Mercurial > hg > graal-jvmci-8
comparison src/share/vm/gc_implementation/g1/g1CollectorPolicy.cpp @ 342:37f87013dfd8
6711316: Open source the Garbage-First garbage collector
Summary: First mercurial integration of the code for the Garbage-First garbage collector.
Reviewed-by: apetrusenko, iveresov, jmasa, sgoldman, tonyp, ysr
author | ysr |
---|---|
date | Thu, 05 Jun 2008 15:57:56 -0700 |
parents | |
children | 0edda524b58c |
comparison
equal
deleted
inserted
replaced
189:0b27f3512f9e | 342:37f87013dfd8 |
---|---|
1 /* | |
2 * Copyright 2001-2007 Sun Microsystems, Inc. All Rights Reserved. | |
3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER. | |
4 * | |
5 * This code is free software; you can redistribute it and/or modify it | |
6 * under the terms of the GNU General Public License version 2 only, as | |
7 * published by the Free Software Foundation. | |
8 * | |
9 * This code is distributed in the hope that it will be useful, but WITHOUT | |
10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or | |
11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License | |
12 * version 2 for more details (a copy is included in the LICENSE file that | |
13 * accompanied this code). | |
14 * | |
15 * You should have received a copy of the GNU General Public License version | |
16 * 2 along with this work; if not, write to the Free Software Foundation, | |
17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA. | |
18 * | |
19 * Please contact Sun Microsystems, Inc., 4150 Network Circle, Santa Clara, | |
20 * CA 95054 USA or visit www.sun.com if you need additional information or | |
21 * have any questions. | |
22 * | |
23 */ | |
24 | |
25 #include "incls/_precompiled.incl" | |
26 #include "incls/_g1CollectorPolicy.cpp.incl" | |
27 | |
28 #define PREDICTIONS_VERBOSE 0 | |
29 | |
30 // <NEW PREDICTION> | |
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 static double cost_per_scan_only_region_ms_defaults[] = { | |
46 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0 | |
47 }; | |
48 | |
49 // all the same | |
50 static double fully_young_cards_per_entry_ratio_defaults[] = { | |
51 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0 | |
52 }; | |
53 | |
54 static double cost_per_entry_ms_defaults[] = { | |
55 0.015, 0.01, 0.01, 0.008, 0.008, 0.0055, 0.0055, 0.005 | |
56 }; | |
57 | |
58 static double cost_per_byte_ms_defaults[] = { | |
59 0.00006, 0.00003, 0.00003, 0.000015, 0.000015, 0.00001, 0.00001, 0.000009 | |
60 }; | |
61 | |
62 // these should be pretty consistent | |
63 static double constant_other_time_ms_defaults[] = { | |
64 5.0, 5.0, 5.0, 5.0, 5.0, 5.0, 5.0, 5.0 | |
65 }; | |
66 | |
67 | |
68 static double young_other_cost_per_region_ms_defaults[] = { | |
69 0.3, 0.2, 0.2, 0.15, 0.15, 0.12, 0.12, 0.1 | |
70 }; | |
71 | |
72 static double non_young_other_cost_per_region_ms_defaults[] = { | |
73 1.0, 0.7, 0.7, 0.5, 0.5, 0.42, 0.42, 0.30 | |
74 }; | |
75 | |
76 // </NEW PREDICTION> | |
77 | |
78 G1CollectorPolicy::G1CollectorPolicy() : | |
79 _parallel_gc_threads((ParallelGCThreads > 0) ? ParallelGCThreads : 1), | |
80 _n_pauses(0), | |
81 _recent_CH_strong_roots_times_ms(new TruncatedSeq(NumPrevPausesForHeuristics)), | |
82 _recent_G1_strong_roots_times_ms(new TruncatedSeq(NumPrevPausesForHeuristics)), | |
83 _recent_evac_times_ms(new TruncatedSeq(NumPrevPausesForHeuristics)), | |
84 _recent_pause_times_ms(new TruncatedSeq(NumPrevPausesForHeuristics)), | |
85 _recent_rs_sizes(new TruncatedSeq(NumPrevPausesForHeuristics)), | |
86 _recent_gc_times_ms(new TruncatedSeq(NumPrevPausesForHeuristics)), | |
87 _all_pause_times_ms(new NumberSeq()), | |
88 _stop_world_start(0.0), | |
89 _all_stop_world_times_ms(new NumberSeq()), | |
90 _all_yield_times_ms(new NumberSeq()), | |
91 | |
92 _all_mod_union_times_ms(new NumberSeq()), | |
93 | |
94 _non_pop_summary(new NonPopSummary()), | |
95 _pop_summary(new PopSummary()), | |
96 _non_pop_abandoned_summary(new NonPopAbandonedSummary()), | |
97 _pop_abandoned_summary(new PopAbandonedSummary()), | |
98 | |
99 _cur_clear_ct_time_ms(0.0), | |
100 | |
101 _region_num_young(0), | |
102 _region_num_tenured(0), | |
103 _prev_region_num_young(0), | |
104 _prev_region_num_tenured(0), | |
105 | |
106 _aux_num(10), | |
107 _all_aux_times_ms(new NumberSeq[_aux_num]), | |
108 _cur_aux_start_times_ms(new double[_aux_num]), | |
109 _cur_aux_times_ms(new double[_aux_num]), | |
110 _cur_aux_times_set(new bool[_aux_num]), | |
111 | |
112 _pop_compute_rc_start(0.0), | |
113 _pop_evac_start(0.0), | |
114 | |
115 _concurrent_mark_init_times_ms(new TruncatedSeq(NumPrevPausesForHeuristics)), | |
116 _concurrent_mark_remark_times_ms(new TruncatedSeq(NumPrevPausesForHeuristics)), | |
117 _concurrent_mark_cleanup_times_ms(new TruncatedSeq(NumPrevPausesForHeuristics)), | |
118 | |
119 // <NEW PREDICTION> | |
120 | |
121 _alloc_rate_ms_seq(new TruncatedSeq(TruncatedSeqLength)), | |
122 _prev_collection_pause_end_ms(0.0), | |
123 _pending_card_diff_seq(new TruncatedSeq(TruncatedSeqLength)), | |
124 _rs_length_diff_seq(new TruncatedSeq(TruncatedSeqLength)), | |
125 _cost_per_card_ms_seq(new TruncatedSeq(TruncatedSeqLength)), | |
126 _cost_per_scan_only_region_ms_seq(new TruncatedSeq(TruncatedSeqLength)), | |
127 _fully_young_cards_per_entry_ratio_seq(new TruncatedSeq(TruncatedSeqLength)), | |
128 _partially_young_cards_per_entry_ratio_seq( | |
129 new TruncatedSeq(TruncatedSeqLength)), | |
130 _cost_per_entry_ms_seq(new TruncatedSeq(TruncatedSeqLength)), | |
131 _partially_young_cost_per_entry_ms_seq(new TruncatedSeq(TruncatedSeqLength)), | |
132 _cost_per_byte_ms_seq(new TruncatedSeq(TruncatedSeqLength)), | |
133 _cost_per_byte_ms_during_cm_seq(new TruncatedSeq(TruncatedSeqLength)), | |
134 _cost_per_scan_only_region_ms_during_cm_seq(new TruncatedSeq(TruncatedSeqLength)), | |
135 _constant_other_time_ms_seq(new TruncatedSeq(TruncatedSeqLength)), | |
136 _young_other_cost_per_region_ms_seq(new TruncatedSeq(TruncatedSeqLength)), | |
137 _non_young_other_cost_per_region_ms_seq( | |
138 new TruncatedSeq(TruncatedSeqLength)), | |
139 | |
140 _pending_cards_seq(new TruncatedSeq(TruncatedSeqLength)), | |
141 _scanned_cards_seq(new TruncatedSeq(TruncatedSeqLength)), | |
142 _rs_lengths_seq(new TruncatedSeq(TruncatedSeqLength)), | |
143 | |
144 _pause_time_target_ms((double) G1MaxPauseTimeMS), | |
145 | |
146 // </NEW PREDICTION> | |
147 | |
148 _in_young_gc_mode(false), | |
149 _full_young_gcs(true), | |
150 _full_young_pause_num(0), | |
151 _partial_young_pause_num(0), | |
152 | |
153 _during_marking(false), | |
154 _in_marking_window(false), | |
155 _in_marking_window_im(false), | |
156 | |
157 _known_garbage_ratio(0.0), | |
158 _known_garbage_bytes(0), | |
159 | |
160 _young_gc_eff_seq(new TruncatedSeq(TruncatedSeqLength)), | |
161 _target_pause_time_ms(-1.0), | |
162 | |
163 _recent_prev_end_times_for_all_gcs_sec(new TruncatedSeq(NumPrevPausesForHeuristics)), | |
164 | |
165 _recent_CS_bytes_used_before(new TruncatedSeq(NumPrevPausesForHeuristics)), | |
166 _recent_CS_bytes_surviving(new TruncatedSeq(NumPrevPausesForHeuristics)), | |
167 | |
168 _recent_avg_pause_time_ratio(0.0), | |
169 _num_markings(0), | |
170 _n_marks(0), | |
171 _n_pauses_at_mark_end(0), | |
172 | |
173 _all_full_gc_times_ms(new NumberSeq()), | |
174 | |
175 _conc_refine_enabled(0), | |
176 _conc_refine_zero_traversals(0), | |
177 _conc_refine_max_traversals(0), | |
178 _conc_refine_current_delta(G1ConcRefineInitialDelta), | |
179 | |
180 // G1PausesBtwnConcMark defaults to -1 | |
181 // so the hack is to do the cast QQQ FIXME | |
182 _pauses_btwn_concurrent_mark((size_t)G1PausesBtwnConcMark), | |
183 _n_marks_since_last_pause(0), | |
184 _conc_mark_initiated(false), | |
185 _should_initiate_conc_mark(false), | |
186 _should_revert_to_full_young_gcs(false), | |
187 _last_full_young_gc(false), | |
188 | |
189 _prev_collection_pause_used_at_end_bytes(0), | |
190 | |
191 _collection_set(NULL), | |
192 #ifdef _MSC_VER // the use of 'this' below gets a warning, make it go away | |
193 #pragma warning( disable:4355 ) // 'this' : used in base member initializer list | |
194 #endif // _MSC_VER | |
195 | |
196 _short_lived_surv_rate_group(new SurvRateGroup(this, "Short Lived", | |
197 G1YoungSurvRateNumRegionsSummary)), | |
198 _survivor_surv_rate_group(new SurvRateGroup(this, "Survivor", | |
199 G1YoungSurvRateNumRegionsSummary)) | |
200 // add here any more surv rate groups | |
201 { | |
202 _recent_prev_end_times_for_all_gcs_sec->add(os::elapsedTime()); | |
203 _prev_collection_pause_end_ms = os::elapsedTime() * 1000.0; | |
204 | |
205 _par_last_ext_root_scan_times_ms = new double[_parallel_gc_threads]; | |
206 _par_last_mark_stack_scan_times_ms = new double[_parallel_gc_threads]; | |
207 _par_last_scan_only_times_ms = new double[_parallel_gc_threads]; | |
208 _par_last_scan_only_regions_scanned = new double[_parallel_gc_threads]; | |
209 | |
210 _par_last_update_rs_start_times_ms = new double[_parallel_gc_threads]; | |
211 _par_last_update_rs_times_ms = new double[_parallel_gc_threads]; | |
212 _par_last_update_rs_processed_buffers = new double[_parallel_gc_threads]; | |
213 | |
214 _par_last_scan_rs_start_times_ms = new double[_parallel_gc_threads]; | |
215 _par_last_scan_rs_times_ms = new double[_parallel_gc_threads]; | |
216 _par_last_scan_new_refs_times_ms = new double[_parallel_gc_threads]; | |
217 | |
218 _par_last_obj_copy_times_ms = new double[_parallel_gc_threads]; | |
219 | |
220 _par_last_termination_times_ms = new double[_parallel_gc_threads]; | |
221 | |
222 // we store the data from the first pass during popularity pauses | |
223 _pop_par_last_update_rs_start_times_ms = new double[_parallel_gc_threads]; | |
224 _pop_par_last_update_rs_times_ms = new double[_parallel_gc_threads]; | |
225 _pop_par_last_update_rs_processed_buffers = new double[_parallel_gc_threads]; | |
226 | |
227 _pop_par_last_scan_rs_start_times_ms = new double[_parallel_gc_threads]; | |
228 _pop_par_last_scan_rs_times_ms = new double[_parallel_gc_threads]; | |
229 | |
230 _pop_par_last_closure_app_times_ms = new double[_parallel_gc_threads]; | |
231 | |
232 // start conservatively | |
233 _expensive_region_limit_ms = 0.5 * (double) G1MaxPauseTimeMS; | |
234 | |
235 // <NEW PREDICTION> | |
236 | |
237 int index; | |
238 if (ParallelGCThreads == 0) | |
239 index = 0; | |
240 else if (ParallelGCThreads > 8) | |
241 index = 7; | |
242 else | |
243 index = ParallelGCThreads - 1; | |
244 | |
245 _pending_card_diff_seq->add(0.0); | |
246 _rs_length_diff_seq->add(rs_length_diff_defaults[index]); | |
247 _cost_per_card_ms_seq->add(cost_per_card_ms_defaults[index]); | |
248 _cost_per_scan_only_region_ms_seq->add( | |
249 cost_per_scan_only_region_ms_defaults[index]); | |
250 _fully_young_cards_per_entry_ratio_seq->add( | |
251 fully_young_cards_per_entry_ratio_defaults[index]); | |
252 _cost_per_entry_ms_seq->add(cost_per_entry_ms_defaults[index]); | |
253 _cost_per_byte_ms_seq->add(cost_per_byte_ms_defaults[index]); | |
254 _constant_other_time_ms_seq->add(constant_other_time_ms_defaults[index]); | |
255 _young_other_cost_per_region_ms_seq->add( | |
256 young_other_cost_per_region_ms_defaults[index]); | |
257 _non_young_other_cost_per_region_ms_seq->add( | |
258 non_young_other_cost_per_region_ms_defaults[index]); | |
259 | |
260 // </NEW PREDICTION> | |
261 | |
262 double time_slice = (double) G1TimeSliceMS / 1000.0; | |
263 double max_gc_time = (double) G1MaxPauseTimeMS / 1000.0; | |
264 guarantee(max_gc_time < time_slice, | |
265 "Max GC time should not be greater than the time slice"); | |
266 _mmu_tracker = new G1MMUTrackerQueue(time_slice, max_gc_time); | |
267 _sigma = (double) G1ConfidencePerc / 100.0; | |
268 | |
269 // start conservatively (around 50ms is about right) | |
270 _concurrent_mark_init_times_ms->add(0.05); | |
271 _concurrent_mark_remark_times_ms->add(0.05); | |
272 _concurrent_mark_cleanup_times_ms->add(0.20); | |
273 _tenuring_threshold = MaxTenuringThreshold; | |
274 | |
275 initialize_all(); | |
276 } | |
277 | |
278 // Increment "i", mod "len" | |
279 static void inc_mod(int& i, int len) { | |
280 i++; if (i == len) i = 0; | |
281 } | |
282 | |
283 void G1CollectorPolicy::initialize_flags() { | |
284 set_min_alignment(HeapRegion::GrainBytes); | |
285 set_max_alignment(GenRemSet::max_alignment_constraint(rem_set_name())); | |
286 CollectorPolicy::initialize_flags(); | |
287 } | |
288 | |
289 void G1CollectorPolicy::init() { | |
290 // Set aside an initial future to_space. | |
291 _g1 = G1CollectedHeap::heap(); | |
292 size_t regions = Universe::heap()->capacity() / HeapRegion::GrainBytes; | |
293 | |
294 assert(Heap_lock->owned_by_self(), "Locking discipline."); | |
295 | |
296 if (G1SteadyStateUsed < 50) { | |
297 vm_exit_during_initialization("G1SteadyStateUsed must be at least 50%."); | |
298 } | |
299 if (UseConcMarkSweepGC) { | |
300 vm_exit_during_initialization("-XX:+UseG1GC is incompatible with " | |
301 "-XX:+UseConcMarkSweepGC."); | |
302 } | |
303 | |
304 if (G1Gen) { | |
305 _in_young_gc_mode = true; | |
306 | |
307 if (G1YoungGenSize == 0) { | |
308 set_adaptive_young_list_length(true); | |
309 _young_list_fixed_length = 0; | |
310 } else { | |
311 set_adaptive_young_list_length(false); | |
312 _young_list_fixed_length = (G1YoungGenSize / HeapRegion::GrainBytes); | |
313 } | |
314 _free_regions_at_end_of_collection = _g1->free_regions(); | |
315 _scan_only_regions_at_end_of_collection = 0; | |
316 calculate_young_list_min_length(); | |
317 guarantee( _young_list_min_length == 0, "invariant, not enough info" ); | |
318 calculate_young_list_target_config(); | |
319 } else { | |
320 _young_list_fixed_length = 0; | |
321 _in_young_gc_mode = false; | |
322 } | |
323 } | |
324 | |
325 void G1CollectorPolicy::calculate_young_list_min_length() { | |
326 _young_list_min_length = 0; | |
327 | |
328 if (!adaptive_young_list_length()) | |
329 return; | |
330 | |
331 if (_alloc_rate_ms_seq->num() > 3) { | |
332 double now_sec = os::elapsedTime(); | |
333 double when_ms = _mmu_tracker->when_max_gc_sec(now_sec) * 1000.0; | |
334 double alloc_rate_ms = predict_alloc_rate_ms(); | |
335 int min_regions = (int) ceil(alloc_rate_ms * when_ms); | |
336 int current_region_num = (int) _g1->young_list_length(); | |
337 _young_list_min_length = min_regions + current_region_num; | |
338 } | |
339 } | |
340 | |
341 void G1CollectorPolicy::calculate_young_list_target_config() { | |
342 if (adaptive_young_list_length()) { | |
343 size_t rs_lengths = (size_t) get_new_prediction(_rs_lengths_seq); | |
344 calculate_young_list_target_config(rs_lengths); | |
345 } else { | |
346 if (full_young_gcs()) | |
347 _young_list_target_length = _young_list_fixed_length; | |
348 else | |
349 _young_list_target_length = _young_list_fixed_length / 2; | |
350 _young_list_target_length = MAX2(_young_list_target_length, (size_t)1); | |
351 size_t so_length = calculate_optimal_so_length(_young_list_target_length); | |
352 guarantee( so_length < _young_list_target_length, "invariant" ); | |
353 _young_list_so_prefix_length = so_length; | |
354 } | |
355 } | |
356 | |
357 // This method calculate the optimal scan-only set for a fixed young | |
358 // gen size. I couldn't work out how to reuse the more elaborate one, | |
359 // i.e. calculate_young_list_target_config(rs_length), as the loops are | |
360 // fundamentally different (the other one finds a config for different | |
361 // S-O lengths, whereas here we need to do the opposite). | |
362 size_t G1CollectorPolicy::calculate_optimal_so_length( | |
363 size_t young_list_length) { | |
364 if (!G1UseScanOnlyPrefix) | |
365 return 0; | |
366 | |
367 if (_all_pause_times_ms->num() < 3) { | |
368 // we won't use a scan-only set at the beginning to allow the rest | |
369 // of the predictors to warm up | |
370 return 0; | |
371 } | |
372 | |
373 if (_cost_per_scan_only_region_ms_seq->num() < 3) { | |
374 // then, we'll only set the S-O set to 1 for a little bit of time, | |
