Tue, 09 Aug 2011 10:16:01 -0700
6593758: RFE: Enhance GC ergonomics to dynamically choose ParallelGCThreads
Summary: Select number of GC threads dynamically based on heap usage and number of Java threads
Reviewed-by: johnc, ysr, jcoomes
1 /*
2 * Copyright (c) 2001, 2011, Oracle and/or its affiliates. All rights reserved.
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 *
5 * This code is free software; you can redistribute it and/or modify it
6 * under the terms of the GNU General Public License version 2 only, as
7 * published by the Free Software Foundation.
8 *
9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 * version 2 for more details (a copy is included in the LICENSE file that
13 * accompanied this code).
14 *
15 * You should have received a copy of the GNU General Public License version
16 * 2 along with this work; if not, write to the Free Software Foundation,
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 *
19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 * or visit www.oracle.com if you need additional information or have any
21 * questions.
22 *
23 */
25 #include "precompiled.hpp"
26 #include "gc_implementation/g1/concurrentG1Refine.hpp"
27 #include "gc_implementation/g1/concurrentMark.hpp"
28 #include "gc_implementation/g1/concurrentMarkThread.inline.hpp"
29 #include "gc_implementation/g1/g1CollectedHeap.inline.hpp"
30 #include "gc_implementation/g1/g1CollectorPolicy.hpp"
31 #include "gc_implementation/g1/g1ErgoVerbose.hpp"
32 #include "gc_implementation/g1/heapRegionRemSet.hpp"
33 #include "gc_implementation/shared/gcPolicyCounters.hpp"
34 #include "runtime/arguments.hpp"
35 #include "runtime/java.hpp"
36 #include "runtime/mutexLocker.hpp"
37 #include "utilities/debug.hpp"
39 // Different defaults for different number of GC threads
40 // They were chosen by running GCOld and SPECjbb on debris with different
41 // numbers of GC threads and choosing them based on the results
43 // all the same
44 static double rs_length_diff_defaults[] = {
45 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0
46 };
48 static double cost_per_card_ms_defaults[] = {
49 0.01, 0.005, 0.005, 0.003, 0.003, 0.002, 0.002, 0.0015
50 };
52 // all the same
53 static double fully_young_cards_per_entry_ratio_defaults[] = {
54 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0
55 };
57 static double cost_per_entry_ms_defaults[] = {
58 0.015, 0.01, 0.01, 0.008, 0.008, 0.0055, 0.0055, 0.005
59 };
61 static double cost_per_byte_ms_defaults[] = {
62 0.00006, 0.00003, 0.00003, 0.000015, 0.000015, 0.00001, 0.00001, 0.000009
63 };
65 // these should be pretty consistent
66 static double constant_other_time_ms_defaults[] = {
67 5.0, 5.0, 5.0, 5.0, 5.0, 5.0, 5.0, 5.0
68 };
71 static double young_other_cost_per_region_ms_defaults[] = {
72 0.3, 0.2, 0.2, 0.15, 0.15, 0.12, 0.12, 0.1
73 };
75 static double non_young_other_cost_per_region_ms_defaults[] = {
76 1.0, 0.7, 0.7, 0.5, 0.5, 0.42, 0.42, 0.30
77 };
79 // Help class for avoiding interleaved logging
80 class LineBuffer: public StackObj {
82 private:
83 static const int BUFFER_LEN = 1024;
84 static const int INDENT_CHARS = 3;
85 char _buffer[BUFFER_LEN];
86 int _indent_level;
87 int _cur;
89 void vappend(const char* format, va_list ap) {
90 int res = vsnprintf(&_buffer[_cur], BUFFER_LEN - _cur, format, ap);
91 if (res != -1) {
92 _cur += res;
93 } else {
94 DEBUG_ONLY(warning("buffer too small in LineBuffer");)
95 _buffer[BUFFER_LEN -1] = 0;
96 _cur = BUFFER_LEN; // vsnprintf above should not add to _buffer if we are called again
97 }
98 }
100 public:
101 explicit LineBuffer(int indent_level): _indent_level(indent_level), _cur(0) {
102 for (; (_cur < BUFFER_LEN && _cur < (_indent_level * INDENT_CHARS)); _cur++) {
103 _buffer[_cur] = ' ';
104 }
105 }
107 #ifndef PRODUCT
108 ~LineBuffer() {
109 assert(_cur == _indent_level * INDENT_CHARS, "pending data in buffer - append_and_print_cr() not called?");
110 }
111 #endif
113 void append(const char* format, ...) {
114 va_list ap;
115 va_start(ap, format);
116 vappend(format, ap);
117 va_end(ap);
118 }
120 void append_and_print_cr(const char* format, ...) {
121 va_list ap;
122 va_start(ap, format);
123 vappend(format, ap);
124 va_end(ap);
125 gclog_or_tty->print_cr("%s", _buffer);
126 _cur = _indent_level * INDENT_CHARS;
127 }
128 };
130 G1CollectorPolicy::G1CollectorPolicy() :
131 _parallel_gc_threads(G1CollectedHeap::use_parallel_gc_threads()
132 ? ParallelGCThreads : 1),
134 _recent_gc_times_ms(new TruncatedSeq(NumPrevPausesForHeuristics)),
135 _all_pause_times_ms(new NumberSeq()),
136 _stop_world_start(0.0),
137 _all_stop_world_times_ms(new NumberSeq()),
138 _all_yield_times_ms(new NumberSeq()),
139 _using_new_ratio_calculations(false),
141 _summary(new Summary()),
143 _cur_clear_ct_time_ms(0.0),
145 _cur_ref_proc_time_ms(0.0),
146 _cur_ref_enq_time_ms(0.0),
148 #ifndef PRODUCT
149 _min_clear_cc_time_ms(-1.0),
150 _max_clear_cc_time_ms(-1.0),
151 _cur_clear_cc_time_ms(0.0),
152 _cum_clear_cc_time_ms(0.0),
153 _num_cc_clears(0L),
154 #endif
156 _aux_num(10),
157 _all_aux_times_ms(new NumberSeq[_aux_num]),
158 _cur_aux_start_times_ms(new double[_aux_num]),
159 _cur_aux_times_ms(new double[_aux_num]),
160 _cur_aux_times_set(new bool[_aux_num]),
162 _concurrent_mark_remark_times_ms(new TruncatedSeq(NumPrevPausesForHeuristics)),
163 _concurrent_mark_cleanup_times_ms(new TruncatedSeq(NumPrevPausesForHeuristics)),
165 _alloc_rate_ms_seq(new TruncatedSeq(TruncatedSeqLength)),
166 _prev_collection_pause_end_ms(0.0),
167 _pending_card_diff_seq(new TruncatedSeq(TruncatedSeqLength)),
168 _rs_length_diff_seq(new TruncatedSeq(TruncatedSeqLength)),
169 _cost_per_card_ms_seq(new TruncatedSeq(TruncatedSeqLength)),
170 _fully_young_cards_per_entry_ratio_seq(new TruncatedSeq(TruncatedSeqLength)),
171 _partially_young_cards_per_entry_ratio_seq(
172 new TruncatedSeq(TruncatedSeqLength)),
173 _cost_per_entry_ms_seq(new TruncatedSeq(TruncatedSeqLength)),
174 _partially_young_cost_per_entry_ms_seq(new TruncatedSeq(TruncatedSeqLength)),
175 _cost_per_byte_ms_seq(new TruncatedSeq(TruncatedSeqLength)),
176 _cost_per_byte_ms_during_cm_seq(new TruncatedSeq(TruncatedSeqLength)),
177 _constant_other_time_ms_seq(new TruncatedSeq(TruncatedSeqLength)),
178 _young_other_cost_per_region_ms_seq(new TruncatedSeq(TruncatedSeqLength)),
179 _non_young_other_cost_per_region_ms_seq(
180 new TruncatedSeq(TruncatedSeqLength)),
182 _pending_cards_seq(new TruncatedSeq(TruncatedSeqLength)),
183 _rs_lengths_seq(new TruncatedSeq(TruncatedSeqLength)),
185 _pause_time_target_ms((double) MaxGCPauseMillis),
187 _full_young_gcs(true),
188 _full_young_pause_num(0),
189 _partial_young_pause_num(0),
191 _during_marking(false),
192 _in_marking_window(false),
193 _in_marking_window_im(false),
195 _known_garbage_ratio(0.0),
196 _known_garbage_bytes(0),
198 _young_gc_eff_seq(new TruncatedSeq(TruncatedSeqLength)),
200 _recent_prev_end_times_for_all_gcs_sec(new TruncatedSeq(NumPrevPausesForHeuristics)),
202 _recent_avg_pause_time_ratio(0.0),
204 _all_full_gc_times_ms(new NumberSeq()),
206 _initiate_conc_mark_if_possible(false),
207 _during_initial_mark_pause(false),
208 _should_revert_to_full_young_gcs(false),
209 _last_full_young_gc(false),
211 _eden_bytes_before_gc(0),
212 _survivor_bytes_before_gc(0),
213 _capacity_before_gc(0),
215 _prev_collection_pause_used_at_end_bytes(0),
217 _eden_cset_region_length(0),
218 _survivor_cset_region_length(0),
219 _old_cset_region_length(0),
221 _collection_set(NULL),
222 _collection_set_bytes_used_before(0),
224 // Incremental CSet attributes
225 _inc_cset_build_state(Inactive),
226 _inc_cset_head(NULL),
227 _inc_cset_tail(NULL),
228 _inc_cset_bytes_used_before(0),
229 _inc_cset_max_finger(NULL),
230 _inc_cset_recorded_rs_lengths(0),
231 _inc_cset_predicted_elapsed_time_ms(0.0),
233 #ifdef _MSC_VER // the use of 'this' below gets a warning, make it go away
234 #pragma warning( disable:4355 ) // 'this' : used in base member initializer list
235 #endif // _MSC_VER
237 _short_lived_surv_rate_group(new SurvRateGroup(this, "Short Lived",
238 G1YoungSurvRateNumRegionsSummary)),
239 _survivor_surv_rate_group(new SurvRateGroup(this, "Survivor",
240 G1YoungSurvRateNumRegionsSummary)),
241 // add here any more surv rate groups
242 _recorded_survivor_regions(0),
243 _recorded_survivor_head(NULL),
244 _recorded_survivor_tail(NULL),
245 _survivors_age_table(true),
247 _gc_overhead_perc(0.0) {
249 // Set up the region size and associated fields. Given that the
250 // policy is created before the heap, we have to set this up here,
251 // so it's done as soon as possible.
252 HeapRegion::setup_heap_region_size(Arguments::min_heap_size());
253 HeapRegionRemSet::setup_remset_size();
255 G1ErgoVerbose::initialize();
256 if (PrintAdaptiveSizePolicy) {
257 // Currently, we only use a single switch for all the heuristics.
258 G1ErgoVerbose::set_enabled(true);
259 // Given that we don't currently have a verboseness level
260 // parameter, we'll hardcode this to high. This can be easily
261 // changed in the future.
262 G1ErgoVerbose::set_level(ErgoHigh);
263 } else {
264 G1ErgoVerbose::set_enabled(false);
265 }
267 // Verify PLAB sizes
268 const size_t region_size = HeapRegion::GrainWords;
269 if (YoungPLABSize > region_size || OldPLABSize > region_size) {
270 char buffer[128];
271 jio_snprintf(buffer, sizeof(buffer), "%sPLABSize should be at most "SIZE_FORMAT,
272 OldPLABSize > region_size ? "Old" : "Young", region_size);
273 vm_exit_during_initialization(buffer);
274 }
276 _recent_prev_end_times_for_all_gcs_sec->add(os::elapsedTime());
277 _prev_collection_pause_end_ms = os::elapsedTime() * 1000.0;
279 _par_last_gc_worker_start_times_ms = new double[_parallel_gc_threads];
280 _par_last_ext_root_scan_times_ms = new double[_parallel_gc_threads];
281 _par_last_mark_stack_scan_times_ms = new double[_parallel_gc_threads];
283 _par_last_update_rs_times_ms = new double[_parallel_gc_threads];
284 _par_last_update_rs_processed_buffers = new double[_parallel_gc_threads];
286 _par_last_scan_rs_times_ms = new double[_parallel_gc_threads];
288 _par_last_obj_copy_times_ms = new double[_parallel_gc_threads];
290 _par_last_termination_times_ms = new double[_parallel_gc_threads];
291 _par_last_termination_attempts = new double[_parallel_gc_threads];
292 _par_last_gc_worker_end_times_ms = new double[_parallel_gc_threads];
293 _par_last_gc_worker_times_ms = new double[_parallel_gc_threads];
294 _par_last_gc_worker_other_times_ms = new double[_parallel_gc_threads];
296 // start conservatively
297 _expensive_region_limit_ms = 0.5 * (double) MaxGCPauseMillis;
299 int index;
300 if (ParallelGCThreads == 0)
301 index = 0;
302 else if (ParallelGCThreads > 8)
303 index = 7;
304 else
305 index = ParallelGCThreads - 1;
307 _pending_card_diff_seq->add(0.0);
308 _rs_length_diff_seq->add(rs_length_diff_defaults[index]);
309 _cost_per_card_ms_seq->add(cost_per_card_ms_defaults[index]);
310 _fully_young_cards_per_entry_ratio_seq->add(
311 fully_young_cards_per_entry_ratio_defaults[index]);
312 _cost_per_entry_ms_seq->add(cost_per_entry_ms_defaults[index]);
313 _cost_per_byte_ms_seq->add(cost_per_byte_ms_defaults[index]);
314 _constant_other_time_ms_seq->add(constant_other_time_ms_defaults[index]);
315 _young_other_cost_per_region_ms_seq->add(
316 young_other_cost_per_region_ms_defaults[index]);
317 _non_young_other_cost_per_region_ms_seq->add(
318 non_young_other_cost_per_region_ms_defaults[index]);
320 // Below, we might need to calculate the pause time target based on
321 // the pause interval. When we do so we are going to give G1 maximum
322 // flexibility and allow it to do pauses when it needs to. So, we'll
323 // arrange that the pause interval to be pause time target + 1 to
324 // ensure that a) the pause time target is maximized with respect to
325 // the pause interval and b) we maintain the invariant that pause
326 // time target < pause interval. If the user does not want this
327 // maximum flexibility, they will have to set the pause interval
328 // explicitly.
330 // First make sure that, if either parameter is set, its value is
331 // reasonable.
332 if (!FLAG_IS_DEFAULT(MaxGCPauseMillis)) {
333 if (MaxGCPauseMillis < 1) {
334 vm_exit_during_initialization("MaxGCPauseMillis should be "
335 "greater than 0");
336 }
337 }
338 if (!FLAG_IS_DEFAULT(GCPauseIntervalMillis)) {
339 if (GCPauseIntervalMillis < 1) {
340 vm_exit_during_initialization("GCPauseIntervalMillis should be "
341 "greater than 0");
342 }
343 }
345 // Then, if the pause time target parameter was not set, set it to
346 // the default value.
