Print this page
rev 2661 : [mq]: g1-reference-processing
Split |
Close |
Expand all |
Collapse all |
--- old/src/share/vm/gc_implementation/g1/g1CollectorPolicy.cpp
+++ new/src/share/vm/gc_implementation/g1/g1CollectorPolicy.cpp
1 1 /*
2 2 * Copyright (c) 2001, 2011, Oracle and/or its affiliates. All rights reserved.
3 3 * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
4 4 *
5 5 * This code is free software; you can redistribute it and/or modify it
6 6 * under the terms of the GNU General Public License version 2 only, as
7 7 * published by the Free Software Foundation.
8 8 *
9 9 * This code is distributed in the hope that it will be useful, but WITHOUT
10 10 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
11 11 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
12 12 * version 2 for more details (a copy is included in the LICENSE file that
13 13 * accompanied this code).
14 14 *
15 15 * You should have received a copy of the GNU General Public License version
16 16 * 2 along with this work; if not, write to the Free Software Foundation,
17 17 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
18 18 *
19 19 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
20 20 * or visit www.oracle.com if you need additional information or have any
21 21 * questions.
22 22 *
23 23 */
24 24
25 25 #include "precompiled.hpp"
26 26 #include "gc_implementation/g1/concurrentG1Refine.hpp"
27 27 #include "gc_implementation/g1/concurrentMark.hpp"
28 28 #include "gc_implementation/g1/concurrentMarkThread.inline.hpp"
29 29 #include "gc_implementation/g1/g1CollectedHeap.inline.hpp"
30 30 #include "gc_implementation/g1/g1CollectorPolicy.hpp"
31 31 #include "gc_implementation/g1/g1ErgoVerbose.hpp"
32 32 #include "gc_implementation/g1/heapRegionRemSet.hpp"
33 33 #include "gc_implementation/shared/gcPolicyCounters.hpp"
34 34 #include "runtime/arguments.hpp"
35 35 #include "runtime/java.hpp"
36 36 #include "runtime/mutexLocker.hpp"
37 37 #include "utilities/debug.hpp"
38 38
39 39 #define PREDICTIONS_VERBOSE 0
40 40
41 41 // <NEW PREDICTION>
42 42
43 43 // Different defaults for different number of GC threads
44 44 // They were chosen by running GCOld and SPECjbb on debris with different
45 45 // numbers of GC threads and choosing them based on the results
46 46
47 47 // all the same
48 48 static double rs_length_diff_defaults[] = {
49 49 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0
50 50 };
51 51
52 52 static double cost_per_card_ms_defaults[] = {
53 53 0.01, 0.005, 0.005, 0.003, 0.003, 0.002, 0.002, 0.0015
54 54 };
55 55
56 56 // all the same
57 57 static double fully_young_cards_per_entry_ratio_defaults[] = {
58 58 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0, 1.0
59 59 };
60 60
61 61 static double cost_per_entry_ms_defaults[] = {
62 62 0.015, 0.01, 0.01, 0.008, 0.008, 0.0055, 0.0055, 0.005
63 63 };
64 64
65 65 static double cost_per_byte_ms_defaults[] = {
66 66 0.00006, 0.00003, 0.00003, 0.000015, 0.000015, 0.00001, 0.00001, 0.000009
67 67 };
68 68
69 69 // these should be pretty consistent
70 70 static double constant_other_time_ms_defaults[] = {
71 71 5.0, 5.0, 5.0, 5.0, 5.0, 5.0, 5.0, 5.0
72 72 };
73 73
74 74
75 75 static double young_other_cost_per_region_ms_defaults[] = {
76 76 0.3, 0.2, 0.2, 0.15, 0.15, 0.12, 0.12, 0.1
77 77 };
78 78
79 79 static double non_young_other_cost_per_region_ms_defaults[] = {
80 80 1.0, 0.7, 0.7, 0.5, 0.5, 0.42, 0.42, 0.30
81 81 };
82 82
83 83 // </NEW PREDICTION>
84 84
85 85 // Help class for avoiding interleaved logging
86 86 class LineBuffer: public StackObj {
87 87
88 88 private:
89 89 static const int BUFFER_LEN = 1024;
90 90 static const int INDENT_CHARS = 3;
91 91 char _buffer[BUFFER_LEN];
92 92 int _indent_level;
93 93 int _cur;
94 94
95 95 void vappend(const char* format, va_list ap) {
96 96 int res = vsnprintf(&_buffer[_cur], BUFFER_LEN - _cur, format, ap);
97 97 if (res != -1) {
98 98 _cur += res;
99 99 } else {
100 100 DEBUG_ONLY(warning("buffer too small in LineBuffer");)
101 101 _buffer[BUFFER_LEN -1] = 0;
102 102 _cur = BUFFER_LEN; // vsnprintf above should not add to _buffer if we are called again
103 103 }
104 104 }
105 105
106 106 public:
107 107 explicit LineBuffer(int indent_level): _indent_level(indent_level), _cur(0) {
108 108 for (; (_cur < BUFFER_LEN && _cur < (_indent_level * INDENT_CHARS)); _cur++) {
109 109 _buffer[_cur] = ' ';
110 110 }
111 111 }
112 112
113 113 #ifndef PRODUCT
114 114 ~LineBuffer() {
115 115 assert(_cur == _indent_level * INDENT_CHARS, "pending data in buffer - append_and_print_cr() not called?");
116 116 }
117 117 #endif
118 118
119 119 void append(const char* format, ...) {
120 120 va_list ap;
121 121 va_start(ap, format);
122 122 vappend(format, ap);
123 123 va_end(ap);
124 124 }
125 125
126 126 void append_and_print_cr(const char* format, ...) {
127 127 va_list ap;
128 128 va_start(ap, format);
129 129 vappend(format, ap);
130 130 va_end(ap);
131 131 gclog_or_tty->print_cr("%s", _buffer);
132 132 _cur = _indent_level * INDENT_CHARS;
133 133 }
134 134 };
135 135
136 136 G1CollectorPolicy::G1CollectorPolicy() :
137 137 _parallel_gc_threads(G1CollectedHeap::use_parallel_gc_threads()
138 138 ? ParallelGCThreads : 1),
139 139
140 140 _n_pauses(0),
141 141 _recent_rs_scan_times_ms(new TruncatedSeq(NumPrevPausesForHeuristics)),
142 142 _recent_pause_times_ms(new TruncatedSeq(NumPrevPausesForHeuristics)),
143 143 _recent_rs_sizes(new TruncatedSeq(NumPrevPausesForHeuristics)),
144 144 _recent_gc_times_ms(new TruncatedSeq(NumPrevPausesForHeuristics)),
↓ open down ↓ |
144 lines elided |
↑ open up ↑ |
145 145 _all_pause_times_ms(new NumberSeq()),
146 146 _stop_world_start(0.0),
147 147 _all_stop_world_times_ms(new NumberSeq()),
148 148 _all_yield_times_ms(new NumberSeq()),
149 149 _using_new_ratio_calculations(false),
150 150
151 151 _all_mod_union_times_ms(new NumberSeq()),
152 152
153 153 _summary(new Summary()),
154 154
155 -#ifndef PRODUCT
156 155 _cur_clear_ct_time_ms(0.0),
156 +
157 + _cur_ref_proc_time_ms(0.0),
158 + _cur_ref_enq_time_ms(0.0),
159 +
160 +#ifndef PRODUCT
157 161 _min_clear_cc_time_ms(-1.0),
158 162 _max_clear_cc_time_ms(-1.0),
159 163 _cur_clear_cc_time_ms(0.0),
160 164 _cum_clear_cc_time_ms(0.0),
161 165 _num_cc_clears(0L),
162 166 #endif
163 167
164 168 _region_num_young(0),
165 169 _region_num_tenured(0),
166 170 _prev_region_num_young(0),
167 171 _prev_region_num_tenured(0),
168 172
169 173 _aux_num(10),
170 174 _all_aux_times_ms(new NumberSeq[_aux_num]),
171 175 _cur_aux_start_times_ms(new double[_aux_num]),
172 176 _cur_aux_times_ms(new double[_aux_num]),
173 177 _cur_aux_times_set(new bool[_aux_num]),
174 178
175 179 _concurrent_mark_remark_times_ms(new TruncatedSeq(NumPrevPausesForHeuristics)),
176 180 _concurrent_mark_cleanup_times_ms(new TruncatedSeq(NumPrevPausesForHeuristics)),
177 181
178 182 // <NEW PREDICTION>
179 183
180 184 _alloc_rate_ms_seq(new TruncatedSeq(TruncatedSeqLength)),
181 185 _prev_collection_pause_end_ms(0.0),
182 186 _pending_card_diff_seq(new TruncatedSeq(TruncatedSeqLength)),
183 187 _rs_length_diff_seq(new TruncatedSeq(TruncatedSeqLength)),
184 188 _cost_per_card_ms_seq(new TruncatedSeq(TruncatedSeqLength)),
185 189 _fully_young_cards_per_entry_ratio_seq(new TruncatedSeq(TruncatedSeqLength)),
186 190 _partially_young_cards_per_entry_ratio_seq(
187 191 new TruncatedSeq(TruncatedSeqLength)),
188 192 _cost_per_entry_ms_seq(new TruncatedSeq(TruncatedSeqLength)),
189 193 _partially_young_cost_per_entry_ms_seq(new TruncatedSeq(TruncatedSeqLength)),
190 194 _cost_per_byte_ms_seq(new TruncatedSeq(TruncatedSeqLength)),
191 195 _cost_per_byte_ms_during_cm_seq(new TruncatedSeq(TruncatedSeqLength)),
192 196 _constant_other_time_ms_seq(new TruncatedSeq(TruncatedSeqLength)),
193 197 _young_other_cost_per_region_ms_seq(new TruncatedSeq(TruncatedSeqLength)),
194 198 _non_young_other_cost_per_region_ms_seq(
195 199 new TruncatedSeq(TruncatedSeqLength)),
196 200
197 201 _pending_cards_seq(new TruncatedSeq(TruncatedSeqLength)),
198 202 _scanned_cards_seq(new TruncatedSeq(TruncatedSeqLength)),
199 203 _rs_lengths_seq(new TruncatedSeq(TruncatedSeqLength)),
200 204
201 205 _pause_time_target_ms((double) MaxGCPauseMillis),
202 206
203 207 // </NEW PREDICTION>
204 208
205 209 _full_young_gcs(true),
206 210 _full_young_pause_num(0),
207 211 _partial_young_pause_num(0),
208 212
209 213 _during_marking(false),
210 214 _in_marking_window(false),
211 215 _in_marking_window_im(false),
212 216
213 217 _known_garbage_ratio(0.0),
214 218 _known_garbage_bytes(0),
215 219
216 220 _young_gc_eff_seq(new TruncatedSeq(TruncatedSeqLength)),
217 221
218 222 _recent_prev_end_times_for_all_gcs_sec(new TruncatedSeq(NumPrevPausesForHeuristics)),
219 223
220 224 _recent_CS_bytes_used_before(new TruncatedSeq(NumPrevPausesForHeuristics)),
221 225 _recent_CS_bytes_surviving(new TruncatedSeq(NumPrevPausesForHeuristics)),
222 226
223 227 _recent_avg_pause_time_ratio(0.0),
224 228 _num_markings(0),
225 229 _n_marks(0),
226 230 _n_pauses_at_mark_end(0),
227 231
228 232 _all_full_gc_times_ms(new NumberSeq()),
229 233
230 234 // G1PausesBtwnConcMark defaults to -1
231 235 // so the hack is to do the cast QQQ FIXME
232 236 _pauses_btwn_concurrent_mark((size_t)G1PausesBtwnConcMark),
233 237 _n_marks_since_last_pause(0),
234 238 _initiate_conc_mark_if_possible(false),
235 239 _during_initial_mark_pause(false),
236 240 _should_revert_to_full_young_gcs(false),
237 241 _last_full_young_gc(false),
238 242
239 243 _eden_bytes_before_gc(0),
240 244 _survivor_bytes_before_gc(0),
241 245 _capacity_before_gc(0),
242 246
243 247 _prev_collection_pause_used_at_end_bytes(0),
244 248
245 249 _collection_set(NULL),
246 250 _collection_set_size(0),
247 251 _collection_set_bytes_used_before(0),
248 252
249 253 // Incremental CSet attributes
250 254 _inc_cset_build_state(Inactive),
251 255 _inc_cset_head(NULL),
252 256 _inc_cset_tail(NULL),
253 257 _inc_cset_size(0),
254 258 _inc_cset_young_index(0),
255 259 _inc_cset_bytes_used_before(0),
256 260 _inc_cset_max_finger(NULL),
257 261 _inc_cset_recorded_young_bytes(0),
258 262 _inc_cset_recorded_rs_lengths(0),
259 263 _inc_cset_predicted_elapsed_time_ms(0.0),
260 264 _inc_cset_predicted_bytes_to_copy(0),
261 265
262 266 #ifdef _MSC_VER // the use of 'this' below gets a warning, make it go away
263 267 #pragma warning( disable:4355 ) // 'this' : used in base member initializer list
264 268 #endif // _MSC_VER
265 269
266 270 _short_lived_surv_rate_group(new SurvRateGroup(this, "Short Lived",
267 271 G1YoungSurvRateNumRegionsSummary)),
268 272 _survivor_surv_rate_group(new SurvRateGroup(this, "Survivor",
269 273 G1YoungSurvRateNumRegionsSummary)),
270 274 // add here any more surv rate groups
271 275 _recorded_survivor_regions(0),
272 276 _recorded_survivor_head(NULL),
273 277 _recorded_survivor_tail(NULL),
274 278 _survivors_age_table(true),
275 279
276 280 _gc_overhead_perc(0.0) {
277 281
278 282 // Set up the region size and associated fields. Given that the
279 283 // policy is created before the heap, we have to set this up here,
280 284 // so it's done as soon as possible.
281 285 HeapRegion::setup_heap_region_size(Arguments::min_heap_size());
282 286 HeapRegionRemSet::setup_remset_size();
283 287
284 288 G1ErgoVerbose::initialize();
285 289 if (PrintAdaptiveSizePolicy) {
286 290 // Currently, we only use a single switch for all the heuristics.
287 291 G1ErgoVerbose::set_enabled(true);
288 292 // Given that we don't currently have a verboseness level
289 293 // parameter, we'll hardcode this to high. This can be easily
290 294 // changed in the future.
291 295 G1ErgoVerbose::set_level(ErgoHigh);
292 296 } else {
293 297 G1ErgoVerbose::set_enabled(false);
294 298 }
295 299
296 300 // Verify PLAB sizes
297 301 const uint region_size = HeapRegion::GrainWords;
298 302 if (YoungPLABSize > region_size || OldPLABSize > region_size) {
299 303 char buffer[128];
300 304 jio_snprintf(buffer, sizeof(buffer), "%sPLABSize should be at most %u",
301 305 OldPLABSize > region_size ? "Old" : "Young", region_size);
302 306 vm_exit_during_initialization(buffer);
303 307 }
304 308
305 309 _recent_prev_end_times_for_all_gcs_sec->add(os::elapsedTime());
306 310 _prev_collection_pause_end_ms = os::elapsedTime() * 1000.0;
307 311
308 312 _par_last_gc_worker_start_times_ms = new double[_parallel_gc_threads];
309 313 _par_last_ext_root_scan_times_ms = new double[_parallel_gc_threads];
310 314 _par_last_mark_stack_scan_times_ms = new double[_parallel_gc_threads];
311 315
312 316 _par_last_update_rs_times_ms = new double[_parallel_gc_threads];
313 317 _par_last_update_rs_processed_buffers = new double[_parallel_gc_threads];
314 318
315 319 _par_last_scan_rs_times_ms = new double[_parallel_gc_threads];
316 320
317 321 _par_last_obj_copy_times_ms = new double[_parallel_gc_threads];
318 322
319 323 _par_last_termination_times_ms = new double[_parallel_gc_threads];
320 324 _par_last_termination_attempts = new double[_parallel_gc_threads];
321 325 _par_last_gc_worker_end_times_ms = new double[_parallel_gc_threads];
322 326 _par_last_gc_worker_times_ms = new double[_parallel_gc_threads];
323 327
324 328 // start conservatively
325 329 _expensive_region_limit_ms = 0.5 * (double) MaxGCPauseMillis;
326 330
327 331 // <NEW PREDICTION>
328 332
329 333 int index;
330 334 if (ParallelGCThreads == 0)
331 335 index = 0;
332 336 else if (ParallelGCThreads > 8)
333 337 index = 7;
334 338 else
335 339 index = ParallelGCThreads - 1;
336 340
337 341 _pending_card_diff_seq->add(0.0);
338 342 _rs_length_diff_seq->add(rs_length_diff_defaults[index]);
339 343 _cost_per_card_ms_seq->add(cost_per_card_ms_defaults[index]);
340 344 _fully_young_cards_per_entry_ratio_seq->add(
341 345 fully_young_cards_per_entry_ratio_defaults[index]);
342 346 _cost_per_entry_ms_seq->add(cost_per_entry_ms_defaults[index]);
343 347 _cost_per_byte_ms_seq->add(cost_per_byte_ms_defaults[index]);
344 348 _constant_other_time_ms_seq->add(constant_other_time_ms_defaults[index]);
345 349 _young_other_cost_per_region_ms_seq->add(
346 350 young_other_cost_per_region_ms_defaults[index]);
347 351 _non_young_other_cost_per_region_ms_seq->add(
348 352 non_young_other_cost_per_region_ms_defaults[index]);
349 353
350 354 // </NEW PREDICTION>
351 355
352 356 // Below, we might need to calculate the pause time target based on
353 357 // the pause interval. When we do so we are going to give G1 maximum
354 358 // flexibility and allow it to do pauses when it needs to. So, we'll
355 359 // arrange that the pause interval to be pause time target + 1 to
356 360 // ensure that a) the pause time target is maximized with respect to
357 361 // the pause interval and b) we maintain the invariant that pause
358 362 // time target < pause interval. If the user does not want this
359 363 // maximum flexibility, they will have to set the pause interval
360 364 // explicitly.
361 365
362 366 // First make sure that, if either parameter is set, its value is
363 367 // reasonable.
364 368 if (!FLAG_IS_DEFAULT(MaxGCPauseMillis)) {
365 369 if (MaxGCPauseMillis < 1) {
366 370 vm_exit_during_initialization("MaxGCPauseMillis should be "
367 371 "greater than 0");
368 372 }
369 373 }
370 374 if (!FLAG_IS_DEFAULT(GCPauseIntervalMillis)) {
371 375 if (GCPauseIntervalMillis < 1) {
372 376 vm_exit_during_initialization("GCPauseIntervalMillis should be "
373 377 "greater than 0");
374 378 }
375 379 }
376 380
377 381 // Then, if the pause time target parameter was not set, set it to
378 382 // the default value.
379 383 if (FLAG_IS_DEFAULT(MaxGCPauseMillis)) {
380 384 if (FLAG_IS_DEFAULT(GCPauseIntervalMillis)) {
381 385 // The default pause time target in G1 is 200ms
382 386 FLAG_SET_DEFAULT(MaxGCPauseMillis, 200);
383 387 } else {
384 388 // We do not allow the pause interval to be set without the
385 389 // pause time target
386 390 vm_exit_during_initialization("GCPauseIntervalMillis cannot be set "
387 391 "without setting MaxGCPauseMillis");
388 392 }
389 393 }
390 394
391 395 // Then, if the interval parameter was not set, set it according to
392 396 // the pause time target (this will also deal with the case when the
393 397 // pause time target is the default value).
394 398 if (FLAG_IS_DEFAULT(GCPauseIntervalMillis)) {
395 399 FLAG_SET_DEFAULT(GCPauseIntervalMillis, MaxGCPauseMillis + 1);
396 400 }
397 401
398 402 // Finally, make sure that the two parameters are consistent.
