1 /*
2 * Copyright (c) 2001, 2016, 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 */
24
25 #include "precompiled.hpp"
26 #include "gc/g1/concurrentG1Refine.hpp"
27 #include "gc/g1/concurrentMarkThread.inline.hpp"
28 #include "gc/g1/g1CollectedHeap.inline.hpp"
29 #include "gc/g1/g1CollectionSet.hpp"
30 #include "gc/g1/g1CollectorPolicy.hpp"
31 #include "gc/g1/g1ConcurrentMark.hpp"
32 #include "gc/g1/g1IHOPControl.hpp"
33 #include "gc/g1/g1GCPhaseTimes.hpp"
34 #include "gc/g1/g1YoungGenSizer.hpp"
35 #include "gc/g1/heapRegion.inline.hpp"
36 #include "gc/g1/heapRegionRemSet.hpp"
37 #include "gc/shared/gcPolicyCounters.hpp"
38 #include "runtime/arguments.hpp"
39 #include "runtime/java.hpp"
40 #include "runtime/mutexLocker.hpp"
41 #include "utilities/debug.hpp"
42 #include "utilities/pair.hpp"
43
44 // Different defaults for different number of GC threads
45 // They were chosen by running GCOld and SPECjbb on debris with different
46 // numbers of GC threads and choosing them based on the results
47
48 // all the same
49 static double rs_length_diff_defaults[] = {
50 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0
51 };
52
53 static double cost_per_card_ms_defaults[] = {
54 0.01, 0.005, 0.005, 0.003, 0.003, 0.002, 0.002, 0.0015
55 };
56
57 // all the same
58 static double young_cards_per_entry_ratio_defaults[] = {
59 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0
60 };
61
62 static double cost_per_entry_ms_defaults[] = {
64 };
65
66 static double cost_per_byte_ms_defaults[] = {
67 0.00006, 0.00003, 0.00003, 0.000015, 0.000015, 0.00001, 0.00001, 0.000009
68 };
69
70 // these should be pretty consistent
71 static double constant_other_time_ms_defaults[] = {
72 5.0, 5.0, 5.0, 5.0, 5.0, 5.0, 5.0, 5.0
73 };
74
75
76 static double young_other_cost_per_region_ms_defaults[] = {
77 0.3, 0.2, 0.2, 0.15, 0.15, 0.12, 0.12, 0.1
78 };
79
80 static double non_young_other_cost_per_region_ms_defaults[] = {
81 1.0, 0.7, 0.7, 0.5, 0.5, 0.42, 0.42, 0.30
82 };
83
84 G1CollectorPolicy::G1CollectorPolicy() :
85 _predictor(G1ConfidencePercent / 100.0),
86
87 _recent_gc_times_ms(new TruncatedSeq(NumPrevPausesForHeuristics)),
88
89 _concurrent_mark_remark_times_ms(new TruncatedSeq(NumPrevPausesForHeuristics)),
90 _concurrent_mark_cleanup_times_ms(new TruncatedSeq(NumPrevPausesForHeuristics)),
91
92 _alloc_rate_ms_seq(new TruncatedSeq(TruncatedSeqLength)),
93 _prev_collection_pause_end_ms(0.0),
94 _rs_length_diff_seq(new TruncatedSeq(TruncatedSeqLength)),
95 _cost_per_card_ms_seq(new TruncatedSeq(TruncatedSeqLength)),
96 _cost_scan_hcc_seq(new TruncatedSeq(TruncatedSeqLength)),
97 _young_cards_per_entry_ratio_seq(new TruncatedSeq(TruncatedSeqLength)),
98 _mixed_cards_per_entry_ratio_seq(new TruncatedSeq(TruncatedSeqLength)),
99 _cost_per_entry_ms_seq(new TruncatedSeq(TruncatedSeqLength)),
100 _mixed_cost_per_entry_ms_seq(new TruncatedSeq(TruncatedSeqLength)),
101 _cost_per_byte_ms_seq(new TruncatedSeq(TruncatedSeqLength)),
102 _cost_per_byte_ms_during_cm_seq(new TruncatedSeq(TruncatedSeqLength)),
103 _constant_other_time_ms_seq(new TruncatedSeq(TruncatedSeqLength)),
104 _young_other_cost_per_region_ms_seq(new TruncatedSeq(TruncatedSeqLength)),
105 _non_young_other_cost_per_region_ms_seq(
106 new TruncatedSeq(TruncatedSeqLength)),
107
108 _pending_cards_seq(new TruncatedSeq(TruncatedSeqLength)),
109 _rs_lengths_seq(new TruncatedSeq(TruncatedSeqLength)),
110
111 _pause_time_target_ms((double) MaxGCPauseMillis),
112
113 _recent_prev_end_times_for_all_gcs_sec(
114 new TruncatedSeq(NumPrevPausesForHeuristics)),
115
116 _recent_avg_pause_time_ratio(0.0),
117 _rs_lengths_prediction(0),
118 _max_survivor_regions(0),
119
120 // add here any more surv rate groups
121 _survivors_age_table(true),
122
123 _gc_overhead_perc(0.0),
124
125 _bytes_allocated_in_old_since_last_gc(0),
126 _ihop_control(NULL),
127 _initial_mark_to_mixed() {
128
129 // SurvRateGroups below must be initialized after the predictor because they
130 // indirectly use it through this object passed to their constructor.
131 _short_lived_surv_rate_group =
132 new SurvRateGroup(&_predictor, "Short Lived", G1YoungSurvRateNumRegionsSummary);
133 _survivor_surv_rate_group =
134 new SurvRateGroup(&_predictor, "Survivor", G1YoungSurvRateNumRegionsSummary);
135
136 // Set up the region size and associated fields. Given that the
137 // policy is created before the heap, we have to set this up here,
138 // so it's done as soon as possible.
139
140 // It would have been natural to pass initial_heap_byte_size() and
141 // max_heap_byte_size() to setup_heap_region_size() but those have
142 // not been set up at this point since they should be aligned with
143 // the region size. So, there is a circular dependency here. We base
144 // the region size on the heap size, but the heap size should be
145 // aligned with the region size. To get around this we use the
146 // unaligned values for the heap.
147 HeapRegion::setup_heap_region_size(InitialHeapSize, MaxHeapSize);
148 HeapRegionRemSet::setup_remset_size();
149
150 _recent_prev_end_times_for_all_gcs_sec->add(os::elapsedTime());
151 _prev_collection_pause_end_ms = os::elapsedTime() * 1000.0;
152 clear_ratio_check_data();
153
154 _phase_times = new G1GCPhaseTimes(ParallelGCThreads);
155
156 int index = MIN2(ParallelGCThreads - 1, 7u);
157
158 _rs_length_diff_seq->add(rs_length_diff_defaults[index]);
159 _cost_per_card_ms_seq->add(cost_per_card_ms_defaults[index]);
160 _cost_scan_hcc_seq->add(0.0);
161 _young_cards_per_entry_ratio_seq->add(
162 young_cards_per_entry_ratio_defaults[index]);
163 _cost_per_entry_ms_seq->add(cost_per_entry_ms_defaults[index]);
164 _cost_per_byte_ms_seq->add(cost_per_byte_ms_defaults[index]);
165 _constant_other_time_ms_seq->add(constant_other_time_ms_defaults[index]);
166 _young_other_cost_per_region_ms_seq->add(
167 young_other_cost_per_region_ms_defaults[index]);
168 _non_young_other_cost_per_region_ms_seq->add(
169 non_young_other_cost_per_region_ms_defaults[index]);
170
171 // Below, we might need to calculate the pause time target based on
172 // the pause interval. When we do so we are going to give G1 maximum
173 // flexibility and allow it to do pauses when it needs to. So, we'll
174 // arrange that the pause interval to be pause time target + 1 to
175 // ensure that a) the pause time target is maximized with respect to
176 // the pause interval and b) we maintain the invariant that pause
177 // time target < pause interval. If the user does not want this
178 // maximum flexibility, they will have to set the pause interval
179 // explicitly.
180
181 // First make sure that, if either parameter is set, its value is
182 // reasonable.
183 if (!FLAG_IS_DEFAULT(MaxGCPauseMillis)) {
184 if (MaxGCPauseMillis < 1) {
185 vm_exit_during_initialization("MaxGCPauseMillis should be "
186 "greater than 0");
187 }
188 }
189 if (!FLAG_IS_DEFAULT(GCPauseIntervalMillis)) {
190 if (GCPauseIntervalMillis < 1) {
191 vm_exit_during_initialization("GCPauseIntervalMillis should be "
192 "greater than 0");
193 }
194 }
195
196 // Then, if the pause time target parameter was not set, set it to
197 // the default value.
198 if (FLAG_IS_DEFAULT(MaxGCPauseMillis)) {
199 if (FLAG_IS_DEFAULT(GCPauseIntervalMillis)) {
200 // The default pause time target in G1 is 200ms
201 FLAG_SET_DEFAULT(MaxGCPauseMillis, 200);
202 } else {
203 // We do not allow the pause interval to be set without the
204 // pause time target
205 vm_exit_during_initialization("GCPauseIntervalMillis cannot be set "
206 "without setting MaxGCPauseMillis");
207 }
208 }
209
210 // Then, if the interval parameter was not set, set it according to
211 // the pause time target (this will also deal with the case when the
212 // pause time target is the default value).
213 if (FLAG_IS_DEFAULT(GCPauseIntervalMillis)) {
214 FLAG_SET_DEFAULT(GCPauseIntervalMillis, MaxGCPauseMillis + 1);
215 }
216
217 // Finally, make sure that the two parameters are consistent.
218 if (MaxGCPauseMillis >= GCPauseIntervalMillis) {
219 char buffer[256];
220 jio_snprintf(buffer, 256,
221 "MaxGCPauseMillis (%u) should be less than "
222 "GCPauseIntervalMillis (%u)",
223 MaxGCPauseMillis, GCPauseIntervalMillis);
224 vm_exit_during_initialization(buffer);
225 }
226
227 double max_gc_time = (double) MaxGCPauseMillis / 1000.0;
228 double time_slice = (double) GCPauseIntervalMillis / 1000.0;
229 _mmu_tracker = new G1MMUTrackerQueue(time_slice, max_gc_time);
230
231 // start conservatively (around 50ms is about right)
232 _concurrent_mark_remark_times_ms->add(0.05);
233 _concurrent_mark_cleanup_times_ms->add(0.20);
234 _tenuring_threshold = MaxTenuringThreshold;
235
236 assert(GCTimeRatio > 0,
237 "we should have set it to a default value set_g1_gc_flags() "
238 "if a user set it to 0");
239 _gc_overhead_perc = 100.0 * (1.0 / (1.0 + GCTimeRatio));
240
241 uintx reserve_perc = G1ReservePercent;
242 // Put an artificial ceiling on this so that it's not set to a silly value.
243 if (reserve_perc > 50) {
244 reserve_perc = 50;
245 warning("G1ReservePercent is set to a value that is too large, "
246 "it's been updated to " UINTX_FORMAT, reserve_perc);
247 }
248 _reserve_factor = (double) reserve_perc / 100.0;
249 // This will be set when the heap is expanded
250 // for the first time during initialization.