347 if (FLAG_IS_DEFAULT(MaxGCPauseMillis)) {
348 if (FLAG_IS_DEFAULT(GCPauseIntervalMillis)) {
349 // The default pause time target in G1 is 200ms
350 FLAG_SET_DEFAULT(MaxGCPauseMillis, 200);
351 } else {
352 // We do not allow the pause interval to be set without the
353 // pause time target
354 vm_exit_during_initialization("GCPauseIntervalMillis cannot be set "
355 "without setting MaxGCPauseMillis");
356 }
357 }
359 // Then, if the interval parameter was not set, set it according to
360 // the pause time target (this will also deal with the case when the
361 // pause time target is the default value).
362 if (FLAG_IS_DEFAULT(GCPauseIntervalMillis)) {
363 FLAG_SET_DEFAULT(GCPauseIntervalMillis, MaxGCPauseMillis + 1);
364 }
366 // Finally, make sure that the two parameters are consistent.
367 if (MaxGCPauseMillis >= GCPauseIntervalMillis) {
368 char buffer[256];
369 jio_snprintf(buffer, 256,
370 "MaxGCPauseMillis (%u) should be less than "
371 "GCPauseIntervalMillis (%u)",
372 MaxGCPauseMillis, GCPauseIntervalMillis);
373 vm_exit_during_initialization(buffer);
374 }
376 double max_gc_time = (double) MaxGCPauseMillis / 1000.0;
377 double time_slice = (double) GCPauseIntervalMillis / 1000.0;
378 _mmu_tracker = new G1MMUTrackerQueue(time_slice, max_gc_time);
379 _sigma = (double) G1ConfidencePercent / 100.0;
381 // start conservatively (around 50ms is about right)
382 _concurrent_mark_remark_times_ms->add(0.05);
383 _concurrent_mark_cleanup_times_ms->add(0.20);
384 _tenuring_threshold = MaxTenuringThreshold;
385 // _max_survivor_regions will be calculated by
386 // update_young_list_target_length() during initialization.
387 _max_survivor_regions = 0;
389 assert(GCTimeRatio > 0,
390 "we should have set it to a default value set_g1_gc_flags() "
391 "if a user set it to 0");
392 _gc_overhead_perc = 100.0 * (1.0 / (1.0 + GCTimeRatio));
394 uintx reserve_perc = G1ReservePercent;
395 // Put an artificial ceiling on this so that it's not set to a silly value.
396 if (reserve_perc > 50) {
397 reserve_perc = 50;
398 warning("G1ReservePercent is set to a value that is too large, "
399 "it's been updated to %u", reserve_perc);
400 }
401 _reserve_factor = (double) reserve_perc / 100.0;
402 // This will be set when the heap is expanded
403 // for the first time during initialization.
404 _reserve_regions = 0;
406 initialize_all();
407 _collectionSetChooser = new CollectionSetChooser();
408 }
410 // Increment "i", mod "len"
411 static void inc_mod(int& i, int len) {
412 i++; if (i == len) i = 0;
413 }
415 void G1CollectorPolicy::initialize_flags() {
416 set_min_alignment(HeapRegion::GrainBytes);
417 set_max_alignment(GenRemSet::max_alignment_constraint(rem_set_name()));
418 if (SurvivorRatio < 1) {
419 vm_exit_during_initialization("Invalid survivor ratio specified");
420 }
421 CollectorPolicy::initialize_flags();
422 }
424 // The easiest way to deal with the parsing of the NewSize /
425 // MaxNewSize / etc. parameteres is to re-use the code in the
426 // TwoGenerationCollectorPolicy class. This is similar to what
427 // ParallelScavenge does with its GenerationSizer class (see
428 // ParallelScavengeHeap::initialize()). We might change this in the
429 // future, but it's a good start.
430 class G1YoungGenSizer : public TwoGenerationCollectorPolicy {
431 private:
432 size_t size_to_region_num(size_t byte_size) {
433 return MAX2((size_t) 1, byte_size / HeapRegion::GrainBytes);
434 }
436 public:
437 G1YoungGenSizer() {
438 initialize_flags();
439 initialize_size_info();
440 }
441 size_t min_young_region_num() {
442 return size_to_region_num(_min_gen0_size);
443 }
444 size_t initial_young_region_num() {
445 return size_to_region_num(_initial_gen0_size);
446 }
447 size_t max_young_region_num() {
448 return size_to_region_num(_max_gen0_size);
449 }
450 };
452 void G1CollectorPolicy::update_young_list_size_using_newratio(size_t number_of_heap_regions) {
453 assert(number_of_heap_regions > 0, "Heap must be initialized");
454 size_t young_size = number_of_heap_regions / (NewRatio + 1);
455 _min_desired_young_length = young_size;
456 _max_desired_young_length = young_size;
457 }
459 void G1CollectorPolicy::init() {
460 // Set aside an initial future to_space.
461 _g1 = G1CollectedHeap::heap();
463 assert(Heap_lock->owned_by_self(), "Locking discipline.");
465 initialize_gc_policy_counters();
467 G1YoungGenSizer sizer;
468 _min_desired_young_length = sizer.min_young_region_num();
469 _max_desired_young_length = sizer.max_young_region_num();
471 if (FLAG_IS_CMDLINE(NewRatio)) {
472 if (FLAG_IS_CMDLINE(NewSize) || FLAG_IS_CMDLINE(MaxNewSize)) {
473 warning("-XX:NewSize and -XX:MaxNewSize override -XX:NewRatio");
474 } else {
475 // Treat NewRatio as a fixed size that is only recalculated when the heap size changes
476 update_young_list_size_using_newratio(_g1->n_regions());
477 _using_new_ratio_calculations = true;
478 }
479 }
481 assert(_min_desired_young_length <= _max_desired_young_length, "Invalid min/max young gen size values");
483 set_adaptive_young_list_length(_min_desired_young_length < _max_desired_young_length);
484 if (adaptive_young_list_length()) {
485 _young_list_fixed_length = 0;
486 } else {
487 assert(_min_desired_young_length == _max_desired_young_length, "Min and max young size differ");
488 _young_list_fixed_length = _min_desired_young_length;
489 }
490 _free_regions_at_end_of_collection = _g1->free_regions();
491 update_young_list_target_length();
492 _prev_eden_capacity = _young_list_target_length * HeapRegion::GrainBytes;
494 // We may immediately start allocating regions and placing them on the
495 // collection set list. Initialize the per-collection set info
496 start_incremental_cset_building();
497 }
499 // Create the jstat counters for the policy.
500 void G1CollectorPolicy::initialize_gc_policy_counters() {
501 _gc_policy_counters = new GCPolicyCounters("GarbageFirst", 1, 3);
502 }
504 bool G1CollectorPolicy::predict_will_fit(size_t young_length,
505 double base_time_ms,
506 size_t base_free_regions,
507 double target_pause_time_ms) {
508 if (young_length >= base_free_regions) {
509 // end condition 1: not enough space for the young regions
510 return false;
511 }
513 double accum_surv_rate = accum_yg_surv_rate_pred((int)(young_length - 1));
514 size_t bytes_to_copy =
515 (size_t) (accum_surv_rate * (double) HeapRegion::GrainBytes);
516 double copy_time_ms = predict_object_copy_time_ms(bytes_to_copy);
517 double young_other_time_ms = predict_young_other_time_ms(young_length);
518 double pause_time_ms = base_time_ms + copy_time_ms + young_other_time_ms;
519 if (pause_time_ms > target_pause_time_ms) {
520 // end condition 2: prediction is over the target pause time
521 return false;
522 }
524 size_t free_bytes =
525 (base_free_regions - young_length) * HeapRegion::GrainBytes;
526 if ((2.0 * sigma()) * (double) bytes_to_copy > (double) free_bytes) {
527 // end condition 3: out-of-space (conservatively!)
528 return false;
529 }
531 // success!
532 return true;
533 }
535 void G1CollectorPolicy::record_new_heap_size(size_t new_number_of_regions) {
536 // re-calculate the necessary reserve
537 double reserve_regions_d = (double) new_number_of_regions * _reserve_factor;
538 // We use ceiling so that if reserve_regions_d is > 0.0 (but
539 // smaller than 1.0) we'll get 1.
540 _reserve_regions = (size_t) ceil(reserve_regions_d);
542 if (_using_new_ratio_calculations) {
543 // -XX:NewRatio was specified so we need to update the
544 // young gen length when the heap size has changed.
545 update_young_list_size_using_newratio(new_number_of_regions);
546 }
547 }
549 size_t G1CollectorPolicy::calculate_young_list_desired_min_length(
550 size_t base_min_length) {
551 size_t desired_min_length = 0;
552 if (adaptive_young_list_length()) {
553 if (_alloc_rate_ms_seq->num() > 3) {
554 double now_sec = os::elapsedTime();
555 double when_ms = _mmu_tracker->when_max_gc_sec(now_sec) * 1000.0;
556 double alloc_rate_ms = predict_alloc_rate_ms();
557 desired_min_length = (size_t) ceil(alloc_rate_ms * when_ms);
558 } else {
559 // otherwise we don't have enough info to make the prediction
560 }
561 }
562 desired_min_length += base_min_length;
563 // make sure we don't go below any user-defined minimum bound
564 return MAX2(_min_desired_young_length, desired_min_length);
565 }
567 size_t G1CollectorPolicy::calculate_young_list_desired_max_length() {
568 // Here, we might want to also take into account any additional
569 // constraints (i.e., user-defined minimum bound). Currently, we
570 // effectively don't set this bound.
571 return _max_desired_young_length;
572 }
574 void G1CollectorPolicy::update_young_list_target_length(size_t rs_lengths) {
575 if (rs_lengths == (size_t) -1) {
576 // if it's set to the default value (-1), we should predict it;
577 // otherwise, use the given value.
578 rs_lengths = (size_t) get_new_prediction(_rs_lengths_seq);
579 }
581 // Calculate the absolute and desired min bounds.
583 // This is how many young regions we already have (currently: the survivors).
584 size_t base_min_length = recorded_survivor_regions();
585 // This is the absolute minimum young length, which ensures that we
586 // can allocate one eden region in the worst-case.
587 size_t absolute_min_length = base_min_length + 1;
588 size_t desired_min_length =
589 calculate_young_list_desired_min_length(base_min_length);
590 if (desired_min_length < absolute_min_length) {
591 desired_min_length = absolute_min_length;
592 }
594 // Calculate the absolute and desired max bounds.
596 // We will try our best not to "eat" into the reserve.
597 size_t absolute_max_length = 0;
598 if (_free_regions_at_end_of_collection > _reserve_regions) {
599 absolute_max_length = _free_regions_at_end_of_collection - _reserve_regions;
600 }
601 size_t desired_max_length = calculate_young_list_desired_max_length();
602 if (desired_max_length > absolute_max_length) {
603 desired_max_length = absolute_max_length;
604 }
606 size_t young_list_target_length = 0;
607 if (adaptive_young_list_length()) {
608 if (full_young_gcs()) {
609 young_list_target_length =
610 calculate_young_list_target_length(rs_lengths,
611 base_min_length,
612 desired_min_length,
613 desired_max_length);
614 _rs_lengths_prediction = rs_lengths;
615 } else {
616 // Don't calculate anything and let the code below bound it to
617 // the desired_min_length, i.e., do the next GC as soon as
618 // possible to maximize how many old regions we can add to it.
619 }
620 } else {
621 if (full_young_gcs()) {
622 young_list_target_length = _young_list_fixed_length;
623 } else {
624 // A bit arbitrary: during partially-young GCs we allocate half
625 // the young regions to try to add old regions to the CSet.
626 young_list_target_length = _young_list_fixed_length / 2;
627 // We choose to accept that we might go under the desired min
628 // length given that we intentionally ask for a smaller young gen.
629 desired_min_length = absolute_min_length;
630 }
631 }
633 // Make sure we don't go over the desired max length, nor under the
634 // desired min length. In case they clash, desired_min_length wins
635 // which is why that test is second.
636 if (young_list_target_length > desired_max_length) {
637 young_list_target_length = desired_max_length;
638 }
639 if (young_list_target_length < desired_min_length) {
640 young_list_target_length = desired_min_length;
641 }
643 assert(young_list_target_length > recorded_survivor_regions(),
644 "we should be able to allocate at least one eden region");
645 assert(young_list_target_length >= absolute_min_length, "post-condition");
646 _young_list_target_length = young_list_target_length;
648 update_max_gc_locker_expansion();
649 }
651 size_t
652 G1CollectorPolicy::calculate_young_list_target_length(size_t rs_lengths,
653 size_t base_min_length,
654 size_t desired_min_length,
655 size_t desired_max_length) {
656 assert(adaptive_young_list_length(), "pre-condition");
657 assert(full_young_gcs(), "only call this for fully-young GCs");
659 // In case some edge-condition makes the desired max length too small...
660 if (desired_max_length <= desired_min_length) {
661 return desired_min_length;
662 }
664 // We'll adjust min_young_length and max_young_length not to include
665 // the already allocated young regions (i.e., so they reflect the
666 // min and max eden regions we'll allocate). The base_min_length
667 // will be reflected in the predictions by the
668 // survivor_regions_evac_time prediction.
669 assert(desired_min_length > base_min_length, "invariant");
670 size_t min_young_length = desired_min_length - base_min_length;
671 assert(desired_max_length > base_min_length, "invariant");
672 size_t max_young_length = desired_max_length - base_min_length;
674 double target_pause_time_ms = _mmu_tracker->max_gc_time() * 1000.0;
675 double survivor_regions_evac_time = predict_survivor_regions_evac_time();
676 size_t pending_cards = (size_t) get_new_prediction(_pending_cards_seq);
677 size_t adj_rs_lengths = rs_lengths + predict_rs_length_diff();
678 size_t scanned_cards = predict_young_card_num(adj_rs_lengths);
679 double base_time_ms =
680 predict_base_elapsed_time_ms(pending_cards, scanned_cards) +
681 survivor_regions_evac_time;
682 size_t available_free_regions = _free_regions_at_end_of_collection;
683 size_t base_free_regions = 0;
684 if (available_free_regions > _reserve_regions) {
685 base_free_regions = available_free_regions - _reserve_regions;
686 }
688 // Here, we will make sure that the shortest young length that
689 // makes sense fits within the target pause time.
691 if (predict_will_fit(min_young_length, base_time_ms,
692 base_free_regions, target_pause_time_ms)) {
693 // The shortest young length will fit into the target pause time;
694 // we'll now check whether the absolute maximum number of young
695 // regions will fit in the target pause time. If not, we'll do
696 // a binary search between min_young_length and max_young_length.
697 if (predict_will_fit(max_young_length, base_time_ms,
698 base_free_regions, target_pause_time_ms)) {
699 // The maximum young length will fit into the target pause time.
700 // We are done so set min young length to the maximum length (as
701 // the result is assumed to be returned in min_young_length).