399 403 if (MaxGCPauseMillis >= GCPauseIntervalMillis) {
400 404 char buffer[256];
401 405 jio_snprintf(buffer, 256,
402 406 "MaxGCPauseMillis (%u) should be less than "
403 407 "GCPauseIntervalMillis (%u)",
404 408 MaxGCPauseMillis, GCPauseIntervalMillis);
405 409 vm_exit_during_initialization(buffer);
406 410 }
407 411
408 412 double max_gc_time = (double) MaxGCPauseMillis / 1000.0;
409 413 double time_slice = (double) GCPauseIntervalMillis / 1000.0;
410 414 _mmu_tracker = new G1MMUTrackerQueue(time_slice, max_gc_time);
411 415 _sigma = (double) G1ConfidencePercent / 100.0;
412 416
413 417 // start conservatively (around 50ms is about right)
414 418 _concurrent_mark_remark_times_ms->add(0.05);
415 419 _concurrent_mark_cleanup_times_ms->add(0.20);
416 420 _tenuring_threshold = MaxTenuringThreshold;
417 421 // _max_survivor_regions will be calculated by
418 422 // update_young_list_target_length() during initialization.
419 423 _max_survivor_regions = 0;
420 424
421 425 assert(GCTimeRatio > 0,
422 426 "we should have set it to a default value set_g1_gc_flags() "
423 427 "if a user set it to 0");
424 428 _gc_overhead_perc = 100.0 * (1.0 / (1.0 + GCTimeRatio));
425 429
426 430 uintx reserve_perc = G1ReservePercent;
427 431 // Put an artificial ceiling on this so that it's not set to a silly value.
428 432 if (reserve_perc > 50) {
429 433 reserve_perc = 50;
430 434 warning("G1ReservePercent is set to a value that is too large, "
431 435 "it's been updated to %u", reserve_perc);
432 436 }
433 437 _reserve_factor = (double) reserve_perc / 100.0;
434 438 // This will be set when the heap is expanded
435 439 // for the first time during initialization.
436 440 _reserve_regions = 0;
437 441
438 442 initialize_all();
439 443 }
440 444
441 445 // Increment "i", mod "len"
442 446 static void inc_mod(int& i, int len) {
443 447 i++; if (i == len) i = 0;
444 448 }
445 449
446 450 void G1CollectorPolicy::initialize_flags() {
447 451 set_min_alignment(HeapRegion::GrainBytes);
448 452 set_max_alignment(GenRemSet::max_alignment_constraint(rem_set_name()));
449 453 if (SurvivorRatio < 1) {
450 454 vm_exit_during_initialization("Invalid survivor ratio specified");
451 455 }
452 456 CollectorPolicy::initialize_flags();
453 457 }
454 458
455 459 // The easiest way to deal with the parsing of the NewSize /
456 460 // MaxNewSize / etc. parameteres is to re-use the code in the
457 461 // TwoGenerationCollectorPolicy class. This is similar to what
458 462 // ParallelScavenge does with its GenerationSizer class (see
459 463 // ParallelScavengeHeap::initialize()). We might change this in the
460 464 // future, but it's a good start.
461 465 class G1YoungGenSizer : public TwoGenerationCollectorPolicy {
462 466
463 467 public:
464 468 G1YoungGenSizer() {
465 469 initialize_flags();
466 470 initialize_size_info();
467 471 }
468 472 size_t size_to_region_num(size_t byte_size) {
469 473 return MAX2((size_t) 1, byte_size / HeapRegion::GrainBytes);
470 474 }
471 475 size_t min_young_region_num() {
472 476 return size_to_region_num(_min_gen0_size);
473 477 }
474 478 size_t initial_young_region_num() {
475 479 return size_to_region_num(_initial_gen0_size);
476 480 }
477 481 size_t max_young_region_num() {
478 482 return size_to_region_num(_max_gen0_size);
479 483 }
480 484 };
481 485
482 486 void G1CollectorPolicy::update_young_list_size_using_newratio(size_t number_of_heap_regions) {
483 487 assert(number_of_heap_regions > 0, "Heap must be initialized");
484 488 size_t young_size = number_of_heap_regions / (NewRatio + 1);
485 489 _min_desired_young_length = young_size;
486 490 _max_desired_young_length = young_size;
487 491 }
488 492
489 493 void G1CollectorPolicy::init() {
490 494 // Set aside an initial future to_space.
491 495 _g1 = G1CollectedHeap::heap();
492 496
493 497 assert(Heap_lock->owned_by_self(), "Locking discipline.");
494 498
495 499 initialize_gc_policy_counters();
496 500
497 501 G1YoungGenSizer sizer;
498 502 size_t initial_region_num = sizer.initial_young_region_num();
499 503 _min_desired_young_length = sizer.min_young_region_num();
500 504 _max_desired_young_length = sizer.max_young_region_num();
501 505
502 506 if (FLAG_IS_CMDLINE(NewRatio)) {
503 507 if (FLAG_IS_CMDLINE(NewSize) || FLAG_IS_CMDLINE(MaxNewSize)) {
504 508 gclog_or_tty->print_cr("-XX:NewSize and -XX:MaxNewSize overrides -XX:NewRatio");
505 509 } else {
506 510 // Treat NewRatio as a fixed size that is only recalculated when the heap size changes
507 511 size_t heap_regions = sizer.size_to_region_num(_g1->n_regions());
508 512 update_young_list_size_using_newratio(heap_regions);
509 513 _using_new_ratio_calculations = true;
510 514 }
511 515 }
512 516
513 517 // GenCollectorPolicy guarantees that min <= initial <= max.
514 518 // Asserting here just to state that we rely on this property.
515 519 assert(_min_desired_young_length <= _max_desired_young_length, "Invalid min/max young gen size values");
516 520 assert(initial_region_num <= _max_desired_young_length, "Initial young gen size too large");
517 521 assert(_min_desired_young_length <= initial_region_num, "Initial young gen size too small");
518 522
519 523 set_adaptive_young_list_length(_min_desired_young_length < _max_desired_young_length);
520 524 if (adaptive_young_list_length()) {
521 525 _young_list_fixed_length = 0;
522 526 } else {
523 527 _young_list_fixed_length = initial_region_num;
524 528 }
525 529 _free_regions_at_end_of_collection = _g1->free_regions();
526 530 update_young_list_target_length();
527 531 _prev_eden_capacity = _young_list_target_length * HeapRegion::GrainBytes;
528 532
529 533 // We may immediately start allocating regions and placing them on the
530 534 // collection set list. Initialize the per-collection set info
531 535 start_incremental_cset_building();
532 536 }
533 537
534 538 // Create the jstat counters for the policy.
535 539 void G1CollectorPolicy::initialize_gc_policy_counters() {
536 540 _gc_policy_counters = new GCPolicyCounters("GarbageFirst", 1, 3);
537 541 }
538 542
539 543 bool G1CollectorPolicy::predict_will_fit(size_t young_length,
540 544 double base_time_ms,
541 545 size_t base_free_regions,
542 546 double target_pause_time_ms) {
543 547 if (young_length >= base_free_regions) {
544 548 // end condition 1: not enough space for the young regions
545 549 return false;
546 550 }
547 551
548 552 double accum_surv_rate = accum_yg_surv_rate_pred((int)(young_length - 1));
549 553 size_t bytes_to_copy =
550 554 (size_t) (accum_surv_rate * (double) HeapRegion::GrainBytes);
551 555 double copy_time_ms = predict_object_copy_time_ms(bytes_to_copy);
552 556 double young_other_time_ms = predict_young_other_time_ms(young_length);
553 557 double pause_time_ms = base_time_ms + copy_time_ms + young_other_time_ms;
554 558 if (pause_time_ms > target_pause_time_ms) {
555 559 // end condition 2: prediction is over the target pause time
556 560 return false;
557 561 }
558 562
559 563 size_t free_bytes =
560 564 (base_free_regions - young_length) * HeapRegion::GrainBytes;
561 565 if ((2.0 * sigma()) * (double) bytes_to_copy > (double) free_bytes) {
562 566 // end condition 3: out-of-space (conservatively!)
563 567 return false;
564 568 }
565 569
566 570 // success!
567 571 return true;
568 572 }
569 573
570 574 void G1CollectorPolicy::record_new_heap_size(size_t new_number_of_regions) {
571 575 // re-calculate the necessary reserve
572 576 double reserve_regions_d = (double) new_number_of_regions * _reserve_factor;
573 577 // We use ceiling so that if reserve_regions_d is > 0.0 (but
574 578 // smaller than 1.0) we'll get 1.
575 579 _reserve_regions = (size_t) ceil(reserve_regions_d);
576 580
577 581 if (_using_new_ratio_calculations) {
578 582 // -XX:NewRatio was specified so we need to update the
579 583 // young gen length when the heap size has changed.
580 584 update_young_list_size_using_newratio(new_number_of_regions);
581 585 }
582 586 }
583 587
584 588 size_t G1CollectorPolicy::calculate_young_list_desired_min_length(
585 589 size_t base_min_length) {
586 590 size_t desired_min_length = 0;
587 591 if (adaptive_young_list_length()) {
588 592 if (_alloc_rate_ms_seq->num() > 3) {
589 593 double now_sec = os::elapsedTime();
590 594 double when_ms = _mmu_tracker->when_max_gc_sec(now_sec) * 1000.0;
591 595 double alloc_rate_ms = predict_alloc_rate_ms();
592 596 desired_min_length = (size_t) ceil(alloc_rate_ms * when_ms);
593 597 } else {
594 598 // otherwise we don't have enough info to make the prediction
595 599 }
596 600 }
597 601 desired_min_length += base_min_length;
598 602 // make sure we don't go below any user-defined minimum bound
599 603 return MAX2(_min_desired_young_length, desired_min_length);
600 604 }
601 605
602 606 size_t G1CollectorPolicy::calculate_young_list_desired_max_length() {
603 607 // Here, we might want to also take into account any additional
604 608 // constraints (i.e., user-defined minimum bound). Currently, we
605 609 // effectively don't set this bound.
606 610 return _max_desired_young_length;
607 611 }
608 612
609 613 void G1CollectorPolicy::update_young_list_target_length(size_t rs_lengths) {
610 614 if (rs_lengths == (size_t) -1) {
611 615 // if it's set to the default value (-1), we should predict it;
612 616 // otherwise, use the given value.
613 617 rs_lengths = (size_t) get_new_prediction(_rs_lengths_seq);
614 618 }
615 619
616 620 // Calculate the absolute and desired min bounds.
617 621
618 622 // This is how many young regions we already have (currently: the survivors).
619 623 size_t base_min_length = recorded_survivor_regions();
620 624 // This is the absolute minimum young length, which ensures that we
621 625 // can allocate one eden region in the worst-case.
622 626 size_t absolute_min_length = base_min_length + 1;
623 627 size_t desired_min_length =
624 628 calculate_young_list_desired_min_length(base_min_length);
625 629 if (desired_min_length < absolute_min_length) {
626 630 desired_min_length = absolute_min_length;
627 631 }
628 632
629 633 // Calculate the absolute and desired max bounds.
630 634
631 635 // We will try our best not to "eat" into the reserve.
632 636 size_t absolute_max_length = 0;
633 637 if (_free_regions_at_end_of_collection > _reserve_regions) {
634 638 absolute_max_length = _free_regions_at_end_of_collection - _reserve_regions;
635 639 }
636 640 size_t desired_max_length = calculate_young_list_desired_max_length();
637 641 if (desired_max_length > absolute_max_length) {
638 642 desired_max_length = absolute_max_length;
639 643 }
640 644
641 645 size_t young_list_target_length = 0;
642 646 if (adaptive_young_list_length()) {
643 647 if (full_young_gcs()) {
644 648 young_list_target_length =
645 649 calculate_young_list_target_length(rs_lengths,
646 650 base_min_length,
647 651 desired_min_length,
648 652 desired_max_length);
649 653 _rs_lengths_prediction = rs_lengths;
650 654 } else {
651 655 // Don't calculate anything and let the code below bound it to
652 656 // the desired_min_length, i.e., do the next GC as soon as
653 657 // possible to maximize how many old regions we can add to it.
654 658 }
655 659 } else {
656 660 if (full_young_gcs()) {
657 661 young_list_target_length = _young_list_fixed_length;
658 662 } else {
659 663 // A bit arbitrary: during partially-young GCs we allocate half
660 664 // the young regions to try to add old regions to the CSet.
661 665 young_list_target_length = _young_list_fixed_length / 2;
662 666 // We choose to accept that we might go under the desired min
663 667 // length given that we intentionally ask for a smaller young gen.
664 668 desired_min_length = absolute_min_length;
665 669 }
666 670 }
667 671
668 672 // Make sure we don't go over the desired max length, nor under the
669 673 // desired min length. In case they clash, desired_min_length wins
670 674 // which is why that test is second.
671 675 if (young_list_target_length > desired_max_length) {
672 676 young_list_target_length = desired_max_length;
673 677 }
674 678 if (young_list_target_length < desired_min_length) {
675 679 young_list_target_length = desired_min_length;
676 680 }
677 681
678 682 assert(young_list_target_length > recorded_survivor_regions(),
679 683 "we should be able to allocate at least one eden region");
680 684 assert(young_list_target_length >= absolute_min_length, "post-condition");
681 685 _young_list_target_length = young_list_target_length;
682 686
683 687 update_max_gc_locker_expansion();
684 688 }
685 689
686 690 size_t
687 691 G1CollectorPolicy::calculate_young_list_target_length(size_t rs_lengths,
688 692 size_t base_min_length,
689 693 size_t desired_min_length,
690 694 size_t desired_max_length) {
691 695 assert(adaptive_young_list_length(), "pre-condition");
692 696 assert(full_young_gcs(), "only call this for fully-young GCs");
693 697
694 698 // In case some edge-condition makes the desired max length too small...
695 699 if (desired_max_length <= desired_min_length) {
696 700 return desired_min_length;
697 701 }
698 702
699 703 // We'll adjust min_young_length and max_young_length not to include
700 704 // the already allocated young regions (i.e., so they reflect the
701 705 // min and max eden regions we'll allocate). The base_min_length
702 706 // will be reflected in the predictions by the
703 707 // survivor_regions_evac_time prediction.
704 708 assert(desired_min_length > base_min_length, "invariant");
705 709 size_t min_young_length = desired_min_length - base_min_length;
706 710 assert(desired_max_length > base_min_length, "invariant");
707 711 size_t max_young_length = desired_max_length - base_min_length;
708 712
709 713 double target_pause_time_ms = _mmu_tracker->max_gc_time() * 1000.0;
710 714 double survivor_regions_evac_time = predict_survivor_regions_evac_time();
711 715 size_t pending_cards = (size_t) get_new_prediction(_pending_cards_seq);
712 716 size_t adj_rs_lengths = rs_lengths + predict_rs_length_diff();
713 717 size_t scanned_cards = predict_young_card_num(adj_rs_lengths);
714 718 double base_time_ms =
715 719 predict_base_elapsed_time_ms(pending_cards, scanned_cards) +
716 720 survivor_regions_evac_time;
717 721 size_t available_free_regions = _free_regions_at_end_of_collection;
718 722 size_t base_free_regions = 0;
719 723 if (available_free_regions > _reserve_regions) {
720 724 base_free_regions = available_free_regions - _reserve_regions;
721 725 }
722 726
723 727 // Here, we will make sure that the shortest young length that
724 728 // makes sense fits within the target pause time.
725 729
726 730 if (predict_will_fit(min_young_length, base_time_ms,
727 731 base_free_regions, target_pause_time_ms)) {
728 732 // The shortest young length will fit into the target pause time;
729 733 // we'll now check whether the absolute maximum number of young
730 734 // regions will fit in the target pause time. If not, we'll do
731 735 // a binary search between min_young_length and max_young_length.
732 736 if (predict_will_fit(max_young_length, base_time_ms,
733 737 base_free_regions, target_pause_time_ms)) {
734 738 // The maximum young length will fit into the target pause time.
735 739 // We are done so set min young length to the maximum length (as
736 740 // the result is assumed to be returned in min_young_length).
737 741 min_young_length = max_young_length;
738 742 } else {
739 743 // The maximum possible number of young regions will not fit within
740 744 // the target pause time so we'll search for the optimal
741 745 // length. The loop invariants are:
742 746 //
743 747 // min_young_length < max_young_length
744 748 // min_young_length is known to fit into the target pause time
745 749 // max_young_length is known not to fit into the target pause time
746 750 //
747 751 // Going into the loop we know the above hold as we've just
748 752 // checked them. Every time around the loop we check whether
749 753 // the middle value between min_young_length and
750 754 // max_young_length fits into the target pause time. If it
751 755 // does, it becomes the new min. If it doesn't, it becomes
752 756 // the new max. This way we maintain the loop invariants.
753 757
754 758 assert(min_young_length < max_young_length, "invariant");
755 759 size_t diff = (max_young_length - min_young_length) / 2;
756 760 while (diff > 0) {
757 761 size_t young_length = min_young_length + diff;
758 762 if (predict_will_fit(young_length, base_time_ms,
759 763 base_free_regions, target_pause_time_ms)) {
760 764 min_young_length = young_length;
761 765 } else {
762 766 max_young_length = young_length;
763 767 }
764 768 assert(min_young_length < max_young_length, "invariant");
765 769 diff = (max_young_length - min_young_length) / 2;
766 770 }
767 771 // The results is min_young_length which, according to the
768 772 // loop invariants, should fit within the target pause time.
769 773
770 774 // These are the post-conditions of the binary search above:
771 775 assert(min_young_length < max_young_length,
772 776 "otherwise we should have discovered that max_young_length "
773 777 "fits into the pause target and not done the binary search");
774 778 assert(predict_will_fit(min_young_length, base_time_ms,
775 779 base_free_regions, target_pause_time_ms),
776 780 "min_young_length, the result of the binary search, should "
777 781 "fit into the pause target");
778 782 assert(!predict_will_fit(min_young_length + 1, base_time_ms,
779 783 base_free_regions, target_pause_time_ms),
780 784 "min_young_length, the result of the binary search, should be "
781 785 "optimal, so no larger length should fit into the pause target");
782 786 }
783 787 } else {
784 788 // Even the minimum length doesn't fit into the pause time
785 789 // target, return it as the result nevertheless.
786 790 }
787 791 return base_min_length + min_young_length;
788 792 }
789 793
790 794 double G1CollectorPolicy::predict_survivor_regions_evac_time() {
791 795 double survivor_regions_evac_time = 0.0;
792 796 for (HeapRegion * r = _recorded_survivor_head;
793 797 r != NULL && r != _recorded_survivor_tail->get_next_young_region();
794 798 r = r->get_next_young_region()) {
795 799 survivor_regions_evac_time += predict_region_elapsed_time_ms(r, true);
796 800 }
797 801 return survivor_regions_evac_time;
798 802 }
799 803
800 804 void G1CollectorPolicy::revise_young_list_target_length_if_necessary() {
801 805 guarantee( adaptive_young_list_length(), "should not call this otherwise" );
802 806
803 807 size_t rs_lengths = _g1->young_list()->sampled_rs_lengths();
804 808 if (rs_lengths > _rs_lengths_prediction) {
805 809 // add 10% to avoid having to recalculate often
806 810 size_t rs_lengths_prediction = rs_lengths * 1100 / 1000;
807 811 update_young_list_target_length(rs_lengths_prediction);
808 812 }
809 813 }
810 814
811 815
812 816
813 817 HeapWord* G1CollectorPolicy::mem_allocate_work(size_t size,
814 818 bool is_tlab,
815 819 bool* gc_overhead_limit_was_exceeded) {
816 820 guarantee(false, "Not using this policy feature yet.");
817 821 return NULL;
818 822 }
819 823
820 824 // This method controls how a collector handles one or more
821 825 // of its generations being fully allocated.