375 // to get enough information on the scanning cost | |
376 return 1; | |
377 } | |
378 | |
379 size_t pending_cards = (size_t) get_new_prediction(_pending_cards_seq); | |
380 size_t rs_lengths = (size_t) get_new_prediction(_rs_lengths_seq); | |
381 size_t adj_rs_lengths = rs_lengths + predict_rs_length_diff(); | |
382 size_t scanned_cards; | |
383 if (full_young_gcs()) | |
384 scanned_cards = predict_young_card_num(adj_rs_lengths); | |
385 else | |
386 scanned_cards = predict_non_young_card_num(adj_rs_lengths); | |
387 double base_time_ms = predict_base_elapsed_time_ms(pending_cards, | |
388 scanned_cards); | |
389 | |
390 size_t so_length = 0; | |
391 double max_gc_eff = 0.0; | |
392 for (size_t i = 0; i < young_list_length; ++i) { | |
393 double gc_eff = 0.0; | |
394 double pause_time_ms = 0.0; | |
395 predict_gc_eff(young_list_length, i, base_time_ms, | |
396 &gc_eff, &pause_time_ms); | |
397 if (gc_eff > max_gc_eff) { | |
398 max_gc_eff = gc_eff; | |
399 so_length = i; | |
400 } | |
401 } | |
402 | |
403 // set it to 95% of the optimal to make sure we sample the "area" | |
404 // around the optimal length to get up-to-date survival rate data | |
405 return so_length * 950 / 1000; | |
406 } | |
407 | |
408 // This is a really cool piece of code! It finds the best | |
409 // target configuration (young length / scan-only prefix length) so | |
410 // that GC efficiency is maximized and that we also meet a pause | |
411 // time. It's a triple nested loop. These loops are explained below | |
412 // from the inside-out :-) | |
413 // | |
414 // (a) The innermost loop will try to find the optimal young length | |
415 // for a fixed S-O length. It uses a binary search to speed up the | |
416 // process. We assume that, for a fixed S-O length, as we add more | |
417 // young regions to the CSet, the GC efficiency will only go up (I'll | |
418 // skip the proof). So, using a binary search to optimize this process | |
419 // makes perfect sense. | |
420 // | |
421 // (b) The middle loop will fix the S-O length before calling the | |
422 // innermost one. It will vary it between two parameters, increasing | |
423 // it by a given increment. | |
424 // | |
425 // (c) The outermost loop will call the middle loop three times. | |
426 // (1) The first time it will explore all possible S-O length values | |
427 // from 0 to as large as it can get, using a coarse increment (to | |
428 // quickly "home in" to where the optimal seems to be). | |
429 // (2) The second time it will explore the values around the optimal | |
430 // that was found by the first iteration using a fine increment. | |
431 // (3) Once the optimal config has been determined by the second | |
432 // iteration, we'll redo the calculation, but setting the S-O length | |
433 // to 95% of the optimal to make sure we sample the "area" | |
434 // around the optimal length to get up-to-date survival rate data | |
435 // | |
436 // Termination conditions for the iterations are several: the pause | |
437 // time is over the limit, we do not have enough to-space, etc. | |
438 | |
439 void G1CollectorPolicy::calculate_young_list_target_config(size_t rs_lengths) { | |
440 guarantee( adaptive_young_list_length(), "pre-condition" ); | |
441 | |
442 double start_time_sec = os::elapsedTime(); | |
443 size_t min_reserve_perc = MAX2((size_t)2, (size_t)G1MinReservePerc); | |
444 min_reserve_perc = MIN2((size_t) 50, min_reserve_perc); | |
445 size_t reserve_regions = | |
446 (size_t) ((double) min_reserve_perc * (double) _g1->n_regions() / 100.0); | |
447 | |
448 if (full_young_gcs() && _free_regions_at_end_of_collection > 0) { | |
449 // we are in fully-young mode and there are free regions in the heap | |
450 | |
451 size_t min_so_length = 0; | |
452 size_t max_so_length = 0; | |
453 | |
454 if (G1UseScanOnlyPrefix) { | |
455 if (_all_pause_times_ms->num() < 3) { | |
456 // we won't use a scan-only set at the beginning to allow the rest | |
457 // of the predictors to warm up | |
458 min_so_length = 0; | |
459 max_so_length = 0; | |
460 } else if (_cost_per_scan_only_region_ms_seq->num() < 3) { | |
461 // then, we'll only set the S-O set to 1 for a little bit of time, | |
462 // to get enough information on the scanning cost | |
463 min_so_length = 1; | |
464 max_so_length = 1; | |
465 } else if (_in_marking_window || _last_full_young_gc) { | |
466 // no S-O prefix during a marking phase either, as at the end | |
467 // of the marking phase we'll have to use a very small young | |
468 // length target to fill up the rest of the CSet with | |
469 // non-young regions and, if we have lots of scan-only regions | |
470 // left-over, we will not be able to add any more non-young | |
471 // regions. | |
472 min_so_length = 0; | |
473 max_so_length = 0; | |
474 } else { | |
475 // this is the common case; we'll never reach the maximum, we | |
476 // one of the end conditions will fire well before that | |
477 // (hopefully!) | |
478 min_so_length = 0; | |
479 max_so_length = _free_regions_at_end_of_collection - 1; | |
480 } | |
481 } else { | |
482 // no S-O prefix, as the switch is not set, but we still need to | |
483 // do one iteration to calculate the best young target that | |
484 // meets the pause time; this way we reuse the same code instead | |
485 // of replicating it | |
486 min_so_length = 0; | |
487 max_so_length = 0; | |
488 } | |
489 | |
490 double target_pause_time_ms = _mmu_tracker->max_gc_time() * 1000.0; | |
491 size_t pending_cards = (size_t) get_new_prediction(_pending_cards_seq); | |
492 size_t adj_rs_lengths = rs_lengths + predict_rs_length_diff(); | |
493 size_t scanned_cards; | |
494 if (full_young_gcs()) | |
495 scanned_cards = predict_young_card_num(adj_rs_lengths); | |
496 else | |
497 scanned_cards = predict_non_young_card_num(adj_rs_lengths); | |
498 // calculate this once, so that we don't have to recalculate it in | |
499 // the innermost loop | |
500 double base_time_ms = predict_base_elapsed_time_ms(pending_cards, | |
501 scanned_cards); | |
502 | |
503 // the result | |
504 size_t final_young_length = 0; | |
505 size_t final_so_length = 0; | |
506 double final_gc_eff = 0.0; | |
507 // we'll also keep track of how many times we go into the inner loop | |
508 // this is for profiling reasons | |
509 size_t calculations = 0; | |
510 | |
511 // this determines which of the three iterations the outer loop is in | |
512 typedef enum { | |
513 pass_type_coarse, | |
514 pass_type_fine, | |
515 pass_type_final | |
516 } pass_type_t; | |
517 | |
518 // range of the outer loop's iteration | |
519 size_t from_so_length = min_so_length; | |
520 size_t to_so_length = max_so_length; | |
521 guarantee( from_so_length <= to_so_length, "invariant" ); | |
522 | |
523 // this will keep the S-O length that's found by the second | |
524 // iteration of the outer loop; we'll keep it just in case the third | |
525 // iteration fails to find something | |
526 size_t fine_so_length = 0; | |
527 | |
528 // the increment step for the coarse (first) iteration | |
529 size_t so_coarse_increments = 5; | |
530 | |
531 // the common case, we'll start with the coarse iteration | |
532 pass_type_t pass = pass_type_coarse; | |
533 size_t so_length_incr = so_coarse_increments; | |
534 | |
535 if (from_so_length == to_so_length) { | |
536 // not point in doing the coarse iteration, we'll go directly into | |
537 // the fine one (we essentially trying to find the optimal young | |
538 // length for a fixed S-O length). | |
539 so_length_incr = 1; | |
540 pass = pass_type_final; | |
541 } else if (to_so_length - from_so_length < 3 * so_coarse_increments) { | |
542 // again, the range is too short so no point in foind the coarse | |
543 // iteration either | |
544 so_length_incr = 1; | |
545 pass = pass_type_fine; | |
546 } | |
547 | |
548 bool done = false; | |
549 // this is the outermost loop | |
550 while (!done) { | |
551 #if 0 | |
552 // leave this in for debugging, just in case | |
553 gclog_or_tty->print_cr("searching between " SIZE_FORMAT " and " SIZE_FORMAT | |
554 ", incr " SIZE_FORMAT ", pass %s", | |
555 from_so_length, to_so_length, so_length_incr, | |
556 (pass == pass_type_coarse) ? "coarse" : | |
557 (pass == pass_type_fine) ? "fine" : "final"); | |
558 #endif // 0 | |
559 | |
560 size_t so_length = from_so_length; | |
561 size_t init_free_regions = | |
562 MAX2((size_t)0, | |
563 _free_regions_at_end_of_collection + | |
564 _scan_only_regions_at_end_of_collection - reserve_regions); | |
565 | |
566 // this determines whether a configuration was found | |
567 bool gc_eff_set = false; | |
568 // this is the middle loop | |
569 while (so_length <= to_so_length) { | |
570 // base time, which excludes region-related time; again we | |
571 // calculate it once to avoid recalculating it in the | |
572 // innermost loop | |
573 double base_time_with_so_ms = | |
574 base_time_ms + predict_scan_only_time_ms(so_length); | |
575 // it's already over the pause target, go around | |
576 if (base_time_with_so_ms > target_pause_time_ms) | |
577 break; | |
578 | |
579 size_t starting_young_length = so_length+1; | |
580 | |
581 // we make sure that the short young length that makes sense | |
582 // (one more than the S-O length) is feasible | |
583 size_t min_young_length = starting_young_length; | |
584 double min_gc_eff; | |
585 bool min_ok; | |
586 ++calculations; | |
587 min_ok = predict_gc_eff(min_young_length, so_length, | |
588 base_time_with_so_ms, | |
589 init_free_regions, target_pause_time_ms, | |
590 &min_gc_eff); | |
591 | |
592 if (min_ok) { | |
593 // the shortest young length is indeed feasible; we'll know | |
594 // set up the max young length and we'll do a binary search | |
595 // between min_young_length and max_young_length | |
596 size_t max_young_length = _free_regions_at_end_of_collection - 1; | |
597 double max_gc_eff = 0.0; | |
598 bool max_ok = false; | |
599 | |
600 // the innermost loop! (finally!) | |
601 while (max_young_length > min_young_length) { | |
602 // we'll make sure that min_young_length is always at a | |
603 // feasible config | |
604 guarantee( min_ok, "invariant" ); | |
605 | |
606 ++calculations; | |
607 max_ok = predict_gc_eff(max_young_length, so_length, | |
608 base_time_with_so_ms, | |
609 init_free_regions, target_pause_time_ms, | |
610 &max_gc_eff); | |
611 | |
612 size_t diff = (max_young_length - min_young_length) / 2; | |
613 if (max_ok) { | |
614 min_young_length = max_young_length; | |
615 min_gc_eff = max_gc_eff; | |
616 min_ok = true; | |
617 } | |
618 max_young_length = min_young_length + diff; | |
619 } | |
620 | |
621 // the innermost loop found a config | |
622 guarantee( min_ok, "invariant" ); | |
623 if (min_gc_eff > final_gc_eff) { | |
624 // it's the best config so far, so we'll keep it | |
625 final_gc_eff = min_gc_eff; | |
626 final_young_length = min_young_length; | |
627 final_so_length = so_length; | |
628 gc_eff_set = true; | |
629 } | |
630 } | |
631 | |
632 // incremental the fixed S-O length and go around | |
633 so_length += so_length_incr; | |
634 } | |
635 | |
636 // this is the end of the outermost loop and we need to decide | |
637 // what to do during the next iteration | |
638 if (pass == pass_type_coarse) { | |
639 // we just did the coarse pass (first iteration) | |
640 | |
641 if (!gc_eff_set) | |
642 // we didn't find a feasible config so we'll just bail out; of | |
643 // course, it might be the case that we missed it; but I'd say | |
644 // it's a bit unlikely | |
645 done = true; | |
646 else { | |
647 // We did find a feasible config with optimal GC eff during | |
648 // the first pass. So the second pass we'll only consider the | |
649 // S-O lengths around that config with a fine increment. | |
650 | |
651 guarantee( so_length_incr == so_coarse_increments, "invariant" ); | |
652 guarantee( final_so_length >= min_so_length, "invariant" ); | |
653 | |
654 #if 0 | |
655 // leave this in for debugging, just in case | |
656 gclog_or_tty->print_cr(" coarse pass: SO length " SIZE_FORMAT, | |
657 final_so_length); | |
658 #endif // 0 | |
659 | |
660 from_so_length = | |
661 (final_so_length - min_so_length > so_coarse_increments) ? | |
662 final_so_length - so_coarse_increments + 1 : min_so_length; | |
663 to_so_length = | |
664 (max_so_length - final_so_length > so_coarse_increments) ? | |
665 final_so_length + so_coarse_increments - 1 : max_so_length; | |
666 | |
667 pass = pass_type_fine; | |
668 so_length_incr = 1; | |
669 } | |
670 } else if (pass == pass_type_fine) { | |
671 // we just finished the second pass | |
672 | |
673 if (!gc_eff_set) { | |
674 // we didn't find a feasible config (yes, it's possible; | |
675 // notice that, sometimes, we go directly into the fine | |
676 // iteration and skip the coarse one) so we bail out | |
677 done = true; | |
678 } else { | |
679 // We did find a feasible config with optimal GC eff | |
680 guarantee( so_length_incr == 1, "invariant" ); | |
681 | |
682 if (final_so_length == 0) { | |
683 // The config is of an empty S-O set, so we'll just bail out | |
684 done = true; | |
685 } else { | |
686 // we'll go around once more, setting the S-O length to 95% | |
687 // of the optimal | |
688 size_t new_so_length = 950 * final_so_length / 1000; | |
689 | |
690 #if 0 | |
691 // leave this in for debugging, just in case | |
692 gclog_or_tty->print_cr(" fine pass: SO length " SIZE_FORMAT | |
693 ", setting it to " SIZE_FORMAT, | |
694 final_so_length, new_so_length); | |
695 #endif // 0 | |
696 | |
697 from_so_length = new_so_length; | |
698 to_so_length = new_so_length; | |
699 fine_so_length = final_so_length; | |
700 | |
701 pass = pass_type_final; | |
702 } | |
703 } | |
704 } else if (pass == pass_type_final) { | |
705 // we just finished the final (third) pass | |
706 | |
707 if (!gc_eff_set) | |
708 // we didn't find a feasible config, so we'll just use the one | |
709 // we found during the second pass, which we saved | |
710 final_so_length = fine_so_length; | |
711 | |
712 // and we're done! | |
713 done = true; | |
714 } else { | |
715 guarantee( false, "should never reach here" ); | |
716 } | |
717 | |
718 // we now go around the outermost loop | |
719 } | |
720 | |
721 // we should have at least one region in the target young length | |
722 _young_list_target_length = MAX2((size_t) 1, final_young_length); | |
723 if (final_so_length >= final_young_length) | |
724 // and we need to ensure that the S-O length is not greater than | |
725 // the target young length (this is being a bit careful) | |
726 final_so_length = 0; | |
727 _young_list_so_prefix_length = final_so_length; | |
728 guarantee( !_in_marking_window || !_last_full_young_gc || | |
729 _young_list_so_prefix_length == 0, "invariant" ); | |
730 | |
731 // let's keep an eye of how long we spend on this calculation | |
732 // right now, I assume that we'll print it when we need it; we | |
733 // should really adde it to the breakdown of a pause | |
734 double end_time_sec = os::elapsedTime(); | |
735 double elapsed_time_ms = (end_time_sec - start_time_sec) * 1000.0; | |
736 | |
737 #if 0 | |
738 // leave this in for debugging, just in case | |
739 gclog_or_tty->print_cr("target = %1.1lf ms, young = " SIZE_FORMAT | |
740 ", SO = " SIZE_FORMAT ", " | |
741 "elapsed %1.2lf ms, calcs: " SIZE_FORMAT " (%s%s) " | |
742 SIZE_FORMAT SIZE_FORMAT, | |
743 target_pause_time_ms, | |
744 _young_list_target_length - _young_list_so_prefix_length, | |
745 _young_list_so_prefix_length, | |
746 elapsed_time_ms, | |
747 calculations, | |
748 full_young_gcs() ? "full" : "partial", | |
749 should_initiate_conc_mark() ? " i-m" : "", | |
750 in_marking_window(), | |
751 in_marking_window_im()); | |
752 #endif // 0 | |
753 | |
754 if (_young_list_target_length < _young_list_min_length) { | |
755 // bummer; this means that, if we do a pause when the optimal | |
756 // config dictates, we'll violate the pause spacing target (the | |
757 // min length was calculate based on the application's current | |
758 // alloc rate); | |
759 | |