702 min_young_length = max_young_length;
703 } else {
704 // The maximum possible number of young regions will not fit within
705 // the target pause time so we'll search for the optimal
706 // length. The loop invariants are:
707 //
708 // min_young_length < max_young_length
709 // min_young_length is known to fit into the target pause time
710 // max_young_length is known not to fit into the target pause time
711 //
712 // Going into the loop we know the above hold as we've just
713 // checked them. Every time around the loop we check whether
714 // the middle value between min_young_length and
715 // max_young_length fits into the target pause time. If it
716 // does, it becomes the new min. If it doesn't, it becomes
717 // the new max. This way we maintain the loop invariants.
719 assert(min_young_length < max_young_length, "invariant");
720 size_t diff = (max_young_length - min_young_length) / 2;
721 while (diff > 0) {
722 size_t young_length = min_young_length + diff;
723 if (predict_will_fit(young_length, base_time_ms,
724 base_free_regions, target_pause_time_ms)) {
725 min_young_length = young_length;
726 } else {
727 max_young_length = young_length;
728 }
729 assert(min_young_length < max_young_length, "invariant");
730 diff = (max_young_length - min_young_length) / 2;
731 }
732 // The results is min_young_length which, according to the
733 // loop invariants, should fit within the target pause time.
735 // These are the post-conditions of the binary search above:
736 assert(min_young_length < max_young_length,
737 "otherwise we should have discovered that max_young_length "
738 "fits into the pause target and not done the binary search");
739 assert(predict_will_fit(min_young_length, base_time_ms,
740 base_free_regions, target_pause_time_ms),
741 "min_young_length, the result of the binary search, should "
742 "fit into the pause target");
743 assert(!predict_will_fit(min_young_length + 1, base_time_ms,
744 base_free_regions, target_pause_time_ms),
745 "min_young_length, the result of the binary search, should be "
746 "optimal, so no larger length should fit into the pause target");
747 }
748 } else {
749 // Even the minimum length doesn't fit into the pause time
750 // target, return it as the result nevertheless.
751 }
752 return base_min_length + min_young_length;
753 }
755 double G1CollectorPolicy::predict_survivor_regions_evac_time() {
756 double survivor_regions_evac_time = 0.0;
757 for (HeapRegion * r = _recorded_survivor_head;
758 r != NULL && r != _recorded_survivor_tail->get_next_young_region();
759 r = r->get_next_young_region()) {
760 survivor_regions_evac_time += predict_region_elapsed_time_ms(r, true);
761 }
762 return survivor_regions_evac_time;
763 }
765 void G1CollectorPolicy::revise_young_list_target_length_if_necessary() {
766 guarantee( adaptive_young_list_length(), "should not call this otherwise" );
768 size_t rs_lengths = _g1->young_list()->sampled_rs_lengths();
769 if (rs_lengths > _rs_lengths_prediction) {
770 // add 10% to avoid having to recalculate often
771 size_t rs_lengths_prediction = rs_lengths * 1100 / 1000;
772 update_young_list_target_length(rs_lengths_prediction);
773 }
774 }
778 HeapWord* G1CollectorPolicy::mem_allocate_work(size_t size,
779 bool is_tlab,
780 bool* gc_overhead_limit_was_exceeded) {
781 guarantee(false, "Not using this policy feature yet.");
782 return NULL;
783 }
785 // This method controls how a collector handles one or more
786 // of its generations being fully allocated.
787 HeapWord* G1CollectorPolicy::satisfy_failed_allocation(size_t size,
788 bool is_tlab) {
789 guarantee(false, "Not using this policy feature yet.");
790 return NULL;
791 }
794 #ifndef PRODUCT
795 bool G1CollectorPolicy::verify_young_ages() {
796 HeapRegion* head = _g1->young_list()->first_region();
797 return
798 verify_young_ages(head, _short_lived_surv_rate_group);
799 // also call verify_young_ages on any additional surv rate groups
800 }
802 bool
803 G1CollectorPolicy::verify_young_ages(HeapRegion* head,
804 SurvRateGroup *surv_rate_group) {
805 guarantee( surv_rate_group != NULL, "pre-condition" );
807 const char* name = surv_rate_group->name();
808 bool ret = true;
809 int prev_age = -1;
811 for (HeapRegion* curr = head;
812 curr != NULL;
813 curr = curr->get_next_young_region()) {
814 SurvRateGroup* group = curr->surv_rate_group();
815 if (group == NULL && !curr->is_survivor()) {
816 gclog_or_tty->print_cr("## %s: encountered NULL surv_rate_group", name);
817 ret = false;
818 }
820 if (surv_rate_group == group) {
821 int age = curr->age_in_surv_rate_group();
823 if (age < 0) {
824 gclog_or_tty->print_cr("## %s: encountered negative age", name);
825 ret = false;
826 }
828 if (age <= prev_age) {
829 gclog_or_tty->print_cr("## %s: region ages are not strictly increasing "
830 "(%d, %d)", name, age, prev_age);
831 ret = false;
832 }
833 prev_age = age;
834 }
835 }
837 return ret;
838 }
839 #endif // PRODUCT
841 void G1CollectorPolicy::record_full_collection_start() {
842 _cur_collection_start_sec = os::elapsedTime();
843 // Release the future to-space so that it is available for compaction into.
844 _g1->set_full_collection();
845 }
847 void G1CollectorPolicy::record_full_collection_end() {
848 // Consider this like a collection pause for the purposes of allocation
849 // since last pause.
850 double end_sec = os::elapsedTime();
851 double full_gc_time_sec = end_sec - _cur_collection_start_sec;
852 double full_gc_time_ms = full_gc_time_sec * 1000.0;
854 _all_full_gc_times_ms->add(full_gc_time_ms);
856 update_recent_gc_times(end_sec, full_gc_time_ms);
858 _g1->clear_full_collection();
860 // "Nuke" the heuristics that control the fully/partially young GC
861 // transitions and make sure we start with fully young GCs after the
862 // Full GC.
863 set_full_young_gcs(true);
864 _last_full_young_gc = false;
865 _should_revert_to_full_young_gcs = false;
866 clear_initiate_conc_mark_if_possible();
867 clear_during_initial_mark_pause();
868 _known_garbage_bytes = 0;
869 _known_garbage_ratio = 0.0;
870 _in_marking_window = false;
871 _in_marking_window_im = false;
873 _short_lived_surv_rate_group->start_adding_regions();
874 // also call this on any additional surv rate groups
876 record_survivor_regions(0, NULL, NULL);
878 _free_regions_at_end_of_collection = _g1->free_regions();
879 // Reset survivors SurvRateGroup.
880 _survivor_surv_rate_group->reset();
881 update_young_list_target_length();
882 _collectionSetChooser->updateAfterFullCollection();
883 }
885 void G1CollectorPolicy::record_stop_world_start() {
886 _stop_world_start = os::elapsedTime();
887 }
889 void G1CollectorPolicy::record_collection_pause_start(double start_time_sec,
890 size_t start_used) {
891 if (PrintGCDetails) {
892 gclog_or_tty->stamp(PrintGCTimeStamps);
893 gclog_or_tty->print("[GC pause");
894 gclog_or_tty->print(" (%s)", full_young_gcs() ? "young" : "partial");
895 }
897 // We only need to do this here as the policy will only be applied
898 // to the GC we're about to start. so, no point is calculating this
899 // every time we calculate / recalculate the target young length.
900 update_survivors_policy();
902 assert(_g1->used() == _g1->recalculate_used(),
903 err_msg("sanity, used: "SIZE_FORMAT" recalculate_used: "SIZE_FORMAT,
904 _g1->used(), _g1->recalculate_used()));
906 double s_w_t_ms = (start_time_sec - _stop_world_start) * 1000.0;
907 _all_stop_world_times_ms->add(s_w_t_ms);
908 _stop_world_start = 0.0;
910 _cur_collection_start_sec = start_time_sec;
911 _cur_collection_pause_used_at_start_bytes = start_used;
912 _cur_collection_pause_used_regions_at_start = _g1->used_regions();
913 _pending_cards = _g1->pending_card_num();
914 _max_pending_cards = _g1->max_pending_card_num();
916 _bytes_in_collection_set_before_gc = 0;
917 _bytes_copied_during_gc = 0;
919 YoungList* young_list = _g1->young_list();
920 _eden_bytes_before_gc = young_list->eden_used_bytes();
921 _survivor_bytes_before_gc = young_list->survivor_used_bytes();
922 _capacity_before_gc = _g1->capacity();
924 #ifdef DEBUG
925 // initialise these to something well known so that we can spot
926 // if they are not set properly
928 for (int i = 0; i < _parallel_gc_threads; ++i) {
929 _par_last_gc_worker_start_times_ms[i] = -1234.0;
930 _par_last_ext_root_scan_times_ms[i] = -1234.0;
931 _par_last_mark_stack_scan_times_ms[i] = -1234.0;
932 _par_last_update_rs_times_ms[i] = -1234.0;
933 _par_last_update_rs_processed_buffers[i] = -1234.0;
934 _par_last_scan_rs_times_ms[i] = -1234.0;
935 _par_last_obj_copy_times_ms[i] = -1234.0;
936 _par_last_termination_times_ms[i] = -1234.0;
937 _par_last_termination_attempts[i] = -1234.0;
938 _par_last_gc_worker_end_times_ms[i] = -1234.0;
939 _par_last_gc_worker_times_ms[i] = -1234.0;
940 _par_last_gc_worker_other_times_ms[i] = -1234.0;
941 }
942 #endif
944 for (int i = 0; i < _aux_num; ++i) {
945 _cur_aux_times_ms[i] = 0.0;
946 _cur_aux_times_set[i] = false;
947 }
949 // These are initialized to zero here and they are set during
950 // the evacuation pause if marking is in progress.
951 _cur_satb_drain_time_ms = 0.0;
952 _last_satb_drain_processed_buffers = 0;
954 _last_young_gc_full = false;
956 // do that for any other surv rate groups
957 _short_lived_surv_rate_group->stop_adding_regions();
958 _survivors_age_table.clear();
960 assert( verify_young_ages(), "region age verification" );
961 }
963 void G1CollectorPolicy::record_mark_closure_time(double mark_closure_time_ms) {
964 _mark_closure_time_ms = mark_closure_time_ms;
965 }
967 void G1CollectorPolicy::record_concurrent_mark_init_end(double
968 mark_init_elapsed_time_ms) {
969 _during_marking = true;
970 assert(!initiate_conc_mark_if_possible(), "we should have cleared it by now");
971 clear_during_initial_mark_pause();
972 _cur_mark_stop_world_time_ms = mark_init_elapsed_time_ms;
973 }
975 void G1CollectorPolicy::record_concurrent_mark_remark_start() {
976 _mark_remark_start_sec = os::elapsedTime();
977 _during_marking = false;
978 }
980 void G1CollectorPolicy::record_concurrent_mark_remark_end() {
981 double end_time_sec = os::elapsedTime();
982 double elapsed_time_ms = (end_time_sec - _mark_remark_start_sec)*1000.0;
983 _concurrent_mark_remark_times_ms->add(elapsed_time_ms);
984 _cur_mark_stop_world_time_ms += elapsed_time_ms;
985 _prev_collection_pause_end_ms += elapsed_time_ms;
987 _mmu_tracker->add_pause(_mark_remark_start_sec, end_time_sec, true);
988 }
990 void G1CollectorPolicy::record_concurrent_mark_cleanup_start() {
991 _mark_cleanup_start_sec = os::elapsedTime();
992 }
994 void G1CollectorPolicy::record_concurrent_mark_cleanup_completed() {
995 _should_revert_to_full_young_gcs = false;
996 _last_full_young_gc = true;
997 _in_marking_window = false;
998 }
1000 void G1CollectorPolicy::record_concurrent_pause() {
1001 if (_stop_world_start > 0.0) {
1002 double yield_ms = (os::elapsedTime() - _stop_world_start) * 1000.0;
1003 _all_yield_times_ms->add(yield_ms);
1004 }
1005 }
1007 void G1CollectorPolicy::record_concurrent_pause_end() {
1008 }
1010 template<class T>
1011 T sum_of(T* sum_arr, int start, int n, int N) {
1012 T sum = (T)0;
1013 for (int i = 0; i < n; i++) {
1014 int j = (start + i) % N;
1015 sum += sum_arr[j];
1016 }
1017 return sum;
1018 }
1020 void G1CollectorPolicy::print_par_stats(int level,
1021 const char* str,
1022 double* data) {
1023 double min = data[0], max = data[0];
1024 double total = 0.0;
1025 LineBuffer buf(level);
1026 buf.append("[%s (ms):", str);
1027 for (uint i = 0; i < no_of_gc_threads(); ++i) {
1028 double val = data[i];
1029 if (val < min)
1030 min = val;
1031 if (val > max)
1032 max = val;
1033 total += val;
1034 buf.append(" %3.1lf", val);
1035 }
1036 buf.append_and_print_cr("");
1037 double avg = total / (double) no_of_gc_threads();
1038 buf.append_and_print_cr(" Avg: %5.1lf, Min: %5.1lf, Max: %5.1lf, Diff: %5.1lf]",
1039 avg, min, max, max - min);
1040 }
1042 void G1CollectorPolicy::print_par_sizes(int level,
1043 const char* str,
1044 double* data) {
1045 double min = data[0], max = data[0];
1046 double total = 0.0;
1047 LineBuffer buf(level);
1048 buf.append("[%s :", str);
1049 for (uint i = 0; i < no_of_gc_threads(); ++i) {
1050 double val = data[i];
1051 if (val < min)
1052 min = val;
1053 if (val > max)
1054 max = val;
1055 total += val;
1056 buf.append(" %d", (int) val);
1057 }
1058 buf.append_and_print_cr("");
1059 double avg = total / (double) no_of_gc_threads();
1060 buf.append_and_print_cr(" Sum: %d, Avg: %d, Min: %d, Max: %d, Diff: %d]",
1061 (int)total, (int)avg, (int)min, (int)max, (int)max - (int)min);
1062 }
1064 void G1CollectorPolicy::print_stats(int level,
1065 const char* str,
1066 double value) {
1067 LineBuffer(level).append_and_print_cr("[%s: %5.1lf ms]", str, value);
1068 }
1070 void G1CollectorPolicy::print_stats(int level,
1071 const char* str,
1072 int value) {
1073 LineBuffer(level).append_and_print_cr("[%s: %d]", str, value);
1074 }
1076 double G1CollectorPolicy::avg_value(double* data) {
1077 if (G1CollectedHeap::use_parallel_gc_threads()) {
1078 double ret = 0.0;
1079 for (uint i = 0; i < no_of_gc_threads(); ++i) {
1080 ret += data[i];
1081 }
1082 return ret / (double) no_of_gc_threads();
1083 } else {
1084 return data[0];
1085 }
1086 }
1088 double G1CollectorPolicy::max_value(double* data) {
1089 if (G1CollectedHeap::use_parallel_gc_threads()) {
1090 double ret = data[0];
1091 for (uint i = 1; i < no_of_gc_threads(); ++i) {
1092 if (data[i] > ret) {
1093 ret = data[i];
1094 }
1095 }
1096 return ret;
1097 } else {
1098 return data[0];
1099 }
1100 }
1102 double G1CollectorPolicy::sum_of_values(double* data) {
1103 if (G1CollectedHeap::use_parallel_gc_threads()) {
1104 double sum = 0.0;
1105 for (uint i = 0; i < no_of_gc_threads(); i++) {
1106 sum += data[i];
1107 }
1108 return sum;
1109 } else {
1110 return data[0];
1111 }
1112 }
1114 double G1CollectorPolicy::max_sum(double* data1, double* data2) {
1115 double ret = data1[0] + data2[0];
1117 if (G1CollectedHeap::use_parallel_gc_threads()) {
1118 for (uint i = 1; i < no_of_gc_threads(); ++i) {
1119 double data = data1[i] + data2[i];
1120 if (data > ret) {
1121 ret = data;
1122 }
1123 }
1124 }
1125 return ret;
1126 }
1128 // Anything below that is considered to be zero
1129 #define MIN_TIMER_GRANULARITY 0.0000001
1131 void G1CollectorPolicy::record_collection_pause_end(int no_of_gc_threads) {
1132 double end_time_sec = os::elapsedTime();
1133 double elapsed_ms = _last_pause_time_ms;
1134 bool parallel = G1CollectedHeap::use_parallel_gc_threads();
1135 assert(_cur_collection_pause_used_regions_at_start >= cset_region_length(),
1136 "otherwise, the subtraction below does not make sense");
1137 size_t rs_size =
1138 _cur_collection_pause_used_regions_at_start - cset_region_length();
1139 size_t cur_used_bytes = _g1->used();
1140 assert(cur_used_bytes == _g1->recalculate_used(), "It should!");
1141 bool last_pause_included_initial_mark = false;
1142 bool update_stats = !_g1->evacuation_failed();
1143 set_no_of_gc_threads(no_of_gc_threads);
1145 #ifndef PRODUCT
1146 if (G1YoungSurvRateVerbose) {
1147 gclog_or_tty->print_cr("");
1148 _short_lived_surv_rate_group->print();
1149 // do that for any other surv rate groups too
1150 }
1151 #endif // PRODUCT
1153 last_pause_included_initial_mark = during_initial_mark_pause();
1154 if (last_pause_included_initial_mark)
1155 record_concurrent_mark_init_end(0.0);
1157 size_t marking_initiating_used_threshold =
1158 (_g1->capacity() / 100) * InitiatingHeapOccupancyPercent;
1160 if (!_g1->mark_in_progress() && !_last_full_young_gc) {
1161 assert(!last_pause_included_initial_mark, "invariant");
1162 if (cur_used_bytes > marking_initiating_used_threshold) {
1163 if (cur_used_bytes > _prev_collection_pause_used_at_end_bytes) {
1164 assert(!during_initial_mark_pause(), "we should not see this here");
1166 ergo_verbose3(ErgoConcCycles,
1167 "request concurrent cycle initiation",
1168 ergo_format_reason("occupancy higher than threshold")
1169 ergo_format_byte("occupancy")
1170 ergo_format_byte_perc("threshold"),
1171 cur_used_bytes,
1172 marking_initiating_used_threshold,
1173 (double) InitiatingHeapOccupancyPercent);
1175 // Note: this might have already been set, if during the last
1176 // pause we decided to start a cycle but at the beginning of
1177 // this pause we decided to postpone it. That's OK.