251 _reserve_regions = 0;
252
253 _ihop_control = create_ihop_control();
254 }
255
256 G1CollectorPolicy::~G1CollectorPolicy() {
257 delete _ihop_control;
258 }
259
260 double G1CollectorPolicy::get_new_prediction(TruncatedSeq const* seq) const {
261 return _predictor.get_new_prediction(seq);
262 }
263
264 size_t G1CollectorPolicy::get_new_size_prediction(TruncatedSeq const* seq) const {
265 return (size_t)get_new_prediction(seq);
266 }
267
268 void G1CollectorPolicy::initialize_alignments() {
269 _space_alignment = HeapRegion::GrainBytes;
270 size_t card_table_alignment = CardTableRS::ct_max_alignment_constraint();
271 size_t page_size = UseLargePages ? os::large_page_size() : os::vm_page_size();
272 _heap_alignment = MAX3(card_table_alignment, _space_alignment, page_size);
273 }
274
275 G1CollectorState* G1CollectorPolicy::collector_state() const { return _g1->collector_state(); }
276
277 void G1CollectorPolicy::post_heap_initialize() {
278 uintx max_regions = G1CollectedHeap::heap()->max_regions();
279 size_t max_young_size = (size_t)_young_gen_sizer->max_young_length(max_regions) * HeapRegion::GrainBytes;
280 if (max_young_size != MaxNewSize) {
281 FLAG_SET_ERGO(size_t, MaxNewSize, max_young_size);
282 }
283 }
284
285 void G1CollectorPolicy::initialize_flags() {
286 if (G1HeapRegionSize != HeapRegion::GrainBytes) {
287 FLAG_SET_ERGO(size_t, G1HeapRegionSize, HeapRegion::GrainBytes);
288 }
289
290 if (SurvivorRatio < 1) {
291 vm_exit_during_initialization("Invalid survivor ratio specified");
292 }
293 CollectorPolicy::initialize_flags();
294 _young_gen_sizer = new G1YoungGenSizer(); // Must be after call to initialize_flags
295 }
296
297
298 void G1CollectorPolicy::init() {
299 // Set aside an initial future to_space.
300 _g1 = G1CollectedHeap::heap();
301 _collection_set = _g1->collection_set();
302 _collection_set->set_policy(this);
303
304 assert(Heap_lock->owned_by_self(), "Locking discipline.");
305
306 initialize_gc_policy_counters();
307
308 if (adaptive_young_list_length()) {
309 _young_list_fixed_length = 0;
310 } else {
311 _young_list_fixed_length = _young_gen_sizer->min_desired_young_length();
312 }
313 _free_regions_at_end_of_collection = _g1->num_free_regions();
314
315 update_young_list_max_and_target_length();
316 // We may immediately start allocating regions and placing them on the
317 // collection set list. Initialize the per-collection set info
318 _collection_set->start_incremental_building();
319 }
320
321 void G1CollectorPolicy::note_gc_start(uint num_active_workers) {
322 phase_times()->note_gc_start(num_active_workers);
323 }
324
325 // Create the jstat counters for the policy.
326 void G1CollectorPolicy::initialize_gc_policy_counters() {
327 _gc_policy_counters = new GCPolicyCounters("GarbageFirst", 1, 3);
328 }
329
330 bool G1CollectorPolicy::predict_will_fit(uint young_length,
331 double base_time_ms,
332 uint base_free_regions,
333 double target_pause_time_ms) const {
334 if (young_length >= base_free_regions) {
335 // end condition 1: not enough space for the young regions
336 return false;
337 }
338
339 double accum_surv_rate = accum_yg_surv_rate_pred((int) young_length - 1);
340 size_t bytes_to_copy =
341 (size_t) (accum_surv_rate * (double) HeapRegion::GrainBytes);
342 double copy_time_ms = predict_object_copy_time_ms(bytes_to_copy);
343 double young_other_time_ms = predict_young_other_time_ms(young_length);
344 double pause_time_ms = base_time_ms + copy_time_ms + young_other_time_ms;
345 if (pause_time_ms > target_pause_time_ms) {
346 // end condition 2: prediction is over the target pause time
347 return false;
348 }
349
350 size_t free_bytes = (base_free_regions - young_length) * HeapRegion::GrainBytes;
351
352 // When copying, we will likely need more bytes free than is live in the region.
353 // Add some safety margin to factor in the confidence of our guess, and the
354 // natural expected waste.
355 // (100.0 / G1ConfidencePercent) is a scale factor that expresses the uncertainty
356 // of the calculation: the lower the confidence, the more headroom.
357 // (100 + TargetPLABWastePct) represents the increase in expected bytes during
358 // copying due to anticipated waste in the PLABs.
359 double safety_factor = (100.0 / G1ConfidencePercent) * (100 + TargetPLABWastePct) / 100.0;
360 size_t expected_bytes_to_copy = (size_t)(safety_factor * bytes_to_copy);
361
362 if (expected_bytes_to_copy > free_bytes) {
363 // end condition 3: out-of-space
364 return false;
365 }
366
367 // success!
368 return true;
369 }
370
371 void G1CollectorPolicy::record_new_heap_size(uint new_number_of_regions) {
372 // re-calculate the necessary reserve
373 double reserve_regions_d = (double) new_number_of_regions * _reserve_factor;
374 // We use ceiling so that if reserve_regions_d is > 0.0 (but
375 // smaller than 1.0) we'll get 1.
376 _reserve_regions = (uint) ceil(reserve_regions_d);
377
378 _young_gen_sizer->heap_size_changed(new_number_of_regions);
379
380 _ihop_control->update_target_occupancy(new_number_of_regions * HeapRegion::GrainBytes);
381 }
382
383 uint G1CollectorPolicy::calculate_young_list_desired_min_length(
384 uint base_min_length) const {
385 uint desired_min_length = 0;
386 if (adaptive_young_list_length()) {
387 if (_alloc_rate_ms_seq->num() > 3) {
388 double now_sec = os::elapsedTime();
389 double when_ms = _mmu_tracker->when_max_gc_sec(now_sec) * 1000.0;
390 double alloc_rate_ms = predict_alloc_rate_ms();
391 desired_min_length = (uint) ceil(alloc_rate_ms * when_ms);
392 } else {
393 // otherwise we don't have enough info to make the prediction
394 }
395 }
396 desired_min_length += base_min_length;
397 // make sure we don't go below any user-defined minimum bound
398 return MAX2(_young_gen_sizer->min_desired_young_length(), desired_min_length);
399 }
400
401 uint G1CollectorPolicy::calculate_young_list_desired_max_length() const {
402 // Here, we might want to also take into account any additional
403 // constraints (i.e., user-defined minimum bound). Currently, we
404 // effectively don't set this bound.
405 return _young_gen_sizer->max_desired_young_length();
406 }
407
408 uint G1CollectorPolicy::update_young_list_max_and_target_length() {
409 return update_young_list_max_and_target_length(predict_rs_lengths());
410 }
411
412 uint G1CollectorPolicy::update_young_list_max_and_target_length(size_t rs_lengths) {
413 uint unbounded_target_length = update_young_list_target_length(rs_lengths);
414 update_max_gc_locker_expansion();
415 return unbounded_target_length;
416 }
417
418 uint G1CollectorPolicy::update_young_list_target_length(size_t rs_lengths) {
419 YoungTargetLengths young_lengths = young_list_target_lengths(rs_lengths);
420 _young_list_target_length = young_lengths.first;
421 return young_lengths.second;
422 }
423
424 G1CollectorPolicy::YoungTargetLengths G1CollectorPolicy::young_list_target_lengths(size_t rs_lengths) const {
425 YoungTargetLengths result;
426
427 // Calculate the absolute and desired min bounds first.
428
429 // This is how many young regions we already have (currently: the survivors).
430 const uint base_min_length = _g1->young_list()->survivor_length();
431 uint desired_min_length = calculate_young_list_desired_min_length(base_min_length);
432 // This is the absolute minimum young length. Ensure that we
433 // will at least have one eden region available for allocation.
434 uint absolute_min_length = base_min_length + MAX2(_g1->young_list()->eden_length(), (uint)1);
435 // If we shrank the young list target it should not shrink below the current size.
436 desired_min_length = MAX2(desired_min_length, absolute_min_length);
437 // Calculate the absolute and desired max bounds.
438
439 uint desired_max_length = calculate_young_list_desired_max_length();
440
441 uint young_list_target_length = 0;
442 if (adaptive_young_list_length()) {
443 if (collector_state()->gcs_are_young()) {
444 young_list_target_length =
445 calculate_young_list_target_length(rs_lengths,
446 base_min_length,
447 desired_min_length,
448 desired_max_length);
449 } else {
450 // Don't calculate anything and let the code below bound it to
451 // the desired_min_length, i.e., do the next GC as soon as
452 // possible to maximize how many old regions we can add to it.
453 }
454 } else {
455 // The user asked for a fixed young gen so we'll fix the young gen
456 // whether the next GC is young or mixed.
457 young_list_target_length = _young_list_fixed_length;
458 }
459
460 result.second = young_list_target_length;
461
462 // We will try our best not to "eat" into the reserve.
463 uint absolute_max_length = 0;
464 if (_free_regions_at_end_of_collection > _reserve_regions) {
465 absolute_max_length = _free_regions_at_end_of_collection - _reserve_regions;
466 }
467 if (desired_max_length > absolute_max_length) {
468 desired_max_length = absolute_max_length;
469 }
470
471 // Make sure we don't go over the desired max length, nor under the
472 // desired min length. In case they clash, desired_min_length wins
473 // which is why that test is second.
474 if (young_list_target_length > desired_max_length) {
475 young_list_target_length = desired_max_length;
476 }
477 if (young_list_target_length < desired_min_length) {
478 young_list_target_length = desired_min_length;
479 }
480
481 assert(young_list_target_length > base_min_length,
482 "we should be able to allocate at least one eden region");
483 assert(young_list_target_length >= absolute_min_length, "post-condition");
484
485 result.first = young_list_target_length;
486 return result;
487 }
488
489 uint
490 G1CollectorPolicy::calculate_young_list_target_length(size_t rs_lengths,
491 uint base_min_length,
492 uint desired_min_length,
493 uint desired_max_length) const {
494 assert(adaptive_young_list_length(), "pre-condition");
495 assert(collector_state()->gcs_are_young(), "only call this for young GCs");
496
497 // In case some edge-condition makes the desired max length too small...
498 if (desired_max_length <= desired_min_length) {
499 return desired_min_length;
500 }
501
502 // We'll adjust min_young_length and max_young_length not to include
503 // the already allocated young regions (i.e., so they reflect the
504 // min and max eden regions we'll allocate). The base_min_length
505 // will be reflected in the predictions by the
506 // survivor_regions_evac_time prediction.
507 assert(desired_min_length > base_min_length, "invariant");
508 uint min_young_length = desired_min_length - base_min_length;
509 assert(desired_max_length > base_min_length, "invariant");
510 uint max_young_length = desired_max_length - base_min_length;
511
512 double target_pause_time_ms = _mmu_tracker->max_gc_time() * 1000.0;
513 double survivor_regions_evac_time = predict_survivor_regions_evac_time();
514 size_t pending_cards = get_new_size_prediction(_pending_cards_seq);
515 size_t adj_rs_lengths = rs_lengths + predict_rs_length_diff();
516 size_t scanned_cards = predict_young_card_num(adj_rs_lengths);
517 double base_time_ms =
518 predict_base_elapsed_time_ms(pending_cards, scanned_cards) +
519 survivor_regions_evac_time;
520 uint available_free_regions = _free_regions_at_end_of_collection;
521 uint base_free_regions = 0;
522 if (available_free_regions > _reserve_regions) {
523 base_free_regions = available_free_regions - _reserve_regions;
524 }
525
526 // Here, we will make sure that the shortest young length that
527 // makes sense fits within the target pause time.
528
529 if (predict_will_fit(min_young_length, base_time_ms,
530 base_free_regions, target_pause_time_ms)) {
531 // The shortest young length will fit into the target pause time;
532 // we'll now check whether the absolute maximum number of young
533 // regions will fit in the target pause time. If not, we'll do
534 // a binary search between min_young_length and max_young_length.
535 if (predict_will_fit(max_young_length, base_time_ms,
536 base_free_regions, target_pause_time_ms)) {
537 // The maximum young length will fit into the target pause time.
538 // We are done so set min young length to the maximum length (as
539 // the result is assumed to be returned in min_young_length).