822 826 HeapWord* G1CollectorPolicy::satisfy_failed_allocation(size_t size,
823 827 bool is_tlab) {
824 828 guarantee(false, "Not using this policy feature yet.");
825 829 return NULL;
826 830 }
827 831
828 832
829 833 #ifndef PRODUCT
830 834 bool G1CollectorPolicy::verify_young_ages() {
831 835 HeapRegion* head = _g1->young_list()->first_region();
832 836 return
833 837 verify_young_ages(head, _short_lived_surv_rate_group);
834 838 // also call verify_young_ages on any additional surv rate groups
835 839 }
836 840
837 841 bool
838 842 G1CollectorPolicy::verify_young_ages(HeapRegion* head,
839 843 SurvRateGroup *surv_rate_group) {
840 844 guarantee( surv_rate_group != NULL, "pre-condition" );
841 845
842 846 const char* name = surv_rate_group->name();
843 847 bool ret = true;
844 848 int prev_age = -1;
845 849
846 850 for (HeapRegion* curr = head;
847 851 curr != NULL;
848 852 curr = curr->get_next_young_region()) {
849 853 SurvRateGroup* group = curr->surv_rate_group();
850 854 if (group == NULL && !curr->is_survivor()) {
851 855 gclog_or_tty->print_cr("## %s: encountered NULL surv_rate_group", name);
852 856 ret = false;
853 857 }
854 858
855 859 if (surv_rate_group == group) {
856 860 int age = curr->age_in_surv_rate_group();
857 861
858 862 if (age < 0) {
859 863 gclog_or_tty->print_cr("## %s: encountered negative age", name);
860 864 ret = false;
861 865 }
862 866
863 867 if (age <= prev_age) {
864 868 gclog_or_tty->print_cr("## %s: region ages are not strictly increasing "
865 869 "(%d, %d)", name, age, prev_age);
866 870 ret = false;
867 871 }
868 872 prev_age = age;
869 873 }
870 874 }
871 875
872 876 return ret;
873 877 }
874 878 #endif // PRODUCT
875 879
876 880 void G1CollectorPolicy::record_full_collection_start() {
877 881 _cur_collection_start_sec = os::elapsedTime();
878 882 // Release the future to-space so that it is available for compaction into.
879 883 _g1->set_full_collection();
880 884 }
881 885
882 886 void G1CollectorPolicy::record_full_collection_end() {
883 887 // Consider this like a collection pause for the purposes of allocation
884 888 // since last pause.
885 889 double end_sec = os::elapsedTime();
886 890 double full_gc_time_sec = end_sec - _cur_collection_start_sec;
887 891 double full_gc_time_ms = full_gc_time_sec * 1000.0;
888 892
889 893 _all_full_gc_times_ms->add(full_gc_time_ms);
890 894
891 895 update_recent_gc_times(end_sec, full_gc_time_ms);
892 896
893 897 _g1->clear_full_collection();
894 898
895 899 // "Nuke" the heuristics that control the fully/partially young GC
896 900 // transitions and make sure we start with fully young GCs after the
897 901 // Full GC.
898 902 set_full_young_gcs(true);
899 903 _last_full_young_gc = false;
900 904 _should_revert_to_full_young_gcs = false;
901 905 clear_initiate_conc_mark_if_possible();
902 906 clear_during_initial_mark_pause();
903 907 _known_garbage_bytes = 0;
904 908 _known_garbage_ratio = 0.0;
905 909 _in_marking_window = false;
906 910 _in_marking_window_im = false;
907 911
908 912 _short_lived_surv_rate_group->start_adding_regions();
909 913 // also call this on any additional surv rate groups
910 914
911 915 record_survivor_regions(0, NULL, NULL);
912 916
913 917 _prev_region_num_young = _region_num_young;
914 918 _prev_region_num_tenured = _region_num_tenured;
915 919
916 920 _free_regions_at_end_of_collection = _g1->free_regions();
917 921 // Reset survivors SurvRateGroup.
918 922 _survivor_surv_rate_group->reset();
919 923 update_young_list_target_length();
920 924 }
921 925
922 926 void G1CollectorPolicy::record_stop_world_start() {
923 927 _stop_world_start = os::elapsedTime();
924 928 }
925 929
926 930 void G1CollectorPolicy::record_collection_pause_start(double start_time_sec,
927 931 size_t start_used) {
928 932 if (PrintGCDetails) {
929 933 gclog_or_tty->stamp(PrintGCTimeStamps);
930 934 gclog_or_tty->print("[GC pause");
931 935 gclog_or_tty->print(" (%s)", full_young_gcs() ? "young" : "partial");
932 936 }
933 937
934 938 // We only need to do this here as the policy will only be applied
935 939 // to the GC we're about to start. so, no point is calculating this
936 940 // every time we calculate / recalculate the target young length.
937 941 update_survivors_policy();
938 942
939 943 assert(_g1->used() == _g1->recalculate_used(),
940 944 err_msg("sanity, used: "SIZE_FORMAT" recalculate_used: "SIZE_FORMAT,
941 945 _g1->used(), _g1->recalculate_used()));
942 946
943 947 double s_w_t_ms = (start_time_sec - _stop_world_start) * 1000.0;
944 948 _all_stop_world_times_ms->add(s_w_t_ms);
945 949 _stop_world_start = 0.0;
946 950
947 951 _cur_collection_start_sec = start_time_sec;
948 952 _cur_collection_pause_used_at_start_bytes = start_used;
949 953 _cur_collection_pause_used_regions_at_start = _g1->used_regions();
950 954 _pending_cards = _g1->pending_card_num();
951 955 _max_pending_cards = _g1->max_pending_card_num();
952 956
953 957 _bytes_in_collection_set_before_gc = 0;
954 958 _bytes_copied_during_gc = 0;
955 959
956 960 YoungList* young_list = _g1->young_list();
957 961 _eden_bytes_before_gc = young_list->eden_used_bytes();
958 962 _survivor_bytes_before_gc = young_list->survivor_used_bytes();
959 963 _capacity_before_gc = _g1->capacity();
960 964
961 965 #ifdef DEBUG
962 966 // initialise these to something well known so that we can spot
963 967 // if they are not set properly
964 968
965 969 for (int i = 0; i < _parallel_gc_threads; ++i) {
966 970 _par_last_gc_worker_start_times_ms[i] = -1234.0;
967 971 _par_last_ext_root_scan_times_ms[i] = -1234.0;
968 972 _par_last_mark_stack_scan_times_ms[i] = -1234.0;
969 973 _par_last_update_rs_times_ms[i] = -1234.0;
970 974 _par_last_update_rs_processed_buffers[i] = -1234.0;
971 975 _par_last_scan_rs_times_ms[i] = -1234.0;
972 976 _par_last_obj_copy_times_ms[i] = -1234.0;
973 977 _par_last_termination_times_ms[i] = -1234.0;
974 978 _par_last_termination_attempts[i] = -1234.0;
975 979 _par_last_gc_worker_end_times_ms[i] = -1234.0;
976 980 _par_last_gc_worker_times_ms[i] = -1234.0;
977 981 }
978 982 #endif
979 983
980 984 for (int i = 0; i < _aux_num; ++i) {
981 985 _cur_aux_times_ms[i] = 0.0;
982 986 _cur_aux_times_set[i] = false;
983 987 }
984 988
985 989 _satb_drain_time_set = false;
986 990 _last_satb_drain_processed_buffers = -1;
987 991
988 992 _last_young_gc_full = false;
989 993
990 994 // do that for any other surv rate groups
991 995 _short_lived_surv_rate_group->stop_adding_regions();
992 996 _survivors_age_table.clear();
993 997
994 998 assert( verify_young_ages(), "region age verification" );
995 999 }
996 1000
997 1001 void G1CollectorPolicy::record_mark_closure_time(double mark_closure_time_ms) {
998 1002 _mark_closure_time_ms = mark_closure_time_ms;
999 1003 }
1000 1004
1001 1005 void G1CollectorPolicy::record_concurrent_mark_init_end(double
1002 1006 mark_init_elapsed_time_ms) {
1003 1007 _during_marking = true;
1004 1008 assert(!initiate_conc_mark_if_possible(), "we should have cleared it by now");
1005 1009 clear_during_initial_mark_pause();
1006 1010 _cur_mark_stop_world_time_ms = mark_init_elapsed_time_ms;
1007 1011 }
1008 1012
1009 1013 void G1CollectorPolicy::record_concurrent_mark_remark_start() {
1010 1014 _mark_remark_start_sec = os::elapsedTime();
1011 1015 _during_marking = false;
1012 1016 }
1013 1017
1014 1018 void G1CollectorPolicy::record_concurrent_mark_remark_end() {
1015 1019 double end_time_sec = os::elapsedTime();
1016 1020 double elapsed_time_ms = (end_time_sec - _mark_remark_start_sec)*1000.0;
1017 1021 _concurrent_mark_remark_times_ms->add(elapsed_time_ms);
1018 1022 _cur_mark_stop_world_time_ms += elapsed_time_ms;
1019 1023 _prev_collection_pause_end_ms += elapsed_time_ms;
1020 1024
1021 1025 _mmu_tracker->add_pause(_mark_remark_start_sec, end_time_sec, true);
1022 1026 }
1023 1027
1024 1028 void G1CollectorPolicy::record_concurrent_mark_cleanup_start() {
1025 1029 _mark_cleanup_start_sec = os::elapsedTime();
1026 1030 }
1027 1031
1028 1032 void
1029 1033 G1CollectorPolicy::record_concurrent_mark_cleanup_end(size_t freed_bytes,
1030 1034 size_t max_live_bytes) {
1031 1035 record_concurrent_mark_cleanup_end_work1(freed_bytes, max_live_bytes);
1032 1036 record_concurrent_mark_cleanup_end_work2();
1033 1037 }
1034 1038
1035 1039 void
1036 1040 G1CollectorPolicy::
1037 1041 record_concurrent_mark_cleanup_end_work1(size_t freed_bytes,
1038 1042 size_t max_live_bytes) {
1039 1043 if (_n_marks < 2) {
1040 1044 _n_marks++;
1041 1045 }
1042 1046 }
1043 1047
1044 1048 // The important thing about this is that it includes "os::elapsedTime".
1045 1049 void G1CollectorPolicy::record_concurrent_mark_cleanup_end_work2() {
1046 1050 double end_time_sec = os::elapsedTime();
1047 1051 double elapsed_time_ms = (end_time_sec - _mark_cleanup_start_sec)*1000.0;
1048 1052 _concurrent_mark_cleanup_times_ms->add(elapsed_time_ms);
1049 1053 _cur_mark_stop_world_time_ms += elapsed_time_ms;
1050 1054 _prev_collection_pause_end_ms += elapsed_time_ms;
1051 1055
1052 1056 _mmu_tracker->add_pause(_mark_cleanup_start_sec, end_time_sec, true);
1053 1057
1054 1058 _num_markings++;
1055 1059 _n_pauses_at_mark_end = _n_pauses;
1056 1060 _n_marks_since_last_pause++;
1057 1061 }
1058 1062
1059 1063 void
1060 1064 G1CollectorPolicy::record_concurrent_mark_cleanup_completed() {
1061 1065 _should_revert_to_full_young_gcs = false;
1062 1066 _last_full_young_gc = true;
1063 1067 _in_marking_window = false;
1064 1068 }
1065 1069
1066 1070 void G1CollectorPolicy::record_concurrent_pause() {
1067 1071 if (_stop_world_start > 0.0) {
1068 1072 double yield_ms = (os::elapsedTime() - _stop_world_start) * 1000.0;
1069 1073 _all_yield_times_ms->add(yield_ms);
1070 1074 }
1071 1075 }
1072 1076
1073 1077 void G1CollectorPolicy::record_concurrent_pause_end() {
1074 1078 }
1075 1079
1076 1080 template<class T>
1077 1081 T sum_of(T* sum_arr, int start, int n, int N) {
1078 1082 T sum = (T)0;
1079 1083 for (int i = 0; i < n; i++) {
1080 1084 int j = (start + i) % N;
1081 1085 sum += sum_arr[j];
1082 1086 }
1083 1087 return sum;
1084 1088 }
1085 1089
1086 1090 void G1CollectorPolicy::print_par_stats(int level,
1087 1091 const char* str,
1088 1092 double* data) {
1089 1093 double min = data[0], max = data[0];
1090 1094 double total = 0.0;
1091 1095 LineBuffer buf(level);
1092 1096 buf.append("[%s (ms):", str);
1093 1097 for (uint i = 0; i < ParallelGCThreads; ++i) {
1094 1098 double val = data[i];
1095 1099 if (val < min)
1096 1100 min = val;
1097 1101 if (val > max)
1098 1102 max = val;
1099 1103 total += val;
1100 1104 buf.append(" %3.1lf", val);
1101 1105 }
1102 1106 buf.append_and_print_cr("");
1103 1107 double avg = total / (double) ParallelGCThreads;
1104 1108 buf.append_and_print_cr(" Avg: %5.1lf, Min: %5.1lf, Max: %5.1lf, Diff: %5.1lf]",
1105 1109 avg, min, max, max - min);
1106 1110 }
1107 1111
1108 1112 void G1CollectorPolicy::print_par_sizes(int level,
1109 1113 const char* str,
1110 1114 double* data) {
1111 1115 double min = data[0], max = data[0];
1112 1116 double total = 0.0;
1113 1117 LineBuffer buf(level);
1114 1118 buf.append("[%s :", str);
1115 1119 for (uint i = 0; i < ParallelGCThreads; ++i) {
1116 1120 double val = data[i];
1117 1121 if (val < min)
1118 1122 min = val;
1119 1123 if (val > max)
1120 1124 max = val;
1121 1125 total += val;
1122 1126 buf.append(" %d", (int) val);
1123 1127 }
1124 1128 buf.append_and_print_cr("");
1125 1129 double avg = total / (double) ParallelGCThreads;
1126 1130 buf.append_and_print_cr(" Sum: %d, Avg: %d, Min: %d, Max: %d, Diff: %d]",
1127 1131 (int)total, (int)avg, (int)min, (int)max, (int)max - (int)min);
1128 1132 }
1129 1133
1130 1134 void G1CollectorPolicy::print_stats (int level,
1131 1135 const char* str,
1132 1136 double value) {
1133 1137 LineBuffer(level).append_and_print_cr("[%s: %5.1lf ms]", str, value);
1134 1138 }
1135 1139
1136 1140 void G1CollectorPolicy::print_stats (int level,
1137 1141 const char* str,
1138 1142 int value) {
1139 1143 LineBuffer(level).append_and_print_cr("[%s: %d]", str, value);
1140 1144 }
1141 1145
1142 1146 double G1CollectorPolicy::avg_value (double* data) {
1143 1147 if (G1CollectedHeap::use_parallel_gc_threads()) {
1144 1148 double ret = 0.0;
1145 1149 for (uint i = 0; i < ParallelGCThreads; ++i)
1146 1150 ret += data[i];
1147 1151 return ret / (double) ParallelGCThreads;
1148 1152 } else {
1149 1153 return data[0];
1150 1154 }
1151 1155 }
1152 1156
1153 1157 double G1CollectorPolicy::max_value (double* data) {
1154 1158 if (G1CollectedHeap::use_parallel_gc_threads()) {
1155 1159 double ret = data[0];
1156 1160 for (uint i = 1; i < ParallelGCThreads; ++i)
1157 1161 if (data[i] > ret)
1158 1162 ret = data[i];
1159 1163 return ret;
1160 1164 } else {
1161 1165 return data[0];
1162 1166 }
1163 1167 }
1164 1168
1165 1169 double G1CollectorPolicy::sum_of_values (double* data) {
1166 1170 if (G1CollectedHeap::use_parallel_gc_threads()) {
1167 1171 double sum = 0.0;
1168 1172 for (uint i = 0; i < ParallelGCThreads; i++)
1169 1173 sum += data[i];
1170 1174 return sum;
1171 1175 } else {
1172 1176 return data[0];
1173 1177 }
1174 1178 }
1175 1179
1176 1180 double G1CollectorPolicy::max_sum (double* data1,
1177 1181 double* data2) {
1178 1182 double ret = data1[0] + data2[0];
1179 1183
1180 1184 if (G1CollectedHeap::use_parallel_gc_threads()) {
1181 1185 for (uint i = 1; i < ParallelGCThreads; ++i) {
1182 1186 double data = data1[i] + data2[i];
1183 1187 if (data > ret)
1184 1188 ret = data;
1185 1189 }
1186 1190 }
1187 1191 return ret;
1188 1192 }
1189 1193
1190 1194 // Anything below that is considered to be zero
1191 1195 #define MIN_TIMER_GRANULARITY 0.0000001
1192 1196
1193 1197 void G1CollectorPolicy::record_collection_pause_end() {
1194 1198 double end_time_sec = os::elapsedTime();
1195 1199 double elapsed_ms = _last_pause_time_ms;
1196 1200 bool parallel = G1CollectedHeap::use_parallel_gc_threads();
1197 1201 size_t rs_size =
1198 1202 _cur_collection_pause_used_regions_at_start - collection_set_size();
1199 1203 size_t cur_used_bytes = _g1->used();
1200 1204 assert(cur_used_bytes == _g1->recalculate_used(), "It should!");
1201 1205 bool last_pause_included_initial_mark = false;
1202 1206 bool update_stats = !_g1->evacuation_failed();
1203 1207
1204 1208 #ifndef PRODUCT
1205 1209 if (G1YoungSurvRateVerbose) {
1206 1210 gclog_or_tty->print_cr("");
1207 1211 _short_lived_surv_rate_group->print();
1208 1212 // do that for any other surv rate groups too
1209 1213 }
1210 1214 #endif // PRODUCT
1211 1215
1212 1216 last_pause_included_initial_mark = during_initial_mark_pause();
1213 1217 if (last_pause_included_initial_mark)
1214 1218 record_concurrent_mark_init_end(0.0);
1215 1219
1216 1220 size_t marking_initiating_used_threshold =
1217 1221 (_g1->capacity() / 100) * InitiatingHeapOccupancyPercent;
1218 1222
1219 1223 if (!_g1->mark_in_progress() && !_last_full_young_gc) {
1220 1224 assert(!last_pause_included_initial_mark, "invariant");
1221 1225 if (cur_used_bytes > marking_initiating_used_threshold) {
1222 1226 if (cur_used_bytes > _prev_collection_pause_used_at_end_bytes) {
1223 1227 assert(!during_initial_mark_pause(), "we should not see this here");
1224 1228
1225 1229 ergo_verbose3(ErgoConcCycles,
1226 1230 "request concurrent cycle initiation",
1227 1231 ergo_format_reason("occupancy higher than threshold")
1228 1232 ergo_format_byte("occupancy")
1229 1233 ergo_format_byte_perc("threshold"),
1230 1234 cur_used_bytes,
1231 1235 marking_initiating_used_threshold,
1232 1236 (double) InitiatingHeapOccupancyPercent);
1233 1237
1234 1238 // Note: this might have already been set, if during the last
1235 1239 // pause we decided to start a cycle but at the beginning of
1236 1240 // this pause we decided to postpone it. That's OK.
1237 1241 set_initiate_conc_mark_if_possible();
1238 1242 } else {
1239 1243 ergo_verbose2(ErgoConcCycles,
1240 1244 "do not request concurrent cycle initiation",
1241 1245 ergo_format_reason("occupancy lower than previous occupancy")
1242 1246 ergo_format_byte("occupancy")
1243 1247 ergo_format_byte("previous occupancy"),
1244 1248 cur_used_bytes,
1245 1249 _prev_collection_pause_used_at_end_bytes);
1246 1250 }
1247 1251 }
1248 1252 }
1249 1253
1250 1254 _prev_collection_pause_used_at_end_bytes = cur_used_bytes;
1251 1255
1252 1256 _mmu_tracker->add_pause(end_time_sec - elapsed_ms/1000.0,
1253 1257 end_time_sec, false);
1254 1258
1255 1259 guarantee(_cur_collection_pause_used_regions_at_start >=
1256 1260 collection_set_size(),
1257 1261 "Negative RS size?");
1258 1262
1259 1263 // This assert is exempted when we're doing parallel collection pauses,
1260 1264 // because the fragmentation caused by the parallel GC allocation buffers
1261 1265 // can lead to more memory being used during collection than was used
1262 1266 // before. Best leave this out until the fragmentation problem is fixed.