1178 set_initiate_conc_mark_if_possible();
1179 } else {
1180 ergo_verbose2(ErgoConcCycles,
1181 "do not request concurrent cycle initiation",
1182 ergo_format_reason("occupancy lower than previous occupancy")
1183 ergo_format_byte("occupancy")
1184 ergo_format_byte("previous occupancy"),
1185 cur_used_bytes,
1186 _prev_collection_pause_used_at_end_bytes);
1187 }
1188 }
1189 }
1191 _prev_collection_pause_used_at_end_bytes = cur_used_bytes;
1193 _mmu_tracker->add_pause(end_time_sec - elapsed_ms/1000.0,
1194 end_time_sec, false);
1196 // This assert is exempted when we're doing parallel collection pauses,
1197 // because the fragmentation caused by the parallel GC allocation buffers
1198 // can lead to more memory being used during collection than was used
1199 // before. Best leave this out until the fragmentation problem is fixed.
1200 // Pauses in which evacuation failed can also lead to negative
1201 // collections, since no space is reclaimed from a region containing an
1202 // object whose evacuation failed.
1203 // Further, we're now always doing parallel collection. But I'm still
1204 // leaving this here as a placeholder for a more precise assertion later.
1205 // (DLD, 10/05.)
1206 assert((true || parallel) // Always using GC LABs now.
1207 || _g1->evacuation_failed()
1208 || _cur_collection_pause_used_at_start_bytes >= cur_used_bytes,
1209 "Negative collection");
1211 size_t freed_bytes =
1212 _cur_collection_pause_used_at_start_bytes - cur_used_bytes;
1213 size_t surviving_bytes = _collection_set_bytes_used_before - freed_bytes;
1215 double survival_fraction =
1216 (double)surviving_bytes/
1217 (double)_collection_set_bytes_used_before;
1219 // These values are used to update the summary information that is
1220 // displayed when TraceGen0Time is enabled, and are output as part
1221 // of the PrintGCDetails output, in the non-parallel case.
1223 double ext_root_scan_time = avg_value(_par_last_ext_root_scan_times_ms);
1224 double mark_stack_scan_time = avg_value(_par_last_mark_stack_scan_times_ms);
1225 double update_rs_time = avg_value(_par_last_update_rs_times_ms);
1226 double update_rs_processed_buffers =
1227 sum_of_values(_par_last_update_rs_processed_buffers);
1228 double scan_rs_time = avg_value(_par_last_scan_rs_times_ms);
1229 double obj_copy_time = avg_value(_par_last_obj_copy_times_ms);
1230 double termination_time = avg_value(_par_last_termination_times_ms);
1232 double known_time = ext_root_scan_time +
1233 mark_stack_scan_time +
1234 update_rs_time +
1235 scan_rs_time +
1236 obj_copy_time;
1238 double other_time_ms = elapsed_ms;
1240 // Subtract the SATB drain time. It's initialized to zero at the
1241 // start of the pause and is updated during the pause if marking
1242 // is in progress.
1243 other_time_ms -= _cur_satb_drain_time_ms;
1245 if (parallel) {
1246 other_time_ms -= _cur_collection_par_time_ms;
1247 } else {
1248 other_time_ms -= known_time;
1249 }
1251 // Subtract the time taken to clean the card table from the
1252 // current value of "other time"
1253 other_time_ms -= _cur_clear_ct_time_ms;
1255 // TraceGen0Time and TraceGen1Time summary info updating.
1256 _all_pause_times_ms->add(elapsed_ms);
1258 if (update_stats) {
1259 _summary->record_total_time_ms(elapsed_ms);
1260 _summary->record_other_time_ms(other_time_ms);
1262 MainBodySummary* body_summary = _summary->main_body_summary();
1263 assert(body_summary != NULL, "should not be null!");
1265 // This will be non-zero iff marking is currently in progress (i.e.
1266 // _g1->mark_in_progress() == true) and the currrent pause was not
1267 // an initial mark pause. Since the body_summary items are NumberSeqs,
1268 // however, they have to be consistent and updated in lock-step with
1269 // each other. Therefore we unconditionally record the SATB drain
1270 // time - even if it's zero.
1271 body_summary->record_satb_drain_time_ms(_cur_satb_drain_time_ms);
1273 body_summary->record_ext_root_scan_time_ms(ext_root_scan_time);
1274 body_summary->record_mark_stack_scan_time_ms(mark_stack_scan_time);
1275 body_summary->record_update_rs_time_ms(update_rs_time);
1276 body_summary->record_scan_rs_time_ms(scan_rs_time);
1277 body_summary->record_obj_copy_time_ms(obj_copy_time);
1279 if (parallel) {
1280 body_summary->record_parallel_time_ms(_cur_collection_par_time_ms);
1281 body_summary->record_termination_time_ms(termination_time);
1283 double parallel_known_time = known_time + termination_time;
1284 double parallel_other_time = _cur_collection_par_time_ms - parallel_known_time;
1285 body_summary->record_parallel_other_time_ms(parallel_other_time);
1286 }
1288 body_summary->record_mark_closure_time_ms(_mark_closure_time_ms);
1289 body_summary->record_clear_ct_time_ms(_cur_clear_ct_time_ms);
1291 // We exempt parallel collection from this check because Alloc Buffer
1292 // fragmentation can produce negative collections. Same with evac
1293 // failure.
1294 // Further, we're now always doing parallel collection. But I'm still
1295 // leaving this here as a placeholder for a more precise assertion later.
1296 // (DLD, 10/05.
1297 assert((true || parallel)
1298 || _g1->evacuation_failed()
1299 || surviving_bytes <= _collection_set_bytes_used_before,
1300 "Or else negative collection!");
1302 // this is where we update the allocation rate of the application
1303 double app_time_ms =
1304 (_cur_collection_start_sec * 1000.0 - _prev_collection_pause_end_ms);
1305 if (app_time_ms < MIN_TIMER_GRANULARITY) {
1306 // This usually happens due to the timer not having the required
1307 // granularity. Some Linuxes are the usual culprits.
1308 // We'll just set it to something (arbitrarily) small.
1309 app_time_ms = 1.0;
1310 }
1311 // We maintain the invariant that all objects allocated by mutator
1312 // threads will be allocated out of eden regions. So, we can use
1313 // the eden region number allocated since the previous GC to
1314 // calculate the application's allocate rate. The only exception
1315 // to that is humongous objects that are allocated separately. But
1316 // given that humongous object allocations do not really affect
1317 // either the pause's duration nor when the next pause will take
1318 // place we can safely ignore them here.
1319 size_t regions_allocated = eden_cset_region_length();
1320 double alloc_rate_ms = (double) regions_allocated / app_time_ms;
1321 _alloc_rate_ms_seq->add(alloc_rate_ms);
1323 double interval_ms =
1324 (end_time_sec - _recent_prev_end_times_for_all_gcs_sec->oldest()) * 1000.0;
1325 update_recent_gc_times(end_time_sec, elapsed_ms);
1326 _recent_avg_pause_time_ratio = _recent_gc_times_ms->sum()/interval_ms;
1327 if (recent_avg_pause_time_ratio() < 0.0 ||
1328 (recent_avg_pause_time_ratio() - 1.0 > 0.0)) {
1329 #ifndef PRODUCT
1330 // Dump info to allow post-facto debugging
1331 gclog_or_tty->print_cr("recent_avg_pause_time_ratio() out of bounds");
1332 gclog_or_tty->print_cr("-------------------------------------------");
1333 gclog_or_tty->print_cr("Recent GC Times (ms):");
1334 _recent_gc_times_ms->dump();
1335 gclog_or_tty->print_cr("(End Time=%3.3f) Recent GC End Times (s):", end_time_sec);
1336 _recent_prev_end_times_for_all_gcs_sec->dump();
1337 gclog_or_tty->print_cr("GC = %3.3f, Interval = %3.3f, Ratio = %3.3f",
1338 _recent_gc_times_ms->sum(), interval_ms, recent_avg_pause_time_ratio());
1339 // In debug mode, terminate the JVM if the user wants to debug at this point.
1340 assert(!G1FailOnFPError, "Debugging data for CR 6898948 has been dumped above");
1341 #endif // !PRODUCT
1342 // Clip ratio between 0.0 and 1.0, and continue. This will be fixed in
1343 // CR 6902692 by redoing the manner in which the ratio is incrementally computed.
1344 if (_recent_avg_pause_time_ratio < 0.0) {
1345 _recent_avg_pause_time_ratio = 0.0;
1346 } else {
1347 assert(_recent_avg_pause_time_ratio - 1.0 > 0.0, "Ctl-point invariant");
1348 _recent_avg_pause_time_ratio = 1.0;
1349 }
1350 }
1351 }
1353 for (int i = 0; i < _aux_num; ++i) {
1354 if (_cur_aux_times_set[i]) {
1355 _all_aux_times_ms[i].add(_cur_aux_times_ms[i]);
1356 }
1357 }
1359 // PrintGCDetails output
1360 if (PrintGCDetails) {
1361 bool print_marking_info =
1362 _g1->mark_in_progress() && !last_pause_included_initial_mark;
1364 gclog_or_tty->print_cr("%s, %1.8lf secs]",
1365 (last_pause_included_initial_mark) ? " (initial-mark)" : "",
1366 elapsed_ms / 1000.0);
1368 if (print_marking_info) {
1369 print_stats(1, "SATB Drain Time", _cur_satb_drain_time_ms);
1370 print_stats(2, "Processed Buffers", _last_satb_drain_processed_buffers);
1371 }
1373 if (parallel) {
1374 print_stats(1, "Parallel Time", _cur_collection_par_time_ms);
1375 print_par_stats(2, "GC Worker Start", _par_last_gc_worker_start_times_ms);
1376 print_par_stats(2, "Ext Root Scanning", _par_last_ext_root_scan_times_ms);
1377 if (print_marking_info) {
1378 print_par_stats(2, "Mark Stack Scanning", _par_last_mark_stack_scan_times_ms);
1379 }
1380 print_par_stats(2, "Update RS", _par_last_update_rs_times_ms);
1381 print_par_sizes(3, "Processed Buffers", _par_last_update_rs_processed_buffers);
1382 print_par_stats(2, "Scan RS", _par_last_scan_rs_times_ms);
1383 print_par_stats(2, "Object Copy", _par_last_obj_copy_times_ms);
1384 print_par_stats(2, "Termination", _par_last_termination_times_ms);
1385 print_par_sizes(3, "Termination Attempts", _par_last_termination_attempts);
1386 print_par_stats(2, "GC Worker End", _par_last_gc_worker_end_times_ms);
1388 for (int i = 0; i < _parallel_gc_threads; i++) {
1389 _par_last_gc_worker_times_ms[i] = _par_last_gc_worker_end_times_ms[i] - _par_last_gc_worker_start_times_ms[i];
1391 double worker_known_time = _par_last_ext_root_scan_times_ms[i] +
1392 _par_last_mark_stack_scan_times_ms[i] +
1393 _par_last_update_rs_times_ms[i] +
1394 _par_last_scan_rs_times_ms[i] +
1395 _par_last_obj_copy_times_ms[i] +
1396 _par_last_termination_times_ms[i];
1398 _par_last_gc_worker_other_times_ms[i] = _cur_collection_par_time_ms - worker_known_time;
1399 }
1400 print_par_stats(2, "GC Worker", _par_last_gc_worker_times_ms);
1401 print_par_stats(2, "GC Worker Other", _par_last_gc_worker_other_times_ms);
1402 } else {
1403 print_stats(1, "Ext Root Scanning", ext_root_scan_time);
1404 if (print_marking_info) {
1405 print_stats(1, "Mark Stack Scanning", mark_stack_scan_time);
1406 }
1407 print_stats(1, "Update RS", update_rs_time);
1408 print_stats(2, "Processed Buffers", (int)update_rs_processed_buffers);
1409 print_stats(1, "Scan RS", scan_rs_time);
1410 print_stats(1, "Object Copying", obj_copy_time);
1411 }
1412 print_stats(1, "Clear CT", _cur_clear_ct_time_ms);
1413 #ifndef PRODUCT
1414 print_stats(1, "Cur Clear CC", _cur_clear_cc_time_ms);
1415 print_stats(1, "Cum Clear CC", _cum_clear_cc_time_ms);
1416 print_stats(1, "Min Clear CC", _min_clear_cc_time_ms);
1417 print_stats(1, "Max Clear CC", _max_clear_cc_time_ms);
1418 if (_num_cc_clears > 0) {
1419 print_stats(1, "Avg Clear CC", _cum_clear_cc_time_ms / ((double)_num_cc_clears));
1420 }
1421 #endif
1422 print_stats(1, "Other", other_time_ms);
1423 print_stats(2, "Choose CSet", _recorded_young_cset_choice_time_ms);
1424 print_stats(2, "Ref Proc", _cur_ref_proc_time_ms);
1425 print_stats(2, "Ref Enq", _cur_ref_enq_time_ms);
1427 for (int i = 0; i < _aux_num; ++i) {
1428 if (_cur_aux_times_set[i]) {
1429 char buffer[96];
1430 sprintf(buffer, "Aux%d", i);
1431 print_stats(1, buffer, _cur_aux_times_ms[i]);
1432 }
1433 }
1434 }
1436 // Update the efficiency-since-mark vars.