540 min_young_length = max_young_length;
541 } else {
542 // The maximum possible number of young regions will not fit within
543 // the target pause time so we'll search for the optimal
544 // length. The loop invariants are:
545 //
546 // min_young_length < max_young_length
547 // min_young_length is known to fit into the target pause time
548 // max_young_length is known not to fit into the target pause time
549 //
550 // Going into the loop we know the above hold as we've just
551 // checked them. Every time around the loop we check whether
552 // the middle value between min_young_length and
553 // max_young_length fits into the target pause time. If it
554 // does, it becomes the new min. If it doesn't, it becomes
555 // the new max. This way we maintain the loop invariants.
556
557 assert(min_young_length < max_young_length, "invariant");
558 uint diff = (max_young_length - min_young_length) / 2;
559 while (diff > 0) {
560 uint young_length = min_young_length + diff;
561 if (predict_will_fit(young_length, base_time_ms,
562 base_free_regions, target_pause_time_ms)) {
563 min_young_length = young_length;
564 } else {
565 max_young_length = young_length;
566 }
567 assert(min_young_length < max_young_length, "invariant");
568 diff = (max_young_length - min_young_length) / 2;
569 }
570 // The results is min_young_length which, according to the
571 // loop invariants, should fit within the target pause time.
572
573 // These are the post-conditions of the binary search above:
574 assert(min_young_length < max_young_length,
575 "otherwise we should have discovered that max_young_length "
576 "fits into the pause target and not done the binary search");
577 assert(predict_will_fit(min_young_length, base_time_ms,
578 base_free_regions, target_pause_time_ms),
579 "min_young_length, the result of the binary search, should "
580 "fit into the pause target");
581 assert(!predict_will_fit(min_young_length + 1, base_time_ms,
582 base_free_regions, target_pause_time_ms),
583 "min_young_length, the result of the binary search, should be "
584 "optimal, so no larger length should fit into the pause target");
585 }
586 } else {
587 // Even the minimum length doesn't fit into the pause time
588 // target, return it as the result nevertheless.
589 }
590 return base_min_length + min_young_length;
591 }
592
593 double G1CollectorPolicy::predict_survivor_regions_evac_time() const {
594 double survivor_regions_evac_time = 0.0;
595 for (HeapRegion * r = _g1->young_list()->first_survivor_region();
596 r != NULL && r != _g1->young_list()->last_survivor_region()->get_next_young_region();
597 r = r->get_next_young_region()) {
598 survivor_regions_evac_time += predict_region_elapsed_time_ms(r, collector_state()->gcs_are_young());
599 }
600 return survivor_regions_evac_time;
601 }
602
603 void G1CollectorPolicy::revise_young_list_target_length_if_necessary(size_t rs_lengths) {
604 guarantee( adaptive_young_list_length(), "should not call this otherwise" );
605
606 if (rs_lengths > _rs_lengths_prediction) {
607 // add 10% to avoid having to recalculate often
608 size_t rs_lengths_prediction = rs_lengths * 1100 / 1000;
609 update_rs_lengths_prediction(rs_lengths_prediction);
610
611 update_young_list_max_and_target_length(rs_lengths_prediction);
612 }
613 }
614
615 void G1CollectorPolicy::update_rs_lengths_prediction() {
616 update_rs_lengths_prediction(predict_rs_lengths());
617 }
618
619 void G1CollectorPolicy::update_rs_lengths_prediction(size_t prediction) {
620 if (collector_state()->gcs_are_young() && adaptive_young_list_length()) {
621 _rs_lengths_prediction = prediction;
622 }
623 }
624
625 #ifndef PRODUCT
626 bool G1CollectorPolicy::verify_young_ages() {
627 HeapRegion* head = _g1->young_list()->first_region();
628 return
629 verify_young_ages(head, _short_lived_surv_rate_group);
630 // also call verify_young_ages on any additional surv rate groups
631 }
632
633 bool
634 G1CollectorPolicy::verify_young_ages(HeapRegion* head,
635 SurvRateGroup *surv_rate_group) {
636 guarantee( surv_rate_group != NULL, "pre-condition" );
637
638 const char* name = surv_rate_group->name();
639 bool ret = true;
640 int prev_age = -1;
641
642 for (HeapRegion* curr = head;
643 curr != NULL;
644 curr = curr->get_next_young_region()) {
645 SurvRateGroup* group = curr->surv_rate_group();
646 if (group == NULL && !curr->is_survivor()) {
647 log_error(gc, verify)("## %s: encountered NULL surv_rate_group", name);
648 ret = false;
649 }
650
651 if (surv_rate_group == group) {
652 int age = curr->age_in_surv_rate_group();
653
654 if (age < 0) {
655 log_error(gc, verify)("## %s: encountered negative age", name);
656 ret = false;
657 }
658
659 if (age <= prev_age) {
660 log_error(gc, verify)("## %s: region ages are not strictly increasing (%d, %d)", name, age, prev_age);
661 ret = false;
662 }
663 prev_age = age;
664 }
665 }
666
667 return ret;
668 }
669 #endif // PRODUCT
670
671 void G1CollectorPolicy::record_full_collection_start() {
672 _full_collection_start_sec = os::elapsedTime();
673 // Release the future to-space so that it is available for compaction into.
674 collector_state()->set_full_collection(true);
675 }
676
677 void G1CollectorPolicy::record_full_collection_end() {
678 // Consider this like a collection pause for the purposes of allocation
679 // since last pause.
680 double end_sec = os::elapsedTime();
681 double full_gc_time_sec = end_sec - _full_collection_start_sec;
682 double full_gc_time_ms = full_gc_time_sec * 1000.0;
683
684 update_recent_gc_times(end_sec, full_gc_time_ms);
685
686 collector_state()->set_full_collection(false);
687
688 // "Nuke" the heuristics that control the young/mixed GC
689 // transitions and make sure we start with young GCs after the Full GC.
690 collector_state()->set_gcs_are_young(true);
691 collector_state()->set_last_young_gc(false);
692 collector_state()->set_initiate_conc_mark_if_possible(need_to_start_conc_mark("end of Full GC", 0));
693 collector_state()->set_during_initial_mark_pause(false);
694 collector_state()->set_in_marking_window(false);
695 collector_state()->set_in_marking_window_im(false);
696
697 _short_lived_surv_rate_group->start_adding_regions();
698 // also call this on any additional surv rate groups
699
700 _free_regions_at_end_of_collection = _g1->num_free_regions();
701 // Reset survivors SurvRateGroup.
702 _survivor_surv_rate_group->reset();
703 update_young_list_max_and_target_length();
704 update_rs_lengths_prediction();
705 cset_chooser()->clear();
706
707 _bytes_allocated_in_old_since_last_gc = 0;
708
709 record_pause(FullGC, _full_collection_start_sec, end_sec);
710 }
711
712 void G1CollectorPolicy::record_collection_pause_start(double start_time_sec) {
713 // We only need to do this here as the policy will only be applied
714 // to the GC we're about to start. so, no point is calculating this
715 // every time we calculate / recalculate the target young length.
716 update_survivors_policy();
717
718 assert(_g1->used() == _g1->recalculate_used(),
719 "sanity, used: " SIZE_FORMAT " recalculate_used: " SIZE_FORMAT,
720 _g1->used(), _g1->recalculate_used());
721
722 phase_times()->record_cur_collection_start_sec(start_time_sec);
723 _pending_cards = _g1->pending_card_num();
724
725 _collection_set->reset_bytes_used_before();
726 _bytes_copied_during_gc = 0;
727
728 collector_state()->set_last_gc_was_young(false);
729
730 // do that for any other surv rate groups
731 _short_lived_surv_rate_group->stop_adding_regions();
732 _survivors_age_table.clear();
733
734 assert( verify_young_ages(), "region age verification" );
735 }
736
737 void G1CollectorPolicy::record_concurrent_mark_init_end(double
738 mark_init_elapsed_time_ms) {
739 collector_state()->set_during_marking(true);
740 assert(!collector_state()->initiate_conc_mark_if_possible(), "we should have cleared it by now");
741 collector_state()->set_during_initial_mark_pause(false);
742 }
743
744 void G1CollectorPolicy::record_concurrent_mark_remark_start() {
745 _mark_remark_start_sec = os::elapsedTime();
746 collector_state()->set_during_marking(false);
747 }
748
749 void G1CollectorPolicy::record_concurrent_mark_remark_end() {
750 double end_time_sec = os::elapsedTime();
751 double elapsed_time_ms = (end_time_sec - _mark_remark_start_sec)*1000.0;
752 _concurrent_mark_remark_times_ms->add(elapsed_time_ms);
753 _prev_collection_pause_end_ms += elapsed_time_ms;
754
755 record_pause(Remark, _mark_remark_start_sec, end_time_sec);
756 }
757
758 void G1CollectorPolicy::record_concurrent_mark_cleanup_start() {
759 _mark_cleanup_start_sec = os::elapsedTime();
760 }
761
762 void G1CollectorPolicy::record_concurrent_mark_cleanup_completed() {
763 bool should_continue_with_reclaim = next_gc_should_be_mixed("request last young-only gc",
764 "skip last young-only gc");
765 collector_state()->set_last_young_gc(should_continue_with_reclaim);
766 // We skip the marking phase.
767 if (!should_continue_with_reclaim) {
768 abort_time_to_mixed_tracking();
769 }
770 collector_state()->set_in_marking_window(false);
771 }
772
773 double G1CollectorPolicy::average_time_ms(G1GCPhaseTimes::GCParPhases phase) const {
774 return phase_times()->average_time_ms(phase);
775 }
776
777 double G1CollectorPolicy::young_other_time_ms() const {
778 return phase_times()->young_cset_choice_time_ms() +
779 phase_times()->young_free_cset_time_ms();
780 }
781
782 double G1CollectorPolicy::non_young_other_time_ms() const {
783 return phase_times()->non_young_cset_choice_time_ms() +
784 phase_times()->non_young_free_cset_time_ms();
785
786 }
787
788 double G1CollectorPolicy::other_time_ms(double pause_time_ms) const {
789 return pause_time_ms -
790 average_time_ms(G1GCPhaseTimes::UpdateRS) -
791 average_time_ms(G1GCPhaseTimes::ScanRS) -
792 average_time_ms(G1GCPhaseTimes::ObjCopy) -
793 average_time_ms(G1GCPhaseTimes::Termination);
794 }
795
796 double G1CollectorPolicy::constant_other_time_ms(double pause_time_ms) const {
797 return other_time_ms(pause_time_ms) - young_other_time_ms() - non_young_other_time_ms();
798 }
799
800 CollectionSetChooser* G1CollectorPolicy::cset_chooser() const {
801 return _collection_set->cset_chooser();
802 }
803
804 bool G1CollectorPolicy::about_to_start_mixed_phase() const {
805 return _g1->concurrent_mark()->cmThread()->during_cycle() || collector_state()->last_young_gc();
806 }
807
808 bool G1CollectorPolicy::need_to_start_conc_mark(const char* source, size_t alloc_word_size) {
809 if (about_to_start_mixed_phase()) {
810 return false;
811 }
812
813 size_t marking_initiating_used_threshold = _ihop_control->get_conc_mark_start_threshold();
814
815 size_t cur_used_bytes = _g1->non_young_capacity_bytes();
816 size_t alloc_byte_size = alloc_word_size * HeapWordSize;
817 size_t marking_request_bytes = cur_used_bytes + alloc_byte_size;
818
819 bool result = false;
820 if (marking_request_bytes > marking_initiating_used_threshold) {
821 result = collector_state()->gcs_are_young() && !collector_state()->last_young_gc();
822 log_debug(gc, ergo, ihop)("%s occupancy: " SIZE_FORMAT "B allocation request: " SIZE_FORMAT "B threshold: " SIZE_FORMAT "B (%1.2f) source: %s",
823 result ? "Request concurrent cycle initiation (occupancy higher than threshold)" : "Do not request concurrent cycle initiation (still doing mixed collections)",
824 cur_used_bytes, alloc_byte_size, marking_initiating_used_threshold, (double) marking_initiating_used_threshold / _g1->capacity() * 100, source);
825 }
826
827 return result;
828 }
829
830 // Anything below that is considered to be zero
831 #define MIN_TIMER_GRANULARITY 0.0000001
832
833 void G1CollectorPolicy::record_collection_pause_end(double pause_time_ms, size_t cards_scanned, size_t heap_used_bytes_before_gc) {
834 double end_time_sec = os::elapsedTime();
835
836 size_t cur_used_bytes = _g1->used();
837 assert(cur_used_bytes == _g1->recalculate_used(), "It should!");
838 bool last_pause_included_initial_mark = false;
839 bool update_stats = !_g1->evacuation_failed();
840
841 NOT_PRODUCT(_short_lived_surv_rate_group->print());
842
843 record_pause(young_gc_pause_kind(), end_time_sec - pause_time_ms / 1000.0, end_time_sec);
844
845 last_pause_included_initial_mark = collector_state()->during_initial_mark_pause();
846 if (last_pause_included_initial_mark) {
847 record_concurrent_mark_init_end(0.0);
848 } else {
849 maybe_start_marking();
850 }
851
852 double app_time_ms = (phase_times()->cur_collection_start_sec() * 1000.0 - _prev_collection_pause_end_ms);
853 if (app_time_ms < MIN_TIMER_GRANULARITY) {
854 // This usually happens due to the timer not having the required
855 // granularity. Some Linuxes are the usual culprits.