1263 1267 // Pauses in which evacuation failed can also lead to negative
1264 1268 // collections, since no space is reclaimed from a region containing an
1265 1269 // object whose evacuation failed.
1266 1270 // Further, we're now always doing parallel collection. But I'm still
1267 1271 // leaving this here as a placeholder for a more precise assertion later.
1268 1272 // (DLD, 10/05.)
1269 1273 assert((true || parallel) // Always using GC LABs now.
1270 1274 || _g1->evacuation_failed()
1271 1275 || _cur_collection_pause_used_at_start_bytes >= cur_used_bytes,
1272 1276 "Negative collection");
1273 1277
1274 1278 size_t freed_bytes =
1275 1279 _cur_collection_pause_used_at_start_bytes - cur_used_bytes;
1276 1280 size_t surviving_bytes = _collection_set_bytes_used_before - freed_bytes;
1277 1281
1278 1282 double survival_fraction =
1279 1283 (double)surviving_bytes/
1280 1284 (double)_collection_set_bytes_used_before;
1281 1285
1282 1286 _n_pauses++;
1283 1287
1284 1288 double ext_root_scan_time = avg_value(_par_last_ext_root_scan_times_ms);
1285 1289 double mark_stack_scan_time = avg_value(_par_last_mark_stack_scan_times_ms);
1286 1290 double update_rs_time = avg_value(_par_last_update_rs_times_ms);
1287 1291 double update_rs_processed_buffers =
1288 1292 sum_of_values(_par_last_update_rs_processed_buffers);
1289 1293 double scan_rs_time = avg_value(_par_last_scan_rs_times_ms);
1290 1294 double obj_copy_time = avg_value(_par_last_obj_copy_times_ms);
1291 1295 double termination_time = avg_value(_par_last_termination_times_ms);
1292 1296
1293 1297 double parallel_known_time = update_rs_time +
1294 1298 ext_root_scan_time +
1295 1299 mark_stack_scan_time +
1296 1300 scan_rs_time +
1297 1301 obj_copy_time +
1298 1302 termination_time;
1299 1303
1300 1304 double parallel_other_time = _cur_collection_par_time_ms - parallel_known_time;
1301 1305
1302 1306 PauseSummary* summary = _summary;
1303 1307
1304 1308 if (update_stats) {
1305 1309 _recent_rs_scan_times_ms->add(scan_rs_time);
1306 1310 _recent_pause_times_ms->add(elapsed_ms);
1307 1311 _recent_rs_sizes->add(rs_size);
1308 1312
1309 1313 MainBodySummary* body_summary = summary->main_body_summary();
1310 1314 guarantee(body_summary != NULL, "should not be null!");
1311 1315
1312 1316 if (_satb_drain_time_set)
1313 1317 body_summary->record_satb_drain_time_ms(_cur_satb_drain_time_ms);
1314 1318 else
1315 1319 body_summary->record_satb_drain_time_ms(0.0);
1316 1320
1317 1321 body_summary->record_ext_root_scan_time_ms(ext_root_scan_time);
1318 1322 body_summary->record_mark_stack_scan_time_ms(mark_stack_scan_time);
1319 1323 body_summary->record_update_rs_time_ms(update_rs_time);
1320 1324 body_summary->record_scan_rs_time_ms(scan_rs_time);
1321 1325 body_summary->record_obj_copy_time_ms(obj_copy_time);
1322 1326 if (parallel) {
1323 1327 body_summary->record_parallel_time_ms(_cur_collection_par_time_ms);
1324 1328 body_summary->record_clear_ct_time_ms(_cur_clear_ct_time_ms);
1325 1329 body_summary->record_termination_time_ms(termination_time);
1326 1330 body_summary->record_parallel_other_time_ms(parallel_other_time);
1327 1331 }
1328 1332 body_summary->record_mark_closure_time_ms(_mark_closure_time_ms);
1329 1333
1330 1334 // We exempt parallel collection from this check because Alloc Buffer
1331 1335 // fragmentation can produce negative collections. Same with evac
1332 1336 // failure.
1333 1337 // Further, we're now always doing parallel collection. But I'm still
1334 1338 // leaving this here as a placeholder for a more precise assertion later.
1335 1339 // (DLD, 10/05.
1336 1340 assert((true || parallel)
1337 1341 || _g1->evacuation_failed()
1338 1342 || surviving_bytes <= _collection_set_bytes_used_before,
1339 1343 "Or else negative collection!");
1340 1344 _recent_CS_bytes_used_before->add(_collection_set_bytes_used_before);
1341 1345 _recent_CS_bytes_surviving->add(surviving_bytes);
1342 1346
1343 1347 // this is where we update the allocation rate of the application
1344 1348 double app_time_ms =
1345 1349 (_cur_collection_start_sec * 1000.0 - _prev_collection_pause_end_ms);
1346 1350 if (app_time_ms < MIN_TIMER_GRANULARITY) {
1347 1351 // This usually happens due to the timer not having the required
1348 1352 // granularity. Some Linuxes are the usual culprits.
1349 1353 // We'll just set it to something (arbitrarily) small.
1350 1354 app_time_ms = 1.0;
1351 1355 }
1352 1356 size_t regions_allocated =
1353 1357 (_region_num_young - _prev_region_num_young) +
1354 1358 (_region_num_tenured - _prev_region_num_tenured);
1355 1359 double alloc_rate_ms = (double) regions_allocated / app_time_ms;
1356 1360 _alloc_rate_ms_seq->add(alloc_rate_ms);
1357 1361 _prev_region_num_young = _region_num_young;
1358 1362 _prev_region_num_tenured = _region_num_tenured;
1359 1363
1360 1364 double interval_ms =
1361 1365 (end_time_sec - _recent_prev_end_times_for_all_gcs_sec->oldest()) * 1000.0;
1362 1366 update_recent_gc_times(end_time_sec, elapsed_ms);
1363 1367 _recent_avg_pause_time_ratio = _recent_gc_times_ms->sum()/interval_ms;
1364 1368 if (recent_avg_pause_time_ratio() < 0.0 ||
1365 1369 (recent_avg_pause_time_ratio() - 1.0 > 0.0)) {
1366 1370 #ifndef PRODUCT
1367 1371 // Dump info to allow post-facto debugging
1368 1372 gclog_or_tty->print_cr("recent_avg_pause_time_ratio() out of bounds");
1369 1373 gclog_or_tty->print_cr("-------------------------------------------");
1370 1374 gclog_or_tty->print_cr("Recent GC Times (ms):");
1371 1375 _recent_gc_times_ms->dump();
1372 1376 gclog_or_tty->print_cr("(End Time=%3.3f) Recent GC End Times (s):", end_time_sec);
1373 1377 _recent_prev_end_times_for_all_gcs_sec->dump();
1374 1378 gclog_or_tty->print_cr("GC = %3.3f, Interval = %3.3f, Ratio = %3.3f",
1375 1379 _recent_gc_times_ms->sum(), interval_ms, recent_avg_pause_time_ratio());
1376 1380 // In debug mode, terminate the JVM if the user wants to debug at this point.
1377 1381 assert(!G1FailOnFPError, "Debugging data for CR 6898948 has been dumped above");
1378 1382 #endif // !PRODUCT
1379 1383 // Clip ratio between 0.0 and 1.0, and continue. This will be fixed in
1380 1384 // CR 6902692 by redoing the manner in which the ratio is incrementally computed.
1381 1385 if (_recent_avg_pause_time_ratio < 0.0) {
1382 1386 _recent_avg_pause_time_ratio = 0.0;
1383 1387 } else {
1384 1388 assert(_recent_avg_pause_time_ratio - 1.0 > 0.0, "Ctl-point invariant");
1385 1389 _recent_avg_pause_time_ratio = 1.0;
1386 1390 }
1387 1391 }
1388 1392 }
1389 1393
1390 1394 if (G1PolicyVerbose > 1) {
1391 1395 gclog_or_tty->print_cr(" Recording collection pause(%d)", _n_pauses);
1392 1396 }
1393 1397
1394 1398 if (G1PolicyVerbose > 1) {
1395 1399 gclog_or_tty->print_cr(" ET: %10.6f ms (avg: %10.6f ms)\n"
1396 1400 " ET-RS: %10.6f ms (avg: %10.6f ms)\n"
1397 1401 " |RS|: " SIZE_FORMAT,
1398 1402 elapsed_ms, recent_avg_time_for_pauses_ms(),
1399 1403 scan_rs_time, recent_avg_time_for_rs_scan_ms(),
1400 1404 rs_size);
1401 1405
1402 1406 gclog_or_tty->print_cr(" Used at start: " SIZE_FORMAT"K"
1403 1407 " At end " SIZE_FORMAT "K\n"
1404 1408 " garbage : " SIZE_FORMAT "K"
1405 1409 " of " SIZE_FORMAT "K\n"
1406 1410 " survival : %6.2f%% (%6.2f%% avg)",
1407 1411 _cur_collection_pause_used_at_start_bytes/K,
1408 1412 _g1->used()/K, freed_bytes/K,
1409 1413 _collection_set_bytes_used_before/K,
1410 1414 survival_fraction*100.0,
1411 1415 recent_avg_survival_fraction()*100.0);
1412 1416 gclog_or_tty->print_cr(" Recent %% gc pause time: %6.2f",
1413 1417 recent_avg_pause_time_ratio() * 100.0);
1414 1418 }
1415 1419
1416 1420 double other_time_ms = elapsed_ms;
1417 1421
1418 1422 if (_satb_drain_time_set) {
1419 1423 other_time_ms -= _cur_satb_drain_time_ms;
1420 1424 }
1421 1425
1422 1426 if (parallel) {
1423 1427 other_time_ms -= _cur_collection_par_time_ms + _cur_clear_ct_time_ms;
1424 1428 } else {
1425 1429 other_time_ms -=
1426 1430 update_rs_time +
1427 1431 ext_root_scan_time + mark_stack_scan_time +
1428 1432 scan_rs_time + obj_copy_time;
1429 1433 }
1430 1434
1431 1435 if (PrintGCDetails) {
1432 1436 gclog_or_tty->print_cr("%s, %1.8lf secs]",
1433 1437 (last_pause_included_initial_mark) ? " (initial-mark)" : "",
1434 1438 elapsed_ms / 1000.0);
1435 1439
1436 1440 if (_satb_drain_time_set) {
1437 1441 print_stats(1, "SATB Drain Time", _cur_satb_drain_time_ms);
1438 1442 }
1439 1443 if (_last_satb_drain_processed_buffers >= 0) {
1440 1444 print_stats(2, "Processed Buffers", _last_satb_drain_processed_buffers);
1441 1445 }
1442 1446 if (parallel) {
1443 1447 print_stats(1, "Parallel Time", _cur_collection_par_time_ms);
1444 1448 print_par_stats(2, "GC Worker Start Time", _par_last_gc_worker_start_times_ms);
1445 1449 print_par_stats(2, "Update RS", _par_last_update_rs_times_ms);
1446 1450 print_par_sizes(3, "Processed Buffers", _par_last_update_rs_processed_buffers);
1447 1451 print_par_stats(2, "Ext Root Scanning", _par_last_ext_root_scan_times_ms);
1448 1452 print_par_stats(2, "Mark Stack Scanning", _par_last_mark_stack_scan_times_ms);
1449 1453 print_par_stats(2, "Scan RS", _par_last_scan_rs_times_ms);
1450 1454 print_par_stats(2, "Object Copy", _par_last_obj_copy_times_ms);
1451 1455 print_par_stats(2, "Termination", _par_last_termination_times_ms);
1452 1456 print_par_sizes(3, "Termination Attempts", _par_last_termination_attempts);
1453 1457 print_par_stats(2, "GC Worker End Time", _par_last_gc_worker_end_times_ms);
1454 1458
1455 1459 for (int i = 0; i < _parallel_gc_threads; i++) {
1456 1460 _par_last_gc_worker_times_ms[i] = _par_last_gc_worker_end_times_ms[i] - _par_last_gc_worker_start_times_ms[i];
1457 1461 }
1458 1462 print_par_stats(2, "GC Worker Times", _par_last_gc_worker_times_ms);
1459 1463
1460 1464 print_stats(2, "Parallel Other", parallel_other_time);
1461 1465 print_stats(1, "Clear CT", _cur_clear_ct_time_ms);
1462 1466 } else {
1463 1467 print_stats(1, "Update RS", update_rs_time);
1464 1468 print_stats(2, "Processed Buffers",
1465 1469 (int)update_rs_processed_buffers);
1466 1470 print_stats(1, "Ext Root Scanning", ext_root_scan_time);
1467 1471 print_stats(1, "Mark Stack Scanning", mark_stack_scan_time);
1468 1472 print_stats(1, "Scan RS", scan_rs_time);
1469 1473 print_stats(1, "Object Copying", obj_copy_time);
1470 1474 }
1471 1475 #ifndef PRODUCT
↓ open down ↓ |
1305 lines elided |
↑ open up ↑ |
1472 1476 print_stats(1, "Cur Clear CC", _cur_clear_cc_time_ms);
1473 1477 print_stats(1, "Cum Clear CC", _cum_clear_cc_time_ms);
1474 1478 print_stats(1, "Min Clear CC", _min_clear_cc_time_ms);
1475 1479 print_stats(1, "Max Clear CC", _max_clear_cc_time_ms);
1476 1480 if (_num_cc_clears > 0) {
1477 1481 print_stats(1, "Avg Clear CC", _cum_clear_cc_time_ms / ((double)_num_cc_clears));
1478 1482 }
1479 1483 #endif
1480 1484 print_stats(1, "Other", other_time_ms);
1481 1485 print_stats(2, "Choose CSet", _recorded_young_cset_choice_time_ms);
1486 + print_stats(2, "Ref Proc", _cur_ref_proc_time_ms);
1487 + print_stats(2, "Ref Enq", _cur_ref_enq_time_ms);
1482 1488
1483 1489 for (int i = 0; i < _aux_num; ++i) {
1484 1490 if (_cur_aux_times_set[i]) {
1485 1491 char buffer[96];
1486 1492 sprintf(buffer, "Aux%d", i);
1487 1493 print_stats(1, buffer, _cur_aux_times_ms[i]);
1488 1494 }
1489 1495 }
1490 1496 }
1491 1497
1492 1498 _all_pause_times_ms->add(elapsed_ms);
1493 1499 if (update_stats) {
1494 1500 summary->record_total_time_ms(elapsed_ms);
1495 1501 summary->record_other_time_ms(other_time_ms);
1496 1502 }
1497 1503 for (int i = 0; i < _aux_num; ++i)
1498 1504 if (_cur_aux_times_set[i])
1499 1505 _all_aux_times_ms[i].add(_cur_aux_times_ms[i]);
1500 1506
1501 1507 // Reset marks-between-pauses counter.
1502 1508 _n_marks_since_last_pause = 0;
1503 1509
1504 1510 // Update the efficiency-since-mark vars.
1505 1511 double proc_ms = elapsed_ms * (double) _parallel_gc_threads;
1506 1512 if (elapsed_ms < MIN_TIMER_GRANULARITY) {
1507 1513 // This usually happens due to the timer not having the required
1508 1514 // granularity. Some Linuxes are the usual culprits.
1509 1515 // We'll just set it to something (arbitrarily) small.