1437 double proc_ms = elapsed_ms * (double) _parallel_gc_threads;
1438 if (elapsed_ms < MIN_TIMER_GRANULARITY) {
1439 // This usually happens due to the timer not having the required
1440 // granularity. Some Linuxes are the usual culprits.
1441 // We'll just set it to something (arbitrarily) small.
1442 proc_ms = 1.0;
1443 }
1444 double cur_efficiency = (double) freed_bytes / proc_ms;
1446 bool new_in_marking_window = _in_marking_window;
1447 bool new_in_marking_window_im = false;
1448 if (during_initial_mark_pause()) {
1449 new_in_marking_window = true;
1450 new_in_marking_window_im = true;
1451 }
1453 if (_last_full_young_gc) {
1454 if (!last_pause_included_initial_mark) {
1455 ergo_verbose2(ErgoPartiallyYoungGCs,
1456 "start partially-young GCs",
1457 ergo_format_byte_perc("known garbage"),
1458 _known_garbage_bytes, _known_garbage_ratio * 100.0);
1459 set_full_young_gcs(false);
1460 } else {
1461 ergo_verbose0(ErgoPartiallyYoungGCs,
1462 "do not start partially-young GCs",
1463 ergo_format_reason("concurrent cycle is about to start"));
1464 }
1465 _last_full_young_gc = false;
1466 }
1468 if ( !_last_young_gc_full ) {
1469 if (_should_revert_to_full_young_gcs) {
1470 ergo_verbose2(ErgoPartiallyYoungGCs,
1471 "end partially-young GCs",
1472 ergo_format_reason("partially-young GCs end requested")
1473 ergo_format_byte_perc("known garbage"),
1474 _known_garbage_bytes, _known_garbage_ratio * 100.0);
1475 set_full_young_gcs(true);
1476 } else if (_known_garbage_ratio < 0.05) {
1477 ergo_verbose3(ErgoPartiallyYoungGCs,
1478 "end partially-young GCs",
1479 ergo_format_reason("known garbage percent lower than threshold")
1480 ergo_format_byte_perc("known garbage")
1481 ergo_format_perc("threshold"),
1482 _known_garbage_bytes, _known_garbage_ratio * 100.0,
1483 0.05 * 100.0);
1484 set_full_young_gcs(true);
1485 } else if (adaptive_young_list_length() &&
1486 (get_gc_eff_factor() * cur_efficiency < predict_young_gc_eff())) {
1487 ergo_verbose5(ErgoPartiallyYoungGCs,
1488 "end partially-young GCs",
1489 ergo_format_reason("current GC efficiency lower than "
1490 "predicted fully-young GC efficiency")
1491 ergo_format_double("GC efficiency factor")
1492 ergo_format_double("current GC efficiency")
1493 ergo_format_double("predicted fully-young GC efficiency")
1494 ergo_format_byte_perc("known garbage"),
1495 get_gc_eff_factor(), cur_efficiency,
1496 predict_young_gc_eff(),
1497 _known_garbage_bytes, _known_garbage_ratio * 100.0);
1498 set_full_young_gcs(true);
1499 }
1500 }
1501 _should_revert_to_full_young_gcs = false;
1503 if (_last_young_gc_full && !_during_marking) {
1504 _young_gc_eff_seq->add(cur_efficiency);
1505 }
1507 _short_lived_surv_rate_group->start_adding_regions();
1508 // do that for any other surv rate groupsx
1510 if (update_stats) {
1511 double pause_time_ms = elapsed_ms;
1513 size_t diff = 0;
1514 if (_max_pending_cards >= _pending_cards)
1515 diff = _max_pending_cards - _pending_cards;
1516 _pending_card_diff_seq->add((double) diff);
1518 double cost_per_card_ms = 0.0;
1519 if (_pending_cards > 0) {
1520 cost_per_card_ms = update_rs_time / (double) _pending_cards;
1521 _cost_per_card_ms_seq->add(cost_per_card_ms);
1522 }
1524 size_t cards_scanned = _g1->cards_scanned();
1526 double cost_per_entry_ms = 0.0;
1527 if (cards_scanned > 10) {
1528 cost_per_entry_ms = scan_rs_time / (double) cards_scanned;
1529 if (_last_young_gc_full)
1530 _cost_per_entry_ms_seq->add(cost_per_entry_ms);
1531 else
1532 _partially_young_cost_per_entry_ms_seq->add(cost_per_entry_ms);
1533 }
1535 if (_max_rs_lengths > 0) {
1536 double cards_per_entry_ratio =
1537 (double) cards_scanned / (double) _max_rs_lengths;
1538 if (_last_young_gc_full)
1539 _fully_young_cards_per_entry_ratio_seq->add(cards_per_entry_ratio);
1540 else
1541 _partially_young_cards_per_entry_ratio_seq->add(cards_per_entry_ratio);
1542 }
1544 size_t rs_length_diff = _max_rs_lengths - _recorded_rs_lengths;
1545 if (rs_length_diff >= 0)
1546 _rs_length_diff_seq->add((double) rs_length_diff);
1548 size_t copied_bytes = surviving_bytes;
1549 double cost_per_byte_ms = 0.0;
1550 if (copied_bytes > 0) {
1551 cost_per_byte_ms = obj_copy_time / (double) copied_bytes;
1552 if (_in_marking_window)
1553 _cost_per_byte_ms_during_cm_seq->add(cost_per_byte_ms);
1554 else
1555 _cost_per_byte_ms_seq->add(cost_per_byte_ms);
1556 }
1558 double all_other_time_ms = pause_time_ms -
1559 (update_rs_time + scan_rs_time + obj_copy_time +
1560 _mark_closure_time_ms + termination_time);
1562 double young_other_time_ms = 0.0;
1563 if (young_cset_region_length() > 0) {
1564 young_other_time_ms =
1565 _recorded_young_cset_choice_time_ms +
1566 _recorded_young_free_cset_time_ms;
1567 _young_other_cost_per_region_ms_seq->add(young_other_time_ms /
1568 (double) young_cset_region_length());
1569 }
1570 double non_young_other_time_ms = 0.0;
1571 if (old_cset_region_length() > 0) {
1572 non_young_other_time_ms =
1573 _recorded_non_young_cset_choice_time_ms +
1574 _recorded_non_young_free_cset_time_ms;
1576 _non_young_other_cost_per_region_ms_seq->add(non_young_other_time_ms /
1577 (double) old_cset_region_length());
1578 }
1580 double constant_other_time_ms = all_other_time_ms -
1581 (young_other_time_ms + non_young_other_time_ms);
1582 _constant_other_time_ms_seq->add(constant_other_time_ms);
1584 double survival_ratio = 0.0;
1585 if (_bytes_in_collection_set_before_gc > 0) {
1586 survival_ratio = (double) _bytes_copied_during_gc /
1587 (double) _bytes_in_collection_set_before_gc;
1588 }
1590 _pending_cards_seq->add((double) _pending_cards);
1591 _rs_lengths_seq->add((double) _max_rs_lengths);
1593 double expensive_region_limit_ms =
1594 (double) MaxGCPauseMillis - predict_constant_other_time_ms();
1595 if (expensive_region_limit_ms < 0.0) {
1596 // this means that the other time was predicted to be longer than
1597 // than the max pause time
1598 expensive_region_limit_ms = (double) MaxGCPauseMillis;
1599 }
1600 _expensive_region_limit_ms = expensive_region_limit_ms;
1601 }
1603 _in_marking_window = new_in_marking_window;
1604 _in_marking_window_im = new_in_marking_window_im;
1605 _free_regions_at_end_of_collection = _g1->free_regions();
1606 update_young_list_target_length();
1608 // Note that _mmu_tracker->max_gc_time() returns the time in seconds.
1609 double update_rs_time_goal_ms = _mmu_tracker->max_gc_time() * MILLIUNITS * G1RSetUpdatingPauseTimePercent / 100.0;
1610 adjust_concurrent_refinement(update_rs_time, update_rs_processed_buffers, update_rs_time_goal_ms);
1612 assert(assertMarkedBytesDataOK(), "Marked regions not OK at pause end.");
1613 }
1615 #define EXT_SIZE_FORMAT "%d%s"
1616 #define EXT_SIZE_PARAMS(bytes) \
1617 byte_size_in_proper_unit((bytes)), \
1618 proper_unit_for_byte_size((bytes))
1620 void G1CollectorPolicy::print_heap_transition() {
1621 if (PrintGCDetails) {
1622 YoungList* young_list = _g1->young_list();
1623 size_t eden_bytes = young_list->eden_used_bytes();
1624 size_t survivor_bytes = young_list->survivor_used_bytes();
1625 size_t used_before_gc = _cur_collection_pause_used_at_start_bytes;
1626 size_t used = _g1->used();
1627 size_t capacity = _g1->capacity();
1628 size_t eden_capacity =
1629 (_young_list_target_length * HeapRegion::GrainBytes) - survivor_bytes;
1631 gclog_or_tty->print_cr(
1632 " [Eden: "EXT_SIZE_FORMAT"("EXT_SIZE_FORMAT")->"EXT_SIZE_FORMAT"("EXT_SIZE_FORMAT") "
1633 "Survivors: "EXT_SIZE_FORMAT"->"EXT_SIZE_FORMAT" "
1634 "Heap: "EXT_SIZE_FORMAT"("EXT_SIZE_FORMAT")->"
1635 EXT_SIZE_FORMAT"("EXT_SIZE_FORMAT")]",
1636 EXT_SIZE_PARAMS(_eden_bytes_before_gc),
1637 EXT_SIZE_PARAMS(_prev_eden_capacity),
1638 EXT_SIZE_PARAMS(eden_bytes),
1639 EXT_SIZE_PARAMS(eden_capacity),
1640 EXT_SIZE_PARAMS(_survivor_bytes_before_gc),
1641 EXT_SIZE_PARAMS(survivor_bytes),
1642 EXT_SIZE_PARAMS(used_before_gc),
1643 EXT_SIZE_PARAMS(_capacity_before_gc),
1644 EXT_SIZE_PARAMS(used),
1645 EXT_SIZE_PARAMS(capacity));
1647 _prev_eden_capacity = eden_capacity;
1648 } else if (PrintGC) {
1649 _g1->print_size_transition(gclog_or_tty,
1650 _cur_collection_pause_used_at_start_bytes,
1651 _g1->used(), _g1->capacity());
1652 }
1653 }
1655 void G1CollectorPolicy::adjust_concurrent_refinement(double update_rs_time,
1656 double update_rs_processed_buffers,
1657 double goal_ms) {
1658 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set();
1659 ConcurrentG1Refine *cg1r = G1CollectedHeap::heap()->concurrent_g1_refine();
1661 if (G1UseAdaptiveConcRefinement) {
1662 const int k_gy = 3, k_gr = 6;
1663 const double inc_k = 1.1, dec_k = 0.9;
1665 int g = cg1r->green_zone();
1666 if (update_rs_time > goal_ms) {
1667 g = (int)(g * dec_k); // Can become 0, that's OK. That would mean a mutator-only processing.