760 // so, we have to bite the bullet, and allocate the minimum | |
761 // number. We'll violate our target, but we just can't meet it. | |
762 | |
763 size_t so_length = 0; | |
764 // a note further up explains why we do not want an S-O length | |
765 // during marking | |
766 if (!_in_marking_window && !_last_full_young_gc) | |
767 // but we can still try to see whether we can find an optimal | |
768 // S-O length | |
769 so_length = calculate_optimal_so_length(_young_list_min_length); | |
770 | |
771 #if 0 | |
772 // leave this in for debugging, just in case | |
773 gclog_or_tty->print_cr("adjusted target length from " | |
774 SIZE_FORMAT " to " SIZE_FORMAT | |
775 ", SO " SIZE_FORMAT, | |
776 _young_list_target_length, _young_list_min_length, | |
777 so_length); | |
778 #endif // 0 | |
779 | |
780 _young_list_target_length = | |
781 MAX2(_young_list_min_length, (size_t)1); | |
782 _young_list_so_prefix_length = so_length; | |
783 } | |
784 } else { | |
785 // we are in a partially-young mode or we've run out of regions (due | |
786 // to evacuation failure) | |
787 | |
788 #if 0 | |
789 // leave this in for debugging, just in case | |
790 gclog_or_tty->print_cr("(partial) setting target to " SIZE_FORMAT | |
791 ", SO " SIZE_FORMAT, | |
792 _young_list_min_length, 0); | |
793 #endif // 0 | |
794 | |
795 // we'll do the pause as soon as possible and with no S-O prefix | |
796 // (see above for the reasons behind the latter) | |
797 _young_list_target_length = | |
798 MAX2(_young_list_min_length, (size_t) 1); | |
799 _young_list_so_prefix_length = 0; | |
800 } | |
801 | |
802 _rs_lengths_prediction = rs_lengths; | |
803 } | |
804 | |
805 // This is used by: calculate_optimal_so_length(length). It returns | |
806 // the GC eff and predicted pause time for a particular config | |
807 void | |
808 G1CollectorPolicy::predict_gc_eff(size_t young_length, | |
809 size_t so_length, | |
810 double base_time_ms, | |
811 double* ret_gc_eff, | |
812 double* ret_pause_time_ms) { | |
813 double so_time_ms = predict_scan_only_time_ms(so_length); | |
814 double accum_surv_rate_adj = 0.0; | |
815 if (so_length > 0) | |
816 accum_surv_rate_adj = accum_yg_surv_rate_pred((int)(so_length - 1)); | |
817 double accum_surv_rate = | |
818 accum_yg_surv_rate_pred((int)(young_length - 1)) - accum_surv_rate_adj; | |
819 size_t bytes_to_copy = | |
820 (size_t) (accum_surv_rate * (double) HeapRegion::GrainBytes); | |
821 double copy_time_ms = predict_object_copy_time_ms(bytes_to_copy); | |
822 double young_other_time_ms = | |
823 predict_young_other_time_ms(young_length - so_length); | |
824 double pause_time_ms = | |
825 base_time_ms + so_time_ms + copy_time_ms + young_other_time_ms; | |
826 size_t reclaimed_bytes = | |
827 (young_length - so_length) * HeapRegion::GrainBytes - bytes_to_copy; | |
828 double gc_eff = (double) reclaimed_bytes / pause_time_ms; | |
829 | |
830 *ret_gc_eff = gc_eff; | |
831 *ret_pause_time_ms = pause_time_ms; | |
832 } | |
833 | |
834 // This is used by: calculate_young_list_target_config(rs_length). It | |
835 // returns the GC eff of a particular config. It returns false if that | |
836 // config violates any of the end conditions of the search in the | |
837 // calling method, or true upon success. The end conditions were put | |
838 // here since it's called twice and it was best not to replicate them | |
839 // in the caller. Also, passing the parameteres avoids having to | |
840 // recalculate them in the innermost loop. | |
841 bool | |
842 G1CollectorPolicy::predict_gc_eff(size_t young_length, | |
843 size_t so_length, | |
844 double base_time_with_so_ms, | |
845 size_t init_free_regions, | |
846 double target_pause_time_ms, | |
847 double* ret_gc_eff) { | |
848 *ret_gc_eff = 0.0; | |
849 | |
850 if (young_length >= init_free_regions) | |
851 // end condition 1: not enough space for the young regions | |
852 return false; | |
853 | |
854 double accum_surv_rate_adj = 0.0; | |
855 if (so_length > 0) | |
856 accum_surv_rate_adj = accum_yg_surv_rate_pred((int)(so_length - 1)); | |
857 double accum_surv_rate = | |
858 accum_yg_surv_rate_pred((int)(young_length - 1)) - accum_surv_rate_adj; | |
859 size_t bytes_to_copy = | |
860 (size_t) (accum_surv_rate * (double) HeapRegion::GrainBytes); | |
861 double copy_time_ms = predict_object_copy_time_ms(bytes_to_copy); | |
862 double young_other_time_ms = | |
863 predict_young_other_time_ms(young_length - so_length); | |
864 double pause_time_ms = | |
865 base_time_with_so_ms + copy_time_ms + young_other_time_ms; | |
866 | |
867 if (pause_time_ms > target_pause_time_ms) | |
868 // end condition 2: over the target pause time | |
869 return false; | |
870 | |
871 size_t reclaimed_bytes = | |
872 (young_length - so_length) * HeapRegion::GrainBytes - bytes_to_copy; | |
873 size_t free_bytes = | |
874 (init_free_regions - young_length) * HeapRegion::GrainBytes; | |
875 | |
876 if ((2.0 + sigma()) * (double) bytes_to_copy > (double) free_bytes) | |
877 // end condition 3: out of to-space (conservatively) | |
878 return false; | |
879 | |
880 // success! | |
881 double gc_eff = (double) reclaimed_bytes / pause_time_ms; | |
882 *ret_gc_eff = gc_eff; | |
883 | |
884 return true; | |
885 } | |
886 | |
887 void G1CollectorPolicy::check_prediction_validity() { | |
888 guarantee( adaptive_young_list_length(), "should not call this otherwise" ); | |
889 | |
890 size_t rs_lengths = _g1->young_list_sampled_rs_lengths(); | |
891 if (rs_lengths > _rs_lengths_prediction) { | |
892 // add 10% to avoid having to recalculate often | |
893 size_t rs_lengths_prediction = rs_lengths * 1100 / 1000; | |
894 calculate_young_list_target_config(rs_lengths_prediction); | |
895 } | |
896 } | |
897 | |
898 HeapWord* G1CollectorPolicy::mem_allocate_work(size_t size, | |
899 bool is_tlab, | |
900 bool* gc_overhead_limit_was_exceeded) { | |
901 guarantee(false, "Not using this policy feature yet."); | |
902 return NULL; | |
903 } | |
904 | |
905 // This method controls how a collector handles one or more | |
906 // of its generations being fully allocated. | |
907 HeapWord* G1CollectorPolicy::satisfy_failed_allocation(size_t size, | |
908 bool is_tlab) { | |
909 guarantee(false, "Not using this policy feature yet."); | |
910 return NULL; | |
911 } | |
912 | |
913 | |
914 #ifndef PRODUCT | |
915 bool G1CollectorPolicy::verify_young_ages() { | |
916 HeapRegion* head = _g1->young_list_first_region(); | |
917 return | |
918 verify_young_ages(head, _short_lived_surv_rate_group); | |
919 // also call verify_young_ages on any additional surv rate groups | |
920 } | |
921 | |
922 bool | |
923 G1CollectorPolicy::verify_young_ages(HeapRegion* head, | |
924 SurvRateGroup *surv_rate_group) { | |
925 guarantee( surv_rate_group != NULL, "pre-condition" ); | |
926 | |
927 const char* name = surv_rate_group->name(); | |
928 bool ret = true; | |
929 int prev_age = -1; | |
930 | |
931 for (HeapRegion* curr = head; | |
932 curr != NULL; | |
933 curr = curr->get_next_young_region()) { | |
934 SurvRateGroup* group = curr->surv_rate_group(); | |
935 if (group == NULL && !curr->is_survivor()) { | |
936 gclog_or_tty->print_cr("## %s: encountered NULL surv_rate_group", name); | |
937 ret = false; | |
938 } | |
939 | |
940 if (surv_rate_group == group) { | |
941 int age = curr->age_in_surv_rate_group(); | |
942 | |
943 if (age < 0) { | |
944 gclog_or_tty->print_cr("## %s: encountered negative age", name); | |
945 ret = false; | |
946 } | |
947 | |
948 if (age <= prev_age) { | |
949 gclog_or_tty->print_cr("## %s: region ages are not strictly increasing " | |
950 "(%d, %d)", name, age, prev_age); | |
951 ret = false; | |
952 } | |
953 prev_age = age; | |
954 } | |
955 } | |
956 | |
957 return ret; | |
958 } | |
959 #endif // PRODUCT | |
960 | |
961 void G1CollectorPolicy::record_full_collection_start() { | |
962 _cur_collection_start_sec = os::elapsedTime(); | |
963 // Release the future to-space so that it is available for compaction into. | |
964 _g1->set_full_collection(); | |
965 } | |
966 | |
967 void G1CollectorPolicy::record_full_collection_end() { | |
968 // Consider this like a collection pause for the purposes of allocation | |
969 // since last pause. | |
970 double end_sec = os::elapsedTime(); | |
971 double full_gc_time_sec = end_sec - _cur_collection_start_sec; | |
972 double full_gc_time_ms = full_gc_time_sec * 1000.0; | |
973 | |
974 checkpoint_conc_overhead(); | |
975 | |
976 _all_full_gc_times_ms->add(full_gc_time_ms); | |
977 | |
978 update_recent_gc_times(end_sec, full_gc_time_sec); | |
979 | |
980 _g1->clear_full_collection(); | |
981 | |
982 // "Nuke" the heuristics that control the fully/partially young GC | |
983 // transitions and make sure we start with fully young GCs after the | |
984 // Full GC. | |
985 set_full_young_gcs(true); | |
986 _last_full_young_gc = false; | |
987 _should_revert_to_full_young_gcs = false; | |
988 _should_initiate_conc_mark = false; | |
989 _known_garbage_bytes = 0; | |
990 _known_garbage_ratio = 0.0; | |
991 _in_marking_window = false; | |
992 _in_marking_window_im = false; | |
993 | |
994 _short_lived_surv_rate_group->record_scan_only_prefix(0); | |
995 _short_lived_surv_rate_group->start_adding_regions(); | |
996 // also call this on any additional surv rate groups | |
997 | |
998 _prev_region_num_young = _region_num_young; | |
999 _prev_region_num_tenured = _region_num_tenured; | |
1000 | |
1001 _free_regions_at_end_of_collection = _g1->free_regions(); | |
1002 _scan_only_regions_at_end_of_collection = 0; | |
1003 calculate_young_list_min_length(); | |
1004 calculate_young_list_target_config(); | |
1005 } | |
1006 | |
1007 void G1CollectorPolicy::record_pop_compute_rc_start() { | |
1008 _pop_compute_rc_start = os::elapsedTime(); | |
1009 } | |
1010 void G1CollectorPolicy::record_pop_compute_rc_end() { | |
1011 double ms = (os::elapsedTime() - _pop_compute_rc_start)*1000.0; | |
1012 _cur_popular_compute_rc_time_ms = ms; | |
1013 _pop_compute_rc_start = 0.0; | |
1014 } | |
1015 void G1CollectorPolicy::record_pop_evac_start() { | |
1016 _pop_evac_start = os::elapsedTime(); | |
1017 } | |
1018 void G1CollectorPolicy::record_pop_evac_end() { | |
1019 double ms = (os::elapsedTime() - _pop_evac_start)*1000.0; | |
1020 _cur_popular_evac_time_ms = ms; | |
1021 _pop_evac_start = 0.0; | |
1022 } | |
1023 | |
1024 void G1CollectorPolicy::record_before_bytes(size_t bytes) { | |
1025 _bytes_in_to_space_before_gc += bytes; | |
1026 } | |
1027 | |
1028 void G1CollectorPolicy::record_after_bytes(size_t bytes) { | |
1029 _bytes_in_to_space_after_gc += bytes; | |
1030 } | |
1031 | |
1032 void G1CollectorPolicy::record_stop_world_start() { | |
1033 _stop_world_start = os::elapsedTime(); | |
1034 } | |
1035 | |
1036 void G1CollectorPolicy::record_collection_pause_start(double start_time_sec, | |
1037 size_t start_used) { | |
1038 if (PrintGCDetails) { | |
1039 gclog_or_tty->stamp(PrintGCTimeStamps); | |
1040 gclog_or_tty->print("[GC pause"); | |
1041 if (in_young_gc_mode()) | |
1042 gclog_or_tty->print(" (%s)", full_young_gcs() ? "young" : "partial"); | |
1043 } | |
1044 | |
1045 assert(_g1->used_regions() == _g1->recalculate_used_regions(), | |
1046 "sanity"); | |
1047 | |
1048 double s_w_t_ms = (start_time_sec - _stop_world_start) * 1000.0; | |
1049 _all_stop_world_times_ms->add(s_w_t_ms); | |
1050 _stop_world_start = 0.0; | |
1051 | |
1052 _cur_collection_start_sec = start_time_sec; | |
1053 _cur_collection_pause_used_at_start_bytes = start_used; | |
1054 _cur_collection_pause_used_regions_at_start = _g1->used_regions(); | |
1055 _pending_cards = _g1->pending_card_num(); | |
1056 _max_pending_cards = _g1->max_pending_card_num(); | |
1057 | |
1058 _bytes_in_to_space_before_gc = 0; | |
1059 _bytes_in_to_space_after_gc = 0; | |
1060 _bytes_in_collection_set_before_gc = 0; | |
1061 | |
1062 #ifdef DEBUG | |
1063 // initialise these to something well known so that we can spot | |
1064 // if they are not set properly | |
1065 | |
1066 for (int i = 0; i < _parallel_gc_threads; ++i) { | |
1067 _par_last_ext_root_scan_times_ms[i] = -666.0; | |
1068 _par_last_mark_stack_scan_times_ms[i] = -666.0; | |
1069 _par_last_scan_only_times_ms[i] = -666.0; | |
1070 _par_last_scan_only_regions_scanned[i] = -666.0; | |
1071 _par_last_update_rs_start_times_ms[i] = -666.0; | |
1072 _par_last_update_rs_times_ms[i] = -666.0; | |
1073 _par_last_update_rs_processed_buffers[i] = -666.0; | |
1074 _par_last_scan_rs_start_times_ms[i] = -666.0; | |
1075 _par_last_scan_rs_times_ms[i] = -666.0; | |
1076 _par_last_scan_new_refs_times_ms[i] = -666.0; | |
1077 _par_last_obj_copy_times_ms[i] = -666.0; | |
1078 _par_last_termination_times_ms[i] = -666.0; | |
1079 | |
1080 _pop_par_last_update_rs_start_times_ms[i] = -666.0; | |
1081 _pop_par_last_update_rs_times_ms[i] = -666.0; | |
1082 _pop_par_last_update_rs_processed_buffers[i] = -666.0; | |
1083 _pop_par_last_scan_rs_start_times_ms[i] = -666.0; | |
1084 _pop_par_last_scan_rs_times_ms[i] = -666.0; | |
1085 _pop_par_last_closure_app_times_ms[i] = -666.0; | |
1086 } | |
1087 #endif | |
1088 | |
1089 for (int i = 0; i < _aux_num; ++i) { | |
1090 _cur_aux_times_ms[i] = 0.0; | |
1091 _cur_aux_times_set[i] = false; | |
1092 } | |
1093 | |
1094 _satb_drain_time_set = false; | |
1095 _last_satb_drain_processed_buffers = -1; | |
1096 | |
1097 if (in_young_gc_mode()) | |
1098 _last_young_gc_full = false; | |
1099 | |
1100 | |
1101 // do that for any other surv rate groups | |
1102 _short_lived_surv_rate_group->stop_adding_regions(); | |
1103 size_t short_lived_so_length = _young_list_so_prefix_length; | |
1104 _short_lived_surv_rate_group->record_scan_only_prefix(short_lived_so_length); | |
1105 tag_scan_only(short_lived_so_length); | |
1106 | |
1107 assert( verify_young_ages(), "region age verification" ); | |
1108 } | |
1109 | |
1110 void G1CollectorPolicy::tag_scan_only(size_t short_lived_scan_only_length) { | |
1111 // done in a way that it can be extended for other surv rate groups too... | |
1112 | |
1113 HeapRegion* head = _g1->young_list_first_region(); | |
1114 bool finished_short_lived = (short_lived_scan_only_length == 0); | |
1115 | |
1116 if (finished_short_lived) | |
1117 return; | |
1118 | |
1119 for (HeapRegion* curr = head; | |
1120 curr != NULL; | |
1121 curr = curr->get_next_young_region()) { | |
1122 SurvRateGroup* surv_rate_group = curr->surv_rate_group(); | |
1123 int age = curr->age_in_surv_rate_group(); | |
1124 | |
1125 if (surv_rate_group == _short_lived_surv_rate_group) { | |
1126 if ((size_t)age < short_lived_scan_only_length) | |
1127 curr->set_scan_only(); | |
1128 else | |
1129 finished_short_lived = true; | |
1130 } | |
1131 | |
1132 | |
1133 if (finished_short_lived) | |
1134 return; | |
1135 } | |
1136 | |
1137 guarantee( false, "we should never reach here" ); | |
1138 } | |
1139 | |
1140 void G1CollectorPolicy::record_popular_pause_preamble_start() { | |
1141 _cur_popular_preamble_start_ms = os::elapsedTime() * 1000.0; | |
1142 } | |
1143 | |
1144 void G1CollectorPolicy::record_popular_pause_preamble_end() { | |
1145 _cur_popular_preamble_time_ms = | |
1146 (os::elapsedTime() * 1000.0) - _cur_popular_preamble_start_ms; | |
1147 | |
1148 // copy the recorded statistics of the first pass to temporary arrays | |
1149 for (int i = 0; i < _parallel_gc_threads; ++i) { | |
1150 _pop_par_last_update_rs_start_times_ms[i] = _par_last_update_rs_start_times_ms[i]; | |
1151 _pop_par_last_update_rs_times_ms[i] = _par_last_update_rs_times_ms[i]; | |
1152 _pop_par_last_update_rs_processed_buffers[i] = _par_last_update_rs_processed_buffers[i]; | |
1153 _pop_par_last_scan_rs_start_times_ms[i] = _par_last_scan_rs_start_times_ms[i]; | |
1154 _pop_par_last_scan_rs_times_ms[i] = _par_last_scan_rs_times_ms[i]; | |
1155 _pop_par_last_closure_app_times_ms[i] = _par_last_obj_copy_times_ms[i]; | |
1156 } | |
1157 } | |
1158 | |
1159 void G1CollectorPolicy::record_mark_closure_time(double mark_closure_time_ms) { | |
1160 _mark_closure_time_ms = mark_closure_time_ms; | |