1510 1516 proc_ms = 1.0;
1511 1517 }
1512 1518 double cur_efficiency = (double) freed_bytes / proc_ms;
1513 1519
1514 1520 bool new_in_marking_window = _in_marking_window;
1515 1521 bool new_in_marking_window_im = false;
1516 1522 if (during_initial_mark_pause()) {
1517 1523 new_in_marking_window = true;
1518 1524 new_in_marking_window_im = true;
1519 1525 }
1520 1526
1521 1527 if (_last_full_young_gc) {
1522 1528 ergo_verbose2(ErgoPartiallyYoungGCs,
1523 1529 "start partially-young GCs",
1524 1530 ergo_format_byte_perc("known garbage"),
1525 1531 _known_garbage_bytes, _known_garbage_ratio * 100.0);
1526 1532 set_full_young_gcs(false);
1527 1533 _last_full_young_gc = false;
1528 1534 }
1529 1535
1530 1536 if ( !_last_young_gc_full ) {
1531 1537 if (_should_revert_to_full_young_gcs) {
1532 1538 ergo_verbose2(ErgoPartiallyYoungGCs,
1533 1539 "end partially-young GCs",
1534 1540 ergo_format_reason("partially-young GCs end requested")
1535 1541 ergo_format_byte_perc("known garbage"),
1536 1542 _known_garbage_bytes, _known_garbage_ratio * 100.0);
1537 1543 set_full_young_gcs(true);
1538 1544 } else if (_known_garbage_ratio < 0.05) {
1539 1545 ergo_verbose3(ErgoPartiallyYoungGCs,
1540 1546 "end partially-young GCs",
1541 1547 ergo_format_reason("known garbage percent lower than threshold")
1542 1548 ergo_format_byte_perc("known garbage")
1543 1549 ergo_format_perc("threshold"),
1544 1550 _known_garbage_bytes, _known_garbage_ratio * 100.0,
1545 1551 0.05 * 100.0);
1546 1552 set_full_young_gcs(true);
1547 1553 } else if (adaptive_young_list_length() &&
1548 1554 (get_gc_eff_factor() * cur_efficiency < predict_young_gc_eff())) {
1549 1555 ergo_verbose5(ErgoPartiallyYoungGCs,
1550 1556 "end partially-young GCs",
1551 1557 ergo_format_reason("current GC efficiency lower than "
1552 1558 "predicted fully-young GC efficiency")
1553 1559 ergo_format_double("GC efficiency factor")
1554 1560 ergo_format_double("current GC efficiency")
1555 1561 ergo_format_double("predicted fully-young GC efficiency")
1556 1562 ergo_format_byte_perc("known garbage"),
1557 1563 get_gc_eff_factor(), cur_efficiency,
1558 1564 predict_young_gc_eff(),
1559 1565 _known_garbage_bytes, _known_garbage_ratio * 100.0);
1560 1566 set_full_young_gcs(true);
1561 1567 }
1562 1568 }
1563 1569 _should_revert_to_full_young_gcs = false;
1564 1570
1565 1571 if (_last_young_gc_full && !_during_marking) {
1566 1572 _young_gc_eff_seq->add(cur_efficiency);
1567 1573 }
1568 1574
1569 1575 _short_lived_surv_rate_group->start_adding_regions();
1570 1576 // do that for any other surv rate groupsx
1571 1577
1572 1578 // <NEW PREDICTION>
1573 1579
1574 1580 if (update_stats) {
1575 1581 double pause_time_ms = elapsed_ms;
1576 1582
1577 1583 size_t diff = 0;
1578 1584 if (_max_pending_cards >= _pending_cards)
1579 1585 diff = _max_pending_cards - _pending_cards;
1580 1586 _pending_card_diff_seq->add((double) diff);
1581 1587
1582 1588 double cost_per_card_ms = 0.0;
1583 1589 if (_pending_cards > 0) {
1584 1590 cost_per_card_ms = update_rs_time / (double) _pending_cards;
1585 1591 _cost_per_card_ms_seq->add(cost_per_card_ms);
1586 1592 }
1587 1593
1588 1594 size_t cards_scanned = _g1->cards_scanned();
1589 1595
1590 1596 double cost_per_entry_ms = 0.0;
1591 1597 if (cards_scanned > 10) {
1592 1598 cost_per_entry_ms = scan_rs_time / (double) cards_scanned;
1593 1599 if (_last_young_gc_full)
1594 1600 _cost_per_entry_ms_seq->add(cost_per_entry_ms);
1595 1601 else
1596 1602 _partially_young_cost_per_entry_ms_seq->add(cost_per_entry_ms);
1597 1603 }
1598 1604
1599 1605 if (_max_rs_lengths > 0) {
1600 1606 double cards_per_entry_ratio =
1601 1607 (double) cards_scanned / (double) _max_rs_lengths;
1602 1608 if (_last_young_gc_full)
1603 1609 _fully_young_cards_per_entry_ratio_seq->add(cards_per_entry_ratio);
1604 1610 else
1605 1611 _partially_young_cards_per_entry_ratio_seq->add(cards_per_entry_ratio);
1606 1612 }
1607 1613
1608 1614 size_t rs_length_diff = _max_rs_lengths - _recorded_rs_lengths;
1609 1615 if (rs_length_diff >= 0)
1610 1616 _rs_length_diff_seq->add((double) rs_length_diff);
1611 1617
1612 1618 size_t copied_bytes = surviving_bytes;
1613 1619 double cost_per_byte_ms = 0.0;
1614 1620 if (copied_bytes > 0) {
1615 1621 cost_per_byte_ms = obj_copy_time / (double) copied_bytes;
1616 1622 if (_in_marking_window)
1617 1623 _cost_per_byte_ms_during_cm_seq->add(cost_per_byte_ms);
1618 1624 else
1619 1625 _cost_per_byte_ms_seq->add(cost_per_byte_ms);
1620 1626 }
1621 1627
1622 1628 double all_other_time_ms = pause_time_ms -
1623 1629 (update_rs_time + scan_rs_time + obj_copy_time +
1624 1630 _mark_closure_time_ms + termination_time);
1625 1631
1626 1632 double young_other_time_ms = 0.0;
1627 1633 if (_recorded_young_regions > 0) {
1628 1634 young_other_time_ms =
1629 1635 _recorded_young_cset_choice_time_ms +
1630 1636 _recorded_young_free_cset_time_ms;
1631 1637 _young_other_cost_per_region_ms_seq->add(young_other_time_ms /
1632 1638 (double) _recorded_young_regions);
1633 1639 }
1634 1640 double non_young_other_time_ms = 0.0;
1635 1641 if (_recorded_non_young_regions > 0) {
1636 1642 non_young_other_time_ms =
1637 1643 _recorded_non_young_cset_choice_time_ms +
1638 1644 _recorded_non_young_free_cset_time_ms;
1639 1645
1640 1646 _non_young_other_cost_per_region_ms_seq->add(non_young_other_time_ms /
1641 1647 (double) _recorded_non_young_regions);
1642 1648 }
1643 1649
1644 1650 double constant_other_time_ms = all_other_time_ms -
1645 1651 (young_other_time_ms + non_young_other_time_ms);
1646 1652 _constant_other_time_ms_seq->add(constant_other_time_ms);
1647 1653
1648 1654 double survival_ratio = 0.0;
1649 1655 if (_bytes_in_collection_set_before_gc > 0) {
1650 1656 survival_ratio = (double) _bytes_copied_during_gc /
1651 1657 (double) _bytes_in_collection_set_before_gc;
1652 1658 }
1653 1659
1654 1660 _pending_cards_seq->add((double) _pending_cards);
1655 1661 _scanned_cards_seq->add((double) cards_scanned);
1656 1662 _rs_lengths_seq->add((double) _max_rs_lengths);
1657 1663
1658 1664 double expensive_region_limit_ms =
1659 1665 (double) MaxGCPauseMillis - predict_constant_other_time_ms();
1660 1666 if (expensive_region_limit_ms < 0.0) {
1661 1667 // this means that the other time was predicted to be longer than
1662 1668 // than the max pause time
1663 1669 expensive_region_limit_ms = (double) MaxGCPauseMillis;
1664 1670 }
1665 1671 _expensive_region_limit_ms = expensive_region_limit_ms;
1666 1672
1667 1673 if (PREDICTIONS_VERBOSE) {
1668 1674 gclog_or_tty->print_cr("");
1669 1675 gclog_or_tty->print_cr("PREDICTIONS %1.4lf %d "
1670 1676 "REGIONS %d %d %d "
1671 1677 "PENDING_CARDS %d %d "
1672 1678 "CARDS_SCANNED %d %d "
1673 1679 "RS_LENGTHS %d %d "
1674 1680 "RS_UPDATE %1.6lf %1.6lf RS_SCAN %1.6lf %1.6lf "
1675 1681 "SURVIVAL_RATIO %1.6lf %1.6lf "
1676 1682 "OBJECT_COPY %1.6lf %1.6lf OTHER_CONSTANT %1.6lf %1.6lf "
1677 1683 "OTHER_YOUNG %1.6lf %1.6lf "
1678 1684 "OTHER_NON_YOUNG %1.6lf %1.6lf "
1679 1685 "VTIME_DIFF %1.6lf TERMINATION %1.6lf "
1680 1686 "ELAPSED %1.6lf %1.6lf ",
1681 1687 _cur_collection_start_sec,
1682 1688 (!_last_young_gc_full) ? 2 :
1683 1689 (last_pause_included_initial_mark) ? 1 : 0,
1684 1690 _recorded_region_num,
1685 1691 _recorded_young_regions,
1686 1692 _recorded_non_young_regions,
1687 1693 _predicted_pending_cards, _pending_cards,
1688 1694 _predicted_cards_scanned, cards_scanned,
1689 1695 _predicted_rs_lengths, _max_rs_lengths,
1690 1696 _predicted_rs_update_time_ms, update_rs_time,
1691 1697 _predicted_rs_scan_time_ms, scan_rs_time,
1692 1698 _predicted_survival_ratio, survival_ratio,
1693 1699 _predicted_object_copy_time_ms, obj_copy_time,
1694 1700 _predicted_constant_other_time_ms, constant_other_time_ms,
1695 1701 _predicted_young_other_time_ms, young_other_time_ms,
1696 1702 _predicted_non_young_other_time_ms,
1697 1703 non_young_other_time_ms,
1698 1704 _vtime_diff_ms, termination_time,
1699 1705 _predicted_pause_time_ms, elapsed_ms);
1700 1706 }
1701 1707
1702 1708 if (G1PolicyVerbose > 0) {
1703 1709 gclog_or_tty->print_cr("Pause Time, predicted: %1.4lfms (predicted %s), actual: %1.4lfms",
1704 1710 _predicted_pause_time_ms,
1705 1711 (_within_target) ? "within" : "outside",
1706 1712 elapsed_ms);
1707 1713 }
1708 1714
1709 1715 }
1710 1716
1711 1717 _in_marking_window = new_in_marking_window;
1712 1718 _in_marking_window_im = new_in_marking_window_im;
1713 1719 _free_regions_at_end_of_collection = _g1->free_regions();
1714 1720 update_young_list_target_length();
1715 1721
1716 1722 // Note that _mmu_tracker->max_gc_time() returns the time in seconds.
1717 1723 double update_rs_time_goal_ms = _mmu_tracker->max_gc_time() * MILLIUNITS * G1RSetUpdatingPauseTimePercent / 100.0;
1718 1724 adjust_concurrent_refinement(update_rs_time, update_rs_processed_buffers, update_rs_time_goal_ms);
1719 1725 // </NEW PREDICTION>
1720 1726 }
1721 1727
1722 1728 #define EXT_SIZE_FORMAT "%d%s"
1723 1729 #define EXT_SIZE_PARAMS(bytes) \
1724 1730 byte_size_in_proper_unit((bytes)), \
1725 1731 proper_unit_for_byte_size((bytes))
1726 1732
1727 1733 void G1CollectorPolicy::print_heap_transition() {
1728 1734 if (PrintGCDetails) {
1729 1735 YoungList* young_list = _g1->young_list();
1730 1736 size_t eden_bytes = young_list->eden_used_bytes();
1731 1737 size_t survivor_bytes = young_list->survivor_used_bytes();
1732 1738 size_t used_before_gc = _cur_collection_pause_used_at_start_bytes;
1733 1739 size_t used = _g1->used();
1734 1740 size_t capacity = _g1->capacity();
1735 1741 size_t eden_capacity =
1736 1742 (_young_list_target_length * HeapRegion::GrainBytes) - survivor_bytes;
1737 1743
1738 1744 gclog_or_tty->print_cr(
1739 1745 " [Eden: "EXT_SIZE_FORMAT"("EXT_SIZE_FORMAT")->"EXT_SIZE_FORMAT"("EXT_SIZE_FORMAT") "
1740 1746 "Survivors: "EXT_SIZE_FORMAT"->"EXT_SIZE_FORMAT" "
1741 1747 "Heap: "EXT_SIZE_FORMAT"("EXT_SIZE_FORMAT")->"
1742 1748 EXT_SIZE_FORMAT"("EXT_SIZE_FORMAT")]",
1743 1749 EXT_SIZE_PARAMS(_eden_bytes_before_gc),
1744 1750 EXT_SIZE_PARAMS(_prev_eden_capacity),
1745 1751 EXT_SIZE_PARAMS(eden_bytes),
1746 1752 EXT_SIZE_PARAMS(eden_capacity),
1747 1753 EXT_SIZE_PARAMS(_survivor_bytes_before_gc),
1748 1754 EXT_SIZE_PARAMS(survivor_bytes),
1749 1755 EXT_SIZE_PARAMS(used_before_gc),
1750 1756 EXT_SIZE_PARAMS(_capacity_before_gc),
1751 1757 EXT_SIZE_PARAMS(used),
1752 1758 EXT_SIZE_PARAMS(capacity));
1753 1759
1754 1760 _prev_eden_capacity = eden_capacity;
1755 1761 } else if (PrintGC) {
1756 1762 _g1->print_size_transition(gclog_or_tty,
1757 1763 _cur_collection_pause_used_at_start_bytes,
1758 1764 _g1->used(), _g1->capacity());
1759 1765 }
1760 1766 }
1761 1767
1762 1768 // <NEW PREDICTION>
1763 1769
1764 1770 void G1CollectorPolicy::adjust_concurrent_refinement(double update_rs_time,
1765 1771 double update_rs_processed_buffers,
1766 1772 double goal_ms) {
1767 1773 DirtyCardQueueSet& dcqs = JavaThread::dirty_card_queue_set();
1768 1774 ConcurrentG1Refine *cg1r = G1CollectedHeap::heap()->concurrent_g1_refine();
1769 1775
1770 1776 if (G1UseAdaptiveConcRefinement) {
1771 1777 const int k_gy = 3, k_gr = 6;
1772 1778 const double inc_k = 1.1, dec_k = 0.9;
1773 1779
1774 1780 int g = cg1r->green_zone();
1775 1781 if (update_rs_time > goal_ms) {
1776 1782 g = (int)(g * dec_k); // Can become 0, that's OK. That would mean a mutator-only processing.
1777 1783 } else {
1778 1784 if (update_rs_time < goal_ms && update_rs_processed_buffers > g) {
1779 1785 g = (int)MAX2(g * inc_k, g + 1.0);
1780 1786 }
1781 1787 }
1782 1788 // Change the refinement threads params
1783 1789 cg1r->set_green_zone(g);
1784 1790 cg1r->set_yellow_zone(g * k_gy);
1785 1791 cg1r->set_red_zone(g * k_gr);
1786 1792 cg1r->reinitialize_threads();
1787 1793
1788 1794 int processing_threshold_delta = MAX2((int)(cg1r->green_zone() * sigma()), 1);
1789 1795 int processing_threshold = MIN2(cg1r->green_zone() + processing_threshold_delta,
1790 1796 cg1r->yellow_zone());
1791 1797 // Change the barrier params
1792 1798 dcqs.set_process_completed_threshold(processing_threshold);
1793 1799 dcqs.set_max_completed_queue(cg1r->red_zone());
1794 1800 }
1795 1801
1796 1802 int curr_queue_size = dcqs.completed_buffers_num();
1797 1803 if (curr_queue_size >= cg1r->yellow_zone()) {
1798 1804 dcqs.set_completed_queue_padding(curr_queue_size);
1799 1805 } else {
1800 1806 dcqs.set_completed_queue_padding(0);
1801 1807 }
1802 1808 dcqs.notify_if_necessary();
1803 1809 }
1804 1810
1805 1811 double
1806 1812 G1CollectorPolicy::
1807 1813 predict_young_collection_elapsed_time_ms(size_t adjustment) {
1808 1814 guarantee( adjustment == 0 || adjustment == 1, "invariant" );
1809 1815
1810 1816 G1CollectedHeap* g1h = G1CollectedHeap::heap();
1811 1817 size_t young_num = g1h->young_list()->length();
1812 1818 if (young_num == 0)
1813 1819 return 0.0;
1814 1820
1815 1821 young_num += adjustment;
1816 1822 size_t pending_cards = predict_pending_cards();
1817 1823 size_t rs_lengths = g1h->young_list()->sampled_rs_lengths() +
1818 1824 predict_rs_length_diff();
1819 1825 size_t card_num;
1820 1826 if (full_young_gcs())
1821 1827 card_num = predict_young_card_num(rs_lengths);
1822 1828 else
1823 1829 card_num = predict_non_young_card_num(rs_lengths);
1824 1830 size_t young_byte_size = young_num * HeapRegion::GrainBytes;
1825 1831 double accum_yg_surv_rate =
1826 1832 _short_lived_surv_rate_group->accum_surv_rate(adjustment);
1827 1833
1828 1834 size_t bytes_to_copy =
1829 1835 (size_t) (accum_yg_surv_rate * (double) HeapRegion::GrainBytes);
1830 1836
1831 1837 return
1832 1838 predict_rs_update_time_ms(pending_cards) +
1833 1839 predict_rs_scan_time_ms(card_num) +
1834 1840 predict_object_copy_time_ms(bytes_to_copy) +
1835 1841 predict_young_other_time_ms(young_num) +
1836 1842 predict_constant_other_time_ms();
1837 1843 }
1838 1844
1839 1845 double
1840 1846 G1CollectorPolicy::predict_base_elapsed_time_ms(size_t pending_cards) {
1841 1847 size_t rs_length = predict_rs_length_diff();
1842 1848 size_t card_num;
1843 1849 if (full_young_gcs())
1844 1850 card_num = predict_young_card_num(rs_length);
1845 1851 else
1846 1852 card_num = predict_non_young_card_num(rs_length);
1847 1853 return predict_base_elapsed_time_ms(pending_cards, card_num);
1848 1854 }
1849 1855
1850 1856 double
1851 1857 G1CollectorPolicy::predict_base_elapsed_time_ms(size_t pending_cards,
1852 1858 size_t scanned_cards) {
1853 1859 return
1854 1860 predict_rs_update_time_ms(pending_cards) +
1855 1861 predict_rs_scan_time_ms(scanned_cards) +
1856 1862 predict_constant_other_time_ms();
1857 1863 }
1858 1864
1859 1865 double
1860 1866 G1CollectorPolicy::predict_region_elapsed_time_ms(HeapRegion* hr,
1861 1867 bool young) {
1862 1868 size_t rs_length = hr->rem_set()->occupied();
1863 1869 size_t card_num;
1864 1870 if (full_young_gcs())
1865 1871 card_num = predict_young_card_num(rs_length);
1866 1872 else
1867 1873 card_num = predict_non_young_card_num(rs_length);
1868 1874 size_t bytes_to_copy = predict_bytes_to_copy(hr);
1869 1875
1870 1876 double region_elapsed_time_ms =
1871 1877 predict_rs_scan_time_ms(card_num) +
1872 1878 predict_object_copy_time_ms(bytes_to_copy);
1873 1879
1874 1880 if (young)
1875 1881 region_elapsed_time_ms += predict_young_other_time_ms(1);
1876 1882 else
1877 1883 region_elapsed_time_ms += predict_non_young_other_time_ms(1);
1878 1884
1879 1885 return region_elapsed_time_ms;
1880 1886 }
1881 1887
1882 1888 size_t
1883 1889 G1CollectorPolicy::predict_bytes_to_copy(HeapRegion* hr) {
1884 1890 size_t bytes_to_copy;
1885 1891 if (hr->is_marked())
1886 1892 bytes_to_copy = hr->max_live_bytes();
1887 1893 else {
1888 1894 guarantee( hr->is_young() && hr->age_in_surv_rate_group() != -1,
1889 1895 "invariant" );
1890 1896 int age = hr->age_in_surv_rate_group();
1891 1897 double yg_surv_rate = predict_yg_surv_rate(age, hr->surv_rate_group());
1892 1898 bytes_to_copy = (size_t) ((double) hr->used() * yg_surv_rate);
1893 1899 }
1894 1900
1895 1901 return bytes_to_copy;
1896 1902 }
1897 1903
1898 1904 void
1899 1905 G1CollectorPolicy::start_recording_regions() {
1900 1906 _recorded_rs_lengths = 0;
1901 1907 _recorded_young_regions = 0;
1902 1908 _recorded_non_young_regions = 0;
1903 1909
1904 1910 #if PREDICTIONS_VERBOSE
1905 1911 _recorded_marked_bytes = 0;
1906 1912 _recorded_young_bytes = 0;
1907 1913 _predicted_bytes_to_copy = 0;
1908 1914 _predicted_rs_lengths = 0;
1909 1915 _predicted_cards_scanned = 0;
1910 1916 #endif // PREDICTIONS_VERBOSE
1911 1917 }
1912 1918
1913 1919 void
1914 1920 G1CollectorPolicy::record_cset_region_info(HeapRegion* hr, bool young) {
1915 1921 #if PREDICTIONS_VERBOSE
1916 1922 if (!young) {
1917 1923 _recorded_marked_bytes += hr->max_live_bytes();
1918 1924 }
1919 1925 _predicted_bytes_to_copy += predict_bytes_to_copy(hr);
1920 1926 #endif // PREDICTIONS_VERBOSE
1921 1927
1922 1928 size_t rs_length = hr->rem_set()->occupied();
1923 1929 _recorded_rs_lengths += rs_length;
1924 1930 }
1925 1931
1926 1932 void
1927 1933 G1CollectorPolicy::record_non_young_cset_region(HeapRegion* hr) {
1928 1934 assert(!hr->is_young(), "should not call this");
1929 1935 ++_recorded_non_young_regions;
1930 1936 record_cset_region_info(hr, false);
1931 1937 }
1932 1938
1933 1939 void
1934 1940 G1CollectorPolicy::set_recorded_young_regions(size_t n_regions) {
1935 1941 _recorded_young_regions = n_regions;
1936 1942 }
1937 1943
1938 1944 void G1CollectorPolicy::set_recorded_young_bytes(size_t bytes) {
1939 1945 #if PREDICTIONS_VERBOSE
1940 1946 _recorded_young_bytes = bytes;
1941 1947 #endif // PREDICTIONS_VERBOSE
1942 1948 }
1943 1949
1944 1950 void G1CollectorPolicy::set_recorded_rs_lengths(size_t rs_lengths) {
1945 1951 _recorded_rs_lengths = rs_lengths;
1946 1952 }
1947 1953
1948 1954 void G1CollectorPolicy::set_predicted_bytes_to_copy(size_t bytes) {
1949 1955 _predicted_bytes_to_copy = bytes;
1950 1956 }
1951 1957
1952 1958 void
1953 1959 G1CollectorPolicy::end_recording_regions() {
1954 1960 // The _predicted_pause_time_ms field is referenced in code
1955 1961 // not under PREDICTIONS_VERBOSE. Let's initialize it.