856 // We'll just set it to something (arbitrarily) small.
857 app_time_ms = 1.0;
858 }
859
860 if (update_stats) {
861 // We maintain the invariant that all objects allocated by mutator
862 // threads will be allocated out of eden regions. So, we can use
863 // the eden region number allocated since the previous GC to
864 // calculate the application's allocate rate. The only exception
865 // to that is humongous objects that are allocated separately. But
866 // given that humongous object allocations do not really affect
867 // either the pause's duration nor when the next pause will take
868 // place we can safely ignore them here.
869 uint regions_allocated = _collection_set->eden_region_length();
870 double alloc_rate_ms = (double) regions_allocated / app_time_ms;
871 _alloc_rate_ms_seq->add(alloc_rate_ms);
872
873 double interval_ms =
874 (end_time_sec - _recent_prev_end_times_for_all_gcs_sec->oldest()) * 1000.0;
875 update_recent_gc_times(end_time_sec, pause_time_ms);
876 _recent_avg_pause_time_ratio = _recent_gc_times_ms->sum()/interval_ms;
877 if (recent_avg_pause_time_ratio() < 0.0 ||
878 (recent_avg_pause_time_ratio() - 1.0 > 0.0)) {
879 // Clip ratio between 0.0 and 1.0, and continue. This will be fixed in
880 // CR 6902692 by redoing the manner in which the ratio is incrementally computed.
881 if (_recent_avg_pause_time_ratio < 0.0) {
882 _recent_avg_pause_time_ratio = 0.0;
883 } else {
884 assert(_recent_avg_pause_time_ratio - 1.0 > 0.0, "Ctl-point invariant");
885 _recent_avg_pause_time_ratio = 1.0;
886 }
887 }
888
889 // Compute the ratio of just this last pause time to the entire time range stored
890 // in the vectors. Comparing this pause to the entire range, rather than only the
891 // most recent interval, has the effect of smoothing over a possible transient 'burst'
892 // of more frequent pauses that don't really reflect a change in heap occupancy.
893 // This reduces the likelihood of a needless heap expansion being triggered.
894 _last_pause_time_ratio =
895 (pause_time_ms * _recent_prev_end_times_for_all_gcs_sec->num()) / interval_ms;
896 }
897
898 bool new_in_marking_window = collector_state()->in_marking_window();
899 bool new_in_marking_window_im = false;
900 if (last_pause_included_initial_mark) {
901 new_in_marking_window = true;
902 new_in_marking_window_im = true;
903 }
904
905 if (collector_state()->last_young_gc()) {
906 // This is supposed to to be the "last young GC" before we start
907 // doing mixed GCs. Here we decide whether to start mixed GCs or not.
908 assert(!last_pause_included_initial_mark, "The last young GC is not allowed to be an initial mark GC");
909
910 if (next_gc_should_be_mixed("start mixed GCs",
911 "do not start mixed GCs")) {
912 collector_state()->set_gcs_are_young(false);
913 } else {
914 // We aborted the mixed GC phase early.
915 abort_time_to_mixed_tracking();
916 }
917
918 collector_state()->set_last_young_gc(false);
919 }
920
921 if (!collector_state()->last_gc_was_young()) {
922 // This is a mixed GC. Here we decide whether to continue doing
923 // mixed GCs or not.
924 if (!next_gc_should_be_mixed("continue mixed GCs",
925 "do not continue mixed GCs")) {
926 collector_state()->set_gcs_are_young(true);
927
928 maybe_start_marking();
929 }
930 }
931
932 _short_lived_surv_rate_group->start_adding_regions();
933 // Do that for any other surv rate groups
934
935 double scan_hcc_time_ms = ConcurrentG1Refine::hot_card_cache_enabled() ? average_time_ms(G1GCPhaseTimes::ScanHCC) : 0.0;
936
937 if (update_stats) {
938 double cost_per_card_ms = 0.0;
939 if (_pending_cards > 0) {
940 cost_per_card_ms = (average_time_ms(G1GCPhaseTimes::UpdateRS) - scan_hcc_time_ms) / (double) _pending_cards;
941 _cost_per_card_ms_seq->add(cost_per_card_ms);
942 }
943 _cost_scan_hcc_seq->add(scan_hcc_time_ms);
944
945 double cost_per_entry_ms = 0.0;
946 if (cards_scanned > 10) {
947 cost_per_entry_ms = average_time_ms(G1GCPhaseTimes::ScanRS) / (double) cards_scanned;
948 if (collector_state()->last_gc_was_young()) {
949 _cost_per_entry_ms_seq->add(cost_per_entry_ms);
950 } else {
951 _mixed_cost_per_entry_ms_seq->add(cost_per_entry_ms);
952 }
953 }
954
955 if (_max_rs_lengths > 0) {
956 double cards_per_entry_ratio =
957 (double) cards_scanned / (double) _max_rs_lengths;
958 if (collector_state()->last_gc_was_young()) {
959 _young_cards_per_entry_ratio_seq->add(cards_per_entry_ratio);
960 } else {
961 _mixed_cards_per_entry_ratio_seq->add(cards_per_entry_ratio);
962 }
963 }
964
965 // This is defensive. For a while _max_rs_lengths could get
966 // smaller than _recorded_rs_lengths which was causing
967 // rs_length_diff to get very large and mess up the RSet length
968 // predictions. The reason was unsafe concurrent updates to the
969 // _inc_cset_recorded_rs_lengths field which the code below guards
970 // against (see CR 7118202). This bug has now been fixed (see CR
971 // 7119027). However, I'm still worried that
972 // _inc_cset_recorded_rs_lengths might still end up somewhat
973 // inaccurate. The concurrent refinement thread calculates an
974 // RSet's length concurrently with other CR threads updating it
975 // which might cause it to calculate the length incorrectly (if,
976 // say, it's in mid-coarsening). So I'll leave in the defensive
977 // conditional below just in case.
978 size_t rs_length_diff = 0;
979 size_t recorded_rs_lengths = _collection_set->recorded_rs_lengths();
980 if (_max_rs_lengths > recorded_rs_lengths) {
981 rs_length_diff = _max_rs_lengths - recorded_rs_lengths;
982 }
983 _rs_length_diff_seq->add((double) rs_length_diff);
984
985 size_t freed_bytes = heap_used_bytes_before_gc - cur_used_bytes;
986 size_t copied_bytes = _collection_set->bytes_used_before() - freed_bytes;
987 double cost_per_byte_ms = 0.0;
988
989 if (copied_bytes > 0) {
990 cost_per_byte_ms = average_time_ms(G1GCPhaseTimes::ObjCopy) / (double) copied_bytes;
991 if (collector_state()->in_marking_window()) {
992 _cost_per_byte_ms_during_cm_seq->add(cost_per_byte_ms);
993 } else {
994 _cost_per_byte_ms_seq->add(cost_per_byte_ms);
995 }
996 }
997
998 if (_collection_set->young_region_length() > 0) {
999 _young_other_cost_per_region_ms_seq->add(young_other_time_ms() /
1000 _collection_set->young_region_length());
1001 }
1002
1003 if (_collection_set->old_region_length() > 0) {
1004 _non_young_other_cost_per_region_ms_seq->add(non_young_other_time_ms() /
1005 _collection_set->old_region_length());
1006 }
1007
1008 _constant_other_time_ms_seq->add(constant_other_time_ms(pause_time_ms));
1009
1010 _pending_cards_seq->add((double) _pending_cards);
1011 _rs_lengths_seq->add((double) _max_rs_lengths);
1012 }
1013
1014 collector_state()->set_in_marking_window(new_in_marking_window);
1015 collector_state()->set_in_marking_window_im(new_in_marking_window_im);
1016 _free_regions_at_end_of_collection = _g1->num_free_regions();
1017 // IHOP control wants to know the expected young gen length if it were not
1018 // restrained by the heap reserve. Using the actual length would make the
1019 // prediction too small and the limit the young gen every time we get to the
1020 // predicted target occupancy.
1021 size_t last_unrestrained_young_length = update_young_list_max_and_target_length();
1022 update_rs_lengths_prediction();
1023
1024 update_ihop_prediction(app_time_ms / 1000.0,
1025 _bytes_allocated_in_old_since_last_gc,
1026 last_unrestrained_young_length * HeapRegion::GrainBytes);
1027 _bytes_allocated_in_old_since_last_gc = 0;
1028
1029 _ihop_control->send_trace_event(_g1->gc_tracer_stw());
1030
1031 // Note that _mmu_tracker->max_gc_time() returns the time in seconds.
1032 double update_rs_time_goal_ms = _mmu_tracker->max_gc_time() * MILLIUNITS * G1RSetUpdatingPauseTimePercent / 100.0;
1033
1034 if (update_rs_time_goal_ms < scan_hcc_time_ms) {
1035 log_debug(gc, ergo, refine)("Adjust concurrent refinement thresholds (scanning the HCC expected to take longer than Update RS time goal)."
1036 "Update RS time goal: %1.2fms Scan HCC time: %1.2fms",
1037 update_rs_time_goal_ms, scan_hcc_time_ms);
1038
1039 update_rs_time_goal_ms = 0;
1040 } else {
1041 update_rs_time_goal_ms -= scan_hcc_time_ms;
1042 }
1043 adjust_concurrent_refinement(average_time_ms(G1GCPhaseTimes::UpdateRS) - scan_hcc_time_ms,
1044 phase_times()->sum_thread_work_items(G1GCPhaseTimes::UpdateRS),
1045 update_rs_time_goal_ms);
1046
1047 cset_chooser()->verify();
1048 }
1049
1050 G1IHOPControl* G1CollectorPolicy::create_ihop_control() const {
1051 if (G1UseAdaptiveIHOP) {
1052 return new G1AdaptiveIHOPControl(InitiatingHeapOccupancyPercent,
1053 &_predictor,
1054 G1ReservePercent,
1055 G1HeapWastePercent);
1056 } else {
1057 return new G1StaticIHOPControl(InitiatingHeapOccupancyPercent);
1058 }
1059 }
1060
1061 void G1CollectorPolicy::update_ihop_prediction(double mutator_time_s,
1062 size_t mutator_alloc_bytes,
1063 size_t young_gen_size) {
1064 // Always try to update IHOP prediction. Even evacuation failures give information
1065 // about e.g. whether to start IHOP earlier next time.
1066
1067 // Avoid using really small application times that might create samples with
1068 // very high or very low values. They may be caused by e.g. back-to-back gcs.
1069 double const min_valid_time = 1e-6;
1070
1071 bool report = false;
1072
1073 double marking_to_mixed_time = -1.0;
1074 if (!collector_state()->last_gc_was_young() && _initial_mark_to_mixed.has_result()) {
1075 marking_to_mixed_time = _initial_mark_to_mixed.last_marking_time();
1076 assert(marking_to_mixed_time > 0.0,
1077 "Initial mark to mixed time must be larger than zero but is %.3f",
1078 marking_to_mixed_time);
1079 if (marking_to_mixed_time > min_valid_time) {
1080 _ihop_control->update_marking_length(marking_to_mixed_time);
1081 report = true;
1082 }
1083 }
1084
1085 // As an approximation for the young gc promotion rates during marking we use
1086 // all of them. In many applications there are only a few if any young gcs during
1087 // marking, which makes any prediction useless. This increases the accuracy of the
1088 // prediction.
1089 if (collector_state()->last_gc_was_young() && mutator_time_s > min_valid_time) {
1090 _ihop_control->update_allocation_info(mutator_time_s, mutator_alloc_bytes, young_gen_size);
1091 report = true;
1092 }
1093
1094 if (report) {
1095 report_ihop_statistics();
1096 }
1097 }
1098
1099 void G1CollectorPolicy::report_ihop_statistics() {
1100 _ihop_control->print();
1101 }
1102
1103 void G1CollectorPolicy::print_phases() {
1104 phase_times()->print();
1105 }
1106
1107 void G1CollectorPolicy::adjust_concurrent_refinement(double update_rs_time,
1108 double update_rs_processed_buffers,
1109 double goal_ms) {
1110 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set();
1111 ConcurrentG1Refine *cg1r = G1CollectedHeap::heap()->concurrent_g1_refine();
1112
1113 if (G1UseAdaptiveConcRefinement) {
1114 const int k_gy = 3, k_gr = 6;
1115 const double inc_k = 1.1, dec_k = 0.9;
1116
1117 size_t g = cg1r->green_zone();
1118 if (update_rs_time > goal_ms) {
1119 g = (size_t)(g * dec_k); // Can become 0, that's OK. That would mean a mutator-only processing.
1120 } else {
1121 if (update_rs_time < goal_ms && update_rs_processed_buffers > g) {
1122 g = (size_t)MAX2(g * inc_k, g + 1.0);
1123 }
1124 }
1125 // Change the refinement threads params
1126 cg1r->set_green_zone(g);
1127 cg1r->set_yellow_zone(g * k_gy);
1128 cg1r->set_red_zone(g * k_gr);
1129 cg1r->reinitialize_threads();
1130
1131 size_t processing_threshold_delta = MAX2<size_t>(cg1r->green_zone() * _predictor.sigma(), 1);
1132 size_t processing_threshold = MIN2(cg1r->green_zone() + processing_threshold_delta,
1133 cg1r->yellow_zone());
1134 // Change the barrier params
1135 dcqs.set_process_completed_threshold((int)processing_threshold);
1136 dcqs.set_max_completed_queue((int)cg1r->red_zone());
1137 }
1138
1139 size_t curr_queue_size = dcqs.completed_buffers_num();
1140 if (curr_queue_size >= cg1r->yellow_zone()) {
1141 dcqs.set_completed_queue_padding(curr_queue_size);
1142 } else {
1143 dcqs.set_completed_queue_padding(0);
1144 }
1145 dcqs.notify_if_necessary();
1146 }
1147
1148 size_t G1CollectorPolicy::predict_rs_lengths() const {
1149 return get_new_size_prediction(_rs_lengths_seq);
1150 }
1151
1152 size_t G1CollectorPolicy::predict_rs_length_diff() const {
1153 return get_new_size_prediction(_rs_length_diff_seq);
1154 }
1155
1156 double G1CollectorPolicy::predict_alloc_rate_ms() const {
1157 return get_new_prediction(_alloc_rate_ms_seq);
1158 }
1159
1160 double G1CollectorPolicy::predict_cost_per_card_ms() const {
1161 return get_new_prediction(_cost_per_card_ms_seq);
1162 }
1163
1164 double G1CollectorPolicy::predict_scan_hcc_ms() const {
1165 return get_new_prediction(_cost_scan_hcc_seq);
1166 }
1167
1168 double G1CollectorPolicy::predict_rs_update_time_ms(size_t pending_cards) const {
1169 return pending_cards * predict_cost_per_card_ms() + predict_scan_hcc_ms();
1170 }
1171
1172 double G1CollectorPolicy::predict_young_cards_per_entry_ratio() const {
1173 return get_new_prediction(_young_cards_per_entry_ratio_seq);
1174 }
1175
1176 double G1CollectorPolicy::predict_mixed_cards_per_entry_ratio() const {
1177 if (_mixed_cards_per_entry_ratio_seq->num() < 2) {
1178 return predict_young_cards_per_entry_ratio();
1179 } else {
1180 return get_new_prediction(_mixed_cards_per_entry_ratio_seq);
1181 }
1182 }
1183
1184 size_t G1CollectorPolicy::predict_young_card_num(size_t rs_length) const {
1185 return (size_t) (rs_length * predict_young_cards_per_entry_ratio());
1186 }
1187
1188 size_t G1CollectorPolicy::predict_non_young_card_num(size_t rs_length) const {
1189 return (size_t)(rs_length * predict_mixed_cards_per_entry_ratio());
1190 }
1191
1192 double G1CollectorPolicy::predict_rs_scan_time_ms(size_t card_num) const {
1193 if (collector_state()->gcs_are_young()) {
1194 return card_num * get_new_prediction(_cost_per_entry_ms_seq);
1195 } else {
1196 return predict_mixed_rs_scan_time_ms(card_num);
1197 }
1198 }
1199
1200 double G1CollectorPolicy::predict_mixed_rs_scan_time_ms(size_t card_num) const {
1201 if (_mixed_cost_per_entry_ms_seq->num() < 3) {
1202 return card_num * get_new_prediction(_cost_per_entry_ms_seq);
1203 } else {
1204 return card_num * get_new_prediction(_mixed_cost_per_entry_ms_seq);
1205 }
1206 }
1207
1208 double G1CollectorPolicy::predict_object_copy_time_ms_during_cm(size_t bytes_to_copy) const {
1209 if (_cost_per_byte_ms_during_cm_seq->num() < 3) {
1210 return (1.1 * bytes_to_copy) * get_new_prediction(_cost_per_byte_ms_seq);
1211 } else {
1212 return bytes_to_copy * get_new_prediction(_cost_per_byte_ms_during_cm_seq);
1213 }
1214 }
1215
1216 double G1CollectorPolicy::predict_object_copy_time_ms(size_t bytes_to_copy) const {
1217 if (collector_state()->during_concurrent_mark()) {
1218 return predict_object_copy_time_ms_during_cm(bytes_to_copy);
1219 } else {
1220 return bytes_to_copy * get_new_prediction(_cost_per_byte_ms_seq);
1221 }
1222 }
1223
1224 double G1CollectorPolicy::predict_constant_other_time_ms() const {
1225 return get_new_prediction(_constant_other_time_ms_seq);
1226 }
1227
1228 double G1CollectorPolicy::predict_young_other_time_ms(size_t young_num) const {
1229 return young_num * get_new_prediction(_young_other_cost_per_region_ms_seq);
1230 }
1231
1232 double G1CollectorPolicy::predict_non_young_other_time_ms(size_t non_young_num) const {
1233 return non_young_num * get_new_prediction(_non_young_other_cost_per_region_ms_seq);
1234 }
1235
1236 double G1CollectorPolicy::predict_remark_time_ms() const {
1237 return get_new_prediction(_concurrent_mark_remark_times_ms);
1238 }
1239
1240 double G1CollectorPolicy::predict_cleanup_time_ms() const {
1241 return get_new_prediction(_concurrent_mark_cleanup_times_ms);
1242 }
1243
1244 double G1CollectorPolicy::predict_yg_surv_rate(int age, SurvRateGroup* surv_rate_group) const {
1245 TruncatedSeq* seq = surv_rate_group->get_seq(age);
1246 guarantee(seq->num() > 0, "There should be some young gen survivor samples available. Tried to access with age %d", age);
1247 double pred = get_new_prediction(seq);
1248 if (pred > 1.0) {
1249 pred = 1.0;
1250 }
1251 return pred;
1252 }
1253
1254 double G1CollectorPolicy::predict_yg_surv_rate(int age) const {
1255 return predict_yg_surv_rate(age, _short_lived_surv_rate_group);
1256 }
1257
1258 double G1CollectorPolicy::accum_yg_surv_rate_pred(int age) const {
1259 return _short_lived_surv_rate_group->accum_surv_rate_pred(age);
1260 }
1261
1262 double G1CollectorPolicy::predict_base_elapsed_time_ms(size_t pending_cards,
1263 size_t scanned_cards) const {
1264 return
1265 predict_rs_update_time_ms(pending_cards) +
1266 predict_rs_scan_time_ms(scanned_cards) +
1267 predict_constant_other_time_ms();
1268 }
1269
1270 double G1CollectorPolicy::predict_base_elapsed_time_ms(size_t pending_cards) const {
1271 size_t rs_length = predict_rs_lengths() + predict_rs_length_diff();
1272 size_t card_num;
1273 if (collector_state()->gcs_are_young()) {
1274 card_num = predict_young_card_num(rs_length);
1275 } else {
1276 card_num = predict_non_young_card_num(rs_length);
1277 }
1278 return predict_base_elapsed_time_ms(pending_cards, card_num);
1279 }
1280
1281 size_t G1CollectorPolicy::predict_bytes_to_copy(HeapRegion* hr) const {
1282 size_t bytes_to_copy;
1283 if (hr->is_marked())
1284 bytes_to_copy = hr->max_live_bytes();
1285 else {
1286 assert(hr->is_young() && hr->age_in_surv_rate_group() != -1, "invariant");
1287 int age = hr->age_in_surv_rate_group();
1288 double yg_surv_rate = predict_yg_surv_rate(age, hr->surv_rate_group());
1289 bytes_to_copy = (size_t) (hr->used() * yg_surv_rate);
1290 }
1291 return bytes_to_copy;
1292 }
1293
1294 double G1CollectorPolicy::predict_region_elapsed_time_ms(HeapRegion* hr,
1295 bool for_young_gc) const {
1296 size_t rs_length = hr->rem_set()->occupied();
1297 size_t card_num;
1298
1299 // Predicting the number of cards is based on which type of GC
1300 // we're predicting for.
1301 if (for_young_gc) {
1302 card_num = predict_young_card_num(rs_length);
1303 } else {
1304 card_num = predict_non_young_card_num(rs_length);
1305 }
1306 size_t bytes_to_copy = predict_bytes_to_copy(hr);
1307
1308 double region_elapsed_time_ms =
1309 predict_rs_scan_time_ms(card_num) +
1310 predict_object_copy_time_ms(bytes_to_copy);
1311
1312 // The prediction of the "other" time for this region is based
1313 // upon the region type and NOT the GC type.
1314 if (hr->is_young()) {
1315 region_elapsed_time_ms += predict_young_other_time_ms(1);
1316 } else {
1317 region_elapsed_time_ms += predict_non_young_other_time_ms(1);
1318 }
1319 return region_elapsed_time_ms;
1320 }
1321
1322 void G1CollectorPolicy::update_recent_gc_times(double end_time_sec,
1323 double elapsed_ms) {
1324 _recent_gc_times_ms->add(elapsed_ms);
1325 _recent_prev_end_times_for_all_gcs_sec->add(end_time_sec);
1326 _prev_collection_pause_end_ms = end_time_sec * 1000.0;
1327 }
1328
1329 void G1CollectorPolicy::clear_ratio_check_data() {
1330 _ratio_over_threshold_count = 0;
1331 _ratio_over_threshold_sum = 0.0;
1332 _pauses_since_start = 0;
1333 }
1334
1335 size_t G1CollectorPolicy::expansion_amount() {
1336 double recent_gc_overhead = recent_avg_pause_time_ratio() * 100.0;
1337 double last_gc_overhead = _last_pause_time_ratio * 100.0;
1338 double threshold = _gc_overhead_perc;
1339 size_t expand_bytes = 0;
1340
1341 // If the heap is at less than half its maximum size, scale the threshold down,
1342 // to a limit of 1. Thus the smaller the heap is, the more likely it is to expand,
1343 // though the scaling code will likely keep the increase small.