1161 } | |
1162 | |
1163 void G1CollectorPolicy::record_concurrent_mark_init_start() { | |
1164 _mark_init_start_sec = os::elapsedTime(); | |
1165 guarantee(!in_young_gc_mode(), "should not do be here in young GC mode"); | |
1166 } | |
1167 | |
1168 void G1CollectorPolicy::record_concurrent_mark_init_end_pre(double | |
1169 mark_init_elapsed_time_ms) { | |
1170 _during_marking = true; | |
1171 _should_initiate_conc_mark = false; | |
1172 _cur_mark_stop_world_time_ms = mark_init_elapsed_time_ms; | |
1173 } | |
1174 | |
1175 void G1CollectorPolicy::record_concurrent_mark_init_end() { | |
1176 double end_time_sec = os::elapsedTime(); | |
1177 double elapsed_time_ms = (end_time_sec - _mark_init_start_sec) * 1000.0; | |
1178 _concurrent_mark_init_times_ms->add(elapsed_time_ms); | |
1179 checkpoint_conc_overhead(); | |
1180 record_concurrent_mark_init_end_pre(elapsed_time_ms); | |
1181 | |
1182 _mmu_tracker->add_pause(_mark_init_start_sec, end_time_sec, true); | |
1183 } | |
1184 | |
1185 void G1CollectorPolicy::record_concurrent_mark_remark_start() { | |
1186 _mark_remark_start_sec = os::elapsedTime(); | |
1187 _during_marking = false; | |
1188 } | |
1189 | |
1190 void G1CollectorPolicy::record_concurrent_mark_remark_end() { | |
1191 double end_time_sec = os::elapsedTime(); | |
1192 double elapsed_time_ms = (end_time_sec - _mark_remark_start_sec)*1000.0; | |
1193 checkpoint_conc_overhead(); | |
1194 _concurrent_mark_remark_times_ms->add(elapsed_time_ms); | |
1195 _cur_mark_stop_world_time_ms += elapsed_time_ms; | |
1196 _prev_collection_pause_end_ms += elapsed_time_ms; | |
1197 | |
1198 _mmu_tracker->add_pause(_mark_remark_start_sec, end_time_sec, true); | |
1199 } | |
1200 | |
1201 void G1CollectorPolicy::record_concurrent_mark_cleanup_start() { | |
1202 _mark_cleanup_start_sec = os::elapsedTime(); | |
1203 } | |
1204 | |
1205 void | |
1206 G1CollectorPolicy::record_concurrent_mark_cleanup_end(size_t freed_bytes, | |
1207 size_t max_live_bytes) { | |
1208 record_concurrent_mark_cleanup_end_work1(freed_bytes, max_live_bytes); | |
1209 record_concurrent_mark_cleanup_end_work2(); | |
1210 } | |
1211 | |
1212 void | |
1213 G1CollectorPolicy:: | |
1214 record_concurrent_mark_cleanup_end_work1(size_t freed_bytes, | |
1215 size_t max_live_bytes) { | |
1216 if (_n_marks < 2) _n_marks++; | |
1217 if (G1PolicyVerbose > 0) | |
1218 gclog_or_tty->print_cr("At end of marking, max_live is " SIZE_FORMAT " MB " | |
1219 " (of " SIZE_FORMAT " MB heap).", | |
1220 max_live_bytes/M, _g1->capacity()/M); | |
1221 } | |
1222 | |
1223 // The important thing about this is that it includes "os::elapsedTime". | |
1224 void G1CollectorPolicy::record_concurrent_mark_cleanup_end_work2() { | |
1225 checkpoint_conc_overhead(); | |
1226 double end_time_sec = os::elapsedTime(); | |
1227 double elapsed_time_ms = (end_time_sec - _mark_cleanup_start_sec)*1000.0; | |
1228 _concurrent_mark_cleanup_times_ms->add(elapsed_time_ms); | |
1229 _cur_mark_stop_world_time_ms += elapsed_time_ms; | |
1230 _prev_collection_pause_end_ms += elapsed_time_ms; | |
1231 | |
1232 _mmu_tracker->add_pause(_mark_cleanup_start_sec, end_time_sec, true); | |
1233 | |
1234 _num_markings++; | |
1235 | |
1236 // We did a marking, so reset the "since_last_mark" variables. | |
1237 double considerConcMarkCost = 1.0; | |
1238 // If there are available processors, concurrent activity is free... | |
1239 if (Threads::number_of_non_daemon_threads() * 2 < | |
1240 os::active_processor_count()) { | |
1241 considerConcMarkCost = 0.0; | |
1242 } | |
1243 _n_pauses_at_mark_end = _n_pauses; | |
1244 _n_marks_since_last_pause++; | |
1245 _conc_mark_initiated = false; | |
1246 } | |
1247 | |
1248 void | |
1249 G1CollectorPolicy::record_concurrent_mark_cleanup_completed() { | |
1250 if (in_young_gc_mode()) { | |
1251 _should_revert_to_full_young_gcs = false; | |
1252 _last_full_young_gc = true; | |
1253 _in_marking_window = false; | |
1254 if (adaptive_young_list_length()) | |
1255 calculate_young_list_target_config(); | |
1256 } | |
1257 } | |
1258 | |
1259 void G1CollectorPolicy::record_concurrent_pause() { | |
1260 if (_stop_world_start > 0.0) { | |
1261 double yield_ms = (os::elapsedTime() - _stop_world_start) * 1000.0; | |
1262 _all_yield_times_ms->add(yield_ms); | |
1263 } | |
1264 } | |
1265 | |
1266 void G1CollectorPolicy::record_concurrent_pause_end() { | |
1267 } | |
1268 | |
1269 void G1CollectorPolicy::record_collection_pause_end_CH_strong_roots() { | |
1270 _cur_CH_strong_roots_end_sec = os::elapsedTime(); | |
1271 _cur_CH_strong_roots_dur_ms = | |
1272 (_cur_CH_strong_roots_end_sec - _cur_collection_start_sec) * 1000.0; | |
1273 } | |
1274 | |
1275 void G1CollectorPolicy::record_collection_pause_end_G1_strong_roots() { | |
1276 _cur_G1_strong_roots_end_sec = os::elapsedTime(); | |
1277 _cur_G1_strong_roots_dur_ms = | |
1278 (_cur_G1_strong_roots_end_sec - _cur_CH_strong_roots_end_sec) * 1000.0; | |
1279 } | |
1280 | |
1281 template<class T> | |
1282 T sum_of(T* sum_arr, int start, int n, int N) { | |
1283 T sum = (T)0; | |
1284 for (int i = 0; i < n; i++) { | |
1285 int j = (start + i) % N; | |
1286 sum += sum_arr[j]; | |
1287 } | |
1288 return sum; | |
1289 } | |
1290 | |
1291 void G1CollectorPolicy::print_par_stats (int level, | |
1292 const char* str, | |
1293 double* data, | |
1294 bool summary) { | |
1295 double min = data[0], max = data[0]; | |
1296 double total = 0.0; | |
1297 int j; | |
1298 for (j = 0; j < level; ++j) | |
1299 gclog_or_tty->print(" "); | |
1300 gclog_or_tty->print("[%s (ms):", str); | |
1301 for (uint i = 0; i < ParallelGCThreads; ++i) { | |
1302 double val = data[i]; | |
1303 if (val < min) | |
1304 min = val; | |
1305 if (val > max) | |
1306 max = val; | |
1307 total += val; | |
1308 gclog_or_tty->print(" %3.1lf", val); | |
1309 } | |
1310 if (summary) { | |
1311 gclog_or_tty->print_cr(""); | |
1312 double avg = total / (double) ParallelGCThreads; | |
1313 gclog_or_tty->print(" "); | |
1314 for (j = 0; j < level; ++j) | |
1315 gclog_or_tty->print(" "); | |
1316 gclog_or_tty->print("Avg: %5.1lf, Min: %5.1lf, Max: %5.1lf", | |
1317 avg, min, max); | |
1318 } | |
1319 gclog_or_tty->print_cr("]"); | |
1320 } | |
1321 | |
1322 void G1CollectorPolicy::print_par_buffers (int level, | |
1323 const char* str, | |
1324 double* data, | |
1325 bool summary) { | |
1326 double min = data[0], max = data[0]; | |
1327 double total = 0.0; | |
1328 int j; | |
1329 for (j = 0; j < level; ++j) | |
1330 gclog_or_tty->print(" "); | |
1331 gclog_or_tty->print("[%s :", str); | |
1332 for (uint i = 0; i < ParallelGCThreads; ++i) { | |
1333 double val = data[i]; | |
1334 if (val < min) | |
1335 min = val; | |
1336 if (val > max) | |
1337 max = val; | |
1338 total += val; | |
1339 gclog_or_tty->print(" %d", (int) val); | |
1340 } | |
1341 if (summary) { | |
1342 gclog_or_tty->print_cr(""); | |
1343 double avg = total / (double) ParallelGCThreads; | |
1344 gclog_or_tty->print(" "); | |
1345 for (j = 0; j < level; ++j) | |
1346 gclog_or_tty->print(" "); | |
1347 gclog_or_tty->print("Sum: %d, Avg: %d, Min: %d, Max: %d", | |
1348 (int)total, (int)avg, (int)min, (int)max); | |
1349 } | |
1350 gclog_or_tty->print_cr("]"); | |
1351 } | |
1352 | |
1353 void G1CollectorPolicy::print_stats (int level, | |
1354 const char* str, | |
1355 double value) { | |
1356 for (int j = 0; j < level; ++j) | |
1357 gclog_or_tty->print(" "); | |
1358 gclog_or_tty->print_cr("[%s: %5.1lf ms]", str, value); | |
1359 } | |
1360 | |
1361 void G1CollectorPolicy::print_stats (int level, | |
1362 const char* str, | |
1363 int value) { | |
1364 for (int j = 0; j < level; ++j) | |
1365 gclog_or_tty->print(" "); | |
1366 gclog_or_tty->print_cr("[%s: %d]", str, value); | |
1367 } | |
1368 | |
1369 double G1CollectorPolicy::avg_value (double* data) { | |
1370 if (ParallelGCThreads > 0) { | |
1371 double ret = 0.0; | |
1372 for (uint i = 0; i < ParallelGCThreads; ++i) | |
1373 ret += data[i]; | |
1374 return ret / (double) ParallelGCThreads; | |
1375 } else { | |
1376 return data[0]; | |
1377 } | |
1378 } | |
1379 | |
1380 double G1CollectorPolicy::max_value (double* data) { | |
1381 if (ParallelGCThreads > 0) { | |
1382 double ret = data[0]; | |
1383 for (uint i = 1; i < ParallelGCThreads; ++i) | |
1384 if (data[i] > ret) | |
1385 ret = data[i]; | |
1386 return ret; | |
1387 } else { | |
1388 return data[0]; | |
1389 } | |
1390 } | |
1391 | |
1392 double G1CollectorPolicy::sum_of_values (double* data) { | |
1393 if (ParallelGCThreads > 0) { | |
1394 double sum = 0.0; | |
1395 for (uint i = 0; i < ParallelGCThreads; i++) | |
1396 sum += data[i]; | |
1397 return sum; | |
1398 } else { | |
1399 return data[0]; | |
1400 } | |
1401 } | |
1402 | |
1403 double G1CollectorPolicy::max_sum (double* data1, | |
1404 double* data2) { | |
1405 double ret = data1[0] + data2[0]; | |
1406 | |
1407 if (ParallelGCThreads > 0) { | |
1408 for (uint i = 1; i < ParallelGCThreads; ++i) { | |
1409 double data = data1[i] + data2[i]; | |
1410 if (data > ret) | |
1411 ret = data; | |
1412 } | |
1413 } | |
1414 return ret; | |
1415 } | |
1416 | |
1417 // Anything below that is considered to be zero | |
1418 #define MIN_TIMER_GRANULARITY 0.0000001 | |
1419 | |
1420 void G1CollectorPolicy::record_collection_pause_end(bool popular, | |
1421 bool abandoned) { | |
1422 double end_time_sec = os::elapsedTime(); | |
1423 double elapsed_ms = _last_pause_time_ms; | |
1424 bool parallel = ParallelGCThreads > 0; | |
1425 double evac_ms = (end_time_sec - _cur_G1_strong_roots_end_sec) * 1000.0; | |
1426 size_t rs_size = | |
1427 _cur_collection_pause_used_regions_at_start - collection_set_size(); | |
1428 size_t cur_used_bytes = _g1->used(); | |
1429 assert(cur_used_bytes == _g1->recalculate_used(), "It should!"); | |
1430 bool last_pause_included_initial_mark = false; | |
1431 | |
1432 #ifndef PRODUCT | |
1433 if (G1YoungSurvRateVerbose) { | |
1434 gclog_or_tty->print_cr(""); | |
1435 _short_lived_surv_rate_group->print(); | |
1436 // do that for any other surv rate groups too | |
1437 } | |
1438 #endif // PRODUCT | |
1439 | |
1440 checkpoint_conc_overhead(); | |
1441 | |
1442 if (in_young_gc_mode()) { | |
1443 last_pause_included_initial_mark = _should_initiate_conc_mark; | |
1444 if (last_pause_included_initial_mark) | |
1445 record_concurrent_mark_init_end_pre(0.0); | |
1446 | |
1447 size_t min_used_targ = | |
1448 (_g1->capacity() / 100) * (G1SteadyStateUsed - G1SteadyStateUsedDelta); | |
1449 | |
1450 if (cur_used_bytes > min_used_targ) { | |
1451 if (cur_used_bytes <= _prev_collection_pause_used_at_end_bytes) { | |
1452 } else if (!_g1->mark_in_progress() && !_last_full_young_gc) { | |
1453 _should_initiate_conc_mark = true; | |
1454 } | |
1455 } | |
1456 | |
1457 _prev_collection_pause_used_at_end_bytes = cur_used_bytes; | |
1458 } | |
1459 | |
1460 _mmu_tracker->add_pause(end_time_sec - elapsed_ms/1000.0, | |
1461 end_time_sec, false); | |
1462 | |
1463 guarantee(_cur_collection_pause_used_regions_at_start >= | |
1464 collection_set_size(), | |
1465 "Negative RS size?"); | |
1466 | |
1467 // This assert is exempted when we're doing parallel collection pauses, | |
1468 // because the fragmentation caused by the parallel GC allocation buffers | |
1469 // can lead to more memory being used during collection than was used | |
1470 // before. Best leave this out until the fragmentation problem is fixed. | |
1471 // Pauses in which evacuation failed can also lead to negative | |
1472 // collections, since no space is reclaimed from a region containing an | |
1473 // object whose evacuation failed. | |
1474 // Further, we're now always doing parallel collection. But I'm still | |
1475 // leaving this here as a placeholder for a more precise assertion later. | |
1476 // (DLD, 10/05.) | |
1477 assert((true || parallel) // Always using GC LABs now. | |
1478 || _g1->evacuation_failed() | |
1479 || _cur_collection_pause_used_at_start_bytes >= cur_used_bytes, | |
1480 "Negative collection"); | |
1481 | |
1482 size_t freed_bytes = | |
1483 _cur_collection_pause_used_at_start_bytes - cur_used_bytes; | |
1484 size_t surviving_bytes = _collection_set_bytes_used_before - freed_bytes; | |
1485 double survival_fraction = | |
1486 (double)surviving_bytes/ | |
1487 (double)_collection_set_bytes_used_before; | |
1488 | |
1489 _n_pauses++; | |
1490 | |
1491 if (!abandoned) { | |
1492 _recent_CH_strong_roots_times_ms->add(_cur_CH_strong_roots_dur_ms); | |
1493 _recent_G1_strong_roots_times_ms->add(_cur_G1_strong_roots_dur_ms); | |
1494 _recent_evac_times_ms->add(evac_ms); | |
1495 _recent_pause_times_ms->add(elapsed_ms); | |
1496 | |
1497 _recent_rs_sizes->add(rs_size); | |
1498 | |
1499 // We exempt parallel collection from this check because Alloc Buffer | |
1500 // fragmentation can produce negative collections. Same with evac | |
1501 // failure. | |
1502 // Further, we're now always doing parallel collection. But I'm still | |
1503 // leaving this here as a placeholder for a more precise assertion later. | |
1504 // (DLD, 10/05. | |
1505 assert((true || parallel) | |
1506 || _g1->evacuation_failed() | |
1507 || surviving_bytes <= _collection_set_bytes_used_before, | |
1508 "Or else negative collection!"); | |
1509 _recent_CS_bytes_used_before->add(_collection_set_bytes_used_before); | |
1510 _recent_CS_bytes_surviving->add(surviving_bytes); | |
1511 | |
1512 // this is where we update the allocation rate of the application | |
1513 double app_time_ms = | |
1514 (_cur_collection_start_sec * 1000.0 - _prev_collection_pause_end_ms); | |
1515 if (app_time_ms < MIN_TIMER_GRANULARITY) { | |
1516 // This usually happens due to the timer not having the required | |
1517 // granularity. Some Linuxes are the usual culprits. | |
1518 // We'll just set it to something (arbitrarily) small. | |
1519 app_time_ms = 1.0; | |
1520 } | |
1521 size_t regions_allocated = | |
1522 (_region_num_young - _prev_region_num_young) + | |
1523 (_region_num_tenured - _prev_region_num_tenured); | |
1524 double alloc_rate_ms = (double) regions_allocated / app_time_ms; | |
1525 _alloc_rate_ms_seq->add(alloc_rate_ms); | |
1526 _prev_region_num_young = _region_num_young; | |
1527 _prev_region_num_tenured = _region_num_tenured; | |
1528 | |
1529 double interval_ms = | |
1530 (end_time_sec - _recent_prev_end_times_for_all_gcs_sec->oldest()) * 1000.0; | |
1531 update_recent_gc_times(end_time_sec, elapsed_ms); | |
1532 _recent_avg_pause_time_ratio = _recent_gc_times_ms->sum()/interval_ms; | |
1533 assert(recent_avg_pause_time_ratio() < 1.00, "All GC?"); | |
1534 } | |
1535 | |
1536 if (G1PolicyVerbose > 1) { | |
1537 gclog_or_tty->print_cr(" Recording collection pause(%d)", _n_pauses); | |
1538 } | |
1539 | |
1540 PauseSummary* summary; | |
1541 if (!abandoned && !popular) | |
1542 summary = _non_pop_summary; | |
1543 else if (!abandoned && popular) | |
1544 summary = _pop_summary; | |
1545 else if (abandoned && !popular) | |
1546 summary = _non_pop_abandoned_summary; | |
1547 else if (abandoned && popular) | |
1548 summary = _pop_abandoned_summary; | |
1549 else | |
1550 guarantee(false, "should not get here!"); | |
1551 | |
1552 double pop_update_rs_time; | |
1553 double pop_update_rs_processed_buffers; | |
1554 double pop_scan_rs_time; | |
1555 double pop_closure_app_time; | |
1556 double pop_other_time; | |
1557 | |
1558 if (popular) { | |
1559 PopPreambleSummary* preamble_summary = summary->pop_preamble_summary(); | |
1560 guarantee(preamble_summary != NULL, "should not be null!"); | |
1561 | |
1562 pop_update_rs_time = avg_value(_pop_par_last_update_rs_times_ms); | |
1563 pop_update_rs_processed_buffers = | |
1564 sum_of_values(_pop_par_last_update_rs_processed_buffers); | |
1565 pop_scan_rs_time = avg_value(_pop_par_last_scan_rs_times_ms); | |
1566 pop_closure_app_time = avg_value(_pop_par_last_closure_app_times_ms); | |
1567 pop_other_time = _cur_popular_preamble_time_ms - | |
1568 (pop_update_rs_time + pop_scan_rs_time + pop_closure_app_time + | |
1569 _cur_popular_evac_time_ms); | |
1570 | |