1956 1962 _predicted_pause_time_ms = -1.0;
1957 1963
1958 1964 #if PREDICTIONS_VERBOSE
1959 1965 _predicted_pending_cards = predict_pending_cards();
1960 1966 _predicted_rs_lengths = _recorded_rs_lengths + predict_rs_length_diff();
1961 1967 if (full_young_gcs())
1962 1968 _predicted_cards_scanned += predict_young_card_num(_predicted_rs_lengths);
1963 1969 else
1964 1970 _predicted_cards_scanned +=
1965 1971 predict_non_young_card_num(_predicted_rs_lengths);
1966 1972 _recorded_region_num = _recorded_young_regions + _recorded_non_young_regions;
1967 1973
1968 1974 _predicted_rs_update_time_ms =
1969 1975 predict_rs_update_time_ms(_g1->pending_card_num());
1970 1976 _predicted_rs_scan_time_ms =
1971 1977 predict_rs_scan_time_ms(_predicted_cards_scanned);
1972 1978 _predicted_object_copy_time_ms =
1973 1979 predict_object_copy_time_ms(_predicted_bytes_to_copy);
1974 1980 _predicted_constant_other_time_ms =
1975 1981 predict_constant_other_time_ms();
1976 1982 _predicted_young_other_time_ms =
1977 1983 predict_young_other_time_ms(_recorded_young_regions);
1978 1984 _predicted_non_young_other_time_ms =
1979 1985 predict_non_young_other_time_ms(_recorded_non_young_regions);
1980 1986
1981 1987 _predicted_pause_time_ms =
1982 1988 _predicted_rs_update_time_ms +
1983 1989 _predicted_rs_scan_time_ms +
1984 1990 _predicted_object_copy_time_ms +
1985 1991 _predicted_constant_other_time_ms +
1986 1992 _predicted_young_other_time_ms +
1987 1993 _predicted_non_young_other_time_ms;
1988 1994 #endif // PREDICTIONS_VERBOSE
1989 1995 }
1990 1996
1991 1997 void G1CollectorPolicy::check_if_region_is_too_expensive(double
1992 1998 predicted_time_ms) {
1993 1999 // I don't think we need to do this when in young GC mode since
1994 2000 // marking will be initiated next time we hit the soft limit anyway...
1995 2001 if (predicted_time_ms > _expensive_region_limit_ms) {
1996 2002 ergo_verbose2(ErgoPartiallyYoungGCs,
1997 2003 "request partially-young GCs end",
1998 2004 ergo_format_reason("predicted region time higher than threshold")
1999 2005 ergo_format_ms("predicted region time")
2000 2006 ergo_format_ms("threshold"),
2001 2007 predicted_time_ms, _expensive_region_limit_ms);
2002 2008 // no point in doing another partial one
2003 2009 _should_revert_to_full_young_gcs = true;
2004 2010 }
2005 2011 }
2006 2012
2007 2013 // </NEW PREDICTION>
2008 2014
2009 2015
2010 2016 void G1CollectorPolicy::update_recent_gc_times(double end_time_sec,
2011 2017 double elapsed_ms) {
2012 2018 _recent_gc_times_ms->add(elapsed_ms);
2013 2019 _recent_prev_end_times_for_all_gcs_sec->add(end_time_sec);
2014 2020 _prev_collection_pause_end_ms = end_time_sec * 1000.0;
2015 2021 }
2016 2022
2017 2023 double G1CollectorPolicy::recent_avg_time_for_pauses_ms() {
2018 2024 if (_recent_pause_times_ms->num() == 0) {
2019 2025 return (double) MaxGCPauseMillis;
2020 2026 }
2021 2027 return _recent_pause_times_ms->avg();
2022 2028 }
2023 2029
2024 2030 double G1CollectorPolicy::recent_avg_time_for_rs_scan_ms() {
2025 2031 if (_recent_rs_scan_times_ms->num() == 0) {
2026 2032 return (double)MaxGCPauseMillis/3.0;
2027 2033 }
2028 2034 return _recent_rs_scan_times_ms->avg();
2029 2035 }
2030 2036
2031 2037 int G1CollectorPolicy::number_of_recent_gcs() {
2032 2038 assert(_recent_rs_scan_times_ms->num() ==
2033 2039 _recent_pause_times_ms->num(), "Sequence out of sync");
2034 2040 assert(_recent_pause_times_ms->num() ==
2035 2041 _recent_CS_bytes_used_before->num(), "Sequence out of sync");
2036 2042 assert(_recent_CS_bytes_used_before->num() ==
2037 2043 _recent_CS_bytes_surviving->num(), "Sequence out of sync");
2038 2044
2039 2045 return _recent_pause_times_ms->num();
2040 2046 }
2041 2047
2042 2048 double G1CollectorPolicy::recent_avg_survival_fraction() {
2043 2049 return recent_avg_survival_fraction_work(_recent_CS_bytes_surviving,
2044 2050 _recent_CS_bytes_used_before);
2045 2051 }
2046 2052
2047 2053 double G1CollectorPolicy::last_survival_fraction() {
2048 2054 return last_survival_fraction_work(_recent_CS_bytes_surviving,
2049 2055 _recent_CS_bytes_used_before);
2050 2056 }
2051 2057
2052 2058 double
2053 2059 G1CollectorPolicy::recent_avg_survival_fraction_work(TruncatedSeq* surviving,
2054 2060 TruncatedSeq* before) {
2055 2061 assert(surviving->num() == before->num(), "Sequence out of sync");
2056 2062 if (before->sum() > 0.0) {
2057 2063 double recent_survival_rate = surviving->sum() / before->sum();
2058 2064 // We exempt parallel collection from this check because Alloc Buffer
2059 2065 // fragmentation can produce negative collections.
2060 2066 // Further, we're now always doing parallel collection. But I'm still
2061 2067 // leaving this here as a placeholder for a more precise assertion later.
2062 2068 // (DLD, 10/05.)
2063 2069 assert((true || G1CollectedHeap::use_parallel_gc_threads()) ||
2064 2070 _g1->evacuation_failed() ||
2065 2071 recent_survival_rate <= 1.0, "Or bad frac");
2066 2072 return recent_survival_rate;
2067 2073 } else {
2068 2074 return 1.0; // Be conservative.
2069 2075 }
2070 2076 }
2071 2077
2072 2078 double
2073 2079 G1CollectorPolicy::last_survival_fraction_work(TruncatedSeq* surviving,
2074 2080 TruncatedSeq* before) {
2075 2081 assert(surviving->num() == before->num(), "Sequence out of sync");
2076 2082 if (surviving->num() > 0 && before->last() > 0.0) {
2077 2083 double last_survival_rate = surviving->last() / before->last();
2078 2084 // We exempt parallel collection from this check because Alloc Buffer
2079 2085 // fragmentation can produce negative collections.
2080 2086 // Further, we're now always doing parallel collection. But I'm still
2081 2087 // leaving this here as a placeholder for a more precise assertion later.
2082 2088 // (DLD, 10/05.)
2083 2089 assert((true || G1CollectedHeap::use_parallel_gc_threads()) ||
2084 2090 last_survival_rate <= 1.0, "Or bad frac");
2085 2091 return last_survival_rate;
2086 2092 } else {
2087 2093 return 1.0;
2088 2094 }
2089 2095 }
2090 2096
2091 2097 static const int survival_min_obs = 5;
2092 2098 static double survival_min_obs_limits[] = { 0.9, 0.7, 0.5, 0.3, 0.1 };
2093 2099 static const double min_survival_rate = 0.1;
2094 2100
2095 2101 double
2096 2102 G1CollectorPolicy::conservative_avg_survival_fraction_work(double avg,
2097 2103 double latest) {
2098 2104 double res = avg;
2099 2105 if (number_of_recent_gcs() < survival_min_obs) {
2100 2106 res = MAX2(res, survival_min_obs_limits[number_of_recent_gcs()]);
2101 2107 }
2102 2108 res = MAX2(res, latest);
2103 2109 res = MAX2(res, min_survival_rate);
2104 2110 // In the parallel case, LAB fragmentation can produce "negative
2105 2111 // collections"; so can evac failure. Cap at 1.0
2106 2112 res = MIN2(res, 1.0);
2107 2113 return res;
2108 2114 }
2109 2115
2110 2116 size_t G1CollectorPolicy::expansion_amount() {
2111 2117 double recent_gc_overhead = recent_avg_pause_time_ratio() * 100.0;
2112 2118 double threshold = _gc_overhead_perc;
2113 2119 if (recent_gc_overhead > threshold) {
2114 2120 // We will double the existing space, or take
2115 2121 // G1ExpandByPercentOfAvailable % of the available expansion
2116 2122 // space, whichever is smaller, bounded below by a minimum
2117 2123 // expansion (unless that's all that's left.)
2118 2124 const size_t min_expand_bytes = 1*M;
2119 2125 size_t reserved_bytes = _g1->max_capacity();
2120 2126 size_t committed_bytes = _g1->capacity();
2121 2127 size_t uncommitted_bytes = reserved_bytes - committed_bytes;
2122 2128 size_t expand_bytes;
2123 2129 size_t expand_bytes_via_pct =
2124 2130 uncommitted_bytes * G1ExpandByPercentOfAvailable / 100;
2125 2131 expand_bytes = MIN2(expand_bytes_via_pct, committed_bytes);
2126 2132 expand_bytes = MAX2(expand_bytes, min_expand_bytes);
2127 2133 expand_bytes = MIN2(expand_bytes, uncommitted_bytes);
2128 2134
2129 2135 ergo_verbose5(ErgoHeapSizing,
2130 2136 "attempt heap expansion",
2131 2137 ergo_format_reason("recent GC overhead higher than "
2132 2138 "threshold after GC")
2133 2139 ergo_format_perc("recent GC overhead")
2134 2140 ergo_format_perc("threshold")
2135 2141 ergo_format_byte("uncommitted")
2136 2142 ergo_format_byte_perc("calculated expansion amount"),
2137 2143 recent_gc_overhead, threshold,
2138 2144 uncommitted_bytes,
2139 2145 expand_bytes_via_pct, (double) G1ExpandByPercentOfAvailable);
2140 2146
2141 2147 return expand_bytes;
2142 2148 } else {
2143 2149 return 0;
2144 2150 }
2145 2151 }
2146 2152
2147 2153 void G1CollectorPolicy::note_start_of_mark_thread() {
2148 2154 _mark_thread_startup_sec = os::elapsedTime();
2149 2155 }
2150 2156
2151 2157 class CountCSClosure: public HeapRegionClosure {
2152 2158 G1CollectorPolicy* _g1_policy;
2153 2159 public:
2154 2160 CountCSClosure(G1CollectorPolicy* g1_policy) :
2155 2161 _g1_policy(g1_policy) {}
2156 2162 bool doHeapRegion(HeapRegion* r) {
2157 2163 _g1_policy->_bytes_in_collection_set_before_gc += r->used();
2158 2164 return false;
2159 2165 }
2160 2166 };
2161 2167
2162 2168 void G1CollectorPolicy::count_CS_bytes_used() {
2163 2169 CountCSClosure cs_closure(this);
2164 2170 _g1->collection_set_iterate(&cs_closure);
2165 2171 }
2166 2172
2167 2173 void G1CollectorPolicy::print_summary (int level,
2168 2174 const char* str,
2169 2175 NumberSeq* seq) const {
2170 2176 double sum = seq->sum();
2171 2177 LineBuffer(level + 1).append_and_print_cr("%-24s = %8.2lf s (avg = %8.2lf ms)",
2172 2178 str, sum / 1000.0, seq->avg());
2173 2179 }
2174 2180
2175 2181 void G1CollectorPolicy::print_summary_sd (int level,
2176 2182 const char* str,
2177 2183 NumberSeq* seq) const {
2178 2184 print_summary(level, str, seq);
2179 2185 LineBuffer(level + 6).append_and_print_cr("(num = %5d, std dev = %8.2lf ms, max = %8.2lf ms)",
2180 2186 seq->num(), seq->sd(), seq->maximum());
2181 2187 }
2182 2188
2183 2189 void G1CollectorPolicy::check_other_times(int level,
2184 2190 NumberSeq* other_times_ms,
2185 2191 NumberSeq* calc_other_times_ms) const {
2186 2192 bool should_print = false;
2187 2193 LineBuffer buf(level + 2);
2188 2194
2189 2195 double max_sum = MAX2(fabs(other_times_ms->sum()),
2190 2196 fabs(calc_other_times_ms->sum()));
2191 2197 double min_sum = MIN2(fabs(other_times_ms->sum()),
2192 2198 fabs(calc_other_times_ms->sum()));
2193 2199 double sum_ratio = max_sum / min_sum;
2194 2200 if (sum_ratio > 1.1) {
2195 2201 should_print = true;
2196 2202 buf.append_and_print_cr("## CALCULATED OTHER SUM DOESN'T MATCH RECORDED ###");
2197 2203 }
2198 2204
2199 2205 double max_avg = MAX2(fabs(other_times_ms->avg()),
2200 2206 fabs(calc_other_times_ms->avg()));
2201 2207 double min_avg = MIN2(fabs(other_times_ms->avg()),
2202 2208 fabs(calc_other_times_ms->avg()));
2203 2209 double avg_ratio = max_avg / min_avg;
2204 2210 if (avg_ratio > 1.1) {
2205 2211 should_print = true;
2206 2212 buf.append_and_print_cr("## CALCULATED OTHER AVG DOESN'T MATCH RECORDED ###");
2207 2213 }
2208 2214
2209 2215 if (other_times_ms->sum() < -0.01) {
2210 2216 buf.append_and_print_cr("## RECORDED OTHER SUM IS NEGATIVE ###");
2211 2217 }
2212 2218
2213 2219 if (other_times_ms->avg() < -0.01) {
2214 2220 buf.append_and_print_cr("## RECORDED OTHER AVG IS NEGATIVE ###");
2215 2221 }
2216 2222
2217 2223 if (calc_other_times_ms->sum() < -0.01) {
2218 2224 should_print = true;
2219 2225 buf.append_and_print_cr("## CALCULATED OTHER SUM IS NEGATIVE ###");
2220 2226 }
2221 2227
2222 2228 if (calc_other_times_ms->avg() < -0.01) {
2223 2229 should_print = true;
2224 2230 buf.append_and_print_cr("## CALCULATED OTHER AVG IS NEGATIVE ###");
2225 2231 }
2226 2232
2227 2233 if (should_print)
2228 2234 print_summary(level, "Other(Calc)", calc_other_times_ms);
2229 2235 }
2230 2236
2231 2237 void G1CollectorPolicy::print_summary(PauseSummary* summary) const {
2232 2238 bool parallel = G1CollectedHeap::use_parallel_gc_threads();
2233 2239 MainBodySummary* body_summary = summary->main_body_summary();
2234 2240 if (summary->get_total_seq()->num() > 0) {
2235 2241 print_summary_sd(0, "Evacuation Pauses", summary->get_total_seq());
2236 2242 if (body_summary != NULL) {
2237 2243 print_summary(1, "SATB Drain", body_summary->get_satb_drain_seq());
2238 2244 if (parallel) {
2239 2245 print_summary(1, "Parallel Time", body_summary->get_parallel_seq());
2240 2246 print_summary(2, "Update RS", body_summary->get_update_rs_seq());
2241 2247 print_summary(2, "Ext Root Scanning",
2242 2248 body_summary->get_ext_root_scan_seq());
2243 2249 print_summary(2, "Mark Stack Scanning",
2244 2250 body_summary->get_mark_stack_scan_seq());
2245 2251 print_summary(2, "Scan RS", body_summary->get_scan_rs_seq());
2246 2252 print_summary(2, "Object Copy", body_summary->get_obj_copy_seq());
2247 2253 print_summary(2, "Termination", body_summary->get_termination_seq());
2248 2254 print_summary(2, "Other", body_summary->get_parallel_other_seq());
2249 2255 {
2250 2256 NumberSeq* other_parts[] = {
2251 2257 body_summary->get_update_rs_seq(),
2252 2258 body_summary->get_ext_root_scan_seq(),
2253 2259 body_summary->get_mark_stack_scan_seq(),
2254 2260 body_summary->get_scan_rs_seq(),
2255 2261 body_summary->get_obj_copy_seq(),
2256 2262 body_summary->get_termination_seq()
2257 2263 };
2258 2264 NumberSeq calc_other_times_ms(body_summary->get_parallel_seq(),
2259 2265 6, other_parts);
2260 2266 check_other_times(2, body_summary->get_parallel_other_seq(),
2261 2267 &calc_other_times_ms);
2262 2268 }
2263 2269 print_summary(1, "Mark Closure", body_summary->get_mark_closure_seq());
2264 2270 print_summary(1, "Clear CT", body_summary->get_clear_ct_seq());
2265 2271 } else {
2266 2272 print_summary(1, "Update RS", body_summary->get_update_rs_seq());
2267 2273 print_summary(1, "Ext Root Scanning",
2268 2274 body_summary->get_ext_root_scan_seq());
2269 2275 print_summary(1, "Mark Stack Scanning",
2270 2276 body_summary->get_mark_stack_scan_seq());
2271 2277 print_summary(1, "Scan RS", body_summary->get_scan_rs_seq());
2272 2278 print_summary(1, "Object Copy", body_summary->get_obj_copy_seq());
2273 2279 }
2274 2280 }
2275 2281 print_summary(1, "Other", summary->get_other_seq());
2276 2282 {
2277 2283 if (body_summary != NULL) {
2278 2284 NumberSeq calc_other_times_ms;
2279 2285 if (parallel) {
2280 2286 // parallel
2281 2287 NumberSeq* other_parts[] = {
2282 2288 body_summary->get_satb_drain_seq(),
2283 2289 body_summary->get_parallel_seq(),
2284 2290 body_summary->get_clear_ct_seq()
2285 2291 };
2286 2292 calc_other_times_ms = NumberSeq(summary->get_total_seq(),
2287 2293 3, other_parts);
2288 2294 } else {
2289 2295 // serial
2290 2296 NumberSeq* other_parts[] = {
2291 2297 body_summary->get_satb_drain_seq(),
2292 2298 body_summary->get_update_rs_seq(),
2293 2299 body_summary->get_ext_root_scan_seq(),
2294 2300 body_summary->get_mark_stack_scan_seq(),
2295 2301 body_summary->get_scan_rs_seq(),
2296 2302 body_summary->get_obj_copy_seq()
2297 2303 };
2298 2304 calc_other_times_ms = NumberSeq(summary->get_total_seq(),
2299 2305 6, other_parts);
2300 2306 }
2301 2307 check_other_times(1, summary->get_other_seq(), &calc_other_times_ms);
2302 2308 }
2303 2309 }
2304 2310 } else {
2305 2311 LineBuffer(1).append_and_print_cr("none");
2306 2312 }
2307 2313 LineBuffer(0).append_and_print_cr("");
2308 2314 }
2309 2315
2310 2316 void G1CollectorPolicy::print_tracing_info() const {
2311 2317 if (TraceGen0Time) {
2312 2318 gclog_or_tty->print_cr("ALL PAUSES");
2313 2319 print_summary_sd(0, "Total", _all_pause_times_ms);
2314 2320 gclog_or_tty->print_cr("");
2315 2321 gclog_or_tty->print_cr("");
2316 2322 gclog_or_tty->print_cr(" Full Young GC Pauses: %8d", _full_young_pause_num);
2317 2323 gclog_or_tty->print_cr(" Partial Young GC Pauses: %8d", _partial_young_pause_num);
2318 2324 gclog_or_tty->print_cr("");
2319 2325
2320 2326 gclog_or_tty->print_cr("EVACUATION PAUSES");
2321 2327 print_summary(_summary);
2322 2328
2323 2329 gclog_or_tty->print_cr("MISC");
2324 2330 print_summary_sd(0, "Stop World", _all_stop_world_times_ms);
2325 2331 print_summary_sd(0, "Yields", _all_yield_times_ms);
2326 2332 for (int i = 0; i < _aux_num; ++i) {
2327 2333 if (_all_aux_times_ms[i].num() > 0) {
2328 2334 char buffer[96];
2329 2335 sprintf(buffer, "Aux%d", i);
2330 2336 print_summary_sd(0, buffer, &_all_aux_times_ms[i]);
2331 2337 }
2332 2338 }
2333 2339
2334 2340 size_t all_region_num = _region_num_young + _region_num_tenured;
2335 2341 gclog_or_tty->print_cr(" New Regions %8d, Young %8d (%6.2lf%%), "
2336 2342 "Tenured %8d (%6.2lf%%)",
2337 2343 all_region_num,
2338 2344 _region_num_young,
2339 2345 (double) _region_num_young / (double) all_region_num * 100.0,
2340 2346 _region_num_tenured,
2341 2347 (double) _region_num_tenured / (double) all_region_num * 100.0);
2342 2348 }
2343 2349 if (TraceGen1Time) {
2344 2350 if (_all_full_gc_times_ms->num() > 0) {
2345 2351 gclog_or_tty->print("\n%4d full_gcs: total time = %8.2f s",
2346 2352 _all_full_gc_times_ms->num(),
2347 2353 _all_full_gc_times_ms->sum() / 1000.0);
2348 2354 gclog_or_tty->print_cr(" (avg = %8.2fms).", _all_full_gc_times_ms->avg());
2349 2355 gclog_or_tty->print_cr(" [std. dev = %8.2f ms, max = %8.2f ms]",
2350 2356 _all_full_gc_times_ms->sd(),
2351 2357 _all_full_gc_times_ms->maximum());
2352 2358 }
2353 2359 }
2354 2360 }
2355 2361
2356 2362 void G1CollectorPolicy::print_yg_surv_rate_info() const {
2357 2363 #ifndef PRODUCT
2358 2364 _short_lived_surv_rate_group->print_surv_rate_summary();
2359 2365 // add this call for any other surv rate groups
2360 2366 #endif // PRODUCT
2361 2367 }
2362 2368
2363 2369 void G1CollectorPolicy::update_region_num(bool young) {
2364 2370 if (young) {
2365 2371 ++_region_num_young;
2366 2372 } else {
2367 2373 ++_region_num_tenured;
2368 2374 }
2369 2375 }
2370 2376
2371 2377 #ifndef PRODUCT
2372 2378 // for debugging, bit of a hack...
