cc_time_ms);
1673 print_stats(1, "Max Clear CC", _max_clear_cc_time_ms);
1674 if (_num_cc_clears > 0) {
1675 print_stats(1, "Avg Clear CC", _cum_clear_cc_time_ms / ((double)_num_cc_clears));
1676 }
1677 #endif
1678 print_stats(1, "Other", other_time_ms);
1679 for (int i = 0; i < _aux_num; ++i) {
1680 if (_cur_aux_times_set[i]) {
1681 char buffer[96];
1682 sprintf(buffer, "Aux%d", i);
1683 print_stats(1, buffer, _cur_aux_times_ms[i]);
1684 }
1685 }
1686 }
1687 if (PrintGCDetails)
1688 gclog_or_tty->print(" [");
1689 if (PrintGC || PrintGCDetails)
1690 _g1->print_size_transition(gclog_or_tty,
1691 _cur_collection_pause_used_at_start_bytes,
1692 _g1->used(), _g1->capacity());
1693 if (PrintGCDetails)
1694 gclog_or_tty->print_cr("]");
1696 _all_pause_times_ms->add(elapsed_ms);
1697 if (update_stats) {
1698 summary->record_total_time_ms(elapsed_ms);
1699 summary->record_other_time_ms(other_time_ms);
1700 }
1701 for (int i = 0; i < _aux_num; ++i)
1702 if (_cur_aux_times_set[i])
1703 _all_aux_times_ms[i].add(_cur_aux_times_ms[i]);
1705 // Reset marks-between-pauses counter.
1706 _n_marks_since_last_pause = 0;
1708 // Update the efficiency-since-mark vars.
1709 double proc_ms = elapsed_ms * (double) _parallel_gc_threads;
1710 if (elapsed_ms < MIN_TIMER_GRANULARITY) {
1711 // This usually happens due to the timer not having the required
1712 // granularity. Some Linuxes are the usual culprits.
1713 // We'll just set it to something (arbitrarily) small.
1714 proc_ms = 1.0;
1715 }
1716 double cur_efficiency = (double) freed_bytes / proc_ms;
1718 bool new_in_marking_window = _in_marking_window;
1719 bool new_in_marking_window_im = false;
1720 if (_should_initiate_conc_mark) {
1721 new_in_marking_window = true;
1722 new_in_marking_window_im = true;
1723 }
1725 if (in_young_gc_mode()) {
1726 if (_last_full_young_gc) {
1727 set_full_young_gcs(false);
1728 _last_full_young_gc = false;
1729 }
1731 if ( !_last_young_gc_full ) {
1732 if ( _should_revert_to_full_young_gcs ||
1733 _known_garbage_ratio < 0.05 ||
1734 (adaptive_young_list_length() &&
1735 (get_gc_eff_factor() * cur_efficiency < predict_young_gc_eff())) ) {
1736 set_full_young_gcs(true);
1737 }
1738 }
1739 _should_revert_to_full_young_gcs = false;
1741 if (_last_young_gc_full && !_during_marking)
1742 _young_gc_eff_seq->add(cur_efficiency);
1743 }
1745 _short_lived_surv_rate_group->start_adding_regions();
1746 // do that for any other surv rate groupsx
1748 // <NEW PREDICTION>
1750 if (update_stats) {
1751 double pause_time_ms = elapsed_ms;
1753 size_t diff = 0;
1754 if (_max_pending_cards >= _pending_cards)
1755 diff = _max_pending_cards - _pending_cards;
1756 _pending_card_diff_seq->add((double) diff);
1758 double cost_per_card_ms = 0.0;
1759 if (_pending_cards > 0) {
1760 cost_per_card_ms = update_rs_time / (double) _pending_cards;
1761 _cost_per_card_ms_seq->add(cost_per_card_ms);
1762 }
1764 double cost_per_scan_only_region_ms = 0.0;
1765 if (scan_only_regions_scanned > 0.0) {
1766 cost_per_scan_only_region_ms =
1767 scan_only_time / scan_only_regions_scanned;
1768 if (_in_marking_window_im)
1769 _cost_per_scan_only_region_ms_during_cm_seq->add(cost_per_scan_only_region_ms);
1770 else
1771 _cost_per_scan_only_region_ms_seq->add(cost_per_scan_only_region_ms);
1772 }
1774 size_t cards_scanned = _g1->cards_scanned();
1776 double cost_per_entry_ms = 0.0;
1777 if (cards_scanned > 10) {
1778 cost_per_entry_ms = scan_rs_time / (double) cards_scanned;
1779 if (_last_young_gc_full)
1780 _cost_per_entry_ms_seq->add(cost_per_entry_ms);
1781 else
1782 _partially_young_cost_per_entry_ms_seq->add(cost_per_entry_ms);
1783 }
1785 if (_max_rs_lengths > 0) {
1786 double cards_per_entry_ratio =
1787 (double) cards_scanned / (double) _max_rs_lengths;
1788 if (_last_young_gc_full)
1789 _fully_young_cards_per_entry_ratio_seq->add(cards_per_entry_ratio);
1790 else
1791 _partially_young_cards_per_entry_ratio_seq->add(cards_per_entry_ratio);
1792 }
1794 size_t rs_length_diff = _max_rs_lengths - _recorded_rs_lengths;
1795 if (rs_length_diff >= 0)
1796 _rs_length_diff_seq->add((double) rs_length_diff);
1798 size_t copied_bytes = surviving_bytes;
1799 double cost_per_byte_ms = 0.0;
1800 if (copied_bytes > 0) {
1801 cost_per_byte_ms = obj_copy_time / (double) copied_bytes;
1802 if (_in_marking_window)
1803 _cost_per_byte_ms_during_cm_seq->add(cost_per_byte_ms);
1804 else
1805 _cost_per_byte_ms_seq->add(cost_per_byte_ms);
1806 }
1808 double all_other_time_ms = pause_time_ms -
1809 (update_rs_time + scan_only_time + scan_rs_time + obj_copy_time +
1810 _mark_closure_time_ms + termination_time);
1812 double young_other_time_ms = 0.0;
1813 if (_recorded_young_regions > 0) {
1814 young_other_time_ms =
1815 _recorded_young_cset_choice_time_ms +
1816 _recorded_young_free_cset_time_ms;
1817 _young_other_cost_per_region_ms_seq->add(young_other_time_ms /
1818 (double) _recorded_young_regions);
1819 }
1820 double non_young_other_time_ms = 0.0;
1821 if (_recorded_non_young_regions > 0) {
1822 non_young_other_time_ms =
1823 _recorded_non_young_cset_choice_time_ms +
1824 _recorded_non_young_free_cset_time_ms;
1826 _non_young_other_cost_per_region_ms_seq->add(non_young_other_time_ms /
1827 (double) _recorded_non_young_regions);
1828 }
1830 double constant_other_time_ms = all_other_time_ms -
1831 (young_other_time_ms + non_young_other_time_ms);