1571 preamble_summary->record_pop_preamble_time_ms(_cur_popular_preamble_time_ms); | |
1572 preamble_summary->record_pop_update_rs_time_ms(pop_update_rs_time); | |
1573 preamble_summary->record_pop_scan_rs_time_ms(pop_scan_rs_time); | |
1574 preamble_summary->record_pop_closure_app_time_ms(pop_closure_app_time); | |
1575 preamble_summary->record_pop_evacuation_time_ms(_cur_popular_evac_time_ms); | |
1576 preamble_summary->record_pop_other_time_ms(pop_other_time); | |
1577 } | |
1578 | |
1579 double ext_root_scan_time = avg_value(_par_last_ext_root_scan_times_ms); | |
1580 double mark_stack_scan_time = avg_value(_par_last_mark_stack_scan_times_ms); | |
1581 double scan_only_time = avg_value(_par_last_scan_only_times_ms); | |
1582 double scan_only_regions_scanned = | |
1583 sum_of_values(_par_last_scan_only_regions_scanned); | |
1584 double update_rs_time = avg_value(_par_last_update_rs_times_ms); | |
1585 double update_rs_processed_buffers = | |
1586 sum_of_values(_par_last_update_rs_processed_buffers); | |
1587 double scan_rs_time = avg_value(_par_last_scan_rs_times_ms); | |
1588 double obj_copy_time = avg_value(_par_last_obj_copy_times_ms); | |
1589 double termination_time = avg_value(_par_last_termination_times_ms); | |
1590 | |
1591 double parallel_other_time; | |
1592 if (!abandoned) { | |
1593 MainBodySummary* body_summary = summary->main_body_summary(); | |
1594 guarantee(body_summary != NULL, "should not be null!"); | |
1595 | |
1596 if (_satb_drain_time_set) | |
1597 body_summary->record_satb_drain_time_ms(_cur_satb_drain_time_ms); | |
1598 else | |
1599 body_summary->record_satb_drain_time_ms(0.0); | |
1600 body_summary->record_ext_root_scan_time_ms(ext_root_scan_time); | |
1601 body_summary->record_mark_stack_scan_time_ms(mark_stack_scan_time); | |
1602 body_summary->record_scan_only_time_ms(scan_only_time); | |
1603 body_summary->record_update_rs_time_ms(update_rs_time); | |
1604 body_summary->record_scan_rs_time_ms(scan_rs_time); | |
1605 body_summary->record_obj_copy_time_ms(obj_copy_time); | |
1606 if (parallel) { | |
1607 body_summary->record_parallel_time_ms(_cur_collection_par_time_ms); | |
1608 body_summary->record_clear_ct_time_ms(_cur_clear_ct_time_ms); | |
1609 body_summary->record_termination_time_ms(termination_time); | |
1610 parallel_other_time = _cur_collection_par_time_ms - | |
1611 (update_rs_time + ext_root_scan_time + mark_stack_scan_time + | |
1612 scan_only_time + scan_rs_time + obj_copy_time + termination_time); | |
1613 body_summary->record_parallel_other_time_ms(parallel_other_time); | |
1614 } | |
1615 body_summary->record_mark_closure_time_ms(_mark_closure_time_ms); | |
1616 } | |
1617 | |
1618 if (G1PolicyVerbose > 1) { | |
1619 gclog_or_tty->print_cr(" ET: %10.6f ms (avg: %10.6f ms)\n" | |
1620 " CH Strong: %10.6f ms (avg: %10.6f ms)\n" | |
1621 " G1 Strong: %10.6f ms (avg: %10.6f ms)\n" | |
1622 " Evac: %10.6f ms (avg: %10.6f ms)\n" | |
1623 " ET-RS: %10.6f ms (avg: %10.6f ms)\n" | |
1624 " |RS|: " SIZE_FORMAT, | |
1625 elapsed_ms, recent_avg_time_for_pauses_ms(), | |
1626 _cur_CH_strong_roots_dur_ms, recent_avg_time_for_CH_strong_ms(), | |
1627 _cur_G1_strong_roots_dur_ms, recent_avg_time_for_G1_strong_ms(), | |
1628 evac_ms, recent_avg_time_for_evac_ms(), | |
1629 scan_rs_time, | |
1630 recent_avg_time_for_pauses_ms() - | |
1631 recent_avg_time_for_G1_strong_ms(), | |
1632 rs_size); | |
1633 | |
1634 gclog_or_tty->print_cr(" Used at start: " SIZE_FORMAT"K" | |
1635 " At end " SIZE_FORMAT "K\n" | |
1636 " garbage : " SIZE_FORMAT "K" | |
1637 " of " SIZE_FORMAT "K\n" | |
1638 " survival : %6.2f%% (%6.2f%% avg)", | |
1639 _cur_collection_pause_used_at_start_bytes/K, | |
1640 _g1->used()/K, freed_bytes/K, | |
1641 _collection_set_bytes_used_before/K, | |
1642 survival_fraction*100.0, | |
1643 recent_avg_survival_fraction()*100.0); | |
1644 gclog_or_tty->print_cr(" Recent %% gc pause time: %6.2f", | |
1645 recent_avg_pause_time_ratio() * 100.0); | |
1646 } | |
1647 | |
1648 double other_time_ms = elapsed_ms; | |
1649 if (popular) | |
1650 other_time_ms -= _cur_popular_preamble_time_ms; | |
1651 | |
1652 if (!abandoned) { | |
1653 if (_satb_drain_time_set) | |
1654 other_time_ms -= _cur_satb_drain_time_ms; | |
1655 | |
1656 if (parallel) | |
1657 other_time_ms -= _cur_collection_par_time_ms + _cur_clear_ct_time_ms; | |
1658 else | |
1659 other_time_ms -= | |
1660 update_rs_time + | |
1661 ext_root_scan_time + mark_stack_scan_time + scan_only_time + | |
1662 scan_rs_time + obj_copy_time; | |
1663 } | |
1664 | |
1665 if (PrintGCDetails) { | |
1666 gclog_or_tty->print_cr("%s%s, %1.8lf secs]", | |
1667 (popular && !abandoned) ? " (popular)" : | |
1668 (!popular && abandoned) ? " (abandoned)" : | |
1669 (popular && abandoned) ? " (popular/abandoned)" : "", | |
1670 (last_pause_included_initial_mark) ? " (initial-mark)" : "", | |
1671 elapsed_ms / 1000.0); | |
1672 | |
1673 if (!abandoned) { | |
1674 if (_satb_drain_time_set) | |
1675 print_stats(1, "SATB Drain Time", _cur_satb_drain_time_ms); | |
1676 if (_last_satb_drain_processed_buffers >= 0) | |
1677 print_stats(2, "Processed Buffers", _last_satb_drain_processed_buffers); | |
1678 } | |
1679 if (popular) | |
1680 print_stats(1, "Popularity Preamble", _cur_popular_preamble_time_ms); | |
1681 if (parallel) { | |
1682 if (popular) { | |
1683 print_par_stats(2, "Update RS (Start)", _pop_par_last_update_rs_start_times_ms, false); | |
1684 print_par_stats(2, "Update RS", _pop_par_last_update_rs_times_ms); | |
1685 if (G1RSBarrierUseQueue) | |
1686 print_par_buffers(3, "Processed Buffers", | |
1687 _pop_par_last_update_rs_processed_buffers, true); | |
1688 print_par_stats(2, "Scan RS", _pop_par_last_scan_rs_times_ms); | |
1689 print_par_stats(2, "Closure app", _pop_par_last_closure_app_times_ms); | |
1690 print_stats(2, "Evacuation", _cur_popular_evac_time_ms); | |
1691 print_stats(2, "Other", pop_other_time); | |
1692 } | |
1693 if (!abandoned) { | |
1694 print_stats(1, "Parallel Time", _cur_collection_par_time_ms); | |
1695 if (!popular) { | |
1696 print_par_stats(2, "Update RS (Start)", _par_last_update_rs_start_times_ms, false); | |
1697 print_par_stats(2, "Update RS", _par_last_update_rs_times_ms); | |
1698 if (G1RSBarrierUseQueue) | |
1699 print_par_buffers(3, "Processed Buffers", | |
1700 _par_last_update_rs_processed_buffers, true); | |
1701 } | |
1702 print_par_stats(2, "Ext Root Scanning", _par_last_ext_root_scan_times_ms); | |
1703 print_par_stats(2, "Mark Stack Scanning", _par_last_mark_stack_scan_times_ms); | |
1704 print_par_stats(2, "Scan-Only Scanning", _par_last_scan_only_times_ms); | |
1705 print_par_buffers(3, "Scan-Only Regions", | |
1706 _par_last_scan_only_regions_scanned, true); | |
1707 print_par_stats(2, "Scan RS", _par_last_scan_rs_times_ms); | |
1708 print_par_stats(2, "Object Copy", _par_last_obj_copy_times_ms); | |
1709 print_par_stats(2, "Termination", _par_last_termination_times_ms); | |
1710 print_stats(2, "Other", parallel_other_time); | |
1711 print_stats(1, "Clear CT", _cur_clear_ct_time_ms); | |
1712 } | |
1713 } else { | |
1714 if (popular) { | |
1715 print_stats(2, "Update RS", pop_update_rs_time); | |
1716 if (G1RSBarrierUseQueue) | |
1717 print_stats(3, "Processed Buffers", | |
1718 (int)pop_update_rs_processed_buffers); | |
1719 print_stats(2, "Scan RS", pop_scan_rs_time); | |
1720 print_stats(2, "Closure App", pop_closure_app_time); | |
1721 print_stats(2, "Evacuation", _cur_popular_evac_time_ms); | |
1722 print_stats(2, "Other", pop_other_time); | |
1723 } | |
1724 if (!abandoned) { | |
1725 if (!popular) { | |
1726 print_stats(1, "Update RS", update_rs_time); | |
1727 if (G1RSBarrierUseQueue) | |
1728 print_stats(2, "Processed Buffers", | |
1729 (int)update_rs_processed_buffers); | |
1730 } | |
1731 print_stats(1, "Ext Root Scanning", ext_root_scan_time); | |
1732 print_stats(1, "Mark Stack Scanning", mark_stack_scan_time); | |
1733 print_stats(1, "Scan-Only Scanning", scan_only_time); | |
1734 print_stats(1, "Scan RS", scan_rs_time); | |
1735 print_stats(1, "Object Copying", obj_copy_time); | |
1736 } | |
1737 } | |
1738 print_stats(1, "Other", other_time_ms); | |
1739 for (int i = 0; i < _aux_num; ++i) { | |
1740 if (_cur_aux_times_set[i]) { | |
1741 char buffer[96]; | |
1742 sprintf(buffer, "Aux%d", i); | |
1743 print_stats(1, buffer, _cur_aux_times_ms[i]); | |
1744 } | |
1745 } | |
1746 } | |
1747 if (PrintGCDetails) | |
1748 gclog_or_tty->print(" ["); | |
1749 if (PrintGC || PrintGCDetails) | |
1750 _g1->print_size_transition(gclog_or_tty, | |
1751 _cur_collection_pause_used_at_start_bytes, | |
1752 _g1->used(), _g1->capacity()); | |
1753 if (PrintGCDetails) | |
1754 gclog_or_tty->print_cr("]"); | |
1755 | |
1756 _all_pause_times_ms->add(elapsed_ms); | |
1757 summary->record_total_time_ms(elapsed_ms); | |
1758 summary->record_other_time_ms(other_time_ms); | |
1759 for (int i = 0; i < _aux_num; ++i) | |
1760 if (_cur_aux_times_set[i]) | |
1761 _all_aux_times_ms[i].add(_cur_aux_times_ms[i]); | |
1762 | |
1763 // Reset marks-between-pauses counter. | |
1764 _n_marks_since_last_pause = 0; | |
1765 | |
1766 // Update the efficiency-since-mark vars. | |
1767 double proc_ms = elapsed_ms * (double) _parallel_gc_threads; | |
1768 if (elapsed_ms < MIN_TIMER_GRANULARITY) { | |
1769 // This usually happens due to the timer not having the required | |
1770 // granularity. Some Linuxes are the usual culprits. | |
1771 // We'll just set it to something (arbitrarily) small. | |
1772 proc_ms = 1.0; | |
1773 } | |
1774 double cur_efficiency = (double) freed_bytes / proc_ms; | |
1775 | |
1776 bool new_in_marking_window = _in_marking_window; | |
1777 bool new_in_marking_window_im = false; | |
1778 if (_should_initiate_conc_mark) { | |
1779 new_in_marking_window = true; | |
1780 new_in_marking_window_im = true; | |
1781 } | |
1782 | |
1783 if (in_young_gc_mode()) { | |
1784 if (_last_full_young_gc) { | |
1785 set_full_young_gcs(false); | |
1786 _last_full_young_gc = false; | |
1787 } | |
1788 | |
1789 if ( !_last_young_gc_full ) { | |
1790 if ( _should_revert_to_full_young_gcs || | |
1791 _known_garbage_ratio < 0.05 || | |
1792 (adaptive_young_list_length() && | |
1793 (get_gc_eff_factor() * cur_efficiency < predict_young_gc_eff())) ) { | |
1794 set_full_young_gcs(true); | |
1795 } | |
1796 } | |
1797 _should_revert_to_full_young_gcs = false; | |
1798 | |
1799 if (_last_young_gc_full && !_during_marking) | |
1800 _young_gc_eff_seq->add(cur_efficiency); | |
1801 } | |
1802 | |
1803 _short_lived_surv_rate_group->start_adding_regions(); | |
1804 // do that for any other surv rate groupsx | |
1805 | |
1806 // <NEW PREDICTION> | |
1807 | |
1808 if (!popular && !abandoned) { | |
1809 double pause_time_ms = elapsed_ms; | |
1810 | |
1811 size_t diff = 0; | |
1812 if (_max_pending_cards >= _pending_cards) | |
1813 diff = _max_pending_cards - _pending_cards; | |
1814 _pending_card_diff_seq->add((double) diff); | |
1815 | |
1816 double cost_per_card_ms = 0.0; | |
1817 if (_pending_cards > 0) { | |
1818 cost_per_card_ms = update_rs_time / (double) _pending_cards; | |
1819 _cost_per_card_ms_seq->add(cost_per_card_ms); | |
1820 } | |
1821 | |
1822 double cost_per_scan_only_region_ms = 0.0; | |
1823 if (scan_only_regions_scanned > 0.0) { | |
1824 cost_per_scan_only_region_ms = | |
1825 scan_only_time / scan_only_regions_scanned; | |
1826 if (_in_marking_window_im) | |
1827 _cost_per_scan_only_region_ms_during_cm_seq->add(cost_per_scan_only_region_ms); | |
1828 else | |
1829 _cost_per_scan_only_region_ms_seq->add(cost_per_scan_only_region_ms); | |
1830 } | |
1831 | |
1832 size_t cards_scanned = _g1->cards_scanned(); | |
1833 | |
1834 double cost_per_entry_ms = 0.0; | |
1835 if (cards_scanned > 10) { | |
1836 cost_per_entry_ms = scan_rs_time / (double) cards_scanned; | |
1837 if (_last_young_gc_full) | |
1838 _cost_per_entry_ms_seq->add(cost_per_entry_ms); | |
1839 else | |
1840 _partially_young_cost_per_entry_ms_seq->add(cost_per_entry_ms); | |
1841 } | |
1842 | |
1843 if (_max_rs_lengths > 0) { | |
1844 double cards_per_entry_ratio = | |
1845 (double) cards_scanned / (double) _max_rs_lengths; | |
1846 if (_last_young_gc_full) | |
1847 _fully_young_cards_per_entry_ratio_seq->add(cards_per_entry_ratio); | |
1848 else | |
1849 _partially_young_cards_per_entry_ratio_seq->add(cards_per_entry_ratio); | |
1850 } | |
1851 | |
1852 size_t rs_length_diff = _max_rs_lengths - _recorded_rs_lengths; | |
1853 if (rs_length_diff >= 0) | |
1854 _rs_length_diff_seq->add((double) rs_length_diff); | |
1855 | |
1856 size_t copied_bytes = surviving_bytes; | |
1857 double cost_per_byte_ms = 0.0; | |
1858 if (copied_bytes > 0) { | |
1859 cost_per_byte_ms = obj_copy_time / (double) copied_bytes; | |
1860 if (_in_marking_window) | |
1861 _cost_per_byte_ms_during_cm_seq->add(cost_per_byte_ms); | |
1862 else | |
1863 _cost_per_byte_ms_seq->add(cost_per_byte_ms); | |
1864 } | |
1865 | |
1866 double all_other_time_ms = pause_time_ms - | |
1867 (update_rs_time + scan_only_time + scan_rs_time + obj_copy_time + | |
1868 _mark_closure_time_ms + termination_time); | |
1869 | |
1870 double young_other_time_ms = 0.0; | |
1871 if (_recorded_young_regions > 0) { | |
1872 young_other_time_ms = | |
1873 _recorded_young_cset_choice_time_ms + | |
1874 _recorded_young_free_cset_time_ms; | |
1875 _young_other_cost_per_region_ms_seq->add(young_other_time_ms / | |
1876 (double) _recorded_young_regions); | |
1877 } | |
1878 double non_young_other_time_ms = 0.0; | |
1879 if (_recorded_non_young_regions > 0) { | |
1880 non_young_other_time_ms = | |
1881 _recorded_non_young_cset_choice_time_ms + | |
1882 _recorded_non_young_free_cset_time_ms; | |
1883 | |
1884 _non_young_other_cost_per_region_ms_seq->add(non_young_other_time_ms / | |
1885 (double) _recorded_non_young_regions); | |
1886 } | |
1887 | |
1888 double constant_other_time_ms = all_other_time_ms - | |
1889 (young_other_time_ms + non_young_other_time_ms); | |
1890 _constant_other_time_ms_seq->add(constant_other_time_ms); | |
1891 | |
1892 double survival_ratio = 0.0; | |
1893 if (_bytes_in_collection_set_before_gc > 0) { | |
1894 survival_ratio = (double) bytes_in_to_space_during_gc() / | |
1895 (double) _bytes_in_collection_set_before_gc; | |
1896 } | |
1897 | |
1898 _pending_cards_seq->add((double) _pending_cards); | |
1899 _scanned_cards_seq->add((double) cards_scanned); | |
1900 _rs_lengths_seq->add((double) _max_rs_lengths); | |
1901 | |
1902 double expensive_region_limit_ms = | |
1903 (double) G1MaxPauseTimeMS - predict_constant_other_time_ms(); | |
1904 if (expensive_region_limit_ms < 0.0) { | |
1905 // this means that the other time was predicted to be longer than | |
1906 // than the max pause time | |
1907 expensive_region_limit_ms = (double) G1MaxPauseTimeMS; | |
1908 } | |
1909 _expensive_region_limit_ms = expensive_region_limit_ms; | |
1910 | |
1911 if (PREDICTIONS_VERBOSE) { | |
1912 gclog_or_tty->print_cr(""); | |
1913 gclog_or_tty->print_cr("PREDICTIONS %1.4lf %d " | |
1914 "REGIONS %d %d %d %d " | |
1915 "PENDING_CARDS %d %d " | |
1916 "CARDS_SCANNED %d %d " | |
1917 "RS_LENGTHS %d %d " | |
1918 "SCAN_ONLY_SCAN %1.6lf %1.6lf " | |
1919 "RS_UPDATE %1.6lf %1.6lf RS_SCAN %1.6lf %1.6lf " | |
1920 "SURVIVAL_RATIO %1.6lf %1.6lf " | |
1921 "OBJECT_COPY %1.6lf %1.6lf OTHER_CONSTANT %1.6lf %1.6lf " | |
1922 "OTHER_YOUNG %1.6lf %1.6lf " | |
1923 "OTHER_NON_YOUNG %1.6lf %1.6lf " | |
1924 "VTIME_DIFF %1.6lf TERMINATION %1.6lf " | |
1925 "ELAPSED %1.6lf %1.6lf ", | |
1926 _cur_collection_start_sec, | |
1927 (!_last_young_gc_full) ? 2 : | |
1928 (last_pause_included_initial_mark) ? 1 : 0, | |
1929 _recorded_region_num, | |
1930 _recorded_young_regions, | |
1931 _recorded_scan_only_regions, | |
1932 _recorded_non_young_regions, | |
1933 _predicted_pending_cards, _pending_cards, | |
1934 _predicted_cards_scanned, cards_scanned, | |
1935 _predicted_rs_lengths, _max_rs_lengths, | |
1936 _predicted_scan_only_scan_time_ms, scan_only_time, | |
1937 _predicted_rs_update_time_ms, update_rs_time, | |
1938 _predicted_rs_scan_time_ms, scan_rs_time, | |
1939 _predicted_survival_ratio, survival_ratio, | |
1940 _predicted_object_copy_time_ms, obj_copy_time, | |
1941 _predicted_constant_other_time_ms, constant_other_time_ms, | |
1942 _predicted_young_other_time_ms, young_other_time_ms, | |
1943 _predicted_non_young_other_time_ms, | |
1944 non_young_other_time_ms, | |
1945 _vtime_diff_ms, termination_time, | |
1946 _predicted_pause_time_ms, elapsed_ms); | |
1947 } | |
1948 | |
1949 if (G1PolicyVerbose > 0) { | |
1950 gclog_or_tty->print_cr("Pause Time, predicted: %1.4lfms (predicted %s), actual: %1.4lfms", | |
1951 _predicted_pause_time_ms, | |
1952 (_within_target) ? "within" : "outside", | |
1953 elapsed_ms); | |
1954 } | |
1955 | |
1956 } | |
1957 | |
1958 _in_marking_window = new_in_marking_window; | |
1959 _in_marking_window_im = new_in_marking_window_im; | |
1960 _free_regions_at_end_of_collection = _g1->free_regions(); | |
1961 _scan_only_regions_at_end_of_collection = _g1->young_list_length(); | |
1962 calculate_young_list_min_length(); | |
1963 calculate_young_list_target_config(); | |
1964 | |
1965 // </NEW PREDICTION> | |
1966 | |
1967 _target_pause_time_ms = -1.0; | |
1968 | |
1969 // TODO: calculate tenuring threshold | |
1970 _tenuring_threshold = MaxTenuringThreshold; | |
1971 } | |
1972 | |
1973 // <NEW PREDICTION> | |
1974 | |
1975 double | |
1976 G1CollectorPolicy:: | |
1977 predict_young_collection_elapsed_time_ms(size_t adjustment) { | |
1978 guarantee( adjustment == 0 || adjustment == 1, "invariant" ); | |
1979 | |
1980 G1CollectedHeap* g1h = G1CollectedHeap::heap(); | |
1981 size_t young_num = g1h->young_list_length(); | |
1982 if (young_num == 0) | |
1983 return 0.0; | |
1984 | |
1985 young_num += adjustment; | |
1986 size_t pending_cards = predict_pending_cards(); | |
1987 size_t rs_lengths = g1h->young_list_sampled_rs_lengths() + | |
1988 predict_rs_length_diff(); | |
1989 size_t card_num; | |
1990 if (full_young_gcs()) | |
1991 card_num = predict_young_card_num(rs_lengths); | |
1992 else | |
1993 card_num = predict_non_young_card_num(rs_lengths); | |
1994 size_t young_byte_size = young_num * HeapRegion::GrainBytes; | |
1995 double accum_yg_surv_rate = | |
1996 _short_lived_surv_rate_group->accum_surv_rate(adjustment); | |
1997 | |
1998 size_t bytes_to_copy = | |
1999 (size_t) (accum_yg_surv_rate * (double) HeapRegion::GrainBytes); | |
2000 | |
2001 return | |
2002 predict_rs_update_time_ms(pending_cards) + | |
2003 predict_rs_scan_time_ms(card_num) + | |
2004 predict_object_copy_time_ms(bytes_to_copy) + | |
2005 predict_young_other_time_ms(young_num) + | |
2006 predict_constant_other_time_ms(); | |
2007 } | |
2008 | |
2009 double | |
2010 G1CollectorPolicy::predict_base_elapsed_time_ms(size_t pending_cards) { | |
2011 size_t rs_length = predict_rs_length_diff(); | |
2012 size_t card_num; | |
2013 if (full_young_gcs()) | |
2014 card_num = predict_young_card_num(rs_length); | |
2015 else | |
2016 card_num = predict_non_young_card_num(rs_length); | |
2017 return predict_base_elapsed_time_ms(pending_cards, card_num); | |
2018 } | |
2019 | |
2020 double | |
2021 G1CollectorPolicy::predict_base_elapsed_time_ms(size_t pending_cards, | |
2022 size_t scanned_cards) { | |
2023 return | |
2024 predict_rs_update_time_ms(pending_cards) + | |
2025 predict_rs_scan_time_ms(scanned_cards) + | |
2026 predict_constant_other_time_ms(); | |
2027 } | |
2028 | |
2029 double | |
2030 G1CollectorPolicy::predict_region_elapsed_time_ms(HeapRegion* hr, | |
2031 bool young) { | |
2032 size_t rs_length = hr->rem_set()->occupied(); | |
2033 size_t card_num; | |
2034 if (full_young_gcs()) | |
2035 card_num = predict_young_card_num(rs_length); | |
2036 else | |
2037 card_num = predict_non_young_card_num(rs_length); | |
2038 size_t bytes_to_copy = predict_bytes_to_copy(hr); | |
2039 | |
2040 double region_elapsed_time_ms = | |
2041 predict_rs_scan_time_ms(card_num) + | |
2042 predict_object_copy_time_ms(bytes_to_copy); | |
2043 | |
2044 if (young) | |
2045 region_elapsed_time_ms += predict_young_other_time_ms(1); | |
2046 else | |
2047 region_elapsed_time_ms += predict_non_young_other_time_ms(1); | |
2048 | |
2049 return region_elapsed_time_ms; | |
2050 } | |
2051 | |
2052 size_t | |
2053 G1CollectorPolicy::predict_bytes_to_copy(HeapRegion* hr) { | |
2054 size_t bytes_to_copy; | |
2055 if (hr->is_marked()) | |
2056 bytes_to_copy = hr->max_live_bytes(); | |
2057 else { | |
2058 guarantee( hr->is_young() && hr->age_in_surv_rate_group() != -1, | |
2059 "invariant" ); | |
2060 int age = hr->age_in_surv_rate_group(); | |
2061 double yg_surv_rate = predict_yg_surv_rate(age); | |
2062 bytes_to_copy = (size_t) ((double) hr->used() * yg_surv_rate); | |
2063 } | |
2064 | |
2065 return bytes_to_copy; | |
2066 } | |
2067 | |
2068 void | |
2069 G1CollectorPolicy::start_recording_regions() { | |
2070 _recorded_rs_lengths = 0; | |
2071 _recorded_scan_only_regions = 0; | |
2072 _recorded_young_regions = 0; | |
2073 _recorded_non_young_regions = 0; | |
2074 | |
2075 #if PREDICTIONS_VERBOSE | |
2076 _predicted_rs_lengths = 0; | |
2077 _predicted_cards_scanned = 0; | |
2078 | |
2079 _recorded_marked_bytes = 0; | |
2080 _recorded_young_bytes = 0; | |
2081 _predicted_bytes_to_copy = 0; | |
2082 #endif // PREDICTIONS_VERBOSE | |
2083 } | |
2084 | |
2085 void | |
2086 G1CollectorPolicy::record_cset_region(HeapRegion* hr, bool young) { | |
2087 if (young) { | |
2088 ++_recorded_young_regions; | |
2089 } else { | |
2090 ++_recorded_non_young_regions; | |
2091 } | |
2092 #if PREDICTIONS_VERBOSE | |
2093 if (young) { | |
2094 _recorded_young_bytes += hr->asSpace()->used(); | |
2095 } else { | |
2096 _recorded_marked_bytes += hr->max_live_bytes(); | |
2097 } | |
2098 _predicted_bytes_to_copy += predict_bytes_to_copy(hr); | |
2099 #endif // PREDICTIONS_VERBOSE | |
2100 | |
2101 size_t rs_length = hr->rem_set()->occupied(); | |
2102 _recorded_rs_lengths += rs_length; | |
2103 } | |
2104 | |
2105 void | |
2106 G1CollectorPolicy::record_scan_only_regions(size_t scan_only_length) { | |
2107 _recorded_scan_only_regions = scan_only_length; | |
2108 } | |
2109 | |
2110 void | |
2111 G1CollectorPolicy::end_recording_regions() { | |
2112 #if PREDICTIONS_VERBOSE | |
2113 _predicted_pending_cards = predict_pending_cards(); | |
2114 _predicted_rs_lengths = _recorded_rs_lengths + predict_rs_length_diff(); | |
2115 if (full_young_gcs()) | |
2116 _predicted_cards_scanned += predict_young_card_num(_predicted_rs_lengths); | |
2117 else | |
2118 _predicted_cards_scanned += | |
2119 predict_non_young_card_num(_predicted_rs_lengths); | |
2120 _recorded_region_num = _recorded_young_regions + _recorded_non_young_regions; | |
2121 | |
2122 _predicted_young_survival_ratio = 0.0; | |
2123 for (int i = 0; i < _recorded_young_regions; ++i) | |
2124 _predicted_young_survival_ratio += predict_yg_surv_rate(i); | |
2125 _predicted_young_survival_ratio /= (double) _recorded_young_regions; | |
2126 | |
2127 _predicted_scan_only_scan_time_ms = | |
2128 predict_scan_only_time_ms(_recorded_scan_only_regions); | |
2129 _predicted_rs_update_time_ms = | |
2130 predict_rs_update_time_ms(_g1->pending_card_num()); | |
2131 _predicted_rs_scan_time_ms = | |
2132 predict_rs_scan_time_ms(_predicted_cards_scanned); | |
2133 _predicted_object_copy_time_ms = | |
2134 predict_object_copy_time_ms(_predicted_bytes_to_copy); | |
2135 _predicted_constant_other_time_ms = | |
2136 predict_constant_other_time_ms(); | |
2137 _predicted_young_other_time_ms = | |
2138 predict_young_other_time_ms(_recorded_young_regions); | |
2139 _predicted_non_young_other_time_ms = | |
2140 predict_non_young_other_time_ms(_recorded_non_young_regions); | |
2141 | |
2142 _predicted_pause_time_ms = | |
2143 _predicted_scan_only_scan_time_ms + | |
2144 _predicted_rs_update_time_ms + | |
2145 _predicted_rs_scan_time_ms + | |
2146 _predicted_object_copy_time_ms + | |
2147 _predicted_constant_other_time_ms + | |
2148 _predicted_young_other_time_ms + | |
2149 _predicted_non_young_other_time_ms; | |
2150 #endif // PREDICTIONS_VERBOSE | |
2151 } | |
2152 | |
2153 void G1CollectorPolicy::check_if_region_is_too_expensive(double | |
2154 predicted_time_ms) { | |
2155 // I don't think we need to do this when in young GC mode since | |
2156 // marking will be initiated next time we hit the soft limit anyway... | |
2157 if (predicted_time_ms > _expensive_region_limit_ms) { | |
2158 if (!in_young_gc_mode()) { | |
2159 set_full_young_gcs(true); | |
2160 _should_initiate_conc_mark = true; | |
2161 } else | |
2162 // no point in doing another partial one | |
2163 _should_revert_to_full_young_gcs = true; | |
2164 } | |
2165 } | |
2166 | |
2167 // </NEW PREDICTION> | |
2168 | |
2169 | |
2170 void G1CollectorPolicy::update_recent_gc_times(double end_time_sec, | |
2171 double elapsed_ms) { | |
2172 _recent_gc_times_ms->add(elapsed_ms); | |
2173 _recent_prev_end_times_for_all_gcs_sec->add(end_time_sec); | |
2174 _prev_collection_pause_end_ms = end_time_sec * 1000.0; | |
2175 } | |
2176 | |
2177 double G1CollectorPolicy::recent_avg_time_for_pauses_ms() { | |
2178 if (_recent_pause_times_ms->num() == 0) return (double) G1MaxPauseTimeMS; | |
2179 else return _recent_pause_times_ms->avg(); | |
2180 } | |
2181 | |
2182 double G1CollectorPolicy::recent_avg_time_for_CH_strong_ms() { | |
2183 if (_recent_CH_strong_roots_times_ms->num() == 0) | |
2184 return (double)G1MaxPauseTimeMS/3.0; | |
2185 else return _recent_CH_strong_roots_times_ms->avg(); | |
2186 } | |
2187 | |
2188 double G1CollectorPolicy::recent_avg_time_for_G1_strong_ms() { | |
2189 if (_recent_G1_strong_roots_times_ms->num() == 0) | |
2190 return (double)G1MaxPauseTimeMS/3.0; | |
2191 else return _recent_G1_strong_roots_times_ms->avg(); | |
2192 } | |
2193 | |
2194 double G1CollectorPolicy::recent_avg_time_for_evac_ms() { | |
2195 if (_recent_evac_times_ms->num() == 0) return (double)G1MaxPauseTimeMS/3.0; | |
2196 else return _recent_evac_times_ms->avg(); | |
2197 } | |
2198 | |
2199 int G1CollectorPolicy::number_of_recent_gcs() { | |
2200 assert(_recent_CH_strong_roots_times_ms->num() == | |
2201 _recent_G1_strong_roots_times_ms->num(), "Sequence out of sync"); | |
2202 assert(_recent_G1_strong_roots_times_ms->num() == | |
2203 _recent_evac_times_ms->num(), "Sequence out of sync"); | |
2204 assert(_recent_evac_times_ms->num() == | |
2205 _recent_pause_times_ms->num(), "Sequence out of sync"); | |
2206 assert(_recent_pause_times_ms->num() == | |
2207 _recent_CS_bytes_used_before->num(), "Sequence out of sync"); | |
2208 assert(_recent_CS_bytes_used_before->num() == | |
2209 _recent_CS_bytes_surviving->num(), "Sequence out of sync"); | |
2210 return _recent_pause_times_ms->num(); | |
2211 } | |
2212 | |
2213 double G1CollectorPolicy::recent_avg_survival_fraction() { | |
2214 return recent_avg_survival_fraction_work(_recent_CS_bytes_surviving, | |
2215 _recent_CS_bytes_used_before); | |
2216 } | |
2217 | |
2218 double G1CollectorPolicy::last_survival_fraction() { | |
2219 return last_survival_fraction_work(_recent_CS_bytes_surviving, | |
2220 _recent_CS_bytes_used_before); | |
2221 } | |
2222 | |
2223 double | |
2224 G1CollectorPolicy::recent_avg_survival_fraction_work(TruncatedSeq* surviving, | |
2225 TruncatedSeq* before) { | |
2226 assert(surviving->num() == before->num(), "Sequence out of sync"); | |
2227 if (before->sum() > 0.0) { | |
2228 double recent_survival_rate = surviving->sum() / before->sum(); | |
2229 // We exempt parallel collection from this check because Alloc Buffer | |
2230 // fragmentation can produce negative collections. | |
2231 // Further, we're now always doing parallel collection. But I'm still | |
2232 // leaving this here as a placeholder for a more precise assertion later. | |
2233 // (DLD, 10/05.) | |
2234 assert((true || ParallelGCThreads > 0) || | |
2235 _g1->evacuation_failed() || | |
2236 recent_survival_rate <= 1.0, "Or bad frac"); | |
2237 return recent_survival_rate; | |
2238 } else { | |
2239 return 1.0; // Be conservative. | |
2240 } | |
2241 } | |
2242 | |
2243 double | |
2244 G1CollectorPolicy::last_survival_fraction_work(TruncatedSeq* surviving, | |
2245 TruncatedSeq* before) { | |
2246 assert(surviving->num() == before->num(), "Sequence out of sync"); | |
2247 if (surviving->num() > 0 && before->last() > 0.0) { | |
2248 double last_survival_rate = surviving->last() / before->last(); | |
2249 // We exempt parallel collection from this check because Alloc Buffer | |
2250 // fragmentation can produce negative collections. | |
2251 // Further, we're now always doing parallel collection. But I'm still | |
2252 // leaving this here as a placeholder for a more precise assertion later. | |
2253 // (DLD, 10/05.) | |
2254 assert((true || ParallelGCThreads > 0) || | |
2255 last_survival_rate <= 1.0, "Or bad frac"); | |
2256 return last_survival_rate; | |
2257 } else { | |
2258 return 1.0; | |
2259 } | |
2260 } | |
2261 | |
2262 static const int survival_min_obs = 5; | |
2263 static double survival_min_obs_limits[] = { 0.9, 0.7, 0.5, 0.3, 0.1 }; | |
2264 static const double min_survival_rate = 0.1; | |
2265 | |
2266 double | |
2267 G1CollectorPolicy::conservative_avg_survival_fraction_work(double avg, | |
2268 double latest) { | |
2269 double res = avg; | |
2270 if (number_of_recent_gcs() < survival_min_obs) { | |
2271 res = MAX2(res, survival_min_obs_limits[number_of_recent_gcs()]); | |
2272 } | |
2273 res = MAX2(res, latest); | |
2274 res = MAX2(res, min_survival_rate); | |
2275 // In the parallel case, LAB fragmentation can produce "negative | |
2276 // collections"; so can evac failure. Cap at 1.0 | |
2277 res = MIN2(res, 1.0); | |
2278 return res; | |
2279 } | |
2280 | |
2281 size_t G1CollectorPolicy::expansion_amount() { | |
2282 if ((int)(recent_avg_pause_time_ratio() * 100.0) > G1GCPct) { | |
2283 // We will double the existing space, or take G1ExpandByPctOfAvail % of | |
2284 // the available expansion space, whichever is smaller, bounded below | |
2285 // by a minimum expansion (unless that's all that's left.) | |
2286 const size_t min_expand_bytes = 1*M; | |
2287 size_t reserved_bytes = _g1->g1_reserved_obj_bytes(); | |
2288 size_t committed_bytes = _g1->capacity(); | |
2289 size_t uncommitted_bytes = reserved_bytes - committed_bytes; | |
2290 size_t expand_bytes; | |
2291 size_t expand_bytes_via_pct = | |
2292 uncommitted_bytes * G1ExpandByPctOfAvail / 100; | |
2293 expand_bytes = MIN2(expand_bytes_via_pct, committed_bytes); | |
2294 expand_bytes = MAX2(expand_bytes, min_expand_bytes); | |
2295 expand_bytes = MIN2(expand_bytes, uncommitted_bytes); | |
2296 if (G1PolicyVerbose > 1) { | |
2297 gclog_or_tty->print("Decided to expand: ratio = %5.2f, " | |
2298 "committed = %d%s, uncommited = %d%s, via pct = %d%s.\n" | |
2299 " Answer = %d.\n", | |
2300 recent_avg_pause_time_ratio(), | |
2301 byte_size_in_proper_unit(committed_bytes), | |
2302 proper_unit_for_byte_size(committed_bytes), | |
2303 byte_size_in_proper_unit(uncommitted_bytes), | |
2304 proper_unit_for_byte_size(uncommitted_bytes), | |
2305 byte_size_in_proper_unit(expand_bytes_via_pct), | |
2306 proper_unit_for_byte_size(expand_bytes_via_pct), | |
2307 byte_size_in_proper_unit(expand_bytes), | |
2308 proper_unit_for_byte_size(expand_bytes)); | |
2309 } | |
2310 return expand_bytes; | |
2311 } else { | |
2312 return 0; | |
2313 } | |
2314 } | |
2315 | |
2316 void G1CollectorPolicy::note_start_of_mark_thread() { | |
2317 _mark_thread_startup_sec = os::elapsedTime(); | |
2318 } | |
2319 | |
2320 class CountCSClosure: public HeapRegionClosure { | |
2321 G1CollectorPolicy* _g1_policy; | |
2322 public: | |
2323 CountCSClosure(G1CollectorPolicy* g1_policy) : | |
2324 _g1_policy(g1_policy) {} | |
2325 bool doHeapRegion(HeapRegion* r) { | |
2326 _g1_policy->_bytes_in_collection_set_before_gc += r->used(); | |
2327 return false; | |
2328 } | |
2329 }; | |
2330 | |
2331 void G1CollectorPolicy::count_CS_bytes_used() { | |
2332 CountCSClosure cs_closure(this); | |
2333 _g1->collection_set_iterate(&cs_closure); | |
2334 } | |
2335 | |
2336 static void print_indent(int level) { | |
2337 for (int j = 0; j < level+1; ++j) | |
2338 gclog_or_tty->print(" "); | |
2339 } | |
2340 | |
2341 void G1CollectorPolicy::print_summary (int level, | |
2342 const char* str, | |
2343 NumberSeq* seq) const { | |
2344 double sum = seq->sum(); | |
2345 print_indent(level); | |
2346 gclog_or_tty->print_cr("%-24s = %8.2lf s (avg = %8.2lf ms)", | |
2347 str, sum / 1000.0, seq->avg()); | |
2348 } | |
2349 | |
2350 void G1CollectorPolicy::print_summary_sd (int level, | |
2351 const char* str, | |
2352 NumberSeq* seq) const { | |
2353 print_summary(level, str, seq); | |
2354 print_indent(level + 5); | |
2355 gclog_or_tty->print_cr("(num = %5d, std dev = %8.2lf ms, max = %8.2lf ms)", | |
2356 seq->num(), seq->sd(), seq->maximum()); | |
2357 } | |
2358 | |
2359 void G1CollectorPolicy::check_other_times(int level, | |
2360 NumberSeq* other_times_ms, | |
2361 NumberSeq* calc_other_times_ms) const { | |
2362 bool should_print = false; | |
2363 | |
2364 double max_sum = MAX2(fabs(other_times_ms->sum()), | |
2365 fabs(calc_other_times_ms->sum())); | |
2366 double min_sum = MIN2(fabs(other_times_ms->sum()), | |
2367 fabs(calc_other_times_ms->sum())); | |
2368 double sum_ratio = max_sum / min_sum; | |
2369 if (sum_ratio > 1.1) { | |
2370 should_print = true; | |
2371 print_indent(level + 1); | |
2372 gclog_or_tty->print_cr("## CALCULATED OTHER SUM DOESN'T MATCH RECORDED ###"); | |
2373 } | |
2374 | |
2375 double max_avg = MAX2(fabs(other_times_ms->avg()), | |
2376 fabs(calc_other_times_ms->avg())); | |
2377 double min_avg = MIN2(fabs(other_times_ms->avg()), | |
2378 fabs(calc_other_times_ms->avg())); | |
2379 double avg_ratio = max_avg / min_avg; | |
2380 if (avg_ratio > 1.1) { | |
2381 should_print = true; | |
2382 print_indent(level + 1); | |
2383 gclog_or_tty->print_cr("## CALCULATED OTHER AVG DOESN'T MATCH RECORDED ###"); | |
2384 } | |
2385 | |
2386 if (other_times_ms->sum() < -0.01) { | |
2387 print_indent(level + 1); | |
2388 gclog_or_tty->print_cr("## RECORDED OTHER SUM IS NEGATIVE ###"); | |
2389 } | |
2390 | |
2391 if (other_times_ms->avg() < -0.01) { | |
2392 print_indent(level + 1); | |
2393 gclog_or_tty->print_cr("## RECORDED OTHER AVG IS NEGATIVE ###"); | |
2394 } | |
2395 | |
2396 if (calc_other_times_ms->sum() < -0.01) { | |
2397 should_print = true; | |
2398 print_indent(level + 1); | |
2399 gclog_or_tty->print_cr("## CALCULATED OTHER SUM IS NEGATIVE ###"); | |
2400 } | |
2401 | |
2402 if (calc_other_times_ms->avg() < -0.01) { | |
2403 should_print = true; | |
2404 print_indent(level + 1); | |
2405 gclog_or_tty->print_cr("## CALCULATED OTHER AVG IS NEGATIVE ###"); | |
2406 } | |
2407 | |
2408 if (should_print) | |
2409 print_summary(level, "Other(Calc)", calc_other_times_ms); | |
2410 } | |
2411 | |
2412 void G1CollectorPolicy::print_summary(PauseSummary* summary) const { | |
2413 bool parallel = ParallelGCThreads > 0; | |
2414 MainBodySummary* body_summary = summary->main_body_summary(); | |
2415 PopPreambleSummary* preamble_summary = summary->pop_preamble_summary(); | |
2416 | |
2417 if (summary->get_total_seq()->num() > 0) { | |
2418 print_summary_sd(0, | |
2419 (preamble_summary == NULL) ? "Non-Popular Pauses" : | |
2420 "Popular Pauses", | |
2421 summary->get_total_seq()); | |
2422 if (preamble_summary != NULL) { | |
2423 print_summary(1, "Popularity Preamble", | |
2424 preamble_summary->get_pop_preamble_seq()); | |
2425 print_summary(2, "Update RS", preamble_summary->get_pop_update_rs_seq()); | |
2426 print_summary(2, "Scan RS", preamble_summary->get_pop_scan_rs_seq()); | |
2427 print_summary(2, "Closure App", | |
2428 preamble_summary->get_pop_closure_app_seq()); | |
2429 print_summary(2, "Evacuation", | |
2430 preamble_summary->get_pop_evacuation_seq()); | |
2431 print_summary(2, "Other", preamble_summary->get_pop_other_seq()); | |
2432 { | |
2433 NumberSeq* other_parts[] = { | |
2434 preamble_summary->get_pop_update_rs_seq(), | |
2435 preamble_summary->get_pop_scan_rs_seq(), | |
2436 preamble_summary->get_pop_closure_app_seq(), | |
2437 preamble_summary->get_pop_evacuation_seq() | |
2438 }; | |
2439 NumberSeq calc_other_times_ms(preamble_summary->get_pop_preamble_seq(), | |
2440 4, other_parts); | |
2441 check_other_times(2, preamble_summary->get_pop_other_seq(), | |
2442 &calc_other_times_ms); | |
2443 } | |
2444 } | |
2445 if (body_summary != NULL) { | |
2446 print_summary(1, "SATB Drain", body_summary->get_satb_drain_seq()); | |
2447 if (parallel) { | |
2448 print_summary(1, "Parallel Time", body_summary->get_parallel_seq()); | |
2449 print_summary(2, "Update RS", body_summary->get_update_rs_seq()); | |
2450 print_summary(2, "Ext Root Scanning", | |
2451 body_summary->get_ext_root_scan_seq()); | |
2452 print_summary(2, "Mark Stack Scanning", | |
2453 body_summary->get_mark_stack_scan_seq()); | |
2454 print_summary(2, "Scan-Only Scanning", | |
2455 body_summary->get_scan_only_seq()); | |
2456 print_summary(2, "Scan RS", body_summary->get_scan_rs_seq()); | |
2457 print_summary(2, "Object Copy", body_summary->get_obj_copy_seq()); | |
2458 print_summary(2, "Termination", body_summary->get_termination_seq()); | |
2459 print_summary(2, "Other", body_summary->get_parallel_other_seq()); | |
2460 { | |
2461 NumberSeq* other_parts[] = { | |
2462 body_summary->get_update_rs_seq(), | |
2463 body_summary->get_ext_root_scan_seq(), | |
2464 body_summary->get_mark_stack_scan_seq(), | |
2465 body_summary->get_scan_only_seq(), | |
2466 body_summary->get_scan_rs_seq(), | |
2467 body_summary->get_obj_copy_seq(), | |
2468 body_summary->get_termination_seq() | |
2469 }; | |
2470 NumberSeq calc_other_times_ms(body_summary->get_parallel_seq(), | |
2471 7, other_parts); | |
2472 check_other_times(2, body_summary->get_parallel_other_seq(), | |
2473 &calc_other_times_ms); | |
2474 } | |
2475 print_summary(1, "Mark Closure", body_summary->get_mark_closure_seq()); | |
2476 print_summary(1, "Clear CT", body_summary->get_clear_ct_seq()); | |
2477 } else { | |
2478 print_summary(1, "Update RS", body_summary->get_update_rs_seq()); | |
2479 print_summary(1, "Ext Root Scanning", | |
2480 body_summary->get_ext_root_scan_seq()); | |
2481 print_summary(1, "Mark Stack Scanning", | |
2482 body_summary->get_mark_stack_scan_seq()); | |
2483 print_summary(1, "Scan-Only Scanning", | |
2484 body_summary->get_scan_only_seq()); | |
2485 print_summary(1, "Scan RS", body_summary->get_scan_rs_seq()); | |
2486 print_summary(1, "Object Copy", body_summary->get_obj_copy_seq()); | |
2487 } | |
2488 } | |
2489 print_summary(1, "Other", summary->get_other_seq()); | |
2490 { | |
2491 NumberSeq calc_other_times_ms; | |
2492 if (body_summary != NULL) { | |
2493 // not abandoned | |
2494 if (parallel) { | |
2495 // parallel | |
2496 NumberSeq* other_parts[] = { | |
2497 body_summary->get_satb_drain_seq(), | |
2498 (preamble_summary == NULL) ? NULL : | |
2499 preamble_summary->get_pop_preamble_seq(), | |
2500 body_summary->get_parallel_seq(), | |
2501 body_summary->get_clear_ct_seq() | |
2502 }; | |
2503 calc_other_times_ms = NumberSeq (summary->get_total_seq(), | |
2504 4, other_parts); | |
2505 } else { | |
2506 // serial | |
2507 NumberSeq* other_parts[] = { | |
2508 body_summary->get_satb_drain_seq(), | |
2509 (preamble_summary == NULL) ? NULL : | |
2510 preamble_summary->get_pop_preamble_seq(), | |
2511 body_summary->get_update_rs_seq(), | |
2512 body_summary->get_ext_root_scan_seq(), | |
2513 body_summary->get_mark_stack_scan_seq(), | |
2514 body_summary->get_scan_only_seq(), | |
2515 body_summary->get_scan_rs_seq(), | |
2516 body_summary->get_obj_copy_seq() | |
2517 }; | |
2518 calc_other_times_ms = NumberSeq(summary->get_total_seq(), | |
2519 8, other_parts); | |
2520 } | |
2521 } else { | |
2522 // abandoned | |
2523 NumberSeq* other_parts[] = { | |
2524 (preamble_summary == NULL) ? NULL : | |
2525 preamble_summary->get_pop_preamble_seq() | |
2526 }; | |
2527 calc_other_times_ms = NumberSeq(summary->get_total_seq(), | |
2528 1, other_parts); | |
2529 } | |
2530 check_other_times(1, summary->get_other_seq(), &calc_other_times_ms); | |
2531 } | |
2532 } else { | |
2533 print_indent(0); | |
2534 gclog_or_tty->print_cr("none"); | |
2535 } | |
2536 gclog_or_tty->print_cr(""); | |
2537 } | |
2538 | |
2539 void | |
2540 G1CollectorPolicy::print_abandoned_summary(PauseSummary* non_pop_summary, | |
2541 PauseSummary* pop_summary) const { | |
2542 bool printed = false; | |
2543 if (non_pop_summary->get_total_seq()->num() > 0) { | |
2544 printed = true; | |
2545 print_summary(non_pop_summary); | |
2546 } | |
2547 if (pop_summary->get_total_seq()->num() > 0) { | |
2548 printed = true; | |
2549 print_summary(pop_summary); | |
2550 } | |
2551 | |
2552 if (!printed) { | |
2553 print_indent(0); | |
2554 gclog_or_tty->print_cr("none"); | |
2555 gclog_or_tty->print_cr(""); | |
2556 } | |
2557 } | |
2558 | |
2559 void G1CollectorPolicy::print_tracing_info() const { | |
2560 if (TraceGen0Time) { | |
2561 gclog_or_tty->print_cr("ALL PAUSES"); | |
2562 print_summary_sd(0, "Total", _all_pause_times_ms); | |
2563 gclog_or_tty->print_cr(""); | |
2564 gclog_or_tty->print_cr(""); | |
2565 gclog_or_tty->print_cr(" Full Young GC Pauses: %8d", _full_young_pause_num); | |
2566 gclog_or_tty->print_cr(" Partial Young GC Pauses: %8d", _partial_young_pause_num); | |
2567 gclog_or_tty->print_cr(""); | |
2568 | |
2569 gclog_or_tty->print_cr("NON-POPULAR PAUSES"); | |
2570 print_summary(_non_pop_summary); | |
2571 | |
2572 gclog_or_tty->print_cr("POPULAR PAUSES"); | |
2573 print_summary(_pop_summary); | |
2574 | |
2575 gclog_or_tty->print_cr("ABANDONED PAUSES"); | |
2576 print_abandoned_summary(_non_pop_abandoned_summary, | |
2577 _pop_abandoned_summary); | |
2578 | |
2579 gclog_or_tty->print_cr("MISC"); | |
2580 print_summary_sd(0, "Stop World", _all_stop_world_times_ms); | |
2581 print_summary_sd(0, "Yields", _all_yield_times_ms); | |
2582 for (int i = 0; i < _aux_num; ++i) { | |
2583 if (_all_aux_times_ms[i].num() > 0) { | |
2584 char buffer[96]; | |
2585 sprintf(buffer, "Aux%d", i); | |
2586 print_summary_sd(0, buffer, &_all_aux_times_ms[i]); | |
2587 } | |
2588 } | |
2589 | |
2590 size_t all_region_num = _region_num_young + _region_num_tenured; | |
2591 gclog_or_tty->print_cr(" New Regions %8d, Young %8d (%6.2lf%%), " | |
2592 "Tenured %8d (%6.2lf%%)", | |
2593 all_region_num, | |
2594 _region_num_young, | |
2595 (double) _region_num_young / (double) all_region_num * 100.0, | |
2596 _region_num_tenured, | |
2597 (double) _region_num_tenured / (double) all_region_num * 100.0); | |
2598 | |
2599 if (!G1RSBarrierUseQueue) { | |
2600 gclog_or_tty->print_cr("Of %d times conc refinement was enabled, %d (%7.2f%%) " | |
2601 "did zero traversals.", | |
2602 _conc_refine_enabled, _conc_refine_zero_traversals, | |
2603 _conc_refine_enabled > 0 ? | |
2604 100.0 * (float)_conc_refine_zero_traversals/ | |
2605 (float)_conc_refine_enabled : 0.0); | |
2606 gclog_or_tty->print_cr(" Max # of traversals = %d.", | |
2607 _conc_refine_max_traversals); | |
2608 gclog_or_tty->print_cr(""); | |
2609 } | |
2610 } | |
2611 if (TraceGen1Time) { | |
2612 if (_all_full_gc_times_ms->num() > 0) { | |
2613 gclog_or_tty->print("\n%4d full_gcs: total time = %8.2f s", | |
2614 _all_full_gc_times_ms->num(), | |
2615 _all_full_gc_times_ms->sum() / 1000.0); | |
2616 gclog_or_tty->print_cr(" (avg = %8.2fms).", _all_full_gc_times_ms->avg()); | |
2617 gclog_or_tty->print_cr(" [std. dev = %8.2f ms, max = %8.2f ms]", | |
2618 _all_full_gc_times_ms->sd(), | |
2619 _all_full_gc_times_ms->maximum()); | |
2620 } | |
2621 } | |
2622 } | |
2623 | |
2624 void G1CollectorPolicy::print_yg_surv_rate_info() const { | |
2625 #ifndef PRODUCT | |
2626 _short_lived_surv_rate_group->print_surv_rate_summary(); | |
2627 // add this call for any other surv rate groups | |
2628 #endif // PRODUCT | |
2629 } | |
2630 | |
2631 void G1CollectorPolicy::update_conc_refine_data() { | |
2632 unsigned traversals = _g1->concurrent_g1_refine()->disable(); | |
2633 if (traversals == 0) _conc_refine_zero_traversals++; | |
2634 _conc_refine_max_traversals = MAX2(_conc_refine_max_traversals, | |
2635 (size_t)traversals); | |
2636 | |
2637 if (G1PolicyVerbose > 1) | |
2638 gclog_or_tty->print_cr("Did a CR traversal series: %d traversals.", traversals); | |
2639 double multiplier = 1.0; | |
2640 if (traversals == 0) { | |
2641 multiplier = 4.0; | |
2642 } else if (traversals > (size_t)G1ConcRefineTargTraversals) { | |
2643 multiplier = 1.0/1.5; | |
2644 } else if (traversals < (size_t)G1ConcRefineTargTraversals) { | |
2645 multiplier = 1.5; | |
2646 } | |
2647 if (G1PolicyVerbose > 1) { | |
2648 gclog_or_tty->print_cr(" Multiplier = %7.2f.", multiplier); | |
2649 gclog_or_tty->print(" Delta went from %d regions to ", | |
2650 _conc_refine_current_delta); | |
2651 } | |
2652 _conc_refine_current_delta = | |
2653 MIN2(_g1->n_regions(), | |
2654 (size_t)(_conc_refine_current_delta * multiplier)); | |
2655 _conc_refine_current_delta = | |
2656 MAX2(_conc_refine_current_delta, (size_t)1); | |
2657 if (G1PolicyVerbose > 1) { | |
2658 gclog_or_tty->print_cr("%d regions.", _conc_refine_current_delta); | |
2659 } | |
2660 _conc_refine_enabled++; | |
2661 } | |
2662 | |
2663 void G1CollectorPolicy::set_single_region_collection_set(HeapRegion* hr) { | |
2664 assert(collection_set() == NULL, "Must be no current CS."); | |
2665 _collection_set_size = 0; | |
2666 _collection_set_bytes_used_before = 0; | |
2667 add_to_collection_set(hr); | |
2668 count_CS_bytes_used(); | |
2669 } | |
2670 | |
2671 bool | |
2672 G1CollectorPolicy::should_add_next_region_to_young_list() { | |
2673 assert(in_young_gc_mode(), "should be in young GC mode"); | |
2674 bool ret; | |
2675 size_t young_list_length = _g1->young_list_length(); | |
2676 | |
2677 if (young_list_length < _young_list_target_length) { | |
2678 ret = true; | |
2679 ++_region_num_young; | |
2680 } else { | |
2681 ret = false; | |
2682 ++_region_num_tenured; | |
2683 } | |
2684 | |
2685 return ret; | |
2686 } | |
2687 | |
2688 #ifndef PRODUCT | |
2689 // for debugging, bit of a hack... | |
2690 static char* | |
2691 region_num_to_mbs(int length) { | |
2692 static char buffer[64]; | |
2693 double bytes = (double) (length * HeapRegion::GrainBytes); | |
2694 double mbs = bytes / (double) (1024 * 1024); | |
2695 sprintf(buffer, "%7.2lfMB", mbs); | |
2696 return buffer; | |
2697 } | |
2698 #endif // PRODUCT | |
2699 | |
2700 void | |
2701 G1CollectorPolicy::checkpoint_conc_overhead() { | |
2702 double conc_overhead = 0.0; | |
2703 if (G1AccountConcurrentOverhead) | |
2704 conc_overhead = COTracker::totalPredConcOverhead(); | |
2705 _mmu_tracker->update_conc_overhead(conc_overhead); | |
2706 #if 0 | |
2707 gclog_or_tty->print(" CO %1.4lf TARGET %1.4lf", | |
2708 conc_overhead, _mmu_tracker->max_gc_time()); | |
2709 #endif | |
2710 } | |
2711 | |
2712 | |
2713 uint G1CollectorPolicy::max_regions(int purpose) { | |
2714 switch (purpose) { | |
2715 case GCAllocForSurvived: | |
2716 return G1MaxSurvivorRegions; | |
2717 case GCAllocForTenured: | |
2718 return UINT_MAX; | |
2719 default: | |
2720 return UINT_MAX; | |
2721 }; | |
2722 } | |
2723 | |
2724 void | |
2725 G1CollectorPolicy_BestRegionsFirst:: | |
2726 set_single_region_collection_set(HeapRegion* hr) { | |
2727 G1CollectorPolicy::set_single_region_collection_set(hr); | |
2728 _collectionSetChooser->removeRegion(hr); | |
2729 } | |
2730 | |
2731 | |
2732 bool | |
2733 G1CollectorPolicy_BestRegionsFirst::should_do_collection_pause(size_t | |
2734 word_size) { | |
2735 assert(_g1->regions_accounted_for(), "Region leakage!"); | |
2736 // Initiate a pause when we reach the steady-state "used" target. | |
2737 size_t used_hard = (_g1->capacity() / 100) * G1SteadyStateUsed; | |
2738 size_t used_soft = | |
2739 MAX2((_g1->capacity() / 100) * (G1SteadyStateUsed - G1SteadyStateUsedDelta), | |
2740 used_hard/2); | |
2741 size_t used = _g1->used(); | |
2742 | |
2743 double max_pause_time_ms = _mmu_tracker->max_gc_time() * 1000.0; | |