2373 2379 static char*
2374 2380 region_num_to_mbs(int length) {
2375 2381 static char buffer[64];
2376 2382 double bytes = (double) (length * HeapRegion::GrainBytes);
2377 2383 double mbs = bytes / (double) (1024 * 1024);
2378 2384 sprintf(buffer, "%7.2lfMB", mbs);
2379 2385 return buffer;
2380 2386 }
2381 2387 #endif // PRODUCT
2382 2388
2383 2389 size_t G1CollectorPolicy::max_regions(int purpose) {
2384 2390 switch (purpose) {
2385 2391 case GCAllocForSurvived:
2386 2392 return _max_survivor_regions;
2387 2393 case GCAllocForTenured:
2388 2394 return REGIONS_UNLIMITED;
2389 2395 default:
2390 2396 ShouldNotReachHere();
2391 2397 return REGIONS_UNLIMITED;
2392 2398 };
2393 2399 }
2394 2400
2395 2401 void G1CollectorPolicy::update_max_gc_locker_expansion() {
2396 2402 size_t expansion_region_num = 0;
2397 2403 if (GCLockerEdenExpansionPercent > 0) {
2398 2404 double perc = (double) GCLockerEdenExpansionPercent / 100.0;
2399 2405 double expansion_region_num_d = perc * (double) _young_list_target_length;
2400 2406 // We use ceiling so that if expansion_region_num_d is > 0.0 (but
2401 2407 // less than 1.0) we'll get 1.
2402 2408 expansion_region_num = (size_t) ceil(expansion_region_num_d);
2403 2409 } else {
2404 2410 assert(expansion_region_num == 0, "sanity");
2405 2411 }
2406 2412 _young_list_max_length = _young_list_target_length + expansion_region_num;
2407 2413 assert(_young_list_target_length <= _young_list_max_length, "post-condition");
2408 2414 }
2409 2415
2410 2416 // Calculates survivor space parameters.
2411 2417 void G1CollectorPolicy::update_survivors_policy() {
2412 2418 double max_survivor_regions_d =
2413 2419 (double) _young_list_target_length / (double) SurvivorRatio;
2414 2420 // We use ceiling so that if max_survivor_regions_d is > 0.0 (but
2415 2421 // smaller than 1.0) we'll get 1.
2416 2422 _max_survivor_regions = (size_t) ceil(max_survivor_regions_d);
2417 2423
2418 2424 _tenuring_threshold = _survivors_age_table.compute_tenuring_threshold(
2419 2425 HeapRegion::GrainWords * _max_survivor_regions);
2420 2426 }
2421 2427
2422 2428 #ifndef PRODUCT
2423 2429 class HRSortIndexIsOKClosure: public HeapRegionClosure {
2424 2430 CollectionSetChooser* _chooser;
2425 2431 public:
2426 2432 HRSortIndexIsOKClosure(CollectionSetChooser* chooser) :
2427 2433 _chooser(chooser) {}
2428 2434
2429 2435 bool doHeapRegion(HeapRegion* r) {
2430 2436 if (!r->continuesHumongous()) {
2431 2437 assert(_chooser->regionProperlyOrdered(r), "Ought to be.");
2432 2438 }
2433 2439 return false;
2434 2440 }
2435 2441 };
2436 2442
2437 2443 bool G1CollectorPolicy_BestRegionsFirst::assertMarkedBytesDataOK() {
2438 2444 HRSortIndexIsOKClosure cl(_collectionSetChooser);
2439 2445 _g1->heap_region_iterate(&cl);
2440 2446 return true;
2441 2447 }
2442 2448 #endif
2443 2449
2444 2450 bool G1CollectorPolicy::force_initial_mark_if_outside_cycle(
2445 2451 GCCause::Cause gc_cause) {
2446 2452 bool during_cycle = _g1->concurrent_mark()->cmThread()->during_cycle();
2447 2453 if (!during_cycle) {
2448 2454 ergo_verbose1(ErgoConcCycles,
2449 2455 "request concurrent cycle initiation",
2450 2456 ergo_format_reason("requested by GC cause")
2451 2457 ergo_format_str("GC cause"),
2452 2458 GCCause::to_string(gc_cause));
2453 2459 set_initiate_conc_mark_if_possible();
2454 2460 return true;
2455 2461 } else {
2456 2462 ergo_verbose1(ErgoConcCycles,
2457 2463 "do not request concurrent cycle initiation",
2458 2464 ergo_format_reason("concurrent cycle already in progress")
2459 2465 ergo_format_str("GC cause"),
2460 2466 GCCause::to_string(gc_cause));
2461 2467 return false;
2462 2468 }
2463 2469 }
2464 2470
2465 2471 void
2466 2472 G1CollectorPolicy::decide_on_conc_mark_initiation() {
2467 2473 // We are about to decide on whether this pause will be an
2468 2474 // initial-mark pause.
2469 2475
2470 2476 // First, during_initial_mark_pause() should not be already set. We
2471 2477 // will set it here if we have to. However, it should be cleared by
2472 2478 // the end of the pause (it's only set for the duration of an
2473 2479 // initial-mark pause).
2474 2480 assert(!during_initial_mark_pause(), "pre-condition");
2475 2481
2476 2482 if (initiate_conc_mark_if_possible()) {
2477 2483 // We had noticed on a previous pause that the heap occupancy has
2478 2484 // gone over the initiating threshold and we should start a
2479 2485 // concurrent marking cycle. So we might initiate one.
2480 2486
2481 2487 bool during_cycle = _g1->concurrent_mark()->cmThread()->during_cycle();
2482 2488 if (!during_cycle) {
2483 2489 // The concurrent marking thread is not "during a cycle", i.e.,
2484 2490 // it has completed the last one. So we can go ahead and
2485 2491 // initiate a new cycle.
2486 2492
2487 2493 set_during_initial_mark_pause();
2488 2494
2489 2495 // And we can now clear initiate_conc_mark_if_possible() as
2490 2496 // we've already acted on it.
2491 2497 clear_initiate_conc_mark_if_possible();
2492 2498
2493 2499 ergo_verbose0(ErgoConcCycles,
2494 2500 "initiate concurrent cycle",
2495 2501 ergo_format_reason("concurrent cycle initiation requested"));
2496 2502 } else {
2497 2503 // The concurrent marking thread is still finishing up the
2498 2504 // previous cycle. If we start one right now the two cycles
2499 2505 // overlap. In particular, the concurrent marking thread might
2500 2506 // be in the process of clearing the next marking bitmap (which
2501 2507 // we will use for the next cycle if we start one). Starting a
2502 2508 // cycle now will be bad given that parts of the marking
2503 2509 // information might get cleared by the marking thread. And we
2504 2510 // cannot wait for the marking thread to finish the cycle as it
2505 2511 // periodically yields while clearing the next marking bitmap
2506 2512 // and, if it's in a yield point, it's waiting for us to
2507 2513 // finish. So, at this point we will not start a cycle and we'll
2508 2514 // let the concurrent marking thread complete the last one.
2509 2515 ergo_verbose0(ErgoConcCycles,
2510 2516 "do not initiate concurrent cycle",
2511 2517 ergo_format_reason("concurrent cycle already in progress"));
2512 2518 }
2513 2519 }
2514 2520 }
2515 2521
2516 2522 void
2517 2523 G1CollectorPolicy_BestRegionsFirst::
2518 2524 record_collection_pause_start(double start_time_sec, size_t start_used) {
2519 2525 G1CollectorPolicy::record_collection_pause_start(start_time_sec, start_used);
2520 2526 }
2521 2527
2522 2528 class KnownGarbageClosure: public HeapRegionClosure {
2523 2529 CollectionSetChooser* _hrSorted;
2524 2530
2525 2531 public:
2526 2532 KnownGarbageClosure(CollectionSetChooser* hrSorted) :
2527 2533 _hrSorted(hrSorted)
2528 2534 {}
2529 2535
2530 2536 bool doHeapRegion(HeapRegion* r) {
2531 2537 // We only include humongous regions in collection
2532 2538 // sets when concurrent mark shows that their contained object is
2533 2539 // unreachable.
2534 2540
2535 2541 // Do we have any marking information for this region?
2536 2542 if (r->is_marked()) {
2537 2543 // We don't include humongous regions in collection
2538 2544 // sets because we collect them immediately at the end of a marking
2539 2545 // cycle. We also don't include young regions because we *must*
2540 2546 // include them in the next collection pause.
2541 2547 if (!r->isHumongous() && !r->is_young()) {
2542 2548 _hrSorted->addMarkedHeapRegion(r);
2543 2549 }
2544 2550 }
2545 2551 return false;
2546 2552 }
2547 2553 };
2548 2554
2549 2555 class ParKnownGarbageHRClosure: public HeapRegionClosure {
2550 2556 CollectionSetChooser* _hrSorted;
2551 2557 jint _marked_regions_added;
2552 2558 jint _chunk_size;
2553 2559 jint _cur_chunk_idx;
2554 2560 jint _cur_chunk_end; // Cur chunk [_cur_chunk_idx, _cur_chunk_end)
2555 2561 int _worker;
2556 2562 int _invokes;
2557 2563
2558 2564 void get_new_chunk() {
2559 2565 _cur_chunk_idx = _hrSorted->getParMarkedHeapRegionChunk(_chunk_size);
2560 2566 _cur_chunk_end = _cur_chunk_idx + _chunk_size;
2561 2567 }
2562 2568 void add_region(HeapRegion* r) {
2563 2569 if (_cur_chunk_idx == _cur_chunk_end) {
2564 2570 get_new_chunk();
2565 2571 }
2566 2572 assert(_cur_chunk_idx < _cur_chunk_end, "postcondition");
2567 2573 _hrSorted->setMarkedHeapRegion(_cur_chunk_idx, r);
2568 2574 _marked_regions_added++;
2569 2575 _cur_chunk_idx++;
2570 2576 }
2571 2577
2572 2578 public:
2573 2579 ParKnownGarbageHRClosure(CollectionSetChooser* hrSorted,
2574 2580 jint chunk_size,
2575 2581 int worker) :
2576 2582 _hrSorted(hrSorted), _chunk_size(chunk_size), _worker(worker),
2577 2583 _marked_regions_added(0), _cur_chunk_idx(0), _cur_chunk_end(0),
2578 2584 _invokes(0)
2579 2585 {}
2580 2586
2581 2587 bool doHeapRegion(HeapRegion* r) {
2582 2588 // We only include humongous regions in collection
2583 2589 // sets when concurrent mark shows that their contained object is
2584 2590 // unreachable.
2585 2591 _invokes++;
2586 2592
2587 2593 // Do we have any marking information for this region?
2588 2594 if (r->is_marked()) {
2589 2595 // We don't include humongous regions in collection
2590 2596 // sets because we collect them immediately at the end of a marking
2591 2597 // cycle.
2592 2598 // We also do not include young regions in collection sets
2593 2599 if (!r->isHumongous() && !r->is_young()) {
2594 2600 add_region(r);
2595 2601 }
2596 2602 }
2597 2603 return false;
2598 2604 }
2599 2605 jint marked_regions_added() { return _marked_regions_added; }
2600 2606 int invokes() { return _invokes; }
2601 2607 };
2602 2608
2603 2609 class ParKnownGarbageTask: public AbstractGangTask {
2604 2610 CollectionSetChooser* _hrSorted;
2605 2611 jint _chunk_size;
2606 2612 G1CollectedHeap* _g1;
2607 2613 public:
2608 2614 ParKnownGarbageTask(CollectionSetChooser* hrSorted, jint chunk_size) :
2609 2615 AbstractGangTask("ParKnownGarbageTask"),
2610 2616 _hrSorted(hrSorted), _chunk_size(chunk_size),
2611 2617 _g1(G1CollectedHeap::heap())
2612 2618 {}
2613 2619
2614 2620 void work(int i) {
2615 2621 ParKnownGarbageHRClosure parKnownGarbageCl(_hrSorted, _chunk_size, i);
2616 2622 // Back to zero for the claim value.
2617 2623 _g1->heap_region_par_iterate_chunked(&parKnownGarbageCl, i,
2618 2624 HeapRegion::InitialClaimValue);
2619 2625 jint regions_added = parKnownGarbageCl.marked_regions_added();
2620 2626 _hrSorted->incNumMarkedHeapRegions(regions_added);
2621 2627 if (G1PrintParCleanupStats) {
2622 2628 gclog_or_tty->print_cr(" Thread %d called %d times, added %d regions to list.",
2623 2629 i, parKnownGarbageCl.invokes(), regions_added);
2624 2630 }
2625 2631 }
2626 2632 };
2627 2633
2628 2634 void
2629 2635 G1CollectorPolicy_BestRegionsFirst::
2630 2636 record_concurrent_mark_cleanup_end(size_t freed_bytes,
2631 2637 size_t max_live_bytes) {
2632 2638 double start;
2633 2639 if (G1PrintParCleanupStats) start = os::elapsedTime();
2634 2640 record_concurrent_mark_cleanup_end_work1(freed_bytes, max_live_bytes);
2635 2641
2636 2642 _collectionSetChooser->clearMarkedHeapRegions();
2637 2643 double clear_marked_end;
2638 2644 if (G1PrintParCleanupStats) {
2639 2645 clear_marked_end = os::elapsedTime();
2640 2646 gclog_or_tty->print_cr(" clear marked regions + work1: %8.3f ms.",
2641 2647 (clear_marked_end - start)*1000.0);
2642 2648 }
2643 2649 if (G1CollectedHeap::use_parallel_gc_threads()) {
2644 2650 const size_t OverpartitionFactor = 4;
2645 2651 const size_t MinWorkUnit = 8;
2646 2652 const size_t WorkUnit =
2647 2653 MAX2(_g1->n_regions() / (ParallelGCThreads * OverpartitionFactor),
2648 2654 MinWorkUnit);
2649 2655 _collectionSetChooser->prepareForAddMarkedHeapRegionsPar(_g1->n_regions(),
2650 2656 WorkUnit);
2651 2657 ParKnownGarbageTask parKnownGarbageTask(_collectionSetChooser,
2652 2658 (int) WorkUnit);
2653 2659 _g1->workers()->run_task(&parKnownGarbageTask);
2654 2660
2655 2661 assert(_g1->check_heap_region_claim_values(HeapRegion::InitialClaimValue),
2656 2662 "sanity check");
2657 2663 } else {
2658 2664 KnownGarbageClosure knownGarbagecl(_collectionSetChooser);
2659 2665 _g1->heap_region_iterate(&knownGarbagecl);
2660 2666 }
2661 2667 double known_garbage_end;
2662 2668 if (G1PrintParCleanupStats) {
2663 2669 known_garbage_end = os::elapsedTime();
2664 2670 gclog_or_tty->print_cr(" compute known garbage: %8.3f ms.",
2665 2671 (known_garbage_end - clear_marked_end)*1000.0);
2666 2672 }
2667 2673 _collectionSetChooser->sortMarkedHeapRegions();
2668 2674 double sort_end;
2669 2675 if (G1PrintParCleanupStats) {
2670 2676 sort_end = os::elapsedTime();
2671 2677 gclog_or_tty->print_cr(" sorting: %8.3f ms.",
2672 2678 (sort_end - known_garbage_end)*1000.0);
2673 2679 }
2674 2680
2675 2681 record_concurrent_mark_cleanup_end_work2();
2676 2682 double work2_end;
2677 2683 if (G1PrintParCleanupStats) {
2678 2684 work2_end = os::elapsedTime();
2679 2685 gclog_or_tty->print_cr(" work2: %8.3f ms.",
2680 2686 (work2_end - sort_end)*1000.0);
2681 2687 }
2682 2688 }
2683 2689
2684 2690 // Add the heap region at the head of the non-incremental collection set
2685 2691 void G1CollectorPolicy::
2686 2692 add_to_collection_set(HeapRegion* hr) {
2687 2693 assert(_inc_cset_build_state == Active, "Precondition");
2688 2694 assert(!hr->is_young(), "non-incremental add of young region");
2689 2695
2690 2696 if (_g1->mark_in_progress())
2691 2697 _g1->concurrent_mark()->registerCSetRegion(hr);
2692 2698
2693 2699 assert(!hr->in_collection_set(), "should not already be in the CSet");
2694 2700 hr->set_in_collection_set(true);
2695 2701 hr->set_next_in_collection_set(_collection_set);
2696 2702 _collection_set = hr;
2697 2703 _collection_set_size++;
2698 2704 _collection_set_bytes_used_before += hr->used();
2699 2705 _g1->register_region_with_in_cset_fast_test(hr);
2700 2706 }
2701 2707
2702 2708 // Initialize the per-collection-set information
2703 2709 void G1CollectorPolicy::start_incremental_cset_building() {
2704 2710 assert(_inc_cset_build_state == Inactive, "Precondition");
2705 2711
2706 2712 _inc_cset_head = NULL;
2707 2713 _inc_cset_tail = NULL;
2708 2714 _inc_cset_size = 0;
2709 2715 _inc_cset_bytes_used_before = 0;
2710 2716
2711 2717 _inc_cset_young_index = 0;
2712 2718
2713 2719 _inc_cset_max_finger = 0;
2714 2720 _inc_cset_recorded_young_bytes = 0;
2715 2721 _inc_cset_recorded_rs_lengths = 0;
2716 2722 _inc_cset_predicted_elapsed_time_ms = 0;
2717 2723 _inc_cset_predicted_bytes_to_copy = 0;
2718 2724 _inc_cset_build_state = Active;
2719 2725 }
2720 2726
2721 2727 void G1CollectorPolicy::add_to_incremental_cset_info(HeapRegion* hr, size_t rs_length) {
2722 2728 // This routine is used when:
2723 2729 // * adding survivor regions to the incremental cset at the end of an
2724 2730 // evacuation pause,
2725 2731 // * adding the current allocation region to the incremental cset
2726 2732 // when it is retired, and
2727 2733 // * updating existing policy information for a region in the
2728 2734 // incremental cset via young list RSet sampling.