1832 _constant_other_time_ms_seq->add(constant_other_time_ms);
1834 double survival_ratio = 0.0;
1835 if (_bytes_in_collection_set_before_gc > 0) {
1836 survival_ratio = (double) bytes_in_to_space_during_gc() /
1837 (double) _bytes_in_collection_set_before_gc;
1838 }
1840 _pending_cards_seq->add((double) _pending_cards);
1841 _scanned_cards_seq->add((double) cards_scanned);
1842 _rs_lengths_seq->add((double) _max_rs_lengths);
1844 double expensive_region_limit_ms =
1845 (double) MaxGCPauseMillis - predict_constant_other_time_ms();
1846 if (expensive_region_limit_ms < 0.0) {
1847 // this means that the other time was predicted to be longer than
1848 // than the max pause time
1849 expensive_region_limit_ms = (double) MaxGCPauseMillis;
1850 }
1851 _expensive_region_limit_ms = expensive_region_limit_ms;
1853 if (PREDICTIONS_VERBOSE) {
1854 gclog_or_tty->print_cr("");
1855 gclog_or_tty->print_cr("PREDICTIONS %1.4lf %d "
1856 "REGIONS %d %d %d %d "
1857 "PENDING_CARDS %d %d "
1858 "CARDS_SCANNED %d %d "
1859 "RS_LENGTHS %d %d "
1860 "SCAN_ONLY_SCAN %1.6lf %1.6lf "
1861 "RS_UPDATE %1.6lf %1.6lf RS_SCAN %1.6lf %1.6lf "
1862 "SURVIVAL_RATIO %1.6lf %1.6lf "
1863 "OBJECT_COPY %1.6lf %1.6lf OTHER_CONSTANT %1.6lf %1.6lf "
1864 "OTHER_YOUNG %1.6lf %1.6lf "
1865 "OTHER_NON_YOUNG %1.6lf %1.6lf "
1866 "VTIME_DIFF %1.6lf TERMINATION %1.6lf "
1867 "ELAPSED %1.6lf %1.6lf ",
1868 _cur_collection_start_sec,
1869 (!_last_young_gc_full) ? 2 :
1870 (last_pause_included_initial_mark) ? 1 : 0,
1871 _recorded_region_num,
1872 _recorded_young_regions,
1873 _recorded_scan_only_regions,
1874 _recorded_non_young_regions,
1875 _predicted_pending_cards, _pending_cards,
1876 _predicted_cards_scanned, cards_scanned,
1877 _predicted_rs_lengths, _max_rs_lengths,
1878 _predicted_scan_only_scan_time_ms, scan_only_time,
1879 _predicted_rs_update_time_ms, update_rs_time,
1880 _predicted_rs_scan_time_ms, scan_rs_time,
1881 _predicted_survival_ratio, survival_ratio,
1882 _predicted_object_copy_time_ms, obj_copy_time,
1883 _predicted_constant_other_time_ms, constant_other_time_ms,
1884 _predicted_young_other_time_ms, young_other_time_ms,
1885 _predicted_non_young_other_time_ms,
1886 non_young_other_time_ms,
1887 _vtime_diff_ms, termination_time,
1888 _predicted_pause_time_ms, elapsed_ms);
1889 }
1891 if (G1PolicyVerbose > 0) {
1892 gclog_or_tty->print_cr("Pause Time, predicted: %1.4lfms (predicted %s), actual: %1.4lfms",
1893 _predicted_pause_time_ms,
1894 (_within_target) ? "within" : "outside",
1895 elapsed_ms);
1896 }
1898 }
1900 _in_marking_window = new_in_marking_window;
1901 _in_marking_window_im = new_in_marking_window_im;
1902 _free_regions_at_end_of_collection = _g1->free_regions();
1903 _scan_only_regions_at_end_of_collection = _g1->young_list_length();
1904 calculate_young_list_min_length();
1905 calculate_young_list_target_config();
1907 // Note that _mmu_tracker->max_gc_time() returns the time in seconds.
1908 double update_rs_time_goal_ms = _mmu_tracker->max_gc_time() * MILLIUNITS * G1RSetUpdatingPauseTimePercent / 100.0;
1909 adjust_concurrent_refinement(update_rs_time, update_rs_processed_buffers, update_rs_time_goal_ms);
1911 // </NEW PREDICTION>
1913 _target_pause_time_ms = -1.0;
1914 }
1916 // <NEW PREDICTION>
1918 void G1CollectorPolicy::adjust_concurrent_refinement(double update_rs_time,
1919 double update_rs_processed_buffers,
1920 double goal_ms) {
1921 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set();
1922 ConcurrentG1Refine *cg1r = G1CollectedHeap::heap()->concurrent_g1_refine();
1924 if (G1UseAdaptiveConcRefinement) {
1925 const int k_gy = 3, k_gr = 6;
1926 const double inc_k = 1.1, dec_k = 0.9;
1928 int g = cg1r->green_zone();
1929 if (update_rs_time > goal_ms) {
1930 g = (int)(g * dec_k); // Can become 0, that's OK. That would mean a mutator-only processing.
1931 } else {
1932 if (update_rs_time < goal_ms && update_rs_processed_buffers > g) {
1933 g = (int)MAX2(g * inc_k, g + 1.0);
1934 }
1935 }
1936 // Change the refinement threads params
1937 cg1r->set_green_zone(g);
1938 cg1r->set_yellow_zone(g * k_gy);
1939 cg1r->set_red_zone(g * k_gr);
1940 cg1r->reinitialize_threads();
1942 int processing_threshold_delta = MAX2((int)(cg1r->green_zone() * sigma()), 1);
1943 int processing_threshold = MIN2(cg1r->green_zone() + processing_threshold_delta,
1944 cg1r->yellow_zone());
1945 // Change the barrier params
1946 dcqs.set_process_completed_threshold(processing_threshold);
1947 dcqs.set_max_completed_queue(cg1r->red_zone());
1948 }
1950 int curr_queue_size = dcqs.completed_buffers_num();
1951 if (curr_queue_size >= cg1r->yellow_zone()) {
1952 dcqs.set_completed_queue_padding(curr_queue_size);
1953 } else {
1954 dcqs.set_completed_queue_padding(0);
1955 }
1956 dcqs.notify_if_necessary();
1957 }
1959 double
1960 G1CollectorPolicy::
1961 predict_young_collection_elapsed_time_ms(size_t adjustment) {
1962 guarantee( adjustment == 0 || adjustment == 1, "invariant" );
1964 G1CollectedHeap* g1h = G1CollectedHeap::heap();
1965 size_t young_num = g1h->young_list_length();
1966 if (young_num == 0)
1967 return 0.0;
1969 young_num += adjustment;
1970 size_t pending_cards = predict_pending_cards();
1971 size_t rs_lengths = g1h->young_list_sampled_rs_lengths() +
1972 predict_rs_length_diff();
1973 size_t card_num;
1974 if (full_young_gcs())
1975 card_num = predict_young_card_num(rs_lengths);
1976 else
1977 card_num = predict_non_young_card_num(rs_lengths);
1978 size_t young_byte_size = young_num * HeapRegion::GrainBytes;
1979 double accum_yg_surv_rate =
1980 _short_lived_surv_rate_group->accum_surv_rate(adjustment);
1982 size_t bytes_to_copy =
1983 (size_t) (accum_yg_surv_rate * (double) HeapRegion::GrainBytes);
1985 return
1986 predict_rs_update_time_ms(pending_cards) +
1987 predict_rs_scan_time_ms(card_num) +
1988 predict_object_copy_time_ms(bytes_to_copy) +
1989 predict_young_other_time_ms(young_num) +
1990 predict_constant_other_time_ms();
1991 }
1993 double
1994 G1CollectorPolicy::predict_base_elapsed_time_ms(size_t pending_cards) {
1995 size_t rs_length = predict_rs_length_diff();
1996 size_t card_num;
1997 if (full_young_gcs())
1998 card_num = predict_young_card_num(rs_length);
1999 else
2000 card_num = predict_non_young_card_num(rs_length);
2001 return predict_base_elapsed_time_ms(pending_cards, card_num);
2002 }
2004 double
2005 G1CollectorPolicy::predict_base_elapsed_time_ms(size_t pending_cards,
2006 size_t scanned_cards) {
2007 return
2008 predict_rs_update_time_ms(pending_cards) +
2009 predict_rs_scan_time_ms(scanned_cards) +
2010 predict_constant_other_time_ms();
2011 }
2013 double
2014 G1CollectorPolicy::predict_region_elapsed_time_ms(HeapRegion* hr,
2015 bool young) {
2016 size_t rs_length = hr->rem_set()->occupied();
2017 size_t card_num;
2018 if (full_young_gcs())
2019 card_num = predict_young_card_num(rs_length);
2020 else
2021 card_num = predict_non_young_card_num(rs_length);
2022 size_t bytes_to_copy = predict_bytes_to_copy(hr);
2024 double region_elapsed_time_ms =
2025 predict_rs_scan_time_ms(card_num) +
2026 predict_object_copy_time_ms(bytes_to_copy);
2028 if (young)
2029 region_elapsed_time_ms += predict_young_other_time_ms(1);
2030 else
2031 region_elapsed_time_ms += predict_non_young_other_time_ms(1);
2033 return region_elapsed_time_ms;
2034 }
2036 size_t
2037 G1CollectorPolicy::predict_bytes_to_copy(HeapRegion* hr) {
2038 size_t bytes_to_copy;
2039 if (hr->is_marked())
2040 bytes_to_copy = hr->max_live_bytes();
2041 else {
2042 guarantee( hr->is_young() && hr->age_in_surv_rate_group() != -1,
2043 "invariant" );
2044 int age = hr->age_in_surv_rate_group();
2045 double yg_surv_rate = predict_yg_surv_rate(age, hr->surv_rate_group());
2046 bytes_to_copy = (size_t) ((double) hr->used() * yg_surv_rate);
2047 }
2049 return bytes_to_copy;
2050 }
2052 void
2053 G1CollectorPolicy::start_recording_regions() {
2054 _recorded_rs_lengths = 0;
2055 _recorded_scan_only_regions = 0;
2056 _recorded_young_regions = 0;
2057 _recorded_non_young_regions = 0;
2059 #if PREDICTIONS_VERBOSE
2060 _predicted_rs_lengths = 0;
2061 _predicted_cards_scanned = 0;
2063 _recorded_marked_bytes = 0;
2064 _recorded_young_bytes = 0;
2065 _predicted_bytes_to_copy = 0;
2066 #endif // PREDICTIONS_VERBOSE
2067 }
2069 void
2070 G1CollectorPolicy::record_cset_region(HeapRegion* hr, bool young) {
2071 if (young) {
2072 ++_recorded_young_regions;
2073 } else {
2074 ++_recorded_non_young_regions;
2075 }
2076 #if PREDICTIONS_VERBOSE
2077 if (young) {
2078 _recorded_young_bytes += hr->used();
2079 } else {
2080 _recorded_marked_bytes += hr->max_live_bytes();
2081 }
2082 _predicted_bytes_to_copy += predict_bytes_to_copy(hr);
2083 #endif // PREDICTIONS_VERBOSE
2085 size_t rs_length = hr->rem_set()->occupied();
2086 _recorded_rs_lengths += rs_length;
2087 }
2089 void
2090 G1CollectorPolicy::record_scan_only_regions(size_t scan_only_length) {
2091 _recorded_scan_only_regions = scan_only_length;
2092 }
2094 void
2095 G1CollectorPolicy::end_recording_regions() {
2096 #if PREDICTIONS_VERBOSE
2097 _predicted_pending_cards = predict_pending_cards();
2098 _predicted_rs_lengths = _recorded_rs_lengths + predict_rs_length_diff();
2099 if (full_young_gcs())
2100 _predicted_cards_scanned += predict_young_card_num(_predicted_rs_lengths);
2101 else
2102 _predicted_cards_scanned +=
2103 predict_non_young_card_num(_predicted_rs_lengths);
2104 _recorded_region_num = _recorded_young_regions + _recorded_non_young_regions;
2106 _predicted_scan_only_scan_time_ms =
2107 predict_scan_only_time_ms(_recorded_scan_only_regions);
2108 _predicted_rs_update_time_ms =
2109 predict_rs_update_time_ms(_g1->pending_card_num());
2110 _predicted_rs_scan_time_ms =
2111 predict_rs_scan_time_ms(_predicted_cards_scanned);
2112 _predicted_object_copy_time_ms =
2113 predict_object_copy_time_ms(_predicted_bytes_to_copy);
2114 _predicted_constant_other_time_ms =
2115 predict_constant_other_time_ms();
2116 _predicted_young_other_time_ms =
2117 predict_young_other_time_ms(_recorded_young_regions);
2118 _predicted_non_young_other_time_ms =
2119 predict_non_young_other_time_ms(_recorded_non_young_regions);
2121 _predicted_pause_time_ms =
2122 _predicted_scan_only_scan_time_ms +
2123 _predicted_rs_update_time_ms +
2124 _predicted_rs_scan_time_ms +
2125 _predicted_object_copy_time_ms +
2126 _predicted_constant_other_time_ms +
2127 _predicted_young_other_time_ms +
2128 _predicted_non_young_other_time_ms;
2129 #endif // PREDICTIONS_VERBOSE
2130 }
2132 void G1CollectorPolicy::check_if_region_is_too_expensive(double
2133 predicted_time_ms) {
2134 // I don't think we need to do this when in young GC mode since
2135 // marking will be initiated next time we hit the soft limit anyway...
2136 if (predicted_time_ms > _expensive_region_limit_ms) {
2137 if (!in_young_gc_mode()) {
2138 set_full_young_gcs(true);
2139 _should_initiate_conc_mark = true;
2140 } else
2141 // no point in doing another partial one
2142 _should_revert_to_full_young_gcs = true;
2143 }
2144 }
2146 // </NEW PREDICTION>
2149 void G1CollectorPolicy::update_recent_gc_times(double end_time_sec,
2150 double elapsed_ms) {
2151 _recent_gc_times_ms->add(elapsed_ms);
2152 _recent_prev_end_times_for_all_gcs_sec->add(end_time_sec);
2153 _prev_collection_pause_end_ms = end_time_sec * 1000.0;
2154 }
2156 double G1CollectorPolicy::recent_avg_time_for_pauses_ms() {
2157 if (_recent_pause_times_ms->num() == 0) return (double) MaxGCPauseMillis;
2158 else return _recent_pause_times_ms->avg();
2159 }
2161 double G1CollectorPolicy::recent_avg_time_for_CH_strong_ms() {
2162 if (_recent_CH_strong_roots_times_ms->num() == 0)
2163 return (double)MaxGCPauseMillis/3.0;
2164 else return _recent_CH_strong_roots_times_ms->avg();
2165 }
2167 double G1CollectorPolicy::recent_avg_time_for_G1_strong_ms() {
2168 if (_recent_G1_strong_roots_times_ms->num() == 0)
2169 return (double)MaxGCPauseMillis/3.0;
2170 else return _recent_G1_strong_roots_times_ms->avg();
2171 }
2173 double G1CollectorPolicy::recent_avg_time_for_evac_ms() {
2174 if (_recent_evac_times_ms->num() == 0) return (double)MaxGCPauseMillis/3.0;
2175 else return _recent_evac_times_ms->avg();
2176 }
2178 int G1CollectorPolicy::number_of_recent_gcs() {
2179 assert(_recent_CH_strong_roots_times_ms->num() ==
2180 _recent_G1_strong_roots_times_ms->num(), "Sequence out of sync");
2181 assert(_recent_G1_strong_roots_times_ms->num() ==
2182 _recent_evac_times_ms->num(), "Sequence out of sync");
2183 assert(_recent_evac_times_ms->num() ==
2184 _recent_pause_times_ms->num(), "Sequence out of sync");
2185 assert(_recent_pause_times_ms->num() ==
2186 _recent_CS_bytes_used_before->num(), "Sequence out of sync");
2187 assert(_recent_CS_bytes_used_before->num() ==
2188 _recent_CS_bytes_surviving->num(), "Sequence out of sync");
2189 return _recent_pause_times_ms->num();
2190 }
2192 double G1CollectorPolicy::recent_avg_survival_fraction() {
2193 return recent_avg_survival_fraction_work(_recent_CS_bytes_surviving,
2194 _recent_CS_bytes_used_before);
2195 }
2197 double G1CollectorPolicy::last_survival_fraction() {
2198 return last_survival_fraction_work(_recent_CS_bytes_surviving,
2199 _recent_CS_bytes_used_before);
2200 }
2202 double
2203 G1CollectorPolicy::recent_avg_survival_fraction_work(TruncatedSeq* surviving,
2204 TruncatedSeq* before) {
2205 assert(surviving->num() == before->num(), "Sequence out of sync");
2206 if (before->sum() > 0.0) {
2207 double recent_survival_rate = surviving->sum() / before->sum();
2208 // We exempt parallel collection from this check because Alloc Buffer
2209 // fragmentation can produce negative collections.
2210 // Further, we're now always doing parallel collection. But I'm still
2211 // leaving this here as a placeholder for a more precise assertion later.
2212 // (DLD, 10/05.)
2213 assert((true || ParallelGCThreads > 0) ||
2214 _g1->evacuation_failed() ||
2215 recent_survival_rate <= 1.0, "Or bad frac");
2216 return recent_survival_rate;
2217 } else {
2218 return 1.0; // Be conservative.
2219 }
2220 }
2222 double
2223 G1CollectorPolicy::last_survival_fraction_work(TruncatedSeq* surviving,
2224 TruncatedSeq* before) {
2225 assert(surviving->num() == before->num(), "Sequence out of sync");
2226 if (surviving->num() > 0 && before->last() > 0.0) {
2227 double last_survival_rate = surviving->last() / before->last();
2228 // We exempt parallel collection from this check because Alloc Buffer
2229 // fragmentation can produce negative collections.
2230 // Further, we're now always doing parallel collection. But I'm still
2231 // leaving this here as a placeholder for a more precise assertion later.
2232 // (DLD, 10/05.)
2233 assert((true || ParallelGCThreads > 0) ||
2234 last_survival_rate <= 1.0, "Or bad frac");
2235 return last_survival_rate;
2236 } else {
2237 return 1.0;
2238 }
2239 }
2241 static const int survival_min_obs = 5;
2242 static double survival_min_obs_limits[] = { 0.9, 0.7, 0.5, 0.3, 0.1 };
2243 static const double min_survival_rate = 0.1;
2245 double
2246 G1CollectorPolicy::conservative_avg_survival_fraction_work(double avg,
2247 double latest) {
2248 double res = avg;
2249 if (number_of_recent_gcs() < survival_min_obs) {
2250 res = MAX2(res, survival_min_obs_limits[number_of_recent_gcs()]);
2251 }
2252 res = MAX2(res, latest);
2253 res = MAX2(res, min_survival_rate);
2254 // In the parallel case, LAB fragmentation can produce "negative
2255 // collections"; so can evac failure. Cap at 1.0
2256 res = MIN2(res, 1.0);
2257 return res;
2258 }
2260 size_t G1CollectorPolicy::expansion_amount() {
2261 if ((int)(recent_avg_pause_time_ratio() * 100.0) > G1GCPercent) {
2262 // We will double the existing space, or take
2263 // G1ExpandByPercentOfAvailable % of the available expansion
2264 // space, whichever is smaller, bounded below by a minimum
2265 // expansion (unless that's all that's left.)
2266 const size_t min_expand_bytes = 1*M;
2267 size_t reserved_bytes = _g1->g1_reserved_obj_bytes();
2268 size_t committed_bytes = _g1->capacity();
2269 size_t uncommitted_bytes = reserved_bytes - committed_bytes;
2270 size_t expand_bytes;
2271 size_t expand_bytes_via_pct =
2272 uncommitted_bytes * G1ExpandByPercentOfAvailable / 100;
2273 expand_bytes = MIN2(expand_bytes_via_pct, committed_bytes);
2274 expand_bytes = MAX2(expand_bytes, min_expand_bytes);
2275 expand_bytes = MIN2(expand_bytes, uncommitted_bytes);
2276 if (G1PolicyVerbose > 1) {
2277 gclog_or_tty->print("Decided to expand: ratio = %5.2f, "
2278 "committed = %d%s, uncommited = %d%s, via pct = %d%s.\n"
2279 " Answer = %d.\n",
2280 recent_avg_pause_time_ratio(),
2281 byte_size_in_proper_unit(committed_bytes),
2282 proper_unit_for_byte_size(committed_bytes),
2283 byte_size_in_proper_unit(uncommitted_bytes),
2284 proper_unit_for_byte_size(uncommitted_bytes),
2285 byte_size_in_proper_unit(expand_bytes_via_pct),
2286 proper_unit_for_byte_size(expand_bytes_via_pct),
2287 byte_size_in_proper_unit(expand_bytes),
2288 proper_unit_for_byte_size(expand_bytes));
2289 }
2290 return expand_bytes;
2291 } else {
2292 return 0;
2293 }
2294 }
2296 void G1CollectorPolicy::note_start_of_mark_thread() {
2297 _mark_thread_startup_sec = os::elapsedTime();
2298 }
2300 class CountCSClosure: public HeapRegionClosure {
2301 G1CollectorPolicy* _g1_policy;
2302 public:
2303 CountCSClosure(G1CollectorPolicy* g1_policy) :
2304 _g1_policy(g1_policy) {}
2305 bool doHeapRegion(HeapRegion* r) {
2306 _g1_policy->_bytes_in_collection_set_before_gc += r->used();
2307 return false;
2308 }
2309 };
2311 void G1CollectorPolicy::count_CS_bytes_used() {
2312 CountCSClosure cs_closure(this);
2313 _g1->collection_set_iterate(&cs_closure);
2314 }
2316 static void print_indent(int level) {
2317 for (int j = 0; j < level+1; ++j)
2318 gclog_or_tty->print(" ");
2319 }
2321 void G1CollectorPolicy::print_summary (int level,
2322 const char* str,
2323 NumberSeq* seq) const {
2324 double sum = seq->sum();
2325 print_indent(level);
2326 gclog_or_tty->print_cr("%-24s = %8.2lf s (avg = %8.2lf ms)",
2327 str, sum / 1000.0, seq->avg());
2328 }
2330 void G1CollectorPolicy::print_summary_sd (int level,
2331 const char* str,
2332 NumberSeq* seq) const {
2333 print_summary(level, str, seq);
2334 print_indent(level + 5);
2335 gclog_or_tty->print_cr("(num = %5d, std dev = %8.2lf ms, max = %8.2lf ms)",
2336 seq->num(), seq->sd(), seq->maximum());
2337 }
2339 void G1CollectorPolicy::check_other_times(int level,
2340 NumberSeq* other_times_ms,
2341 NumberSeq* calc_other_times_ms) const {
2342 bool should_print = false;
2344 double max_sum = MAX2(fabs(other_times_ms->sum()),
2345 fabs(calc_other_times_ms->sum()));
2346 double min_sum = MIN2(fabs(other_times_ms->sum()),
2347 fabs(calc_other_times_ms->sum()));
2348 double sum_ratio = max_sum / min_sum;
2349 if (sum_ratio > 1.1) {
2350 should_print = true;
2351 print_indent(level + 1);
2352 gclog_or_tty->print_cr("## CALCULATED OTHER SUM DOESN'T MATCH RECORDED ###");
2353 }
2355 double max_avg = MAX2(fabs(other_times_ms->avg()),
2356 fabs(calc_other_times_ms->avg()));
2357 double min_avg = MIN2(fabs(other_times_ms->avg()),
2358 fabs(calc_other_times_ms->avg()));
2359 double avg_ratio = max_avg / min_avg;
2360 if (avg_ratio > 1.1) {
2361 should_print = true;
2362 print_indent(level + 1);
2363 gclog_or_tty->print_cr("## CALCULATED OTHER AVG DOESN'T MATCH RECORDED ###");
2364 }
2366 if (other_times_ms->sum() < -0.01) {
2367 print_indent(level + 1);
2368 gclog_or_tty->print_cr("## RECORDED OTHER SUM IS NEGATIVE ###");
2369 }
2371 if (other_times_ms->avg() < -0.01) {
2372 print_indent(level + 1);
2373 gclog_or_tty->print_cr("## RECORDED OTHER AVG IS NEGATIVE ###");
2374 }
2376 if (calc_other_times_ms->sum() < -0.01) {
2377 should_print = true;
2378 print_indent(level + 1);
2379 gclog_or_tty->print_cr("## CALCULATED OTHER SUM IS NEGATIVE ###");
2380 }
2382 if (calc_other_times_ms->avg() < -0.01) {
2383 should_print = true;
2384 print_indent(level + 1);
2385 gclog_or_tty->print_cr("## CALCULATED OTHER AVG IS NEGATIVE ###");
2386 }
2388 if (should_print)
2389 print_summary(level, "Other(Calc)", calc_other_times_ms);
2390 }
2392 void G1CollectorPolicy::print_summary(PauseSummary* summary) const {
2393 bool parallel = ParallelGCThreads > 0;
2394 MainBodySummary* body_summary = summary->main_body_summary();
2395 if (summary->get_total_seq()->num() > 0) {
2396 print_summary_sd(0, "Evacuation Pauses", summary->get_total_seq());
2397 if (body_summary != NULL) {
2398 print_summary(1, "SATB Drain", body_summary->get_satb_drain_seq());
2399 if (parallel) {
2400 print_summary(1, "Parallel Time", body_summary->get_parallel_seq());
2401 print_summary(2, "Update RS", body_summary->get_update_rs_seq());
2402 print_summary(2, "Ext Root Scanning",
2403 body_summary->get_ext_root_scan_seq());
2404 print_summary(2, "Mark Stack Scanning",
2405 body_summary->get_mark_stack_scan_seq());
2406 print_summary(2, "Scan-Only Scanning",
2407 body_summary->get_scan_only_seq());
2408 print_summary(2, "Scan RS", body_summary->get_scan_rs_seq());
2409 print_summary(2, "Object Copy", body_summary->get_obj_copy_seq());
2410 print_summary(2, "Termination", body_summary->get_termination_seq());
2411 print_summary(2, "Other", body_summary->get_parallel_other_seq());
2412 {
2413 NumberSeq* other_parts[] = {
2414 body_summary->get_update_rs_seq(),
2415 body_summary->get_ext_root_scan_seq(),
2416 body_summary->get_mark_stack_scan_seq(),
2417 body_summary->get_scan_only_seq(),
2418 body_summary->get_scan_rs_seq(),
2419 body_summary->get_obj_copy_seq(),
2420 body_summary->get_termination_seq()
2421 };
2422 NumberSeq calc_other_times_ms(body_summary->get_parallel_seq(),
2423 7, other_parts);
2424 check_other_times(2, body_summary->get_parallel_other_seq(),
2425 &calc_other_times_ms);
2426 }
2427 print_summary(1, "Mark Closure", body_summary->get_mark_closure_seq());
2428 print_summary(1, "Clear CT", body_summary->get_clear_ct_seq());
2429 } else {
2430 print_summary(1, "Update RS", body_summary->get_update_rs_seq());
2431 print_summary(1, "Ext Root Scanning",
2432 body_summary->get_ext_root_scan_seq());
2433 print_summary(1, "Mark Stack Scanning",
2434 body_summary->get_mark_stack_scan_seq());
2435 print_summary(1, "Scan-Only Scanning",
2436 body_summary->get_scan_only_seq());
2437 print_summary(1, "Scan RS", body_summary->get_scan_rs_seq());
2438 print_summary(1, "Object Copy", body_summary->get_obj_copy_seq());
2439 }
2440 }
2441 print_summary(1, "Other", summary->get_other_seq());
2442 {
2443 NumberSeq calc_other_times_ms;
2444 if (body_summary != NULL) {
2445 // not abandoned
2446 if (parallel) {
2447 // parallel
2448 NumberSeq* other_parts[] = {
2449 body_summary->get_satb_drain_seq(),
2450 body_summary->get_parallel_seq(),
2451 body_summary->get_clear_ct_seq()
2452 };
2453 calc_other_times_ms = NumberSeq(summary->get_total_seq(),
2454 3, other_parts);
2455 } else {
2456 // serial
2457 NumberSeq* other_parts[] = {
2458 body_summary->get_satb_drain_seq(),
2459 body_summary->get_update_rs_seq(),
2460 body_summary->get_ext_root_scan_seq(),
2461 body_summary->get_mark_stack_scan_seq(),
2462 body_summary->get_scan_only_seq(),
2463 body_summary->get_scan_rs_seq(),
2464 body_summary->get_obj_copy_seq()
2465 };
2466 calc_other_times_ms = NumberSeq(summary->get_total_seq(),
2467 7, other_parts);
2468 }
2469 } else {
2470 // abandoned
2471 calc_other_times_ms = NumberSeq();
2472 }
2473 check_other_times(1, summary->get_other_seq(), &calc_other_times_ms);
2474 }
2475 } else {
2476 print_indent(0);
2477 gclog_or_tty->print_cr("none");
2478 }
2479 gclog_or_tty->print_cr("");
2480 }
2482 void
2483 G1CollectorPolicy::print_abandoned_summary(PauseSummary* summary) const {
2484 bool printed = false;
2485 if (summary->get_total_seq()->num() > 0) {
2486 printed = true;
2487 print_summary(summary);
2488 }
2489 if (!printed) {
2490 print_indent(0);
2491 gclog_or_tty->print_cr("none");
2492 gclog_or_tty->print_cr("");
2493 }
2494 }
2496 void G1CollectorPolicy::print_tracing_info() const {
2497 if (TraceGen0Time) {
2498 gclog_or_tty->print_cr("ALL PAUSES");
2499 print_summary_sd(0, "Total", _all_pause_times_ms);
2500 gclog_or_tty->print_cr("");
2501 gclog_or_tty->print_cr("");
2502 gclog_or_tty->print_cr(" Full Young GC Pauses: %8d", _full_young_pause_num);
2503 gclog_or_tty->print_cr(" Partial Young GC Pauses: %8d", _partial_young_pause_num);
2504 gclog_or_tty->print_cr("");
2506 gclog_or_tty->print_cr("EVACUATION PAUSES");
2507 print_summary(_summary);
2509 gclog_or_tty->print_cr("ABANDONED PAUSES");
2510 print_abandoned_summary(_abandoned_summary);
2512 gclog_or_tty->print_cr("MISC");
2513 print_summary_sd(0, "Stop World", _all_stop_world_times_ms);
2514 print_summary_sd(0, "Yields", _all_yield_times_ms);
2515 for (int i = 0; i < _aux_num; ++i) {
2516 if (_all_aux_times_ms[i].num() > 0) {
2517 char buffer[96];
2518 sprintf(buffer, "Aux%d", i);
2519 print_summary_sd(0, buffer, &_all_aux_times_ms[i]);
2520 }
2521 }
2523 size_t all_region_num = _region_num_young + _region_num_tenured;
2524 gclog_or_tty->print_cr(" New Regions %8d, Young %8d (%6.2lf%%), "
2525 "Tenured %8d (%6.2lf%%)",
2526 all_region_num,
2527 _region_num_young,
2528 (double) _region_num_young / (double) all_region_num * 100.0,
2529 _region_num_tenured,
2530 (double) _region_num_tenured / (double) all_region_num * 100.0);
2531 }
2532 if (TraceGen1Time) {
2533 if (_all_full_gc_times_ms->num() > 0) {
2534 gclog_or_tty->print("\n%4d full_gcs: total time = %8.2f s",
2535 _all_full_gc_times_ms->num(),
2536 _all_full_gc_times_ms->sum() / 1000.0);
2537 gclog_or_tty->print_cr(" (avg = %8.2fms).", _all_full_gc_times_ms->avg());
2538 gclog_or_tty->print_cr(" [std. dev = %8.2f ms, max = %8.2f ms]",
2539 _all_full_gc_times_ms->sd(),
2540 _all_full_gc_times_ms->maximum());
2541 }
2542 }
2543 }
2545 void G1CollectorPolicy::print_yg_surv_rate_info() const {
2546 #ifndef PRODUCT
2547 _short_lived_surv_rate_group->print_surv_rate_summary();
2548 // add this call for any other surv rate groups
2549 #endif // PRODUCT
2550 }
2552 bool
2553 G1CollectorPolicy::should_add_next_region_to_young_list() {
2554 assert(in_young_gc_mode(), "should be in young GC mode");
2555 bool ret;
2556 size_t young_list_length = _g1->young_list_length();
2557 size_t young_list_max_length = _young_list_target_length;
2558 if (G1FixedEdenSize) {
2559 young_list_max_length -= _max_survivor_regions;
2560 }
2561 if (young_list_length < young_list_max_length) {
2562 ret = true;
2563 ++_region_num_young;
2564 } else {
2565 ret = false;
2566 ++_region_num_tenured;
2567 }
2569 return ret;
2570 }
2572 #ifndef PRODUCT
2573 // for debugging, bit of a hack...
2574 static char*
2575 region_num_to_mbs(int length) {
2576 static char buffer[64];
2577 double bytes = (double) (length * HeapRegion::GrainBytes);
2578 double mbs = bytes / (double) (1024 * 1024);
2579 sprintf(buffer, "%7.2lfMB", mbs);
2580 return buffer;
2581 }
2582 #endif // PRODUCT
2584 size_t G1CollectorPolicy::max_regions(int purpose) {
2585 switch (purpose) {
2586 case GCAllocForSurvived:
2587 return _max_survivor_regions;
2588 case GCAllocForTenured:
2589 return REGIONS_UNLIMITED;
2590 default:
2591 ShouldNotReachHere();
2592 return REGIONS_UNLIMITED;
2593 };
2594 }
2596 // Calculates survivor space parameters.
2597 void G1CollectorPolicy::calculate_survivors_policy()
2598 {
2599 if (G1FixedSurvivorSpaceSize == 0) {
2600 _max_survivor_regions = _young_list_target_length / SurvivorRatio;
2601 } else {
2602 _max_survivor_regions = G1FixedSurvivorSpaceSize / HeapRegion::GrainBytes;
2603 }
2605 if (G1FixedTenuringThreshold) {
2606 _tenuring_threshold = MaxTenuringThreshold;
2607 } else {
2608 _tenuring_threshold = _survivors_age_table.compute_tenuring_threshold(
2609 HeapRegion::GrainWords * _max_survivor_regions);
2610 }
2611 }
2613 bool
2614 G1CollectorPolicy_BestRegionsFirst::should_do_collection_pause(size_t
2615 word_size) {
2616 assert(_g1->regions_accounted_for(), "Region leakage!");
2617 double max_pause_time_ms = _mmu_tracker->max_gc_time() * 1000.0;
2619 size_t young_list_length = _g1->young_list_length();
2620 size_t young_list_max_length = _young_list_target_length;
2621 if (G1FixedEdenSize) {
2622 young_list_max_length -= _max_survivor_regions;
2623 }
2624 bool reached_target_length = young_list_length >= young_list_max_length;
2626 if (in_young_gc_mode()) {
2627 if (reached_target_length) {
2628 assert( young_list_length > 0 && _g1->young_list_length() > 0,
2629 "invariant" );
2630 _target_pause_time_ms = max_pause_time_ms;
2631 return true;
2632 }
2633 } else {
2634 guarantee( false, "should not reach here" );
2635 }
2637 return false;
2638 }
2640 #ifndef PRODUCT
2641 class HRSortIndexIsOKClosure: public HeapRegionClosure {
2642 CollectionSetChooser* _chooser;
2643 public:
2644 HRSortIndexIsOKClosure(CollectionSetChooser* chooser) :
2645 _chooser(chooser) {}
2647 bool doHeapRegion(HeapRegion* r) {
2648 if (!r->continuesHumongous()) {
2649 assert(_chooser->regionProperlyOrdered(r), "Ought to be.");
2650 }
2651 return false;
2652 }
2653 };
2655 bool G1CollectorPolicy_BestRegionsFirst::assertMarkedBytesDataOK() {
2656 HRSortIndexIsOKClosure cl(_collectionSetChooser);
2657 _g1->heap_region_iterate(&cl);
2658 return true;
2659 }
2660 #endif
2662 void
2663 G1CollectorPolicy_BestRegionsFirst::
2664 record_collection_pause_start(double start_time_sec, size_t start_used) {
2665 G1CollectorPolicy::record_collection_pause_start(start_time_sec, start_used);
2666 }
2668 class NextNonCSElemFinder: public HeapRegionClosure {
2669 HeapRegion* _res;
2670 public:
2671 NextNonCSElemFinder(): _res(NULL) {}
2672 bool doHeapRegion(HeapRegion* r) {
2673 if (!r->in_collection_set()) {
2674 _res = r;
2675 return true;
2676 } else {
2677 return false;
2678 }
2679 }
2680 HeapRegion* res() { return _res; }
2681 };
2683 class KnownGarbageClosure: public HeapRegionClosure {
2684 CollectionSetChooser* _hrSorted;
2686 public:
2687 KnownGarbageClosure(CollectionSetChooser* hrSorted) :
2688 _hrSorted(hrSorted)
2689 {}
2691 bool doHeapRegion(HeapRegion* r) {
2692 // We only include humongous regions in collection
2693 // sets when concurrent mark shows that their contained object is
2694 // unreachable.