1668 } else {
1669 if (update_rs_time < goal_ms && update_rs_processed_buffers > g) {
1670 g = (int)MAX2(g * inc_k, g + 1.0);
1671 }
1672 }
1673 // Change the refinement threads params
1674 cg1r->set_green_zone(g);
1675 cg1r->set_yellow_zone(g * k_gy);
1676 cg1r->set_red_zone(g * k_gr);
1677 cg1r->reinitialize_threads();
1679 int processing_threshold_delta = MAX2((int)(cg1r->green_zone() * sigma()), 1);
1680 int processing_threshold = MIN2(cg1r->green_zone() + processing_threshold_delta,
1681 cg1r->yellow_zone());
1682 // Change the barrier params
1683 dcqs.set_process_completed_threshold(processing_threshold);
1684 dcqs.set_max_completed_queue(cg1r->red_zone());
1685 }
1687 int curr_queue_size = dcqs.completed_buffers_num();
1688 if (curr_queue_size >= cg1r->yellow_zone()) {
1689 dcqs.set_completed_queue_padding(curr_queue_size);
1690 } else {
1691 dcqs.set_completed_queue_padding(0);
1692 }
1693 dcqs.notify_if_necessary();
1694 }
1696 double
1697 G1CollectorPolicy::
1698 predict_young_collection_elapsed_time_ms(size_t adjustment) {
1699 guarantee( adjustment == 0 || adjustment == 1, "invariant" );
1701 G1CollectedHeap* g1h = G1CollectedHeap::heap();
1702 size_t young_num = g1h->young_list()->length();
1703 if (young_num == 0)
1704 return 0.0;
1706 young_num += adjustment;
1707 size_t pending_cards = predict_pending_cards();
1708 size_t rs_lengths = g1h->young_list()->sampled_rs_lengths() +
1709 predict_rs_length_diff();
1710 size_t card_num;
1711 if (full_young_gcs())
1712 card_num = predict_young_card_num(rs_lengths);
1713 else
1714 card_num = predict_non_young_card_num(rs_lengths);
1715 size_t young_byte_size = young_num * HeapRegion::GrainBytes;
1716 double accum_yg_surv_rate =
1717 _short_lived_surv_rate_group->accum_surv_rate(adjustment);
1719 size_t bytes_to_copy =
1720 (size_t) (accum_yg_surv_rate * (double) HeapRegion::GrainBytes);
1722 return
1723 predict_rs_update_time_ms(pending_cards) +
1724 predict_rs_scan_time_ms(card_num) +
1725 predict_object_copy_time_ms(bytes_to_copy) +
1726 predict_young_other_time_ms(young_num) +
1727 predict_constant_other_time_ms();
1728 }
1730 double
1731 G1CollectorPolicy::predict_base_elapsed_time_ms(size_t pending_cards) {
1732 size_t rs_length = predict_rs_length_diff();
1733 size_t card_num;
1734 if (full_young_gcs())
1735 card_num = predict_young_card_num(rs_length);
1736 else
1737 card_num = predict_non_young_card_num(rs_length);
1738 return predict_base_elapsed_time_ms(pending_cards, card_num);
1739 }
1741 double
1742 G1CollectorPolicy::predict_base_elapsed_time_ms(size_t pending_cards,
1743 size_t scanned_cards) {
1744 return
1745 predict_rs_update_time_ms(pending_cards) +
1746 predict_rs_scan_time_ms(scanned_cards) +
1747 predict_constant_other_time_ms();
1748 }
1750 double
1751 G1CollectorPolicy::predict_region_elapsed_time_ms(HeapRegion* hr,
1752 bool young) {
1753 size_t rs_length = hr->rem_set()->occupied();
1754 size_t card_num;
1755 if (full_young_gcs())
1756 card_num = predict_young_card_num(rs_length);
1757 else
1758 card_num = predict_non_young_card_num(rs_length);
1759 size_t bytes_to_copy = predict_bytes_to_copy(hr);
1761 double region_elapsed_time_ms =
1762 predict_rs_scan_time_ms(card_num) +
1763 predict_object_copy_time_ms(bytes_to_copy);
1765 if (young)
1766 region_elapsed_time_ms += predict_young_other_time_ms(1);
1767 else
1768 region_elapsed_time_ms += predict_non_young_other_time_ms(1);
1770 return region_elapsed_time_ms;
1771 }
1773 size_t
1774 G1CollectorPolicy::predict_bytes_to_copy(HeapRegion* hr) {
1775 size_t bytes_to_copy;
1776 if (hr->is_marked())
1777 bytes_to_copy = hr->max_live_bytes();
1778 else {
1779 guarantee( hr->is_young() && hr->age_in_surv_rate_group() != -1,
1780 "invariant" );
1781 int age = hr->age_in_surv_rate_group();
1782 double yg_surv_rate = predict_yg_surv_rate(age, hr->surv_rate_group());
1783 bytes_to_copy = (size_t) ((double) hr->used() * yg_surv_rate);
1784 }
1786 return bytes_to_copy;
1787 }
1789 void
1790 G1CollectorPolicy::init_cset_region_lengths(size_t eden_cset_region_length,
1791 size_t survivor_cset_region_length) {
1792 _eden_cset_region_length = eden_cset_region_length;
1793 _survivor_cset_region_length = survivor_cset_region_length;
1794 _old_cset_region_length = 0;
1795 }
1797 void G1CollectorPolicy::set_recorded_rs_lengths(size_t rs_lengths) {
1798 _recorded_rs_lengths = rs_lengths;
1799 }
1801 void G1CollectorPolicy::check_if_region_is_too_expensive(double
1802 predicted_time_ms) {
1803 // I don't think we need to do this when in young GC mode since
1804 // marking will be initiated next time we hit the soft limit anyway...
1805 if (predicted_time_ms > _expensive_region_limit_ms) {
1806 ergo_verbose2(ErgoPartiallyYoungGCs,
1807 "request partially-young GCs end",
1808 ergo_format_reason("predicted region time higher than threshold")
1809 ergo_format_ms("predicted region time")
1810 ergo_format_ms("threshold"),
1811 predicted_time_ms, _expensive_region_limit_ms);
1812 // no point in doing another partial one
1813 _should_revert_to_full_young_gcs = true;
1814 }
1815 }
1817 void G1CollectorPolicy::update_recent_gc_times(double end_time_sec,
1818 double elapsed_ms) {
1819 _recent_gc_times_ms->add(elapsed_ms);
1820 _recent_prev_end_times_for_all_gcs_sec->add(end_time_sec);
1821 _prev_collection_pause_end_ms = end_time_sec * 1000.0;
1822 }
1824 size_t G1CollectorPolicy::expansion_amount() {
1825 double recent_gc_overhead = recent_avg_pause_time_ratio() * 100.0;
1826 double threshold = _gc_overhead_perc;
1827 if (recent_gc_overhead > threshold) {
1828 // We will double the existing space, or take
1829 // G1ExpandByPercentOfAvailable % of the available expansion
1830 // space, whichever is smaller, bounded below by a minimum
1831 // expansion (unless that's all that's left.)
1832 const size_t min_expand_bytes = 1*M;
1833 size_t reserved_bytes = _g1->max_capacity();
1834 size_t committed_bytes = _g1->capacity();
1835 size_t uncommitted_bytes = reserved_bytes - committed_bytes;
1836 size_t expand_bytes;
1837 size_t expand_bytes_via_pct =
1838 uncommitted_bytes * G1ExpandByPercentOfAvailable / 100;
1839 expand_bytes = MIN2(expand_bytes_via_pct, committed_bytes);
1840 expand_bytes = MAX2(expand_bytes, min_expand_bytes);
1841 expand_bytes = MIN2(expand_bytes, uncommitted_bytes);
1843 ergo_verbose5(ErgoHeapSizing,
1844 "attempt heap expansion",
1845 ergo_format_reason("recent GC overhead higher than "
1846 "threshold after GC")
1847 ergo_format_perc("recent GC overhead")
1848 ergo_format_perc("threshold")
1849 ergo_format_byte("uncommitted")
1850 ergo_format_byte_perc("calculated expansion amount"),
1851 recent_gc_overhead, threshold,
1852 uncommitted_bytes,
1853 expand_bytes_via_pct, (double) G1ExpandByPercentOfAvailable);
1855 return expand_bytes;
1856 } else {
1857 return 0;
1858 }
1859 }
1861 class CountCSClosure: public HeapRegionClosure {
1862 G1CollectorPolicy* _g1_policy;
1863 public:
1864 CountCSClosure(G1CollectorPolicy* g1_policy) :
1865 _g1_policy(g1_policy) {}
1866 bool doHeapRegion(HeapRegion* r) {
1867 _g1_policy->_bytes_in_collection_set_before_gc += r->used();
1868 return false;
1869 }
1870 };
1872 void G1CollectorPolicy::count_CS_bytes_used() {
1873 CountCSClosure cs_closure(this);
1874 _g1->collection_set_iterate(&cs_closure);
1875 }
1877 void G1CollectorPolicy::print_summary(int level,
1878 const char* str,
1879 NumberSeq* seq) const {
1880 double sum = seq->sum();
1881 LineBuffer(level + 1).append_and_print_cr("%-24s = %8.2lf s (avg = %8.2lf ms)",
1882 str, sum / 1000.0, seq->avg());
1883 }
1885 void G1CollectorPolicy::print_summary_sd(int level,
1886 const char* str,
1887 NumberSeq* seq) const {
1888 print_summary(level, str, seq);
1889 LineBuffer(level + 6).append_and_print_cr("(num = %5d, std dev = %8.2lf ms, max = %8.2lf ms)",
1890 seq->num(), seq->sd(), seq->maximum());
1891 }
1893 void G1CollectorPolicy::check_other_times(int level,
1894 NumberSeq* other_times_ms,
1895 NumberSeq* calc_other_times_ms) const {
1896 bool should_print = false;
1897 LineBuffer buf(level + 2);
1899 double max_sum = MAX2(fabs(other_times_ms->sum()),
1900 fabs(calc_other_times_ms->sum()));
1901 double min_sum = MIN2(fabs(other_times_ms->sum()),
1902 fabs(calc_other_times_ms->sum()));
1903 double sum_ratio = max_sum / min_sum;
1904 if (sum_ratio > 1.1) {
1905 should_print = true;
1906 buf.append_and_print_cr("## CALCULATED OTHER SUM DOESN'T MATCH RECORDED ###");
1907 }
1909 double max_avg = MAX2(fabs(other_times_ms->avg()),
1910 fabs(calc_other_times_ms->avg()));
1911 double min_avg = MIN2(fabs(other_times_ms->avg()),
1912 fabs(calc_other_times_ms->avg()));
1913 double avg_ratio = max_avg / min_avg;
1914 if (avg_ratio > 1.1) {
1915 should_print = true;
1916 buf.append_and_print_cr("## CALCULATED OTHER AVG DOESN'T MATCH RECORDED ###");
1917 }
1919 if (other_times_ms->sum() < -0.01) {
1920 buf.append_and_print_cr("## RECORDED OTHER SUM IS NEGATIVE ###");
1921 }
1923 if (other_times_ms->avg() < -0.01) {
1924 buf.append_and_print_cr("## RECORDED OTHER AVG IS NEGATIVE ###");
1925 }
1927 if (calc_other_times_ms->sum() < -0.01) {
1928 should_print = true;
1929 buf.append_and_print_cr("## CALCULATED OTHER SUM IS NEGATIVE ###");
1930 }
1932 if (calc_other_times_ms->avg() < -0.01) {
1933 should_print = true;
1934 buf.append_and_print_cr("## CALCULATED OTHER AVG IS NEGATIVE ###");
1935 }
1937 if (should_print)
1938 print_summary(level, "Other(Calc)", calc_other_times_ms);
1939 }
1941 void G1CollectorPolicy::print_summary(PauseSummary* summary) const {
1942 bool parallel = G1CollectedHeap::use_parallel_gc_threads();
1943 MainBodySummary* body_summary = summary->main_body_summary();
1944 if (summary->get_total_seq()->num() > 0) {
1945 print_summary_sd(0, "Evacuation Pauses", summary->get_total_seq());
1946 if (body_summary != NULL) {
1947 print_summary(1, "SATB Drain", body_summary->get_satb_drain_seq());
1948 if (parallel) {
1949 print_summary(1, "Parallel Time", body_summary->get_parallel_seq());
1950 print_summary(2, "Ext Root Scanning", body_summary->get_ext_root_scan_seq());
1951 print_summary(2, "Mark Stack Scanning", body_summary->get_mark_stack_scan_seq());
1952 print_summary(2, "Update RS", body_summary->get_update_rs_seq());
1953 print_summary(2, "Scan RS", body_summary->get_scan_rs_seq());
1954 print_summary(2, "Object Copy", body_summary->get_obj_copy_seq());
1955 print_summary(2, "Termination", body_summary->get_termination_seq());
1956 print_summary(2, "Parallel Other", body_summary->get_parallel_other_seq());
1957 {
1958 NumberSeq* other_parts[] = {
1959 body_summary->get_ext_root_scan_seq(),
1960 body_summary->get_mark_stack_scan_seq(),
1961 body_summary->get_update_rs_seq(),
1962 body_summary->get_scan_rs_seq(),
1963 body_summary->get_obj_copy_seq(),
1964 body_summary->get_termination_seq()
1965 };
1966 NumberSeq calc_other_times_ms(body_summary->get_parallel_seq(),
1967 6, other_parts);
1968 check_other_times(2, body_summary->get_parallel_other_seq(),
1969 &calc_other_times_ms);
1970 }
1971 } else {
1972 print_summary(1, "Ext Root Scanning", body_summary->get_ext_root_scan_seq());
1973 print_summary(1, "Mark Stack Scanning", body_summary->get_mark_stack_scan_seq());
1974 print_summary(1, "Update RS", body_summary->get_update_rs_seq());
1975 print_summary(1, "Scan RS", body_summary->get_scan_rs_seq());
1976 print_summary(1, "Object Copy", body_summary->get_obj_copy_seq());
1977 }
1978 }
1979 print_summary(1, "Mark Closure", body_summary->get_mark_closure_seq());
1980 print_summary(1, "Clear CT", body_summary->get_clear_ct_seq());
1981 print_summary(1, "Other", summary->get_other_seq());
1982 {
1983 if (body_summary != NULL) {
1984 NumberSeq calc_other_times_ms;
1985 if (parallel) {
1986 // parallel
1987 NumberSeq* other_parts[] = {
1988 body_summary->get_satb_drain_seq(),
1989 body_summary->get_parallel_seq(),
1990 body_summary->get_clear_ct_seq()
1991 };
1992 calc_other_times_ms = NumberSeq(summary->get_total_seq(),
1993 3, other_parts);
1994 } else {
1995 // serial
1996 NumberSeq* other_parts[] = {
1997 body_summary->get_satb_drain_seq(),
1998 body_summary->get_update_rs_seq(),
1999 body_summary->get_ext_root_scan_seq(),
2000 body_summary->get_mark_stack_scan_seq(),
2001 body_summary->get_scan_rs_seq(),
2002 body_summary->get_obj_copy_seq()
2003 };
2004 calc_other_times_ms = NumberSeq(summary->get_total_seq(),
2005 6, other_parts);
2006 }
2007 check_other_times(1, summary->get_other_seq(), &calc_other_times_ms);
2008 }
2009 }
2010 } else {
2011 LineBuffer(1).append_and_print_cr("none");
2012 }
2013 LineBuffer(0).append_and_print_cr("");
2014 }
2016 void G1CollectorPolicy::print_tracing_info() const {
2017 if (TraceGen0Time) {
2018 gclog_or_tty->print_cr("ALL PAUSES");
2019 print_summary_sd(0, "Total", _all_pause_times_ms);
2020 gclog_or_tty->print_cr("");
2021 gclog_or_tty->print_cr("");
2022 gclog_or_tty->print_cr(" Full Young GC Pauses: %8d", _full_young_pause_num);
2023 gclog_or_tty->print_cr(" Partial Young GC Pauses: %8d", _partial_young_pause_num);
2024 gclog_or_tty->print_cr("");
2026 gclog_or_tty->print_cr("EVACUATION PAUSES");
2027 print_summary(_summary);
2029 gclog_or_tty->print_cr("MISC");
2030 print_summary_sd(0, "Stop World", _all_stop_world_times_ms);
2031 print_summary_sd(0, "Yields", _all_yield_times_ms);
2032 for (int i = 0; i < _aux_num; ++i) {
2033 if (_all_aux_times_ms[i].num() > 0) {
2034 char buffer[96];
2035 sprintf(buffer, "Aux%d", i);
2036 print_summary_sd(0, buffer, &_all_aux_times_ms[i]);
2037 }
2038 }
2039 }
2040 if (TraceGen1Time) {
2041 if (_all_full_gc_times_ms->num() > 0) {
2042 gclog_or_tty->print("\n%4d full_gcs: total time = %8.2f s",
2043 _all_full_gc_times_ms->num(),
2044 _all_full_gc_times_ms->sum() / 1000.0);
2045 gclog_or_tty->print_cr(" (avg = %8.2fms).", _all_full_gc_times_ms->avg());
2046 gclog_or_tty->print_cr(" [std. dev = %8.2f ms, max = %8.2f ms]",
2047 _all_full_gc_times_ms->sd(),
2048 _all_full_gc_times_ms->maximum());
2049 }
2050 }
2051 }
2053 void G1CollectorPolicy::print_yg_surv_rate_info() const {
2054 #ifndef PRODUCT
2055 _short_lived_surv_rate_group->print_surv_rate_summary();
2056 // add this call for any other surv rate groups
2057 #endif // PRODUCT
2058 }
2060 #ifndef PRODUCT
2061 // for debugging, bit of a hack...
2062 static char*
2063 region_num_to_mbs(int length) {
2064 static char buffer[64];
2065 double bytes = (double) (length * HeapRegion::GrainBytes);
2066 double mbs = bytes / (double) (1024 * 1024);
2067 sprintf(buffer, "%7.2lfMB", mbs);
2068 return buffer;
2069 }
2070 #endif // PRODUCT
2072 size_t G1CollectorPolicy::max_regions(int purpose) {
2073 switch (purpose) {
2074 case GCAllocForSurvived:
2075 return _max_survivor_regions;
2076 case GCAllocForTenured:
2077 return REGIONS_UNLIMITED;
2078 default:
2079 ShouldNotReachHere();
2080 return REGIONS_UNLIMITED;
2081 };
2082 }
2084 void G1CollectorPolicy::update_max_gc_locker_expansion() {
2085 size_t expansion_region_num = 0;
2086 if (GCLockerEdenExpansionPercent > 0) {
2087 double perc = (double) GCLockerEdenExpansionPercent / 100.0;
2088 double expansion_region_num_d = perc * (double) _young_list_target_length;
2089 // We use ceiling so that if expansion_region_num_d is > 0.0 (but
2090 // less than 1.0) we'll get 1.
2091 expansion_region_num = (size_t) ceil(expansion_region_num_d);
2092 } else {
2093 assert(expansion_region_num == 0, "sanity");
2094 }
2095 _young_list_max_length = _young_list_target_length + expansion_region_num;
2096 assert(_young_list_target_length <= _young_list_max_length, "post-condition");
2097 }
2099 // Calculates survivor space parameters.
2100 void G1CollectorPolicy::update_survivors_policy() {
2101 double max_survivor_regions_d =
2102 (double) _young_list_target_length / (double) SurvivorRatio;
2103 // We use ceiling so that if max_survivor_regions_d is > 0.0 (but
2104 // smaller than 1.0) we'll get 1.
2105 _max_survivor_regions = (size_t) ceil(max_survivor_regions_d);
2107 _tenuring_threshold = _survivors_age_table.compute_tenuring_threshold(
2108 HeapRegion::GrainWords * _max_survivor_regions);
2109 }
2111 #ifndef PRODUCT
2112 class HRSortIndexIsOKClosure: public HeapRegionClosure {
2113 CollectionSetChooser* _chooser;
2114 public:
2115 HRSortIndexIsOKClosure(CollectionSetChooser* chooser) :
2116 _chooser(chooser) {}
2118 bool doHeapRegion(HeapRegion* r) {
2119 if (!r->continuesHumongous()) {
2120 assert(_chooser->regionProperlyOrdered(r), "Ought to be.");
2121 }
2122 return false;
2123 }
2124 };
2126 bool G1CollectorPolicy::assertMarkedBytesDataOK() {
2127 HRSortIndexIsOKClosure cl(_collectionSetChooser);
2128 _g1->heap_region_iterate(&cl);
2129 return true;
2130 }
2131 #endif
2133 bool G1CollectorPolicy::force_initial_mark_if_outside_cycle(
2134 GCCause::Cause gc_cause) {
2135 bool during_cycle = _g1->concurrent_mark()->cmThread()->during_cycle();
2136 if (!during_cycle) {
2137 ergo_verbose1(ErgoConcCycles,
2138 "request concurrent cycle initiation",
2139 ergo_format_reason("requested by GC cause")
2140 ergo_format_str("GC cause"),
2141 GCCause::to_string(gc_cause));
2142 set_initiate_conc_mark_if_possible();
2143 return true;
2144 } else {
2145 ergo_verbose1(ErgoConcCycles,
2146 "do not request concurrent cycle initiation",
2147 ergo_format_reason("concurrent cycle already in progress")
2148 ergo_format_str("GC cause"),
2149 GCCause::to_string(gc_cause));
2150 return false;
2151 }
2152 }
2154 void
2155 G1CollectorPolicy::decide_on_conc_mark_initiation() {
2156 // We are about to decide on whether this pause will be an
2157 // initial-mark pause.
2159 // First, during_initial_mark_pause() should not be already set. We
2160 // will set it here if we have to. However, it should be cleared by
2161 // the end of the pause (it's only set for the duration of an
2162 // initial-mark pause).
2163 assert(!during_initial_mark_pause(), "pre-condition");
2165 if (initiate_conc_mark_if_possible()) {
2166 // We had noticed on a previous pause that the heap occupancy has
2167 // gone over the initiating threshold and we should start a
2168 // concurrent marking cycle. So we might initiate one.
2170 bool during_cycle = _g1->concurrent_mark()->cmThread()->during_cycle();
2171 if (!during_cycle) {
2172 // The concurrent marking thread is not "during a cycle", i.e.,
2173 // it has completed the last one. So we can go ahead and
2174 // initiate a new cycle.
2176 set_during_initial_mark_pause();
2177 // We do not allow non-full young GCs during marking.
2178 if (!full_young_gcs()) {
2179 set_full_young_gcs(true);
2180 ergo_verbose0(ErgoPartiallyYoungGCs,
2181 "end partially-young GCs",
2182 ergo_format_reason("concurrent cycle is about to start"));
2183 }
2185 // And we can now clear initiate_conc_mark_if_possible() as
2186 // we've already acted on it.
2187 clear_initiate_conc_mark_if_possible();
2189 ergo_verbose0(ErgoConcCycles,
2190 "initiate concurrent cycle",
2191 ergo_format_reason("concurrent cycle initiation requested"));
2192 } else {
2193 // The concurrent marking thread is still finishing up the
2194 // previous cycle. If we start one right now the two cycles
2195 // overlap. In particular, the concurrent marking thread might
2196 // be in the process of clearing the next marking bitmap (which
2197 // we will use for the next cycle if we start one). Starting a
2198 // cycle now will be bad given that parts of the marking
2199 // information might get cleared by the marking thread. And we
2200 // cannot wait for the marking thread to finish the cycle as it
2201 // periodically yields while clearing the next marking bitmap
2202 // and, if it's in a yield point, it's waiting for us to
2203 // finish. So, at this point we will not start a cycle and we'll
2204 // let the concurrent marking thread complete the last one.
2205 ergo_verbose0(ErgoConcCycles,
2206 "do not initiate concurrent cycle",
2207 ergo_format_reason("concurrent cycle already in progress"));
2208 }
2209 }
2210 }
2212 class KnownGarbageClosure: public HeapRegionClosure {
2213 CollectionSetChooser* _hrSorted;
2215 public:
2216 KnownGarbageClosure(CollectionSetChooser* hrSorted) :
2217 _hrSorted(hrSorted)
2218 {}
2220 bool doHeapRegion(HeapRegion* r) {
2221 // We only include humongous regions in collection
2222 // sets when concurrent mark shows that their contained object is
2223 // unreachable.
2225 // Do we have any marking information for this region?
2226 if (r->is_marked()) {
2227 // We don't include humongous regions in collection
2228 // sets because we collect them immediately at the end of a marking
2229 // cycle. We also don't include young regions because we *must*
2230 // include them in the next collection pause.
2231 if (!r->isHumongous() && !r->is_young()) {
2232 _hrSorted->addMarkedHeapRegion(r);
2233 }
2234 }
2235 return false;
2236 }
2237 };
2239 class ParKnownGarbageHRClosure: public HeapRegionClosure {
2240 CollectionSetChooser* _hrSorted;
2241 jint _marked_regions_added;
2242 jint _chunk_size;
2243 jint _cur_chunk_idx;
2244 jint _cur_chunk_end; // Cur chunk [_cur_chunk_idx, _cur_chunk_end)
2245 int _worker;
2246 int _invokes;
2248 void get_new_chunk() {
2249 _cur_chunk_idx = _hrSorted->getParMarkedHeapRegionChunk(_chunk_size);
2250 _cur_chunk_end = _cur_chunk_idx + _chunk_size;
2251 }
2252 void add_region(HeapRegion* r) {
2253 if (_cur_chunk_idx == _cur_chunk_end) {
2254 get_new_chunk();
2255 }
2256 assert(_cur_chunk_idx < _cur_chunk_end, "postcondition");
2257 _hrSorted->setMarkedHeapRegion(_cur_chunk_idx, r);
2258 _marked_regions_added++;
2259 _cur_chunk_idx++;
2260 }
2262 public:
2263 ParKnownGarbageHRClosure(CollectionSetChooser* hrSorted,
2264 jint chunk_size,
2265 int worker) :
2266 _hrSorted(hrSorted), _chunk_size(chunk_size), _worker(worker),
2267 _marked_regions_added(0), _cur_chunk_idx(0), _cur_chunk_end(0),
2268 _invokes(0)
2269 {}
2271 bool doHeapRegion(HeapRegion* r) {
2272 // We only include humongous regions in collection
2273 // sets when concurrent mark shows that their contained object is
2274 // unreachable.
2275 _invokes++;
2277 // Do we have any marking information for this region?
2278 if (r->is_marked()) {
2279 // We don't include humongous regions in collection
2280 // sets because we collect them immediately at the end of a marking
2281 // cycle.
2282 // We also do not include young regions in collection sets
2283 if (!r->isHumongous() && !r->is_young()) {
2284 add_region(r);
2285 }
2286 }
2287 return false;
2288 }
2289 jint marked_regions_added() { return _marked_regions_added; }
2290 int invokes() { return _invokes; }
2291 };
2293 class ParKnownGarbageTask: public AbstractGangTask {
2294 CollectionSetChooser* _hrSorted;
2295 jint _chunk_size;
2296 G1CollectedHeap* _g1;
2297 public:
2298 ParKnownGarbageTask(CollectionSetChooser* hrSorted, jint chunk_size) :
2299 AbstractGangTask("ParKnownGarbageTask"),
2300 _hrSorted(hrSorted), _chunk_size(chunk_size),
2301 _g1(G1CollectedHeap::heap())
2302 {}
2304 void work(int i) {
2305 ParKnownGarbageHRClosure parKnownGarbageCl(_hrSorted, _chunk_size, i);
2306 // Back to zero for the claim value.
2307 _g1->heap_region_par_iterate_chunked(&parKnownGarbageCl, i,
2308 _g1->workers()->active_workers(),
2309 HeapRegion::InitialClaimValue);
2310 jint regions_added = parKnownGarbageCl.marked_regions_added();
2311 _hrSorted->incNumMarkedHeapRegions(regions_added);
2312 if (G1PrintParCleanupStats) {
2313 gclog_or_tty->print_cr(" Thread %d called %d times, added %d regions to list.",
2314 i, parKnownGarbageCl.invokes(), regions_added);
2315 }
2316 }
2317 };
2319 void
2320 G1CollectorPolicy::record_concurrent_mark_cleanup_end(int no_of_gc_threads) {
2321 double start_sec;
2322 if (G1PrintParCleanupStats) {
2323 start_sec = os::elapsedTime();
2324 }
2326 _collectionSetChooser->clearMarkedHeapRegions();
2327 double clear_marked_end_sec;
2328 if (G1PrintParCleanupStats) {
2329 clear_marked_end_sec = os::elapsedTime();
2330 gclog_or_tty->print_cr(" clear marked regions: %8.3f ms.",
2331 (clear_marked_end_sec - start_sec) * 1000.0);
2332 }
2334 if (G1CollectedHeap::use_parallel_gc_threads()) {
2335 const size_t OverpartitionFactor = 4;
2336 size_t WorkUnit;
2337 // The use of MinChunkSize = 8 in the original code
2338 // causes some assertion failures when the total number of
2339 // region is less than 8. The code here tries to fix that.
2340 // Should the original code also be fixed?
2341 if (no_of_gc_threads > 0) {
2342 const size_t MinWorkUnit =
2343 MAX2(_g1->n_regions() / no_of_gc_threads, (size_t) 1U);
2344 WorkUnit =
2345 MAX2(_g1->n_regions() / (no_of_gc_threads * OverpartitionFactor),
2346 MinWorkUnit);
2347 } else {
2348 assert(no_of_gc_threads > 0,
2349 "The active gc workers should be greater than 0");
2350 // In a product build do something reasonable to avoid a crash.
2351 const size_t MinWorkUnit =
2352 MAX2(_g1->n_regions() / ParallelGCThreads, (size_t) 1U);
2353 WorkUnit =
2354 MAX2(_g1->n_regions() / (ParallelGCThreads * OverpartitionFactor),
2355 MinWorkUnit);
2356 }
2357 _collectionSetChooser->prepareForAddMarkedHeapRegionsPar(_g1->n_regions(),
2358 WorkUnit);
2359 ParKnownGarbageTask parKnownGarbageTask(_collectionSetChooser,
2360 (int) WorkUnit);
2361 _g1->workers()->run_task(&parKnownGarbageTask);
2363 assert(_g1->check_heap_region_claim_values(HeapRegion::InitialClaimValue),
2364 "sanity check");
2365 } else {
2366 KnownGarbageClosure knownGarbagecl(_collectionSetChooser);
2367 _g1->heap_region_iterate(&knownGarbagecl);
2368 }
2369 double known_garbage_end_sec;
2370 if (G1PrintParCleanupStats) {
2371 known_garbage_end_sec = os::elapsedTime();
2372 gclog_or_tty->print_cr(" compute known garbage: %8.3f ms.",
2373 (known_garbage_end_sec - clear_marked_end_sec) * 1000.0);
2374 }
2376 _collectionSetChooser->sortMarkedHeapRegions();
2377 double end_sec = os::elapsedTime();
2378 if (G1PrintParCleanupStats) {
2379 gclog_or_tty->print_cr(" sorting: %8.3f ms.",
2380 (end_sec - known_garbage_end_sec) * 1000.0);
2381 }
2383 double elapsed_time_ms = (end_sec - _mark_cleanup_start_sec) * 1000.0;
2384 _concurrent_mark_cleanup_times_ms->add(elapsed_time_ms);
2385 _cur_mark_stop_world_time_ms += elapsed_time_ms;
2386 _prev_collection_pause_end_ms += elapsed_time_ms;
2387 _mmu_tracker->add_pause(_mark_cleanup_start_sec, end_sec, true);
2388 }
2390 // Add the heap region at the head of the non-incremental collection set
2391 void G1CollectorPolicy::add_old_region_to_cset(HeapRegion* hr) {
2392 assert(_inc_cset_build_state == Active, "Precondition");
2393 assert(!hr->is_young(), "non-incremental add of young region");
2395 if (_g1->mark_in_progress())
2396 _g1->concurrent_mark()->registerCSetRegion(hr);
2398 assert(!hr->in_collection_set(), "should not already be in the CSet");
2399 hr->set_in_collection_set(true);
2400 hr->set_next_in_collection_set(_collection_set);
2401 _collection_set = hr;
2402 _collection_set_bytes_used_before += hr->used();
2403 _g1->register_region_with_in_cset_fast_test(hr);
2404 size_t rs_length = hr->rem_set()->occupied();
2405 _recorded_rs_lengths += rs_length;
2406 _old_cset_region_length += 1;
2407 }
2409 // Initialize the per-collection-set information
2410 void G1CollectorPolicy::start_incremental_cset_building() {
2411 assert(_inc_cset_build_state == Inactive, "Precondition");
2413 _inc_cset_head = NULL;
2414 _inc_cset_tail = NULL;
2415 _inc_cset_bytes_used_before = 0;
2417 _inc_cset_max_finger = 0;
2418 _inc_cset_recorded_rs_lengths = 0;
2419 _inc_cset_predicted_elapsed_time_ms = 0;
2420 _inc_cset_build_state = Active;
2421 }
2423 void G1CollectorPolicy::add_to_incremental_cset_info(HeapRegion* hr, size_t rs_length) {
2424 // This routine is used when:
2425 // * adding survivor regions to the incremental cset at the end of an
2426 // evacuation pause,
2427 // * adding the current allocation region to the incremental cset
2428 // when it is retired, and
2429 // * updating existing policy information for a region in the
2430 // incremental cset via young list RSet sampling.
2431 // Therefore this routine may be called at a safepoint by the
2432 // VM thread, or in-between safepoints by mutator threads (when
2433 // retiring the current allocation region) or a concurrent
2434 // refine thread (RSet sampling).
2436 double region_elapsed_time_ms = predict_region_elapsed_time_ms(hr, true);
2437 size_t used_bytes = hr->used();
2439 _inc_cset_recorded_rs_lengths += rs_length;
2440 _inc_cset_predicted_elapsed_time_ms += region_elapsed_time_ms;
2442 _inc_cset_bytes_used_before += used_bytes;
2444 // Cache the values we have added to the aggregated informtion
2445 // in the heap region in case we have to remove this region from
2446 // the incremental collection set, or it is updated by the
2447 // rset sampling code
2448 hr->set_recorded_rs_length(rs_length);
2449 hr->set_predicted_elapsed_time_ms(region_elapsed_time_ms);
2450 }
2452 void G1CollectorPolicy::remove_from_incremental_cset_info(HeapRegion* hr) {
2453 // This routine is currently only called as part of the updating of
2454 // existing policy information for regions in the incremental cset that
2455 // is performed by the concurrent refine thread(s) as part of young list
2456 // RSet sampling. Therefore we should not be at a safepoint.
2458 assert(!SafepointSynchronize::is_at_safepoint(), "should not be at safepoint");
2459 assert(hr->is_young(), "it should be");
2461 size_t used_bytes = hr->used();
2462 size_t old_rs_length = hr->recorded_rs_length();
2463 double old_elapsed_time_ms = hr->predicted_elapsed_time_ms();
2465 // Subtract the old recorded/predicted policy information for
2466 // the given heap region from the collection set info.
2467 _inc_cset_recorded_rs_lengths -= old_rs_length;
2468 _inc_cset_predicted_elapsed_time_ms -= old_elapsed_time_ms;
2470 _inc_cset_bytes_used_before -= used_bytes;
2472 // Clear the values cached in the heap region
2473 hr->set_recorded_rs_length(0);
2474 hr->set_predicted_elapsed_time_ms(0);
2475 }
2477 void G1CollectorPolicy::update_incremental_cset_info(HeapRegion* hr, size_t new_rs_length) {
2478 // Update the collection set information that is dependent on the new RS length
2479 assert(hr->is_young(), "Precondition");
2481 remove_from_incremental_cset_info(hr);
2482 add_to_incremental_cset_info(hr, new_rs_length);
2483 }
2485 void G1CollectorPolicy::add_region_to_incremental_cset_common(HeapRegion* hr) {
2486 assert(hr->is_young(), "invariant");
2487 assert(hr->young_index_in_cset() > -1, "should have already been set");
2488 assert(_inc_cset_build_state == Active, "Precondition");
2490 // We need to clear and set the cached recorded/cached collection set
2491 // information in the heap region here (before the region gets added
2492 // to the collection set). An individual heap region's cached values
2493 // are calculated, aggregated with the policy collection set info,
2494 // and cached in the heap region here (initially) and (subsequently)
2495 // by the Young List sampling code.
2497 size_t rs_length = hr->rem_set()->occupied();
2498 add_to_incremental_cset_info(hr, rs_length);
2500 HeapWord* hr_end = hr->end();
2501 _inc_cset_max_finger = MAX2(_inc_cset_max_finger, hr_end);
2503 assert(!hr->in_collection_set(), "invariant");
2504 hr->set_in_collection_set(true);
2505 assert( hr->next_in_collection_set() == NULL, "invariant");
2507 _g1->register_region_with_in_cset_fast_test(hr);
2508 }
2510 // Add the region at the RHS of the incremental cset
2511 void G1CollectorPolicy::add_region_to_incremental_cset_rhs(HeapRegion* hr) {
2512 // We should only ever be appending survivors at the end of a pause
2513 assert( hr->is_survivor(), "Logic");
2515 // Do the 'common' stuff
2516 add_region_to_incremental_cset_common(hr);
2518 // Now add the region at the right hand side
2519 if (_inc_cset_tail == NULL) {
2520 assert(_inc_cset_head == NULL, "invariant");
2521 _inc_cset_head = hr;
2522 } else {
2523 _inc_cset_tail->set_next_in_collection_set(hr);
2524 }
2525 _inc_cset_tail = hr;
2526 }
2528 // Add the region to the LHS of the incremental cset
2529 void G1CollectorPolicy::add_region_to_incremental_cset_lhs(HeapRegion* hr) {
2530 // Survivors should be added to the RHS at the end of a pause
2531 assert(!hr->is_survivor(), "Logic");
2533 // Do the 'common' stuff
2534 add_region_to_incremental_cset_common(hr);
2536 // Add the region at the left hand side
2537 hr->set_next_in_collection_set(_inc_cset_head);
2538 if (_inc_cset_head == NULL) {
2539 assert(_inc_cset_tail == NULL, "Invariant");
2540 _inc_cset_tail = hr;
2541 }
2542 _inc_cset_head = hr;
2543 }
2545 #ifndef PRODUCT
2546 void G1CollectorPolicy::print_collection_set(HeapRegion* list_head, outputStream* st) {
2547 assert(list_head == inc_cset_head() || list_head == collection_set(), "must be");
2549 st->print_cr("\nCollection_set:");
2550 HeapRegion* csr = list_head;
2551 while (csr != NULL) {
2552 HeapRegion* next = csr->next_in_collection_set();
2553 assert(csr->in_collection_set(), "bad CS");
2554 st->print_cr(" [%08x-%08x], t: %08x, P: %08x, N: %08x, C: %08x, "
2555 "age: %4d, y: %d, surv: %d",
2556 csr->bottom(), csr->end(),
2557 csr->top(),
2558 csr->prev_top_at_mark_start(),
2559 csr->next_top_at_mark_start(),
2560 csr->top_at_conc_mark_count(),
2561 csr->age_in_surv_rate_group_cond(),
2562 csr->is_young(),
2563 csr->is_survivor());
2564 csr = next;
2565 }
2566 }
2567 #endif // !PRODUCT
2569 void G1CollectorPolicy::choose_collection_set(double target_pause_time_ms) {
2570 // Set this here - in case we're not doing young collections.
2571 double non_young_start_time_sec = os::elapsedTime();
2573 YoungList* young_list = _g1->young_list();
2575 guarantee(target_pause_time_ms > 0.0,
2576 err_msg("target_pause_time_ms = %1.6lf should be positive",
2577 target_pause_time_ms));
2578 guarantee(_collection_set == NULL, "Precondition");
2580 double base_time_ms = predict_base_elapsed_time_ms(_pending_cards);
2581 double predicted_pause_time_ms = base_time_ms;
2583 double time_remaining_ms = target_pause_time_ms - base_time_ms;
2585 ergo_verbose3(ErgoCSetConstruction | ErgoHigh,
2586 "start choosing CSet",
2587 ergo_format_ms("predicted base time")
2588 ergo_format_ms("remaining time")
2589 ergo_format_ms("target pause time"),
2590 base_time_ms, time_remaining_ms, target_pause_time_ms);
2592 // the 10% and 50% values are arbitrary...
2593 double threshold = 0.10 * target_pause_time_ms;
2594 if (time_remaining_ms < threshold) {
2595 double prev_time_remaining_ms = time_remaining_ms;
2596 time_remaining_ms = 0.50 * target_pause_time_ms;
2597 ergo_verbose3(ErgoCSetConstruction,
2598 "adjust remaining time",
2599 ergo_format_reason("remaining time lower than threshold")
2600 ergo_format_ms("remaining time")
2601 ergo_format_ms("threshold")
2602 ergo_format_ms("adjusted remaining time"),
2603 prev_time_remaining_ms, threshold, time_remaining_ms);
2604 }
2606 size_t expansion_bytes = _g1->expansion_regions() * HeapRegion::GrainBytes;
2608 HeapRegion* hr;
2609 double young_start_time_sec = os::elapsedTime();
2611 _collection_set_bytes_used_before = 0;
2612 _last_young_gc_full = full_young_gcs() ? true : false;
2614 if (_last_young_gc_full) {
2615 ++_full_young_pause_num;
2616 } else {
2617 ++_partial_young_pause_num;
2618 }
2620 // The young list is laid with the survivor regions from the previous
2621 // pause are appended to the RHS of the young list, i.e.
2622 // [Newly Young Regions ++ Survivors from last pause].
2624 size_t survivor_region_length = young_list->survivor_length();
2625 size_t eden_region_length = young_list->length() - survivor_region_length;
2626 init_cset_region_lengths(eden_region_length, survivor_region_length);
2627 hr = young_list->first_survivor_region();
2628 while (hr != NULL) {
2629 assert(hr->is_survivor(), "badly formed young list");
2630 hr->set_young();
2631 hr = hr->get_next_young_region();
2632 }
2634 // Clear the fields that point to the survivor list - they are all young now.
2635 young_list->clear_survivors();
2637 if (_g1->mark_in_progress())
2638 _g1->concurrent_mark()->register_collection_set_finger(_inc_cset_max_finger);
2640 _collection_set = _inc_cset_head;
2641 _collection_set_bytes_used_before = _inc_cset_bytes_used_before;
2642 time_remaining_ms -= _inc_cset_predicted_elapsed_time_ms;
2643 predicted_pause_time_ms += _inc_cset_predicted_elapsed_time_ms;
2645 ergo_verbose3(ErgoCSetConstruction | ErgoHigh,
2646 "add young regions to CSet",
2647 ergo_format_region("eden")
2648 ergo_format_region("survivors")
2649 ergo_format_ms("predicted young region time"),
2650 eden_region_length, survivor_region_length,
2651 _inc_cset_predicted_elapsed_time_ms);
2653 // The number of recorded young regions is the incremental
2654 // collection set's current size
2655 set_recorded_rs_lengths(_inc_cset_recorded_rs_lengths);
2657 double young_end_time_sec = os::elapsedTime();
2658 _recorded_young_cset_choice_time_ms =
2659 (young_end_time_sec - young_start_time_sec) * 1000.0;
2661 // We are doing young collections so reset this.
2662 non_young_start_time_sec = young_end_time_sec;
2664 if (!full_young_gcs()) {
2665 bool should_continue = true;
2666 NumberSeq seq;
2667 double avg_prediction = 100000000000000000.0; // something very large
2669 double prev_predicted_pause_time_ms = predicted_pause_time_ms;
2670 do {
2671 // Note that add_old_region_to_cset() increments the
2672 // _old_cset_region_length field and cset_region_length() returns the
2673 // sum of _eden_cset_region_length, _survivor_cset_region_length, and
2674 // _old_cset_region_length. So, as old regions are added to the
2675 // CSet, _old_cset_region_length will be incremented and
2676 // cset_region_length(), which is used below, will always reflect
2677 // the the total number of regions added up to this point to the CSet.
2679 hr = _collectionSetChooser->getNextMarkedRegion(time_remaining_ms,
2680 avg_prediction);
2681 if (hr != NULL) {
2682 _g1->old_set_remove(hr);
2683 double predicted_time_ms = predict_region_elapsed_time_ms(hr, false);
2684 time_remaining_ms -= predicted_time_ms;
2685 predicted_pause_time_ms += predicted_time_ms;
2686 add_old_region_to_cset(hr);
2687 seq.add(predicted_time_ms);
2688 avg_prediction = seq.avg() + seq.sd();
2689 }
2691 should_continue = true;
2692 if (hr == NULL) {
2693 // No need for an ergo verbose message here,
2694 // getNextMarkRegion() does this when it returns NULL.
2695 should_continue = false;
2696 } else {
2697 if (adaptive_young_list_length()) {
2698 if (time_remaining_ms < 0.0) {
2699 ergo_verbose1(ErgoCSetConstruction,
2700 "stop adding old regions to CSet",
2701 ergo_format_reason("remaining time is lower than 0")
2702 ergo_format_ms("remaining time"),
2703 time_remaining_ms);
2704 should_continue = false;
2705 }
2706 } else {
2707 if (cset_region_length() >= _young_list_fixed_length) {
2708 ergo_verbose2(ErgoCSetConstruction,
2709 "stop adding old regions to CSet",
2710 ergo_format_reason("CSet length reached target")
2711 ergo_format_region("CSet")
2712 ergo_format_region("young target"),
2713 cset_region_length(), _young_list_fixed_length);
2714 should_continue = false;
2715 }
2716 }
2717 }
2718 } while (should_continue);
2720 if (!adaptive_young_list_length() &&
2721 cset_region_length() < _young_list_fixed_length) {
2722 ergo_verbose2(ErgoCSetConstruction,
2723 "request partially-young GCs end",
2724 ergo_format_reason("CSet length lower than target")
2725 ergo_format_region("CSet")
2726 ergo_format_region("young target"),
2727 cset_region_length(), _young_list_fixed_length);
2728 _should_revert_to_full_young_gcs = true;
2729 }
2731 ergo_verbose2(ErgoCSetConstruction | ErgoHigh,
2732 "add old regions to CSet",
2733 ergo_format_region("old")
2734 ergo_format_ms("predicted old region time"),
2735 old_cset_region_length(),
2736 predicted_pause_time_ms - prev_predicted_pause_time_ms);
2737 }
2739 stop_incremental_cset_building();
2741 count_CS_bytes_used();
2743 ergo_verbose5(ErgoCSetConstruction,
2744 "finish choosing CSet",
2745 ergo_format_region("eden")
2746 ergo_format_region("survivors")
2747 ergo_format_region("old")
2748 ergo_format_ms("predicted pause time")
2749 ergo_format_ms("target pause time"),
2750 eden_region_length, survivor_region_length,
2751 old_cset_region_length(),
2752 predicted_pause_time_ms, target_pause_time_ms);
2754 double non_young_end_time_sec = os::elapsedTime();
2755 _recorded_non_young_cset_choice_time_ms =
2756 (non_young_end_time_sec - non_young_start_time_sec) * 1000.0;
2757 }