1344 if (_g1->capacity() <= _g1->max_capacity() / 2) {
1345 threshold *= (double)_g1->capacity() / (double)(_g1->max_capacity() / 2);
1346 threshold = MAX2(threshold, 1.0);
1347 }
1348
1349 // If the last GC time ratio is over the threshold, increment the count of
1350 // times it has been exceeded, and add this ratio to the sum of exceeded
1351 // ratios.
1352 if (last_gc_overhead > threshold) {
1353 _ratio_over_threshold_count++;
1354 _ratio_over_threshold_sum += last_gc_overhead;
1355 }
1356
1357 // Check if we've had enough GC time ratio checks that were over the
1358 // threshold to trigger an expansion. We'll also expand if we've
1359 // reached the end of the history buffer and the average of all entries
1360 // is still over the threshold. This indicates a smaller number of GCs were
1361 // long enough to make the average exceed the threshold.
1362 bool filled_history_buffer = _pauses_since_start == NumPrevPausesForHeuristics;
1363 if ((_ratio_over_threshold_count == MinOverThresholdForGrowth) ||
1364 (filled_history_buffer && (recent_gc_overhead > threshold))) {
1365 size_t min_expand_bytes = HeapRegion::GrainBytes;
1366 size_t reserved_bytes = _g1->max_capacity();
1367 size_t committed_bytes = _g1->capacity();
1368 size_t uncommitted_bytes = reserved_bytes - committed_bytes;
1369 size_t expand_bytes_via_pct =
1370 uncommitted_bytes * G1ExpandByPercentOfAvailable / 100;
1371 double scale_factor = 1.0;
1372
1373 // If the current size is less than 1/4 of the Initial heap size, expand
1374 // by half of the delta between the current and Initial sizes. IE, grow
1375 // back quickly.
1376 //
1377 // Otherwise, take the current size, or G1ExpandByPercentOfAvailable % of
1378 // the available expansion space, whichever is smaller, as the base
1379 // expansion size. Then possibly scale this size according to how much the
1380 // threshold has (on average) been exceeded by. If the delta is small
1381 // (less than the StartScaleDownAt value), scale the size down linearly, but
1382 // not by less than MinScaleDownFactor. If the delta is large (greater than
1383 // the StartScaleUpAt value), scale up, but adding no more than MaxScaleUpFactor
1384 // times the base size. The scaling will be linear in the range from
1385 // StartScaleUpAt to (StartScaleUpAt + ScaleUpRange). In other words,
1386 // ScaleUpRange sets the rate of scaling up.
1387 if (committed_bytes < InitialHeapSize / 4) {
1388 expand_bytes = (InitialHeapSize - committed_bytes) / 2;
1389 } else {
1390 double const MinScaleDownFactor = 0.2;
1391 double const MaxScaleUpFactor = 2;
1392 double const StartScaleDownAt = _gc_overhead_perc;
1393 double const StartScaleUpAt = _gc_overhead_perc * 1.5;
1394 double const ScaleUpRange = _gc_overhead_perc * 2.0;
1395
1396 double ratio_delta;
1397 if (filled_history_buffer) {
1398 ratio_delta = recent_gc_overhead - threshold;
1399 } else {
1400 ratio_delta = (_ratio_over_threshold_sum/_ratio_over_threshold_count) - threshold;
1401 }
1402
1403 expand_bytes = MIN2(expand_bytes_via_pct, committed_bytes);
1404 if (ratio_delta < StartScaleDownAt) {
1405 scale_factor = ratio_delta / StartScaleDownAt;
1406 scale_factor = MAX2(scale_factor, MinScaleDownFactor);
1407 } else if (ratio_delta > StartScaleUpAt) {
1408 scale_factor = 1 + ((ratio_delta - StartScaleUpAt) / ScaleUpRange);
1409 scale_factor = MIN2(scale_factor, MaxScaleUpFactor);
1410 }
1411 }
1412
1413 log_debug(gc, ergo, heap)("Attempt heap expansion (recent GC overhead higher than threshold after GC) "
1414 "recent GC overhead: %1.2f %% threshold: %1.2f %% uncommitted: " SIZE_FORMAT "B base expansion amount and scale: " SIZE_FORMAT "B (%1.2f%%)",
1415 recent_gc_overhead, threshold, uncommitted_bytes, expand_bytes, scale_factor * 100);
1416
1417 expand_bytes = static_cast<size_t>(expand_bytes * scale_factor);
1418
1419 // Ensure the expansion size is at least the minimum growth amount
1420 // and at most the remaining uncommitted byte size.
1421 expand_bytes = MAX2(expand_bytes, min_expand_bytes);
1422 expand_bytes = MIN2(expand_bytes, uncommitted_bytes);
1423
1424 clear_ratio_check_data();
1425 } else {
1426 // An expansion was not triggered. If we've started counting, increment
1427 // the number of checks we've made in the current window. If we've
1428 // reached the end of the window without resizing, clear the counters to
1429 // start again the next time we see a ratio above the threshold.
1430 if (_ratio_over_threshold_count > 0) {
1431 _pauses_since_start++;
1432 if (_pauses_since_start > NumPrevPausesForHeuristics) {
1433 clear_ratio_check_data();
1434 }
1435 }
1436 }
1437
1438 return expand_bytes;
1439 }
1440
1441 void G1CollectorPolicy::print_yg_surv_rate_info() const {
1442 #ifndef PRODUCT
1443 _short_lived_surv_rate_group->print_surv_rate_summary();
1444 // add this call for any other surv rate groups
1445 #endif // PRODUCT
1446 }
1447
1448 bool G1CollectorPolicy::is_young_list_full() const {
1449 uint young_list_length = _g1->young_list()->length();
1450 uint young_list_target_length = _young_list_target_length;
1451 return young_list_length >= young_list_target_length;
1452 }
1453
1454 bool G1CollectorPolicy::can_expand_young_list() const {
1455 uint young_list_length = _g1->young_list()->length();
1456 uint young_list_max_length = _young_list_max_length;
1457 return young_list_length < young_list_max_length;
1458 }
1459
1460 bool G1CollectorPolicy::adaptive_young_list_length() const {
1461 return _young_gen_sizer->adaptive_young_list_length();
1462 }
1463
1464 void G1CollectorPolicy::update_max_gc_locker_expansion() {
1465 uint expansion_region_num = 0;
1466 if (GCLockerEdenExpansionPercent > 0) {
1467 double perc = (double) GCLockerEdenExpansionPercent / 100.0;
1468 double expansion_region_num_d = perc * (double) _young_list_target_length;
1469 // We use ceiling so that if expansion_region_num_d is > 0.0 (but
1470 // less than 1.0) we'll get 1.
1471 expansion_region_num = (uint) ceil(expansion_region_num_d);
1472 } else {
1473 assert(expansion_region_num == 0, "sanity");
1474 }
1475 _young_list_max_length = _young_list_target_length + expansion_region_num;
1476 assert(_young_list_target_length <= _young_list_max_length, "post-condition");
1477 }
1478
1479 // Calculates survivor space parameters.
1480 void G1CollectorPolicy::update_survivors_policy() {
1481 double max_survivor_regions_d =
1482 (double) _young_list_target_length / (double) SurvivorRatio;
1483 // We use ceiling so that if max_survivor_regions_d is > 0.0 (but
1484 // smaller than 1.0) we'll get 1.
1485 _max_survivor_regions = (uint) ceil(max_survivor_regions_d);
1486
1487 _tenuring_threshold = _survivors_age_table.compute_tenuring_threshold(
1488 HeapRegion::GrainWords * _max_survivor_regions, counters());
1489 }
1490
1491 bool G1CollectorPolicy::force_initial_mark_if_outside_cycle(GCCause::Cause gc_cause) {
1492 // We actually check whether we are marking here and not if we are in a
1493 // reclamation phase. This means that we will schedule a concurrent mark
1494 // even while we are still in the process of reclaiming memory.
1495 bool during_cycle = _g1->concurrent_mark()->cmThread()->during_cycle();
1496 if (!during_cycle) {
1497 log_debug(gc, ergo)("Request concurrent cycle initiation (requested by GC cause). GC cause: %s", GCCause::to_string(gc_cause));
1498 collector_state()->set_initiate_conc_mark_if_possible(true);
1499 return true;
1500 } else {
1501 log_debug(gc, ergo)("Do not request concurrent cycle initiation (concurrent cycle already in progress). GC cause: %s", GCCause::to_string(gc_cause));
1502 return false;
1503 }
1504 }
1505
1506 void G1CollectorPolicy::initiate_conc_mark() {
1507 collector_state()->set_during_initial_mark_pause(true);
1508 collector_state()->set_initiate_conc_mark_if_possible(false);
1509 }
1510
1511 void G1CollectorPolicy::decide_on_conc_mark_initiation() {
1512 // We are about to decide on whether this pause will be an
1513 // initial-mark pause.
1514
1515 // First, collector_state()->during_initial_mark_pause() should not be already set. We
1516 // will set it here if we have to. However, it should be cleared by
1517 // the end of the pause (it's only set for the duration of an
1518 // initial-mark pause).
1519 assert(!collector_state()->during_initial_mark_pause(), "pre-condition");
1520
1521 if (collector_state()->initiate_conc_mark_if_possible()) {
1522 // We had noticed on a previous pause that the heap occupancy has
1523 // gone over the initiating threshold and we should start a
1524 // concurrent marking cycle. So we might initiate one.
1525
1526 if (!about_to_start_mixed_phase() && collector_state()->gcs_are_young()) {
1527 // Initiate a new initial mark if there is no marking or reclamation going on.
1528 initiate_conc_mark();
1529 log_debug(gc, ergo)("Initiate concurrent cycle (concurrent cycle initiation requested)");
1530 } else if (_g1->is_user_requested_concurrent_full_gc(_g1->gc_cause())) {
1531 // Initiate a user requested initial mark. An initial mark must be young only
1532 // GC, so the collector state must be updated to reflect this.
1533 collector_state()->set_gcs_are_young(true);
1534 collector_state()->set_last_young_gc(false);
1535
1536 abort_time_to_mixed_tracking();
1537 initiate_conc_mark();
1538 log_debug(gc, ergo)("Initiate concurrent cycle (user requested concurrent cycle)");
1539 } else {
1540 // The concurrent marking thread is still finishing up the
1541 // previous cycle. If we start one right now the two cycles
1542 // overlap. In particular, the concurrent marking thread might
1543 // be in the process of clearing the next marking bitmap (which
1544 // we will use for the next cycle if we start one). Starting a
1545 // cycle now will be bad given that parts of the marking
1546 // information might get cleared by the marking thread. And we
1547 // cannot wait for the marking thread to finish the cycle as it
1548 // periodically yields while clearing the next marking bitmap
1549 // and, if it's in a yield point, it's waiting for us to
1550 // finish. So, at this point we will not start a cycle and we'll
1551 // let the concurrent marking thread complete the last one.
1552 log_debug(gc, ergo)("Do not initiate concurrent cycle (concurrent cycle already in progress)");
1553 }
1554 }
1555 }
1556
1557 class ParKnownGarbageHRClosure: public HeapRegionClosure {
1558 G1CollectedHeap* _g1h;
1559 CSetChooserParUpdater _cset_updater;
1560
1561 public:
1562 ParKnownGarbageHRClosure(CollectionSetChooser* hrSorted,
1563 uint chunk_size) :
1564 _g1h(G1CollectedHeap::heap()),
1565 _cset_updater(hrSorted, true /* parallel */, chunk_size) { }
1566
1567 bool doHeapRegion(HeapRegion* r) {
1568 // Do we have any marking information for this region?
1569 if (r->is_marked()) {
1570 // We will skip any region that's currently used as an old GC
1571 // alloc region (we should not consider those for collection
1572 // before we fill them up).
1573 if (_cset_updater.should_add(r) && !_g1h->is_old_gc_alloc_region(r)) {
1574 _cset_updater.add_region(r);
1575 }
1576 }
1577 return false;
1578 }
1579 };
1580
1581 class ParKnownGarbageTask: public AbstractGangTask {
1582 CollectionSetChooser* _hrSorted;
1583 uint _chunk_size;
1584 G1CollectedHeap* _g1;
1585 HeapRegionClaimer _hrclaimer;
1586
1587 public:
1588 ParKnownGarbageTask(CollectionSetChooser* hrSorted, uint chunk_size, uint n_workers) :
1589 AbstractGangTask("ParKnownGarbageTask"),
1590 _hrSorted(hrSorted), _chunk_size(chunk_size),
1591 _g1(G1CollectedHeap::heap()), _hrclaimer(n_workers) {}
1592
1593 void work(uint worker_id) {
1594 ParKnownGarbageHRClosure parKnownGarbageCl(_hrSorted, _chunk_size);
1595 _g1->heap_region_par_iterate(&parKnownGarbageCl, worker_id, &_hrclaimer);
1596 }
1597 };
1598
1599 uint G1CollectorPolicy::calculate_parallel_work_chunk_size(uint n_workers, uint n_regions) const {
1600 assert(n_workers > 0, "Active gc workers should be greater than 0");
1601 const uint overpartition_factor = 4;
1602 const uint min_chunk_size = MAX2(n_regions / n_workers, 1U);
1603 return MAX2(n_regions / (n_workers * overpartition_factor), min_chunk_size);
1604 }
1605
1606 void G1CollectorPolicy::record_concurrent_mark_cleanup_end() {
1607 cset_chooser()->clear();
1608
1609 WorkGang* workers = _g1->workers();
1610 uint n_workers = workers->active_workers();
1611
1612 uint n_regions = _g1->num_regions();
1613 uint chunk_size = calculate_parallel_work_chunk_size(n_workers, n_regions);
1614 cset_chooser()->prepare_for_par_region_addition(n_workers, n_regions, chunk_size);
1615 ParKnownGarbageTask par_known_garbage_task(cset_chooser(), chunk_size, n_workers);
1616 workers->run_task(&par_known_garbage_task);
1617
1618 cset_chooser()->sort_regions();
1619
1620 double end_sec = os::elapsedTime();
1621 double elapsed_time_ms = (end_sec - _mark_cleanup_start_sec) * 1000.0;
1622 _concurrent_mark_cleanup_times_ms->add(elapsed_time_ms);
1623 _prev_collection_pause_end_ms += elapsed_time_ms;
1624
1625 record_pause(Cleanup, _mark_cleanup_start_sec, end_sec);
1626 }
1627
1628 double G1CollectorPolicy::reclaimable_bytes_perc(size_t reclaimable_bytes) const {
1629 // Returns the given amount of reclaimable bytes (that represents
1630 // the amount of reclaimable space still to be collected) as a
1631 // percentage of the current heap capacity.
1632 size_t capacity_bytes = _g1->capacity();
1633 return (double) reclaimable_bytes * 100.0 / (double) capacity_bytes;
1634 }
1635
1636 void G1CollectorPolicy::maybe_start_marking() {
1637 if (need_to_start_conc_mark("end of GC")) {
1638 // Note: this might have already been set, if during the last
1639 // pause we decided to start a cycle but at the beginning of
1640 // this pause we decided to postpone it. That's OK.
1641 collector_state()->set_initiate_conc_mark_if_possible(true);
1642 }
1643 }
1644
1645 G1CollectorPolicy::PauseKind G1CollectorPolicy::young_gc_pause_kind() const {
1646 assert(!collector_state()->full_collection(), "must be");
1647 if (collector_state()->during_initial_mark_pause()) {
1648 assert(collector_state()->last_gc_was_young(), "must be");
1649 assert(!collector_state()->last_young_gc(), "must be");
1650 return InitialMarkGC;
1651 } else if (collector_state()->last_young_gc()) {
1652 assert(!collector_state()->during_initial_mark_pause(), "must be");
1653 assert(collector_state()->last_gc_was_young(), "must be");
1654 return LastYoungGC;
1655 } else if (!collector_state()->last_gc_was_young()) {
1656 assert(!collector_state()->during_initial_mark_pause(), "must be");
1657 assert(!collector_state()->last_young_gc(), "must be");
1658 return MixedGC;
1659 } else {
1660 assert(collector_state()->last_gc_was_young(), "must be");
1661 assert(!collector_state()->during_initial_mark_pause(), "must be");
1662 assert(!collector_state()->last_young_gc(), "must be");
1663 return YoungOnlyGC;
1664 }
1665 }
1666
1667 void G1CollectorPolicy::record_pause(PauseKind kind, double start, double end) {
1668 // Manage the MMU tracker. For some reason it ignores Full GCs.
1669 if (kind != FullGC) {
1670 _mmu_tracker->add_pause(start, end);
1671 }
1672 // Manage the mutator time tracking from initial mark to first mixed gc.
1673 switch (kind) {
1674 case FullGC:
1675 abort_time_to_mixed_tracking();
1676 break;
1677 case Cleanup:
1678 case Remark:
1679 case YoungOnlyGC:
1680 case LastYoungGC:
1681 _initial_mark_to_mixed.add_pause(end - start);
1682 break;
1683 case InitialMarkGC:
1684 _initial_mark_to_mixed.record_initial_mark_end(end);
1685 break;
1686 case MixedGC:
1687 _initial_mark_to_mixed.record_mixed_gc_start(start);
1688 break;
1689 default:
1690 ShouldNotReachHere();
1691 }
1692 }
1693
1694 void G1CollectorPolicy::abort_time_to_mixed_tracking() {
1695 _initial_mark_to_mixed.reset();
1696 }
1697
1698 bool G1CollectorPolicy::next_gc_should_be_mixed(const char* true_action_str,
1699 const char* false_action_str) const {
1700 if (cset_chooser()->is_empty()) {
1701 log_debug(gc, ergo)("%s (candidate old regions not available)", false_action_str);
1702 return false;
1703 }
1704
1705 // Is the amount of uncollected reclaimable space above G1HeapWastePercent?
1706 size_t reclaimable_bytes = cset_chooser()->remaining_reclaimable_bytes();
1707 double reclaimable_perc = reclaimable_bytes_perc(reclaimable_bytes);
1708 double threshold = (double) G1HeapWastePercent;
1709 if (reclaimable_perc <= threshold) {
1710 log_debug(gc, ergo)("%s (reclaimable percentage not over threshold). candidate old regions: %u reclaimable: " SIZE_FORMAT " (%1.2f) threshold: " UINTX_FORMAT,
1711 false_action_str, cset_chooser()->remaining_regions(), reclaimable_bytes, reclaimable_perc, G1HeapWastePercent);
1712 return false;
1713 }
1714 log_debug(gc, ergo)("%s (candidate old regions available). candidate old regions: %u reclaimable: " SIZE_FORMAT " (%1.2f) threshold: " UINTX_FORMAT,
1715 true_action_str, cset_chooser()->remaining_regions(), reclaimable_bytes, reclaimable_perc, G1HeapWastePercent);
1716 return true;
1717 }
1718
1719 uint G1CollectorPolicy::calc_min_old_cset_length() const {
1720 // The min old CSet region bound is based on the maximum desired
1721 // number of mixed GCs after a cycle. I.e., even if some old regions
1722 // look expensive, we should add them to the CSet anyway to make
1723 // sure we go through the available old regions in no more than the
1724 // maximum desired number of mixed GCs.
1725 //
1726 // The calculation is based on the number of marked regions we added
1727 // to the CSet chooser in the first place, not how many remain, so
1728 // that the result is the same during all mixed GCs that follow a cycle.
1729
1730 const size_t region_num = (size_t) cset_chooser()->length();
1731 const size_t gc_num = (size_t) MAX2(G1MixedGCCountTarget, (uintx) 1);
1732 size_t result = region_num / gc_num;
1733 // emulate ceiling
1734 if (result * gc_num < region_num) {
1735 result += 1;
1736 }
1737 return (uint) result;
1738 }
1739
1740 uint G1CollectorPolicy::calc_max_old_cset_length() const {
1741 // The max old CSet region bound is based on the threshold expressed
1742 // as a percentage of the heap size. I.e., it should bound the
1743 // number of old regions added to the CSet irrespective of how many
1744 // of them are available.
1745
1746 const G1CollectedHeap* g1h = G1CollectedHeap::heap();
1747 const size_t region_num = g1h->num_regions();
1748 const size_t perc = (size_t) G1OldCSetRegionThresholdPercent;
1749 size_t result = region_num * perc / 100;
1750 // emulate ceiling
1751 if (100 * result < region_num * perc) {
1752 result += 1;
1753 }
1754 return (uint) result;
1755 }
1756
1757 void G1CollectorPolicy::finalize_collection_set(double target_pause_time_ms) {
1758 double time_remaining_ms = _collection_set->finalize_young_part(target_pause_time_ms);
1759 _collection_set->finalize_old_part(time_remaining_ms);
1760 }
1761
|
1 /*
2 * Copyright (c) 2016, 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 */
24
25 #include "precompiled.hpp"
26 #include "gc/g1/g1Measurements.hpp"
27 #include "gc/g1/g1Predictions.hpp"
28 #include "runtime/os.hpp"
29 #include "utilities/debug.hpp"
30 #include "utilities/numberSeq.hpp"
31
32 // Different defaults for different number of GC threads
33 // They were chosen by running GCOld and SPECjbb on debris with different
34 // numbers of GC threads and choosing them based on the results
35
36 // all the same
37 static double rs_length_diff_defaults[] = {
38 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0
39 };
40
41 static double cost_per_card_ms_defaults[] = {
42 0.01, 0.005, 0.005, 0.003, 0.003, 0.002, 0.002, 0.0015
43 };
44
45 // all the same
46 static double young_cards_per_entry_ratio_defaults[] = {
47 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0
48 };
49
50 static double cost_per_entry_ms_defaults[] = {
52 };
53
54 static double cost_per_byte_ms_defaults[] = {
55 0.00006, 0.00003, 0.00003, 0.000015, 0.000015, 0.00001, 0.00001, 0.000009
56 };
57
58 // these should be pretty consistent
59 static double constant_other_time_ms_defaults[] = {
60 5.0, 5.0, 5.0, 5.0, 5.0, 5.0, 5.0, 5.0
61 };
62
63
64 static double young_other_cost_per_region_ms_defaults[] = {
65 0.3, 0.2, 0.2, 0.15, 0.15, 0.12, 0.12, 0.1
66 };
67
68 static double non_young_other_cost_per_region_ms_defaults[] = {
69 1.0, 0.7, 0.7, 0.5, 0.5, 0.42, 0.42, 0.30
70 };
71
72 G1Measurements::G1Measurements(G1Predictions* predictor) :
73 _predictor(predictor),
74 _recent_gc_times_ms(new TruncatedSeq(NumPrevPausesForHeuristics)),
75 _concurrent_mark_remark_times_ms(new TruncatedSeq(NumPrevPausesForHeuristics)),
76 _concurrent_mark_cleanup_times_ms(new TruncatedSeq(NumPrevPausesForHeuristics)),
77 _alloc_rate_ms_seq(new TruncatedSeq(TruncatedSeqLength)),
78 _prev_collection_pause_end_ms(0.0),
79 _rs_length_diff_seq(new TruncatedSeq(TruncatedSeqLength)),
80 _cost_per_card_ms_seq(new TruncatedSeq(TruncatedSeqLength)),
81 _cost_scan_hcc_seq(new TruncatedSeq(TruncatedSeqLength)),
82 _young_cards_per_entry_ratio_seq(new TruncatedSeq(TruncatedSeqLength)),
83 _mixed_cards_per_entry_ratio_seq(new TruncatedSeq(TruncatedSeqLength)),
84 _cost_per_entry_ms_seq(new TruncatedSeq(TruncatedSeqLength)),
85 _mixed_cost_per_entry_ms_seq(new TruncatedSeq(TruncatedSeqLength)),
86 _cost_per_byte_ms_seq(new TruncatedSeq(TruncatedSeqLength)),
87 _cost_per_byte_ms_during_cm_seq(new TruncatedSeq(TruncatedSeqLength)),
88 _constant_other_time_ms_seq(new TruncatedSeq(TruncatedSeqLength)),
89 _young_other_cost_per_region_ms_seq(new TruncatedSeq(TruncatedSeqLength)),
90 _non_young_other_cost_per_region_ms_seq(new TruncatedSeq(TruncatedSeqLength)),
91 _pending_cards_seq(new TruncatedSeq(TruncatedSeqLength)),
92 _rs_lengths_seq(new TruncatedSeq(TruncatedSeqLength)),
93 _recent_prev_end_times_for_all_gcs_sec(new TruncatedSeq(NumPrevPausesForHeuristics)) {
94
95 // Seed sequences with initial values.
96 _recent_prev_end_times_for_all_gcs_sec->add(os::elapsedTime());
97 _prev_collection_pause_end_ms = os::elapsedTime() * 1000.0;
98
99 int index = MIN2(ParallelGCThreads - 1, 7u);
100
101 _rs_length_diff_seq->add(rs_length_diff_defaults[index]);
102 _cost_per_card_ms_seq->add(cost_per_card_ms_defaults[index]);
103 _cost_scan_hcc_seq->add(0.0);
104 _young_cards_per_entry_ratio_seq->add(
105 young_cards_per_entry_ratio_defaults[index]);
106 _cost_per_entry_ms_seq->add(cost_per_entry_ms_defaults[index]);
107 _cost_per_byte_ms_seq->add(cost_per_byte_ms_defaults[index]);
108 _constant_other_time_ms_seq->add(constant_other_time_ms_defaults[index]);
109 _young_other_cost_per_region_ms_seq->add(
110 young_other_cost_per_region_ms_defaults[index]);
111 _non_young_other_cost_per_region_ms_seq->add(
112 non_young_other_cost_per_region_ms_defaults[index]);
113
114 // start conservatively (around 50ms is about right)
115 _concurrent_mark_remark_times_ms->add(0.05);
116 _concurrent_mark_cleanup_times_ms->add(0.20);
117 }
118
119 double G1Measurements::get_new_prediction(TruncatedSeq const* seq) const {
120 return _predictor->get_new_prediction(seq);
121 }
122
123 size_t G1Measurements::get_new_size_prediction(TruncatedSeq const* seq) const {
124 return (size_t)get_new_prediction(seq);
125 }
126
127 int G1Measurements::num_alloc_rate_ms() const {
128 return _alloc_rate_ms_seq->num();
129 }
130
131 void G1Measurements::report_concurrent_mark_remark_times_ms(double ms) {
132 _concurrent_mark_remark_times_ms->add(ms);
133 }
134
135 void G1Measurements::report_alloc_rate_ms(double alloc_rate) {
136 _alloc_rate_ms_seq->add(alloc_rate);
137 }
138
139 void G1Measurements::compute_pause_time_ratio(double interval_ms, double pause_time_ms) {
140 _recent_avg_pause_time_ratio = _recent_gc_times_ms->sum() / interval_ms;
141 if (_recent_avg_pause_time_ratio < 0.0 ||
142 (_recent_avg_pause_time_ratio - 1.0 > 0.0)) {
143 // Clip ratio between 0.0 and 1.0, and continue. This will be fixed in
144 // CR 6902692 by redoing the manner in which the ratio is incrementally computed.
145 if (_recent_avg_pause_time_ratio < 0.0) {
146 _recent_avg_pause_time_ratio = 0.0;
147 } else {
148 assert(_recent_avg_pause_time_ratio - 1.0 > 0.0, "Ctl-point invariant");
149 _recent_avg_pause_time_ratio = 1.0;
150 }
151 }
152
153 // Compute the ratio of just this last pause time to the entire time range stored
154 // in the vectors. Comparing this pause to the entire range, rather than only the
155 // most recent interval, has the effect of smoothing over a possible transient 'burst'
156 // of more frequent pauses that don't really reflect a change in heap occupancy.
157 // This reduces the likelihood of a needless heap expansion being triggered.
158 _last_pause_time_ratio =
159 (pause_time_ms * _recent_prev_end_times_for_all_gcs_sec->num()) / interval_ms;
160 }
161
162 void G1Measurements::report_cost_per_card_ms(double cost_per_card_ms) {
163 _cost_per_card_ms_seq->add(cost_per_card_ms);
164 }
165
166 void G1Measurements::report_cost_scan_hcc(double cost_scan_hcc) {
167 _cost_scan_hcc_seq->add(cost_scan_hcc);
168 }
169
170 void G1Measurements::report_cost_per_entry_ms(double cost_per_entry_ms, bool last_gc_was_young) {
171 if (last_gc_was_young) {
172 _cost_per_entry_ms_seq->add(cost_per_entry_ms);
173 } else {
174 _mixed_cost_per_entry_ms_seq->add(cost_per_entry_ms);
175 }
176 }
177
178 void G1Measurements::report_cards_per_entry_ratio(double cards_per_entry_ratio, bool last_gc_was_young) {
179 if (last_gc_was_young) {
180 _young_cards_per_entry_ratio_seq->add(cards_per_entry_ratio);
181 } else {
182 _mixed_cards_per_entry_ratio_seq->add(cards_per_entry_ratio);
183 }
184 }
185
186 void G1Measurements::report_rs_length_diff(double rs_length_diff) {
187 _rs_length_diff_seq->add(rs_length_diff);
188 }
189
190 void G1Measurements::report_cost_per_byte_ms(double cost_per_byte_ms, bool in_marking_window) {
191 if (in_marking_window) {
192 _cost_per_byte_ms_during_cm_seq->add(cost_per_byte_ms);
193 } else {
194 _cost_per_byte_ms_seq->add(cost_per_byte_ms);
195 }
196 }
197
198 void G1Measurements::report_young_other_cost_per_region_ms(double other_cost_per_region_ms) {
199 _young_other_cost_per_region_ms_seq->add(other_cost_per_region_ms);
200 }
201
202 void G1Measurements::report_non_young_other_cost_per_region_ms(double other_cost_per_region_ms) {
203 _non_young_other_cost_per_region_ms_seq->add(other_cost_per_region_ms);
204 }
205
206 void G1Measurements::report_constant_other_time_ms(double constant_other_time_ms) {
207 _constant_other_time_ms_seq->add(constant_other_time_ms);
208 }
209
210 void G1Measurements::report_pending_cards(double pending_cards) {
211 _pending_cards_seq->add(pending_cards);
212 }
213
214 void G1Measurements::report_rs_lengths(double rs_lengths) {
215 _rs_lengths_seq->add(rs_lengths);
216 }
217
218 size_t G1Measurements::predict_rs_length_diff() const {
219 return get_new_size_prediction(_rs_length_diff_seq);
220 }
221
222 double G1Measurements::predict_alloc_rate_ms() const {
223 return get_new_prediction(_alloc_rate_ms_seq);
224 }
225
226 double G1Measurements::predict_cost_per_card_ms() const {
227 return get_new_prediction(_cost_per_card_ms_seq);
228 }
229
230 double G1Measurements::predict_scan_hcc_ms() const {
231 return get_new_prediction(_cost_scan_hcc_seq);
232 }
233
234 double G1Measurements::predict_rs_update_time_ms(size_t pending_cards) const {
235 return pending_cards * predict_cost_per_card_ms() + predict_scan_hcc_ms();
236 }
237
238 double G1Measurements::predict_young_cards_per_entry_ratio() const {
239 return get_new_prediction(_young_cards_per_entry_ratio_seq);
240 }
241
242 double G1Measurements::predict_mixed_cards_per_entry_ratio() const {
243 if (_mixed_cards_per_entry_ratio_seq->num() < 2) {
244 return predict_young_cards_per_entry_ratio();
245 } else {
246 return get_new_prediction(_mixed_cards_per_entry_ratio_seq);
247 }
248 }
249
250 size_t G1Measurements::predict_card_num(size_t rs_length, bool gcs_are_young) const {
251 if (gcs_are_young) {
252 return (size_t) (rs_length * predict_young_cards_per_entry_ratio());
253 } else {
254 return (size_t) (rs_length * predict_mixed_cards_per_entry_ratio());
255 }
256 }
257
258 double G1Measurements::predict_rs_scan_time_ms(size_t card_num, bool gcs_are_young) const {
259 if (gcs_are_young) {
260 return card_num * get_new_prediction(_cost_per_entry_ms_seq);
261 } else {
262 return predict_mixed_rs_scan_time_ms(card_num);
263 }
264 }
265
266 double G1Measurements::predict_mixed_rs_scan_time_ms(size_t card_num) const {
267 if (_mixed_cost_per_entry_ms_seq->num() < 3) {
268 return card_num * get_new_prediction(_cost_per_entry_ms_seq);
269 } else {
270 return card_num * get_new_prediction(_mixed_cost_per_entry_ms_seq);
271 }
272 }
273
274 double G1Measurements::predict_object_copy_time_ms_during_cm(size_t bytes_to_copy) const {
275 if (_cost_per_byte_ms_during_cm_seq->num() < 3) {
276 return (1.1 * bytes_to_copy) * get_new_prediction(_cost_per_byte_ms_seq);
277 } else {
278 return bytes_to_copy * get_new_prediction(_cost_per_byte_ms_during_cm_seq);
279 }
280 }
281
282 double G1Measurements::predict_object_copy_time_ms(size_t bytes_to_copy, bool during_concurrent_mark) const {
283 if (during_concurrent_mark) {
284 return predict_object_copy_time_ms_during_cm(bytes_to_copy);
285 } else {
286 return bytes_to_copy * get_new_prediction(_cost_per_byte_ms_seq);
287 }
288 }
289
290 double G1Measurements::predict_constant_other_time_ms() const {
291 return get_new_prediction(_constant_other_time_ms_seq);
292 }
293
294 double G1Measurements::predict_young_other_time_ms(size_t young_num) const {
295 return young_num * get_new_prediction(_young_other_cost_per_region_ms_seq);
296 }
297
298 double G1Measurements::predict_non_young_other_time_ms(size_t non_young_num) const {
299 return non_young_num * get_new_prediction(_non_young_other_cost_per_region_ms_seq);
300 }
301
302 double G1Measurements::predict_remark_time_ms() const {
303 return get_new_prediction(_concurrent_mark_remark_times_ms);
304 }
305
306 double G1Measurements::predict_cleanup_time_ms() const {
307 return get_new_prediction(_concurrent_mark_cleanup_times_ms);
308 }
309
310 size_t G1Measurements::predict_rs_lengths() const {
311 return get_new_size_prediction(_rs_lengths_seq);
312 }
313
314 size_t G1Measurements::predict_pending_cards() const {
315 return get_new_size_prediction(_pending_cards_seq);
316 }
317
318 void G1Measurements::update_recent_gc_times(double end_time_sec,
319 double pause_time_ms) {
320 double interval_ms =
321 (end_time_sec - _recent_prev_end_times_for_all_gcs_sec->oldest()) * 1000.0;
322 _recent_gc_times_ms->add(pause_time_ms);
323 _recent_prev_end_times_for_all_gcs_sec->add(end_time_sec);
324 _prev_collection_pause_end_ms = end_time_sec * 1000.0;
325
326 compute_pause_time_ratio(interval_ms, pause_time_ms);
327 }
328
329 void G1Measurements::report_concurrent_mark_cleanup_times_ms(double ms) {
330 _concurrent_mark_cleanup_times_ms->add(ms);
331 }
332
|