2744 | |
2745 size_t young_list_length = _g1->young_list_length(); | |
2746 bool reached_target_length = young_list_length >= _young_list_target_length; | |
2747 | |
2748 if (in_young_gc_mode()) { | |
2749 if (reached_target_length) { | |
2750 assert( young_list_length > 0 && _g1->young_list_length() > 0, | |
2751 "invariant" ); | |
2752 _target_pause_time_ms = max_pause_time_ms; | |
2753 return true; | |
2754 } | |
2755 } else { | |
2756 guarantee( false, "should not reach here" ); | |
2757 } | |
2758 | |
2759 return false; | |
2760 } | |
2761 | |
2762 #ifndef PRODUCT | |
2763 class HRSortIndexIsOKClosure: public HeapRegionClosure { | |
2764 CollectionSetChooser* _chooser; | |
2765 public: | |
2766 HRSortIndexIsOKClosure(CollectionSetChooser* chooser) : | |
2767 _chooser(chooser) {} | |
2768 | |
2769 bool doHeapRegion(HeapRegion* r) { | |
2770 if (!r->continuesHumongous()) { | |
2771 assert(_chooser->regionProperlyOrdered(r), "Ought to be."); | |
2772 } | |
2773 return false; | |
2774 } | |
2775 }; | |
2776 | |
2777 bool G1CollectorPolicy_BestRegionsFirst::assertMarkedBytesDataOK() { | |
2778 HRSortIndexIsOKClosure cl(_collectionSetChooser); | |
2779 _g1->heap_region_iterate(&cl); | |
2780 return true; | |
2781 } | |
2782 #endif | |
2783 | |
2784 void | |
2785 G1CollectorPolicy_BestRegionsFirst:: | |
2786 record_collection_pause_start(double start_time_sec, size_t start_used) { | |
2787 G1CollectorPolicy::record_collection_pause_start(start_time_sec, start_used); | |
2788 } | |
2789 | |
2790 class NextNonCSElemFinder: public HeapRegionClosure { | |
2791 HeapRegion* _res; | |
2792 public: | |
2793 NextNonCSElemFinder(): _res(NULL) {} | |
2794 bool doHeapRegion(HeapRegion* r) { | |
2795 if (!r->in_collection_set()) { | |
2796 _res = r; | |
2797 return true; | |
2798 } else { | |
2799 return false; | |
2800 } | |
2801 } | |
2802 HeapRegion* res() { return _res; } | |
2803 }; | |
2804 | |
2805 class KnownGarbageClosure: public HeapRegionClosure { | |
2806 CollectionSetChooser* _hrSorted; | |
2807 | |
2808 public: | |
2809 KnownGarbageClosure(CollectionSetChooser* hrSorted) : | |
2810 _hrSorted(hrSorted) | |
2811 {} | |
2812 | |
2813 bool doHeapRegion(HeapRegion* r) { | |
2814 // We only include humongous regions in collection | |
2815 // sets when concurrent mark shows that their contained object is | |
2816 // unreachable. | |
2817 | |
2818 // Do we have any marking information for this region? | |
2819 if (r->is_marked()) { | |
2820 // We don't include humongous regions in collection | |
2821 // sets because we collect them immediately at the end of a marking | |
2822 // cycle. We also don't include young regions because we *must* | |
2823 // include them in the next collection pause. | |
2824 if (!r->isHumongous() && !r->is_young()) { | |
2825 _hrSorted->addMarkedHeapRegion(r); | |
2826 } | |
2827 } | |
2828 return false; | |
2829 } | |
2830 }; | |
2831 | |
2832 class ParKnownGarbageHRClosure: public HeapRegionClosure { | |
2833 CollectionSetChooser* _hrSorted; | |
2834 jint _marked_regions_added; | |
2835 jint _chunk_size; | |
2836 jint _cur_chunk_idx; | |
2837 jint _cur_chunk_end; // Cur chunk [_cur_chunk_idx, _cur_chunk_end) | |
2838 int _worker; | |
2839 int _invokes; | |
2840 | |
2841 void get_new_chunk() { | |
2842 _cur_chunk_idx = _hrSorted->getParMarkedHeapRegionChunk(_chunk_size); | |
2843 _cur_chunk_end = _cur_chunk_idx + _chunk_size; | |
2844 } | |
2845 void add_region(HeapRegion* r) { | |
2846 if (_cur_chunk_idx == _cur_chunk_end) { | |
2847 get_new_chunk(); | |
2848 } | |
2849 assert(_cur_chunk_idx < _cur_chunk_end, "postcondition"); | |
2850 _hrSorted->setMarkedHeapRegion(_cur_chunk_idx, r); | |
2851 _marked_regions_added++; | |
2852 _cur_chunk_idx++; | |
2853 } | |
2854 | |
2855 public: | |
2856 ParKnownGarbageHRClosure(CollectionSetChooser* hrSorted, | |
2857 jint chunk_size, | |
2858 int worker) : | |
2859 _hrSorted(hrSorted), _chunk_size(chunk_size), _worker(worker), | |
2860 _marked_regions_added(0), _cur_chunk_idx(0), _cur_chunk_end(0), | |
2861 _invokes(0) | |
2862 {} | |
2863 | |
2864 bool doHeapRegion(HeapRegion* r) { | |
2865 // We only include humongous regions in collection | |
2866 // sets when concurrent mark shows that their contained object is | |
2867 // unreachable. | |
2868 _invokes++; | |
2869 | |
2870 // Do we have any marking information for this region? | |
2871 if (r->is_marked()) { | |
2872 // We don't include humongous regions in collection | |
2873 // sets because we collect them immediately at the end of a marking | |
2874 // cycle. | |
2875 // We also do not include young regions in collection sets | |
2876 if (!r->isHumongous() && !r->is_young()) { | |
2877 add_region(r); | |
2878 } | |
2879 } | |
2880 return false; | |
2881 } | |
2882 jint marked_regions_added() { return _marked_regions_added; } | |
2883 int invokes() { return _invokes; } | |
2884 }; | |
2885 | |
2886 class ParKnownGarbageTask: public AbstractGangTask { | |
2887 CollectionSetChooser* _hrSorted; | |
2888 jint _chunk_size; | |
2889 G1CollectedHeap* _g1; | |
2890 public: | |
2891 ParKnownGarbageTask(CollectionSetChooser* hrSorted, jint chunk_size) : | |
2892 AbstractGangTask("ParKnownGarbageTask"), | |
2893 _hrSorted(hrSorted), _chunk_size(chunk_size), | |
2894 _g1(G1CollectedHeap::heap()) | |
2895 {} | |
2896 | |
2897 void work(int i) { | |
2898 ParKnownGarbageHRClosure parKnownGarbageCl(_hrSorted, _chunk_size, i); | |
2899 // Back to zero for the claim value. | |
2900 _g1->heap_region_par_iterate_chunked(&parKnownGarbageCl, i, 0); | |
2901 jint regions_added = parKnownGarbageCl.marked_regions_added(); | |
2902 _hrSorted->incNumMarkedHeapRegions(regions_added); | |
2903 if (G1PrintParCleanupStats) { | |
2904 gclog_or_tty->print(" Thread %d called %d times, added %d regions to list.\n", | |
2905 i, parKnownGarbageCl.invokes(), regions_added); | |
2906 } | |
2907 } | |
2908 }; | |
2909 | |
2910 void | |
2911 G1CollectorPolicy_BestRegionsFirst:: | |
2912 record_concurrent_mark_cleanup_end(size_t freed_bytes, | |
2913 size_t max_live_bytes) { | |
2914 double start; | |
2915 if (G1PrintParCleanupStats) start = os::elapsedTime(); | |
2916 record_concurrent_mark_cleanup_end_work1(freed_bytes, max_live_bytes); | |
2917 | |
2918 _collectionSetChooser->clearMarkedHeapRegions(); | |
2919 double clear_marked_end; | |
2920 if (G1PrintParCleanupStats) { | |
2921 clear_marked_end = os::elapsedTime(); | |
2922 gclog_or_tty->print_cr(" clear marked regions + work1: %8.3f ms.", | |
2923 (clear_marked_end - start)*1000.0); | |
2924 } | |
2925 if (ParallelGCThreads > 0) { | |
2926 const size_t OverpartitionFactor = 4; | |
2927 const size_t MinChunkSize = 8; | |
2928 const size_t ChunkSize = | |
2929 MAX2(_g1->n_regions() / (ParallelGCThreads * OverpartitionFactor), | |
2930 MinChunkSize); | |
2931 _collectionSetChooser->prepareForAddMarkedHeapRegionsPar(_g1->n_regions(), | |
2932 ChunkSize); | |
2933 ParKnownGarbageTask parKnownGarbageTask(_collectionSetChooser, | |
2934 (int) ChunkSize); | |
2935 _g1->workers()->run_task(&parKnownGarbageTask); | |
2936 } else { | |
2937 KnownGarbageClosure knownGarbagecl(_collectionSetChooser); | |
2938 _g1->heap_region_iterate(&knownGarbagecl); | |
2939 } | |
2940 double known_garbage_end; | |
2941 if (G1PrintParCleanupStats) { | |
2942 known_garbage_end = os::elapsedTime(); | |
2943 gclog_or_tty->print_cr(" compute known garbage: %8.3f ms.", | |
2944 (known_garbage_end - clear_marked_end)*1000.0); | |
2945 } | |
2946 _collectionSetChooser->sortMarkedHeapRegions(); | |
2947 double sort_end; | |
2948 if (G1PrintParCleanupStats) { | |
2949 sort_end = os::elapsedTime(); | |
2950 gclog_or_tty->print_cr(" sorting: %8.3f ms.", | |
2951 (sort_end - known_garbage_end)*1000.0); | |
2952 } | |
2953 | |
2954 record_concurrent_mark_cleanup_end_work2(); | |
2955 double work2_end; | |
2956 if (G1PrintParCleanupStats) { | |
2957 work2_end = os::elapsedTime(); | |
2958 gclog_or_tty->print_cr(" work2: %8.3f ms.", | |
2959 (work2_end - sort_end)*1000.0); | |
2960 } | |
2961 } | |
2962 | |
2963 // Add the heap region to the collection set and return the conservative | |
2964 // estimate of the number of live bytes. | |
2965 void G1CollectorPolicy:: | |
2966 add_to_collection_set(HeapRegion* hr) { | |
2967 if (G1TraceRegions) { | |
2968 gclog_or_tty->print_cr("added region to cset %d:["PTR_FORMAT", "PTR_FORMAT"], " | |
2969 "top "PTR_FORMAT", young %s", | |
2970 hr->hrs_index(), hr->bottom(), hr->end(), | |
2971 hr->top(), (hr->is_young()) ? "YES" : "NO"); | |
2972 } | |
2973 | |
2974 if (_g1->mark_in_progress()) | |
2975 _g1->concurrent_mark()->registerCSetRegion(hr); | |
2976 | |
2977 assert(!hr->in_collection_set(), | |
2978 "should not already be in the CSet"); | |
2979 hr->set_in_collection_set(true); | |
2980 hr->set_next_in_collection_set(_collection_set); | |
2981 _collection_set = hr; | |
2982 _collection_set_size++; | |
2983 _collection_set_bytes_used_before += hr->used(); | |
2984 } | |
2985 | |
2986 void | |
2987 G1CollectorPolicy_BestRegionsFirst:: | |
2988 choose_collection_set(HeapRegion* pop_region) { | |
2989 double non_young_start_time_sec; | |
2990 start_recording_regions(); | |
2991 | |
2992 if (pop_region != NULL) { | |
2993 _target_pause_time_ms = (double) G1MaxPauseTimeMS; | |
2994 } else { | |
2995 guarantee(_target_pause_time_ms > -1.0, | |
2996 "_target_pause_time_ms should have been set!"); | |
2997 } | |
2998 | |
2999 // pop region is either null (and so is CS), or else it *is* the CS. | |
3000 assert(_collection_set == pop_region, "Precondition"); | |
3001 | |
3002 double base_time_ms = predict_base_elapsed_time_ms(_pending_cards); | |
3003 double predicted_pause_time_ms = base_time_ms; | |
3004 | |
3005 double target_time_ms = _target_pause_time_ms; | |
3006 double time_remaining_ms = target_time_ms - base_time_ms; | |
3007 | |
3008 // the 10% and 50% values are arbitrary... | |
3009 if (time_remaining_ms < 0.10*target_time_ms) { | |
3010 time_remaining_ms = 0.50 * target_time_ms; | |
3011 _within_target = false; | |
3012 } else { | |
3013 _within_target = true; | |
3014 } | |
3015 | |
3016 // We figure out the number of bytes available for future to-space. | |
3017 // For new regions without marking information, we must assume the | |
3018 // worst-case of complete survival. If we have marking information for a | |
3019 // region, we can bound the amount of live data. We can add a number of | |
3020 // such regions, as long as the sum of the live data bounds does not | |
3021 // exceed the available evacuation space. | |
3022 size_t max_live_bytes = _g1->free_regions() * HeapRegion::GrainBytes; | |
3023 | |
3024 size_t expansion_bytes = | |
3025 _g1->expansion_regions() * HeapRegion::GrainBytes; | |
3026 | |
3027 if (pop_region == NULL) { | |
3028 _collection_set_bytes_used_before = 0; | |
3029 _collection_set_size = 0; | |
3030 } | |
3031 | |
3032 // Adjust for expansion and slop. | |
3033 max_live_bytes = max_live_bytes + expansion_bytes; | |
3034 | |
3035 assert(pop_region != NULL || _g1->regions_accounted_for(), "Region leakage!"); | |
3036 | |
3037 HeapRegion* hr; | |
3038 if (in_young_gc_mode()) { | |
3039 double young_start_time_sec = os::elapsedTime(); | |
3040 | |
3041 if (G1PolicyVerbose > 0) { | |
3042 gclog_or_tty->print_cr("Adding %d young regions to the CSet", | |
3043 _g1->young_list_length()); | |
3044 } | |
3045 _young_cset_length = 0; | |
3046 _last_young_gc_full = full_young_gcs() ? true : false; | |
3047 if (_last_young_gc_full) | |
3048 ++_full_young_pause_num; | |
3049 else | |
3050 ++_partial_young_pause_num; | |
3051 hr = _g1->pop_region_from_young_list(); | |
3052 while (hr != NULL) { | |
3053 | |
3054 assert( hr->young_index_in_cset() == -1, "invariant" ); | |
3055 assert( hr->age_in_surv_rate_group() != -1, "invariant" ); | |
3056 hr->set_young_index_in_cset((int) _young_cset_length); | |
3057 | |
3058 ++_young_cset_length; | |
3059 double predicted_time_ms = predict_region_elapsed_time_ms(hr, true); | |
3060 time_remaining_ms -= predicted_time_ms; | |
3061 predicted_pause_time_ms += predicted_time_ms; | |
3062 if (hr == pop_region) { | |
3063 // The popular region was young. Skip over it. | |
3064 assert(hr->in_collection_set(), "It's the pop region."); | |
3065 } else { | |
3066 assert(!hr->in_collection_set(), "It's not the pop region."); | |
3067 add_to_collection_set(hr); | |
3068 record_cset_region(hr, true); | |
3069 } | |
3070 max_live_bytes -= MIN2(hr->max_live_bytes(), max_live_bytes); | |
3071 if (G1PolicyVerbose > 0) { | |
3072 gclog_or_tty->print_cr(" Added [" PTR_FORMAT ", " PTR_FORMAT") to CS.", | |
3073 hr->bottom(), hr->end()); | |
3074 gclog_or_tty->print_cr(" (" SIZE_FORMAT " KB left in heap.)", | |
3075 max_live_bytes/K); | |
3076 } | |
3077 hr = _g1->pop_region_from_young_list(); | |
3078 } | |
3079 | |
3080 record_scan_only_regions(_g1->young_list_scan_only_length()); | |
3081 | |
3082 double young_end_time_sec = os::elapsedTime(); | |
3083 _recorded_young_cset_choice_time_ms = | |
3084 (young_end_time_sec - young_start_time_sec) * 1000.0; | |
3085 | |
3086 non_young_start_time_sec = os::elapsedTime(); | |
3087 | |
3088 if (_young_cset_length > 0 && _last_young_gc_full) { | |
3089 // don't bother adding more regions... | |
3090 goto choose_collection_set_end; | |
3091 } | |
3092 } else if (pop_region != NULL) { | |
3093 // We're not in young mode, and we chose a popular region; don't choose | |
3094 // any more. | |
3095 return; | |
3096 } | |
3097 | |
3098 if (!in_young_gc_mode() || !full_young_gcs()) { | |
3099 bool should_continue = true; | |
3100 NumberSeq seq; | |
3101 double avg_prediction = 100000000000000000.0; // something very large | |
3102 do { | |
3103 hr = _collectionSetChooser->getNextMarkedRegion(time_remaining_ms, | |
3104 avg_prediction); | |
3105 if (hr != NULL && !hr->popular()) { | |
3106 double predicted_time_ms = predict_region_elapsed_time_ms(hr, false); | |
3107 time_remaining_ms -= predicted_time_ms; | |
3108 predicted_pause_time_ms += predicted_time_ms; | |
3109 add_to_collection_set(hr); | |
3110 record_cset_region(hr, false); | |
3111 max_live_bytes -= MIN2(hr->max_live_bytes(), max_live_bytes); | |
3112 if (G1PolicyVerbose > 0) { | |
3113 gclog_or_tty->print_cr(" (" SIZE_FORMAT " KB left in heap.)", | |
3114 max_live_bytes/K); | |
3115 } | |
3116 seq.add(predicted_time_ms); | |
3117 avg_prediction = seq.avg() + seq.sd(); | |
3118 } | |
3119 should_continue = | |
3120 ( hr != NULL) && | |
3121 ( (adaptive_young_list_length()) ? time_remaining_ms > 0.0 | |
3122 : _collection_set_size < _young_list_fixed_length ); | |
3123 } while (should_continue); | |
3124 | |
3125 if (!adaptive_young_list_length() && | |
3126 _collection_set_size < _young_list_fixed_length) | |
3127 _should_revert_to_full_young_gcs = true; | |
3128 } | |
3129 | |
3130 choose_collection_set_end: | |
3131 count_CS_bytes_used(); | |
3132 | |
3133 end_recording_regions(); | |
3134 | |
3135 double non_young_end_time_sec = os::elapsedTime(); | |
3136 _recorded_non_young_cset_choice_time_ms = | |
3137 (non_young_end_time_sec - non_young_start_time_sec) * 1000.0; | |
3138 } | |
3139 | |
3140 void G1CollectorPolicy_BestRegionsFirst::record_full_collection_end() { | |
3141 G1CollectorPolicy::record_full_collection_end(); | |
3142 _collectionSetChooser->updateAfterFullCollection(); | |
3143 } | |
3144 | |
3145 void G1CollectorPolicy_BestRegionsFirst:: | |
3146 expand_if_possible(size_t numRegions) { | |
3147 size_t expansion_bytes = numRegions * HeapRegion::GrainBytes; | |
3148 _g1->expand(expansion_bytes); | |
3149 } | |
3150 | |
3151 void G1CollectorPolicy_BestRegionsFirst:: | |
3152 record_collection_pause_end(bool popular, bool abandoned) { | |
3153 G1CollectorPolicy::record_collection_pause_end(popular, abandoned); | |
3154 assert(assertMarkedBytesDataOK(), "Marked regions not OK at pause end."); | |
3155 } | |
3156 | |
3157 // Local Variables: *** | |
3158 // c-indentation-style: gnu *** | |
3159 // End: *** |