2696 // Do we have any marking information for this region?
2697 if (r->is_marked()) {
2698 // We don't include humongous regions in collection
2699 // sets because we collect them immediately at the end of a marking
2700 // cycle. We also don't include young regions because we *must*
2701 // include them in the next collection pause.
2702 if (!r->isHumongous() && !r->is_young()) {
2703 _hrSorted->addMarkedHeapRegion(r);
2704 }
2705 }
2706 return false;
2707 }
2708 };
2710 class ParKnownGarbageHRClosure: public HeapRegionClosure {
2711 CollectionSetChooser* _hrSorted;
2712 jint _marked_regions_added;
2713 jint _chunk_size;
2714 jint _cur_chunk_idx;
2715 jint _cur_chunk_end; // Cur chunk [_cur_chunk_idx, _cur_chunk_end)
2716 int _worker;
2717 int _invokes;
2719 void get_new_chunk() {
2720 _cur_chunk_idx = _hrSorted->getParMarkedHeapRegionChunk(_chunk_size);
2721 _cur_chunk_end = _cur_chunk_idx + _chunk_size;
2722 }
2723 void add_region(HeapRegion* r) {
2724 if (_cur_chunk_idx == _cur_chunk_end) {
2725 get_new_chunk();
2726 }
2727 assert(_cur_chunk_idx < _cur_chunk_end, "postcondition");
2728 _hrSorted->setMarkedHeapRegion(_cur_chunk_idx, r);
2729 _marked_regions_added++;
2730 _cur_chunk_idx++;
2731 }
2733 public:
2734 ParKnownGarbageHRClosure(CollectionSetChooser* hrSorted,
2735 jint chunk_size,
2736 int worker) :
2737 _hrSorted(hrSorted), _chunk_size(chunk_size), _worker(worker),
2738 _marked_regions_added(0), _cur_chunk_idx(0), _cur_chunk_end(0),
2739 _invokes(0)
2740 {}
2742 bool doHeapRegion(HeapRegion* r) {
2743 // We only include humongous regions in collection
2744 // sets when concurrent mark shows that their contained object is
2745 // unreachable.
2746 _invokes++;
2748 // Do we have any marking information for this region?
2749 if (r->is_marked()) {
2750 // We don't include humongous regions in collection
2751 // sets because we collect them immediately at the end of a marking
2752 // cycle.
2753 // We also do not include young regions in collection sets
2754 if (!r->isHumongous() && !r->is_young()) {
2755 add_region(r);
2756 }
2757 }
2758 return false;
2759 }
2760 jint marked_regions_added() { return _marked_regions_added; }
2761 int invokes() { return _invokes; }
2762 };
2764 class ParKnownGarbageTask: public AbstractGangTask {
2765 CollectionSetChooser* _hrSorted;
2766 jint _chunk_size;
2767 G1CollectedHeap* _g1;
2768 public:
2769 ParKnownGarbageTask(CollectionSetChooser* hrSorted, jint chunk_size) :
2770 AbstractGangTask("ParKnownGarbageTask"),
2771 _hrSorted(hrSorted), _chunk_size(chunk_size),
2772 _g1(G1CollectedHeap::heap())
2773 {}
2775 void work(int i) {
2776 ParKnownGarbageHRClosure parKnownGarbageCl(_hrSorted, _chunk_size, i);
2777 // Back to zero for the claim value.
2778 _g1->heap_region_par_iterate_chunked(&parKnownGarbageCl, i,
2779 HeapRegion::InitialClaimValue);
2780 jint regions_added = parKnownGarbageCl.marked_regions_added();
2781 _hrSorted->incNumMarkedHeapRegions(regions_added);
2782 if (G1PrintParCleanupStats) {
2783 gclog_or_tty->print(" Thread %d called %d times, added %d regions to list.\n",
2784 i, parKnownGarbageCl.invokes(), regions_added);
2785 }
2786 }
2787 };
2789 void
2790 G1CollectorPolicy_BestRegionsFirst::
2791 record_concurrent_mark_cleanup_end(size_t freed_bytes,
2792 size_t max_live_bytes) {
2793 double start;
2794 if (G1PrintParCleanupStats) start = os::elapsedTime();
2795 record_concurrent_mark_cleanup_end_work1(freed_bytes, max_live_bytes);
2797 _collectionSetChooser->clearMarkedHeapRegions();
2798 double clear_marked_end;
2799 if (G1PrintParCleanupStats) {
2800 clear_marked_end = os::elapsedTime();
2801 gclog_or_tty->print_cr(" clear marked regions + work1: %8.3f ms.",
2802 (clear_marked_end - start)*1000.0);
2803 }
2804 if (ParallelGCThreads > 0) {
2805 const size_t OverpartitionFactor = 4;
2806 const size_t MinChunkSize = 8;
2807 const size_t ChunkSize =
2808 MAX2(_g1->n_regions() / (ParallelGCThreads * OverpartitionFactor),
2809 MinChunkSize);
2810 _collectionSetChooser->prepareForAddMarkedHeapRegionsPar(_g1->n_regions(),
2811 ChunkSize);
2812 ParKnownGarbageTask parKnownGarbageTask(_collectionSetChooser,
2813 (int) ChunkSize);
2814 _g1->workers()->run_task(&parKnownGarbageTask);
2816 assert(_g1->check_heap_region_claim_values(HeapRegion::InitialClaimValue),
2817 "sanity check");
2818 } else {
2819 KnownGarbageClosure knownGarbagecl(_collectionSetChooser);
2820 _g1->heap_region_iterate(&knownGarbagecl);
2821 }
2822 double known_garbage_end;
2823 if (G1PrintParCleanupStats) {
2824 known_garbage_end = os::elapsedTime();
2825 gclog_or_tty->print_cr(" compute known garbage: %8.3f ms.",
2826 (known_garbage_end - clear_marked_end)*1000.0);
2827 }
2828 _collectionSetChooser->sortMarkedHeapRegions();
2829 double sort_end;
2830 if (G1PrintParCleanupStats) {
2831 sort_end = os::elapsedTime();
2832 gclog_or_tty->print_cr(" sorting: %8.3f ms.",
2833 (sort_end - known_garbage_end)*1000.0);
2834 }
2836 record_concurrent_mark_cleanup_end_work2();
2837 double work2_end;
2838 if (G1PrintParCleanupStats) {
2839 work2_end = os::elapsedTime();
2840 gclog_or_tty->print_cr(" work2: %8.3f ms.",
2841 (work2_end - sort_end)*1000.0);
2842 }
2843 }
2845 // Add the heap region to the collection set and return the conservative
2846 // estimate of the number of live bytes.
2847 void G1CollectorPolicy::
2848 add_to_collection_set(HeapRegion* hr) {
2849 if (G1PrintHeapRegions) {
2850 gclog_or_tty->print_cr("added region to cset %d:["PTR_FORMAT", "PTR_FORMAT"], "
2851 "top "PTR_FORMAT", young %s",
2852 hr->hrs_index(), hr->bottom(), hr->end(),
2853 hr->top(), (hr->is_young()) ? "YES" : "NO");
2854 }
2856 if (_g1->mark_in_progress())
2857 _g1->concurrent_mark()->registerCSetRegion(hr);
2859 assert(!hr->in_collection_set(),
2860 "should not already be in the CSet");
2861 hr->set_in_collection_set(true);
2862 hr->set_next_in_collection_set(_collection_set);
2863 _collection_set = hr;
2864 _collection_set_size++;
2865 _collection_set_bytes_used_before += hr->used();
2866 _g1->register_region_with_in_cset_fast_test(hr);
2867 }
2869 void
2870 G1CollectorPolicy_BestRegionsFirst::
2871 choose_collection_set() {
2872 double non_young_start_time_sec;
2873 start_recording_regions();
2875 guarantee(_target_pause_time_ms > -1.0
2876 NOT_PRODUCT(|| Universe::heap()->gc_cause() == GCCause::_scavenge_alot),
2877 "_target_pause_time_ms should have been set!");
2878 #ifndef PRODUCT
2879 if (_target_pause_time_ms <= -1.0) {
2880 assert(ScavengeALot && Universe::heap()->gc_cause() == GCCause::_scavenge_alot, "Error");
2881 _target_pause_time_ms = _mmu_tracker->max_gc_time() * 1000.0;
2882 }
2883 #endif
2884 assert(_collection_set == NULL, "Precondition");
2886 double base_time_ms = predict_base_elapsed_time_ms(_pending_cards);
2887 double predicted_pause_time_ms = base_time_ms;
2889 double target_time_ms = _target_pause_time_ms;
2890 double time_remaining_ms = target_time_ms - base_time_ms;
2892 // the 10% and 50% values are arbitrary...
2893 if (time_remaining_ms < 0.10*target_time_ms) {
2894 time_remaining_ms = 0.50 * target_time_ms;
2895 _within_target = false;
2896 } else {
2897 _within_target = true;
2898 }
2900 // We figure out the number of bytes available for future to-space.
2901 // For new regions without marking information, we must assume the
2902 // worst-case of complete survival. If we have marking information for a
2903 // region, we can bound the amount of live data. We can add a number of
2904 // such regions, as long as the sum of the live data bounds does not
2905 // exceed the available evacuation space.
2906 size_t max_live_bytes = _g1->free_regions() * HeapRegion::GrainBytes;
2908 size_t expansion_bytes =
2909 _g1->expansion_regions() * HeapRegion::GrainBytes;
2911 _collection_set_bytes_used_before = 0;
2912 _collection_set_size = 0;
2914 // Adjust for expansion and slop.
2915 max_live_bytes = max_live_bytes + expansion_bytes;
2917 assert(_g1->regions_accounted_for(), "Region leakage!");
2919 HeapRegion* hr;
2920 if (in_young_gc_mode()) {
2921 double young_start_time_sec = os::elapsedTime();
2923 if (G1PolicyVerbose > 0) {
2924 gclog_or_tty->print_cr("Adding %d young regions to the CSet",
2925 _g1->young_list_length());
2926 }
2927 _young_cset_length = 0;
2928 _last_young_gc_full = full_young_gcs() ? true : false;
2929 if (_last_young_gc_full)
2930 ++_full_young_pause_num;
2931 else
2932 ++_partial_young_pause_num;
2933 hr = _g1->pop_region_from_young_list();
2934 while (hr != NULL) {
2936 assert( hr->young_index_in_cset() == -1, "invariant" );
2937 assert( hr->age_in_surv_rate_group() != -1, "invariant" );
2938 hr->set_young_index_in_cset((int) _young_cset_length);
2940 ++_young_cset_length;
2941 double predicted_time_ms = predict_region_elapsed_time_ms(hr, true);
2942 time_remaining_ms -= predicted_time_ms;
2943 predicted_pause_time_ms += predicted_time_ms;
2944 assert(!hr->in_collection_set(), "invariant");
2945 add_to_collection_set(hr);
2946 record_cset_region(hr, true);
2947 max_live_bytes -= MIN2(hr->max_live_bytes(), max_live_bytes);
2948 if (G1PolicyVerbose > 0) {
2949 gclog_or_tty->print_cr(" Added [" PTR_FORMAT ", " PTR_FORMAT") to CS.",
2950 hr->bottom(), hr->end());
2951 gclog_or_tty->print_cr(" (" SIZE_FORMAT " KB left in heap.)",
2952 max_live_bytes/K);
2953 }
2954 hr = _g1->pop_region_from_young_list();
2955 }
2957 record_scan_only_regions(_g1->young_list_scan_only_length());
2959 double young_end_time_sec = os::elapsedTime();
2960 _recorded_young_cset_choice_time_ms =
2961 (young_end_time_sec - young_start_time_sec) * 1000.0;
2963 non_young_start_time_sec = os::elapsedTime();
2965 if (_young_cset_length > 0 && _last_young_gc_full) {
2966 // don't bother adding more regions...
2967 goto choose_collection_set_end;
2968 }
2969 }
2971 if (!in_young_gc_mode() || !full_young_gcs()) {
2972 bool should_continue = true;
2973 NumberSeq seq;
2974 double avg_prediction = 100000000000000000.0; // something very large
2975 do {
2976 hr = _collectionSetChooser->getNextMarkedRegion(time_remaining_ms,
2977 avg_prediction);
2978 if (hr != NULL) {
2979 double predicted_time_ms = predict_region_elapsed_time_ms(hr, false);
2980 time_remaining_ms -= predicted_time_ms;
2981 predicted_pause_time_ms += predicted_time_ms;
2982 add_to_collection_set(hr);
2983 record_cset_region(hr, false);
2984 max_live_bytes -= MIN2(hr->max_live_bytes(), max_live_bytes);
2985 if (G1PolicyVerbose > 0) {
2986 gclog_or_tty->print_cr(" (" SIZE_FORMAT " KB left in heap.)",
2987 max_live_bytes/K);
2988 }
2989 seq.add(predicted_time_ms);
2990 avg_prediction = seq.avg() + seq.sd();
2991 }
2992 should_continue =
2993 ( hr != NULL) &&
2994 ( (adaptive_young_list_length()) ? time_remaining_ms > 0.0
2995 : _collection_set_size < _young_list_fixed_length );
2996 } while (should_continue);
2998 if (!adaptive_young_list_length() &&
2999 _collection_set_size < _young_list_fixed_length)
3000 _should_revert_to_full_young_gcs = true;
3001 }
3003 choose_collection_set_end:
3004 count_CS_bytes_used();
3006 end_recording_regions();
3008 double non_young_end_time_sec = os::elapsedTime();
3009 _recorded_non_young_cset_choice_time_ms =
3010 (non_young_end_time_sec - non_young_start_time_sec) * 1000.0;
3011 }
3013 void G1CollectorPolicy_BestRegionsFirst::record_full_collection_end() {
3014 G1CollectorPolicy::record_full_collection_end();
3015 _collectionSetChooser->updateAfterFullCollection();
3016 }
3018 void G1CollectorPolicy_BestRegionsFirst::
3019 expand_if_possible(size_t numRegions) {
3020 size_t expansion_bytes = numRegions * HeapRegion::GrainBytes;
3021 _g1->expand(expansion_bytes);
3022 }
3024 void G1CollectorPolicy_BestRegionsFirst::
3025 record_collection_pause_end(bool abandoned) {
3026 G1CollectorPolicy::record_collection_pause_end(abandoned);
3027 assert(assertMarkedBytesDataOK(), "Marked regions not OK at pause end.");
3028 }