2729 2735 // Therefore this routine may be called at a safepoint by the
2730 2736 // VM thread, or in-between safepoints by mutator threads (when
2731 2737 // retiring the current allocation region) or a concurrent
2732 2738 // refine thread (RSet sampling).
2733 2739
2734 2740 double region_elapsed_time_ms = predict_region_elapsed_time_ms(hr, true);
2735 2741 size_t used_bytes = hr->used();
2736 2742
2737 2743 _inc_cset_recorded_rs_lengths += rs_length;
2738 2744 _inc_cset_predicted_elapsed_time_ms += region_elapsed_time_ms;
2739 2745
2740 2746 _inc_cset_bytes_used_before += used_bytes;
2741 2747
2742 2748 // Cache the values we have added to the aggregated informtion
2743 2749 // in the heap region in case we have to remove this region from
2744 2750 // the incremental collection set, or it is updated by the
2745 2751 // rset sampling code
2746 2752 hr->set_recorded_rs_length(rs_length);
2747 2753 hr->set_predicted_elapsed_time_ms(region_elapsed_time_ms);
2748 2754
2749 2755 #if PREDICTIONS_VERBOSE
2750 2756 size_t bytes_to_copy = predict_bytes_to_copy(hr);
2751 2757 _inc_cset_predicted_bytes_to_copy += bytes_to_copy;
2752 2758
2753 2759 // Record the number of bytes used in this region
2754 2760 _inc_cset_recorded_young_bytes += used_bytes;
2755 2761
2756 2762 // Cache the values we have added to the aggregated informtion
2757 2763 // in the heap region in case we have to remove this region from
2758 2764 // the incremental collection set, or it is updated by the
2759 2765 // rset sampling code
2760 2766 hr->set_predicted_bytes_to_copy(bytes_to_copy);
2761 2767 #endif // PREDICTIONS_VERBOSE
2762 2768 }
2763 2769
2764 2770 void G1CollectorPolicy::remove_from_incremental_cset_info(HeapRegion* hr) {
2765 2771 // This routine is currently only called as part of the updating of
2766 2772 // existing policy information for regions in the incremental cset that
2767 2773 // is performed by the concurrent refine thread(s) as part of young list
2768 2774 // RSet sampling. Therefore we should not be at a safepoint.
2769 2775
2770 2776 assert(!SafepointSynchronize::is_at_safepoint(), "should not be at safepoint");
2771 2777 assert(hr->is_young(), "it should be");
2772 2778
2773 2779 size_t used_bytes = hr->used();
2774 2780 size_t old_rs_length = hr->recorded_rs_length();
2775 2781 double old_elapsed_time_ms = hr->predicted_elapsed_time_ms();
2776 2782
2777 2783 // Subtract the old recorded/predicted policy information for
2778 2784 // the given heap region from the collection set info.
2779 2785 _inc_cset_recorded_rs_lengths -= old_rs_length;
2780 2786 _inc_cset_predicted_elapsed_time_ms -= old_elapsed_time_ms;
2781 2787
2782 2788 _inc_cset_bytes_used_before -= used_bytes;
2783 2789
2784 2790 // Clear the values cached in the heap region
2785 2791 hr->set_recorded_rs_length(0);
2786 2792 hr->set_predicted_elapsed_time_ms(0);
2787 2793
2788 2794 #if PREDICTIONS_VERBOSE
2789 2795 size_t old_predicted_bytes_to_copy = hr->predicted_bytes_to_copy();
2790 2796 _inc_cset_predicted_bytes_to_copy -= old_predicted_bytes_to_copy;
2791 2797
2792 2798 // Subtract the number of bytes used in this region
2793 2799 _inc_cset_recorded_young_bytes -= used_bytes;
2794 2800
2795 2801 // Clear the values cached in the heap region
2796 2802 hr->set_predicted_bytes_to_copy(0);
2797 2803 #endif // PREDICTIONS_VERBOSE
2798 2804 }
2799 2805
2800 2806 void G1CollectorPolicy::update_incremental_cset_info(HeapRegion* hr, size_t new_rs_length) {
2801 2807 // Update the collection set information that is dependent on the new RS length
2802 2808 assert(hr->is_young(), "Precondition");
2803 2809
2804 2810 remove_from_incremental_cset_info(hr);
2805 2811 add_to_incremental_cset_info(hr, new_rs_length);
2806 2812 }
2807 2813
2808 2814 void G1CollectorPolicy::add_region_to_incremental_cset_common(HeapRegion* hr) {
2809 2815 assert( hr->is_young(), "invariant");
2810 2816 assert( hr->young_index_in_cset() == -1, "invariant" );
2811 2817 assert(_inc_cset_build_state == Active, "Precondition");
2812 2818
2813 2819 // We need to clear and set the cached recorded/cached collection set
2814 2820 // information in the heap region here (before the region gets added
2815 2821 // to the collection set). An individual heap region's cached values
2816 2822 // are calculated, aggregated with the policy collection set info,
2817 2823 // and cached in the heap region here (initially) and (subsequently)
2818 2824 // by the Young List sampling code.
2819 2825
2820 2826 size_t rs_length = hr->rem_set()->occupied();
2821 2827 add_to_incremental_cset_info(hr, rs_length);
2822 2828
2823 2829 HeapWord* hr_end = hr->end();
2824 2830 _inc_cset_max_finger = MAX2(_inc_cset_max_finger, hr_end);
2825 2831
2826 2832 assert(!hr->in_collection_set(), "invariant");
2827 2833 hr->set_in_collection_set(true);
2828 2834 assert( hr->next_in_collection_set() == NULL, "invariant");
2829 2835
2830 2836 _inc_cset_size++;
2831 2837 _g1->register_region_with_in_cset_fast_test(hr);
2832 2838
2833 2839 hr->set_young_index_in_cset((int) _inc_cset_young_index);
2834 2840 ++_inc_cset_young_index;
2835 2841 }
2836 2842
2837 2843 // Add the region at the RHS of the incremental cset
2838 2844 void G1CollectorPolicy::add_region_to_incremental_cset_rhs(HeapRegion* hr) {
2839 2845 // We should only ever be appending survivors at the end of a pause
2840 2846 assert( hr->is_survivor(), "Logic");
2841 2847
2842 2848 // Do the 'common' stuff
2843 2849 add_region_to_incremental_cset_common(hr);
2844 2850
2845 2851 // Now add the region at the right hand side
2846 2852 if (_inc_cset_tail == NULL) {
2847 2853 assert(_inc_cset_head == NULL, "invariant");
2848 2854 _inc_cset_head = hr;
2849 2855 } else {
2850 2856 _inc_cset_tail->set_next_in_collection_set(hr);
2851 2857 }
2852 2858 _inc_cset_tail = hr;
2853 2859 }
2854 2860
2855 2861 // Add the region to the LHS of the incremental cset
2856 2862 void G1CollectorPolicy::add_region_to_incremental_cset_lhs(HeapRegion* hr) {
2857 2863 // Survivors should be added to the RHS at the end of a pause
2858 2864 assert(!hr->is_survivor(), "Logic");
2859 2865
2860 2866 // Do the 'common' stuff
2861 2867 add_region_to_incremental_cset_common(hr);
2862 2868
2863 2869 // Add the region at the left hand side
2864 2870 hr->set_next_in_collection_set(_inc_cset_head);
2865 2871 if (_inc_cset_head == NULL) {
2866 2872 assert(_inc_cset_tail == NULL, "Invariant");
2867 2873 _inc_cset_tail = hr;
2868 2874 }
2869 2875 _inc_cset_head = hr;
2870 2876 }
2871 2877
2872 2878 #ifndef PRODUCT
2873 2879 void G1CollectorPolicy::print_collection_set(HeapRegion* list_head, outputStream* st) {
2874 2880 assert(list_head == inc_cset_head() || list_head == collection_set(), "must be");
2875 2881
2876 2882 st->print_cr("\nCollection_set:");
2877 2883 HeapRegion* csr = list_head;
2878 2884 while (csr != NULL) {
2879 2885 HeapRegion* next = csr->next_in_collection_set();
2880 2886 assert(csr->in_collection_set(), "bad CS");
2881 2887 st->print_cr(" [%08x-%08x], t: %08x, P: %08x, N: %08x, C: %08x, "
2882 2888 "age: %4d, y: %d, surv: %d",
2883 2889 csr->bottom(), csr->end(),
2884 2890 csr->top(),
2885 2891 csr->prev_top_at_mark_start(),
2886 2892 csr->next_top_at_mark_start(),
2887 2893 csr->top_at_conc_mark_count(),
2888 2894 csr->age_in_surv_rate_group_cond(),
2889 2895 csr->is_young(),
2890 2896 csr->is_survivor());
2891 2897 csr = next;
2892 2898 }
2893 2899 }
2894 2900 #endif // !PRODUCT
2895 2901
2896 2902 void
2897 2903 G1CollectorPolicy_BestRegionsFirst::choose_collection_set(
2898 2904 double target_pause_time_ms) {
2899 2905 // Set this here - in case we're not doing young collections.
2900 2906 double non_young_start_time_sec = os::elapsedTime();
2901 2907
2902 2908 YoungList* young_list = _g1->young_list();
2903 2909
2904 2910 start_recording_regions();
2905 2911
2906 2912 guarantee(target_pause_time_ms > 0.0,
2907 2913 err_msg("target_pause_time_ms = %1.6lf should be positive",
2908 2914 target_pause_time_ms));
2909 2915 guarantee(_collection_set == NULL, "Precondition");
2910 2916
2911 2917 double base_time_ms = predict_base_elapsed_time_ms(_pending_cards);
2912 2918 double predicted_pause_time_ms = base_time_ms;
2913 2919
2914 2920 double time_remaining_ms = target_pause_time_ms - base_time_ms;
2915 2921
2916 2922 ergo_verbose3(ErgoCSetConstruction | ErgoHigh,
2917 2923 "start choosing CSet",
2918 2924 ergo_format_ms("predicted base time")
2919 2925 ergo_format_ms("remaining time")
2920 2926 ergo_format_ms("target pause time"),
2921 2927 base_time_ms, time_remaining_ms, target_pause_time_ms);
2922 2928
2923 2929 // the 10% and 50% values are arbitrary...
2924 2930 double threshold = 0.10 * target_pause_time_ms;
2925 2931 if (time_remaining_ms < threshold) {
2926 2932 double prev_time_remaining_ms = time_remaining_ms;
2927 2933 time_remaining_ms = 0.50 * target_pause_time_ms;
2928 2934 _within_target = false;
2929 2935 ergo_verbose3(ErgoCSetConstruction,
2930 2936 "adjust remaining time",
2931 2937 ergo_format_reason("remaining time lower than threshold")
2932 2938 ergo_format_ms("remaining time")
2933 2939 ergo_format_ms("threshold")
2934 2940 ergo_format_ms("adjusted remaining time"),
2935 2941 prev_time_remaining_ms, threshold, time_remaining_ms);
2936 2942 } else {
2937 2943 _within_target = true;
2938 2944 }
2939 2945
2940 2946 size_t expansion_bytes = _g1->expansion_regions() * HeapRegion::GrainBytes;
2941 2947
2942 2948 HeapRegion* hr;
2943 2949 double young_start_time_sec = os::elapsedTime();
2944 2950
2945 2951 _collection_set_bytes_used_before = 0;
2946 2952 _collection_set_size = 0;
2947 2953 _young_cset_length = 0;
2948 2954 _last_young_gc_full = full_young_gcs() ? true : false;
2949 2955
2950 2956 if (_last_young_gc_full) {
2951 2957 ++_full_young_pause_num;
2952 2958 } else {
2953 2959 ++_partial_young_pause_num;
2954 2960 }
2955 2961
2956 2962 // The young list is laid with the survivor regions from the previous
2957 2963 // pause are appended to the RHS of the young list, i.e.
2958 2964 // [Newly Young Regions ++ Survivors from last pause].
2959 2965
2960 2966 size_t survivor_region_num = young_list->survivor_length();
2961 2967 size_t eden_region_num = young_list->length() - survivor_region_num;
2962 2968 size_t old_region_num = 0;
2963 2969 hr = young_list->first_survivor_region();
2964 2970 while (hr != NULL) {
2965 2971 assert(hr->is_survivor(), "badly formed young list");
2966 2972 hr->set_young();
2967 2973 hr = hr->get_next_young_region();
2968 2974 }
2969 2975
2970 2976 // Clear the fields that point to the survivor list - they are all young now.
2971 2977 young_list->clear_survivors();
2972 2978
2973 2979 if (_g1->mark_in_progress())
2974 2980 _g1->concurrent_mark()->register_collection_set_finger(_inc_cset_max_finger);
2975 2981
2976 2982 _young_cset_length = _inc_cset_young_index;
2977 2983 _collection_set = _inc_cset_head;
2978 2984 _collection_set_size = _inc_cset_size;
2979 2985 _collection_set_bytes_used_before = _inc_cset_bytes_used_before;
2980 2986 time_remaining_ms -= _inc_cset_predicted_elapsed_time_ms;
2981 2987 predicted_pause_time_ms += _inc_cset_predicted_elapsed_time_ms;
2982 2988
2983 2989 ergo_verbose3(ErgoCSetConstruction | ErgoHigh,
2984 2990 "add young regions to CSet",
2985 2991 ergo_format_region("eden")
2986 2992 ergo_format_region("survivors")
2987 2993 ergo_format_ms("predicted young region time"),
2988 2994 eden_region_num, survivor_region_num,
2989 2995 _inc_cset_predicted_elapsed_time_ms);
2990 2996
2991 2997 // The number of recorded young regions is the incremental
2992 2998 // collection set's current size
2993 2999 set_recorded_young_regions(_inc_cset_size);
2994 3000 set_recorded_rs_lengths(_inc_cset_recorded_rs_lengths);
2995 3001 set_recorded_young_bytes(_inc_cset_recorded_young_bytes);
2996 3002 #if PREDICTIONS_VERBOSE
2997 3003 set_predicted_bytes_to_copy(_inc_cset_predicted_bytes_to_copy);
2998 3004 #endif // PREDICTIONS_VERBOSE
2999 3005
3000 3006 assert(_inc_cset_size == young_list->length(), "Invariant");
3001 3007
3002 3008 double young_end_time_sec = os::elapsedTime();
3003 3009 _recorded_young_cset_choice_time_ms =
3004 3010 (young_end_time_sec - young_start_time_sec) * 1000.0;
3005 3011
3006 3012 // We are doing young collections so reset this.
3007 3013 non_young_start_time_sec = young_end_time_sec;
3008 3014
3009 3015 if (!full_young_gcs()) {
3010 3016 bool should_continue = true;
3011 3017 NumberSeq seq;
3012 3018 double avg_prediction = 100000000000000000.0; // something very large
3013 3019
3014 3020 size_t prev_collection_set_size = _collection_set_size;
3015 3021 double prev_predicted_pause_time_ms = predicted_pause_time_ms;
3016 3022 do {
3017 3023 hr = _collectionSetChooser->getNextMarkedRegion(time_remaining_ms,
3018 3024 avg_prediction);
3019 3025 if (hr != NULL) {
3020 3026 double predicted_time_ms = predict_region_elapsed_time_ms(hr, false);
3021 3027 time_remaining_ms -= predicted_time_ms;
3022 3028 predicted_pause_time_ms += predicted_time_ms;
3023 3029 add_to_collection_set(hr);
3024 3030 record_non_young_cset_region(hr);
3025 3031 seq.add(predicted_time_ms);
3026 3032 avg_prediction = seq.avg() + seq.sd();
3027 3033 }
3028 3034
3029 3035 should_continue = true;
3030 3036 if (hr == NULL) {
3031 3037 // No need for an ergo verbose message here,
3032 3038 // getNextMarkRegion() does this when it returns NULL.
3033 3039 should_continue = false;
3034 3040 } else {
3035 3041 if (adaptive_young_list_length()) {
3036 3042 if (time_remaining_ms < 0.0) {
3037 3043 ergo_verbose1(ErgoCSetConstruction,
3038 3044 "stop adding old regions to CSet",
3039 3045 ergo_format_reason("remaining time is lower than 0")
3040 3046 ergo_format_ms("remaining time"),
3041 3047 time_remaining_ms);
3042 3048 should_continue = false;
3043 3049 }
3044 3050 } else {
3045 3051 if (_collection_set_size < _young_list_fixed_length) {
3046 3052 ergo_verbose2(ErgoCSetConstruction,
3047 3053 "stop adding old regions to CSet",
3048 3054 ergo_format_reason("CSet length lower than target")
3049 3055 ergo_format_region("CSet")
3050 3056 ergo_format_region("young target"),
3051 3057 _collection_set_size, _young_list_fixed_length);
3052 3058 should_continue = false;
3053 3059 }
3054 3060 }
3055 3061 }
3056 3062 } while (should_continue);
3057 3063
3058 3064 if (!adaptive_young_list_length() &&
3059 3065 _collection_set_size < _young_list_fixed_length) {
3060 3066 ergo_verbose2(ErgoCSetConstruction,
3061 3067 "request partially-young GCs end",
3062 3068 ergo_format_reason("CSet length lower than target")
3063 3069 ergo_format_region("CSet")
3064 3070 ergo_format_region("young target"),
3065 3071 _collection_set_size, _young_list_fixed_length);
3066 3072 _should_revert_to_full_young_gcs = true;
3067 3073 }
3068 3074
3069 3075 old_region_num = _collection_set_size - prev_collection_set_size;
3070 3076
3071 3077 ergo_verbose2(ErgoCSetConstruction | ErgoHigh,
3072 3078 "add old regions to CSet",
3073 3079 ergo_format_region("old")
3074 3080 ergo_format_ms("predicted old region time"),
3075 3081 old_region_num,
3076 3082 predicted_pause_time_ms - prev_predicted_pause_time_ms);
3077 3083 }
3078 3084
3079 3085 stop_incremental_cset_building();
3080 3086
3081 3087 count_CS_bytes_used();
3082 3088
3083 3089 end_recording_regions();
3084 3090
3085 3091 ergo_verbose5(ErgoCSetConstruction,
3086 3092 "finish choosing CSet",
3087 3093 ergo_format_region("eden")
3088 3094 ergo_format_region("survivors")
3089 3095 ergo_format_region("old")
3090 3096 ergo_format_ms("predicted pause time")
3091 3097 ergo_format_ms("target pause time"),
3092 3098 eden_region_num, survivor_region_num, old_region_num,
3093 3099 predicted_pause_time_ms, target_pause_time_ms);
3094 3100
3095 3101 double non_young_end_time_sec = os::elapsedTime();
3096 3102 _recorded_non_young_cset_choice_time_ms =
3097 3103 (non_young_end_time_sec - non_young_start_time_sec) * 1000.0;
3098 3104 }
3099 3105
3100 3106 void G1CollectorPolicy_BestRegionsFirst::record_full_collection_end() {
3101 3107 G1CollectorPolicy::record_full_collection_end();
3102 3108 _collectionSetChooser->updateAfterFullCollection();
3103 3109 }
3104 3110
3105 3111 void G1CollectorPolicy_BestRegionsFirst::
3106 3112 record_collection_pause_end() {
3107 3113 G1CollectorPolicy::record_collection_pause_end();
3108 3114 assert(assertMarkedBytesDataOK(), "Marked regions not OK at pause end.");
3109 3115 }
↓ open down ↓ |
1618 lines elided |
↑ open up ↑ |
XXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXXX