1 /*
   2  * Copyright (c) 2004, 2018, Oracle and/or its affiliates. All rights reserved.
   3  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
   4  *
   5  * This code is free software; you can redistribute it and/or modify it
   6  * under the terms of the GNU General Public License version 2 only, as
   7  * published by the Free Software Foundation.
   8  *
   9  * This code is distributed in the hope that it will be useful, but WITHOUT
  10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
  11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
  12  * version 2 for more details (a copy is included in the LICENSE file that
  13  * accompanied this code).
  14  *
  15  * You should have received a copy of the GNU General Public License version
  16  * 2 along with this work; if not, write to the Free Software Foundation,
  17  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
  18  *
  19  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
  20  * or visit www.oracle.com if you need additional information or have any
  21  * questions.
  22  *
  23  */
  24 
  25 #include "precompiled.hpp"
  26 #include "gc/shared/adaptiveSizePolicy.hpp"
  27 #include "gc/shared/gcCause.hpp"
  28 #include "gc/shared/gcUtil.inline.hpp"
  29 #include "gc/shared/softRefPolicy.hpp"
  30 #include "logging/log.hpp"
  31 #include "runtime/timer.hpp"
  32 
  33 elapsedTimer AdaptiveSizePolicy::_minor_timer;
  34 elapsedTimer AdaptiveSizePolicy::_major_timer;
  35 
  36 // The throughput goal is implemented as
  37 //      _throughput_goal = 1 - ( 1 / (1 + gc_cost_ratio))
  38 // gc_cost_ratio is the ratio
  39 //      application cost / gc cost
  40 // For example a gc_cost_ratio of 4 translates into a
  41 // throughput goal of .80
  42 
  43 AdaptiveSizePolicy::AdaptiveSizePolicy(size_t init_eden_size,
  44                                        size_t init_promo_size,
  45                                        size_t init_survivor_size,
  46                                        double gc_pause_goal_sec,
  47                                        uint gc_cost_ratio) :
  48     _throughput_goal(1.0 - double(1.0 / (1.0 + (double) gc_cost_ratio))),
  49     _eden_size(init_eden_size),
  50     _promo_size(init_promo_size),
  51     _survivor_size(init_survivor_size),
  52     _gc_overhead_limit_exceeded(false),
  53     _print_gc_overhead_limit_would_be_exceeded(false),
  54     _gc_overhead_limit_count(0),
  55     _latest_minor_mutator_interval_seconds(0),
  56     _threshold_tolerance_percent(1.0 + ThresholdTolerance/100.0),
  57     _gc_pause_goal_sec(gc_pause_goal_sec),
  58     _young_gen_change_for_minor_throughput(0),
  59     _old_gen_change_for_major_throughput(0) {
  60   assert(AdaptiveSizePolicyGCTimeLimitThreshold > 0,
  61     "No opportunity to clear SoftReferences before GC overhead limit");
  62   _avg_minor_pause    =
  63     new AdaptivePaddedAverage(AdaptiveTimeWeight, PausePadding);
  64   _avg_minor_interval = new AdaptiveWeightedAverage(AdaptiveTimeWeight);
  65   _avg_minor_gc_cost  = new AdaptiveWeightedAverage(AdaptiveTimeWeight);
  66   _avg_major_gc_cost  = new AdaptiveWeightedAverage(AdaptiveTimeWeight);
  67 
  68   _avg_young_live     = new AdaptiveWeightedAverage(AdaptiveSizePolicyWeight);
  69   _avg_old_live       = new AdaptiveWeightedAverage(AdaptiveSizePolicyWeight);
  70   _avg_eden_live      = new AdaptiveWeightedAverage(AdaptiveSizePolicyWeight);
  71 
  72   _avg_survived       = new AdaptivePaddedAverage(AdaptiveSizePolicyWeight,
  73                                                   SurvivorPadding);
  74   _avg_pretenured     = new AdaptivePaddedNoZeroDevAverage(
  75                                                   AdaptiveSizePolicyWeight,
  76                                                   SurvivorPadding);
  77 
  78   _minor_pause_old_estimator =
  79     new LinearLeastSquareFit(AdaptiveSizePolicyWeight);
  80   _minor_pause_young_estimator =
  81     new LinearLeastSquareFit(AdaptiveSizePolicyWeight);
  82   _minor_collection_estimator =
  83     new LinearLeastSquareFit(AdaptiveSizePolicyWeight);
  84   _major_collection_estimator =
  85     new LinearLeastSquareFit(AdaptiveSizePolicyWeight);
  86 
  87   // Start the timers
  88   _minor_timer.start();
  89 
  90   _young_gen_policy_is_ready = false;
  91 }
  92 
  93 bool AdaptiveSizePolicy::tenuring_threshold_change() const {
  94   return decrement_tenuring_threshold_for_gc_cost() ||
  95          increment_tenuring_threshold_for_gc_cost() ||
  96          decrement_tenuring_threshold_for_survivor_limit();
  97 }
  98 
  99 void AdaptiveSizePolicy::minor_collection_begin() {
 100   // Update the interval time
 101   _minor_timer.stop();
 102   // Save most recent collection time
 103   _latest_minor_mutator_interval_seconds = _minor_timer.seconds();
 104   _minor_timer.reset();
 105   _minor_timer.start();
 106 }
 107 
 108 void AdaptiveSizePolicy::update_minor_pause_young_estimator(
 109     double minor_pause_in_ms) {
 110   double eden_size_in_mbytes = ((double)_eden_size)/((double)M);
 111   _minor_pause_young_estimator->update(eden_size_in_mbytes,
 112     minor_pause_in_ms);
 113 }
 114 
 115 void AdaptiveSizePolicy::minor_collection_end(GCCause::Cause gc_cause) {
 116   // Update the pause time.
 117   _minor_timer.stop();
 118 
 119   if (!GCCause::is_user_requested_gc(gc_cause) ||
 120       UseAdaptiveSizePolicyWithSystemGC) {
 121     double minor_pause_in_seconds = _minor_timer.seconds();
 122     double minor_pause_in_ms = minor_pause_in_seconds * MILLIUNITS;
 123 
 124     // Sample for performance counter
 125     _avg_minor_pause->sample(minor_pause_in_seconds);
 126 
 127     // Cost of collection (unit-less)
 128     double collection_cost = 0.0;
 129     if ((_latest_minor_mutator_interval_seconds > 0.0) &&
 130         (minor_pause_in_seconds > 0.0)) {
 131       double interval_in_seconds =
 132         _latest_minor_mutator_interval_seconds + minor_pause_in_seconds;
 133       collection_cost =
 134         minor_pause_in_seconds / interval_in_seconds;
 135       _avg_minor_gc_cost->sample(collection_cost);
 136       // Sample for performance counter
 137       _avg_minor_interval->sample(interval_in_seconds);
 138     }
 139 
 140     // The policy does not have enough data until at least some
 141     // young collections have been done.
 142     _young_gen_policy_is_ready =
 143       (_avg_minor_gc_cost->count() >= AdaptiveSizePolicyReadyThreshold);
 144 
 145     // Calculate variables used to estimate pause time vs. gen sizes
 146     double eden_size_in_mbytes = ((double)_eden_size) / ((double)M);
 147     update_minor_pause_young_estimator(minor_pause_in_ms);
 148     update_minor_pause_old_estimator(minor_pause_in_ms);
 149 
 150     log_trace(gc, ergo)("AdaptiveSizePolicy::minor_collection_end: minor gc cost: %f  average: %f",
 151                         collection_cost, _avg_minor_gc_cost->average());
 152     log_trace(gc, ergo)("  minor pause: %f minor period %f",
 153                         minor_pause_in_ms, _latest_minor_mutator_interval_seconds * MILLIUNITS);
 154 
 155     // Calculate variable used to estimate collection cost vs. gen sizes
 156     assert(collection_cost >= 0.0, "Expected to be non-negative");
 157     _minor_collection_estimator->update(eden_size_in_mbytes, collection_cost);
 158   }
 159 
 160   // Interval times use this timer to measure the mutator time.
 161   // Reset the timer after the GC pause.
 162   _minor_timer.reset();
 163   _minor_timer.start();
 164 }
 165 
 166 size_t AdaptiveSizePolicy::eden_increment(size_t cur_eden, uint percent_change) {
 167   size_t eden_heap_delta;
 168   eden_heap_delta = cur_eden / 100 * percent_change;
 169   return eden_heap_delta;
 170 }
 171 
 172 size_t AdaptiveSizePolicy::eden_increment(size_t cur_eden) {
 173   return eden_increment(cur_eden, YoungGenerationSizeIncrement);
 174 }
 175 
 176 size_t AdaptiveSizePolicy::eden_decrement(size_t cur_eden) {
 177   size_t eden_heap_delta = eden_increment(cur_eden) /
 178     AdaptiveSizeDecrementScaleFactor;
 179   return eden_heap_delta;
 180 }
 181 
 182 size_t AdaptiveSizePolicy::promo_increment(size_t cur_promo, uint percent_change) {
 183   size_t promo_heap_delta;
 184   promo_heap_delta = cur_promo / 100 * percent_change;
 185   return promo_heap_delta;
 186 }
 187 
 188 size_t AdaptiveSizePolicy::promo_increment(size_t cur_promo) {
 189   return promo_increment(cur_promo, TenuredGenerationSizeIncrement);
 190 }
 191 
 192 size_t AdaptiveSizePolicy::promo_decrement(size_t cur_promo) {
 193   size_t promo_heap_delta = promo_increment(cur_promo);
 194   promo_heap_delta = promo_heap_delta / AdaptiveSizeDecrementScaleFactor;
 195   return promo_heap_delta;
 196 }
 197 
 198 double AdaptiveSizePolicy::time_since_major_gc() const {
 199   _major_timer.stop();
 200   double result = _major_timer.seconds();
 201   _major_timer.start();
 202   return result;
 203 }
 204 
 205 // Linear decay of major gc cost
 206 double AdaptiveSizePolicy::decaying_major_gc_cost() const {
 207   double major_interval = major_gc_interval_average_for_decay();
 208   double major_gc_cost_average = major_gc_cost();
 209   double decayed_major_gc_cost = major_gc_cost_average;
 210   if(time_since_major_gc() > 0.0) {
 211     decayed_major_gc_cost = major_gc_cost() *
 212       (((double) AdaptiveSizeMajorGCDecayTimeScale) * major_interval)
 213       / time_since_major_gc();
 214   }
 215 
 216   // The decayed cost should always be smaller than the
 217   // average cost but the vagaries of finite arithmetic could
 218   // produce a larger value in decayed_major_gc_cost so protect
 219   // against that.
 220   return MIN2(major_gc_cost_average, decayed_major_gc_cost);
 221 }
 222 
 223 // Use a value of the major gc cost that has been decayed
 224 // by the factor
 225 //
 226 //      average-interval-between-major-gc * AdaptiveSizeMajorGCDecayTimeScale /
 227 //        time-since-last-major-gc
 228 //
 229 // if the average-interval-between-major-gc * AdaptiveSizeMajorGCDecayTimeScale
 230 // is less than time-since-last-major-gc.
 231 //
 232 // In cases where there are initial major gc's that
 233 // are of a relatively high cost but no later major
 234 // gc's, the total gc cost can remain high because
 235 // the major gc cost remains unchanged (since there are no major
 236 // gc's).  In such a situation the value of the unchanging
 237 // major gc cost can keep the mutator throughput below
 238 // the goal when in fact the major gc cost is becoming diminishingly
 239 // small.  Use the decaying gc cost only to decide whether to
 240 // adjust for throughput.  Using it also to determine the adjustment
 241 // to be made for throughput also seems reasonable but there is
 242 // no test case to use to decide if it is the right thing to do
 243 // don't do it yet.
 244 
 245 double AdaptiveSizePolicy::decaying_gc_cost() const {
 246   double decayed_major_gc_cost = major_gc_cost();
 247   double avg_major_interval = major_gc_interval_average_for_decay();
 248   if (UseAdaptiveSizeDecayMajorGCCost &&
 249       (AdaptiveSizeMajorGCDecayTimeScale > 0) &&
 250       (avg_major_interval > 0.00)) {
 251     double time_since_last_major_gc = time_since_major_gc();
 252 
 253     // Decay the major gc cost?
 254     if (time_since_last_major_gc >
 255         ((double) AdaptiveSizeMajorGCDecayTimeScale) * avg_major_interval) {
 256 
 257       // Decay using the time-since-last-major-gc
 258       decayed_major_gc_cost = decaying_major_gc_cost();
 259       log_trace(gc, ergo)("decaying_gc_cost: major interval average: %f  time since last major gc: %f",
 260                     avg_major_interval, time_since_last_major_gc);
 261       log_trace(gc, ergo)("  major gc cost: %f  decayed major gc cost: %f",
 262                     major_gc_cost(), decayed_major_gc_cost);
 263     }
 264   }
 265   double result = MIN2(1.0, decayed_major_gc_cost + minor_gc_cost());
 266   return result;
 267 }
 268 
 269 
 270 void AdaptiveSizePolicy::clear_generation_free_space_flags() {
 271   set_change_young_gen_for_min_pauses(0);
 272   set_change_old_gen_for_maj_pauses(0);
 273 
 274   set_change_old_gen_for_throughput(0);
 275   set_change_young_gen_for_throughput(0);
 276   set_decrease_for_footprint(0);
 277   set_decide_at_full_gc(0);
 278 }
 279 
 280 void AdaptiveSizePolicy::check_gc_overhead_limit(
 281                                           size_t young_live,
 282                                           size_t eden_live,
 283                                           size_t max_old_gen_size,
 284                                           size_t max_eden_size,
 285                                           bool   is_full_gc,
 286                                           GCCause::Cause gc_cause,
 287                                           SoftRefPolicy* soft_ref_policy) {
 288 
 289   // Ignore explicit GC's.  Exiting here does not set the flag and
 290   // does not reset the count.  Updating of the averages for system
 291   // GC's is still controlled by UseAdaptiveSizePolicyWithSystemGC.
 292   if (GCCause::is_user_requested_gc(gc_cause) ||
 293       GCCause::is_serviceability_requested_gc(gc_cause)) {
 294     return;
 295   }
 296   // eden_limit is the upper limit on the size of eden based on
 297   // the maximum size of the young generation and the sizes
 298   // of the survivor space.
 299   // The question being asked is whether the gc costs are high
 300   // and the space being recovered by a collection is low.
 301   // free_in_young_gen is the free space in the young generation
 302   // after a collection and promo_live is the free space in the old
 303   // generation after a collection.
 304   //
 305   // Use the minimum of the current value of the live in the
 306   // young gen or the average of the live in the young gen.
 307   // If the current value drops quickly, that should be taken
 308   // into account (i.e., don't trigger if the amount of free
 309   // space has suddenly jumped up).  If the current is much
 310   // higher than the average, use the average since it represents
 311   // the longer term behavior.
 312   const size_t live_in_eden =
 313     MIN2(eden_live, (size_t) avg_eden_live()->average());
 314   const size_t free_in_eden = max_eden_size > live_in_eden ?
 315     max_eden_size - live_in_eden : 0;
 316   const size_t free_in_old_gen = (size_t)(max_old_gen_size - avg_old_live()->average());
 317   const size_t total_free_limit = free_in_old_gen + free_in_eden;
 318   const size_t total_mem = max_old_gen_size + max_eden_size;
 319   const double mem_free_limit = total_mem * (GCHeapFreeLimit/100.0);
 320   const double mem_free_old_limit = max_old_gen_size * (GCHeapFreeLimit/100.0);
 321   const double mem_free_eden_limit = max_eden_size * (GCHeapFreeLimit/100.0);
 322   const double gc_cost_limit = GCTimeLimit/100.0;
 323   size_t promo_limit = (size_t)(max_old_gen_size - avg_old_live()->average());
 324   // But don't force a promo size below the current promo size. Otherwise,
 325   // the promo size will shrink for no good reason.
 326   promo_limit = MAX2(promo_limit, _promo_size);
 327 
 328 
 329   log_trace(gc, ergo)(
 330         "PSAdaptiveSizePolicy::check_gc_overhead_limit:"
 331         " promo_limit: " SIZE_FORMAT
 332         " max_eden_size: " SIZE_FORMAT
 333         " total_free_limit: " SIZE_FORMAT
 334         " max_old_gen_size: " SIZE_FORMAT
 335         " max_eden_size: " SIZE_FORMAT
 336         " mem_free_limit: " SIZE_FORMAT,
 337         promo_limit, max_eden_size, total_free_limit,
 338         max_old_gen_size, max_eden_size,
 339         (size_t) mem_free_limit);
 340 
 341   bool print_gc_overhead_limit_would_be_exceeded = false;
 342   if (is_full_gc) {
 343     if (gc_cost() > gc_cost_limit &&
 344       free_in_old_gen < (size_t) mem_free_old_limit &&
 345       free_in_eden < (size_t) mem_free_eden_limit) {
 346       // Collections, on average, are taking too much time, and
 347       //      gc_cost() > gc_cost_limit
 348       // we have too little space available after a full gc.
 349       //      total_free_limit < mem_free_limit
 350       // where
 351       //   total_free_limit is the free space available in
 352       //     both generations
 353       //   total_mem is the total space available for allocation
 354       //     in both generations (survivor spaces are not included
 355       //     just as they are not included in eden_limit).
 356       //   mem_free_limit is a fraction of total_mem judged to be an
 357       //     acceptable amount that is still unused.
 358       // The heap can ask for the value of this variable when deciding
 359       // whether to thrown an OutOfMemory error.
 360       // Note that the gc time limit test only works for the collections
 361       // of the young gen + tenured gen and not for collections of the
 362       // permanent gen.  That is because the calculation of the space
 363       // freed by the collection is the free space in the young gen +
 364       // tenured gen.
 365       // At this point the GC overhead limit is being exceeded.
 366       inc_gc_overhead_limit_count();
 367       if (UseGCOverheadLimit) {
 368         if (gc_overhead_limit_count() >=
 369             AdaptiveSizePolicyGCTimeLimitThreshold){
 370           // All conditions have been met for throwing an out-of-memory
 371           set_gc_overhead_limit_exceeded(true);
 372           // Avoid consecutive OOM due to the gc time limit by resetting
 373           // the counter.
 374           reset_gc_overhead_limit_count();
 375         } else {
 376           // The required consecutive collections which exceed the
 377           // GC time limit may or may not have been reached. We
 378           // are approaching that condition and so as not to
 379           // throw an out-of-memory before all SoftRef's have been
 380           // cleared, set _should_clear_all_soft_refs in CollectorPolicy.
 381           // The clearing will be done on the next GC.
 382           bool near_limit = gc_overhead_limit_near();
 383           if (near_limit) {
 384             soft_ref_policy->set_should_clear_all_soft_refs(true);
 385             log_trace(gc, ergo)("Nearing GC overhead limit, will be clearing all SoftReference");
 386           }
 387         }
 388       }
 389       // Set this even when the overhead limit will not
 390       // cause an out-of-memory.  Diagnostic message indicating
 391       // that the overhead limit is being exceeded is sometimes
 392       // printed.
 393       print_gc_overhead_limit_would_be_exceeded = true;
 394 
 395     } else {
 396       // Did not exceed overhead limits
 397       reset_gc_overhead_limit_count();
 398     }
 399   }
 400 
 401   if (UseGCOverheadLimit) {
 402     if (gc_overhead_limit_exceeded()) {
 403       log_trace(gc, ergo)("GC is exceeding overhead limit of " UINTX_FORMAT "%%", GCTimeLimit);
 404       reset_gc_overhead_limit_count();
 405     } else if (print_gc_overhead_limit_would_be_exceeded) {
 406       assert(gc_overhead_limit_count() > 0, "Should not be printing");
 407       log_trace(gc, ergo)("GC would exceed overhead limit of " UINTX_FORMAT "%% %d consecutive time(s)",
 408                           GCTimeLimit, gc_overhead_limit_count());
 409     }
 410   }
 411 }
 412 // Printing
 413 
 414 bool AdaptiveSizePolicy::print() const {
 415   assert(UseAdaptiveSizePolicy, "UseAdaptiveSizePolicy need to be enabled.");
 416 
 417   if (!log_is_enabled(Debug, gc, ergo)) {
 418     return false;
 419   }
 420 
 421   // Print goal for which action is needed.
 422   char* action = NULL;
 423   bool change_for_pause = false;
 424   if ((change_old_gen_for_maj_pauses() ==
 425          decrease_old_gen_for_maj_pauses_true) ||
 426       (change_young_gen_for_min_pauses() ==
 427          decrease_young_gen_for_min_pauses_true)) {
 428     action = (char*) " *** pause time goal ***";
 429     change_for_pause = true;
 430   } else if ((change_old_gen_for_throughput() ==
 431                increase_old_gen_for_throughput_true) ||
 432             (change_young_gen_for_throughput() ==
 433                increase_young_gen_for_througput_true)) {
 434     action = (char*) " *** throughput goal ***";
 435   } else if (decrease_for_footprint()) {
 436     action = (char*) " *** reduced footprint ***";
 437   } else {
 438     // No actions were taken.  This can legitimately be the
 439     // situation if not enough data has been gathered to make
 440     // decisions.
 441     return false;
 442   }
 443 
 444   // Pauses
 445   // Currently the size of the old gen is only adjusted to
 446   // change the major pause times.
 447   char* young_gen_action = NULL;
 448   char* tenured_gen_action = NULL;
 449 
 450   char* shrink_msg = (char*) "(attempted to shrink)";
 451   char* grow_msg = (char*) "(attempted to grow)";
 452   char* no_change_msg = (char*) "(no change)";
 453   if (change_young_gen_for_min_pauses() ==
 454       decrease_young_gen_for_min_pauses_true) {
 455     young_gen_action = shrink_msg;
 456   } else if (change_for_pause) {
 457     young_gen_action = no_change_msg;
 458   }
 459 
 460   if (change_old_gen_for_maj_pauses() == decrease_old_gen_for_maj_pauses_true) {
 461     tenured_gen_action = shrink_msg;
 462   } else if (change_for_pause) {
 463     tenured_gen_action = no_change_msg;
 464   }
 465 
 466   // Throughput
 467   if (change_old_gen_for_throughput() == increase_old_gen_for_throughput_true) {
 468     assert(change_young_gen_for_throughput() ==
 469            increase_young_gen_for_througput_true,
 470            "Both generations should be growing");
 471     young_gen_action = grow_msg;
 472     tenured_gen_action = grow_msg;
 473   } else if (change_young_gen_for_throughput() ==
 474              increase_young_gen_for_througput_true) {
 475     // Only the young generation may grow at start up (before
 476     // enough full collections have been done to grow the old generation).
 477     young_gen_action = grow_msg;
 478     tenured_gen_action = no_change_msg;
 479   }
 480 
 481   // Minimum footprint
 482   if (decrease_for_footprint() != 0) {
 483     young_gen_action = shrink_msg;
 484     tenured_gen_action = shrink_msg;
 485   }
 486 
 487   log_debug(gc, ergo)("UseAdaptiveSizePolicy actions to meet %s", action);
 488   log_debug(gc, ergo)("                       GC overhead (%%)");
 489   log_debug(gc, ergo)("    Young generation:     %7.2f\t  %s",
 490                       100.0 * avg_minor_gc_cost()->average(), young_gen_action);
 491   log_debug(gc, ergo)("    Tenured generation:   %7.2f\t  %s",
 492                       100.0 * avg_major_gc_cost()->average(), tenured_gen_action);
 493   return true;
 494 }
 495 
 496 void AdaptiveSizePolicy::print_tenuring_threshold( uint new_tenuring_threshold_arg) const {
 497   // Tenuring threshold
 498   if (decrement_tenuring_threshold_for_survivor_limit()) {
 499     log_debug(gc, ergo)("Tenuring threshold: (attempted to decrease to avoid survivor space overflow) = %u", new_tenuring_threshold_arg);
 500   } else if (decrement_tenuring_threshold_for_gc_cost()) {
 501     log_debug(gc, ergo)("Tenuring threshold: (attempted to decrease to balance GC costs) = %u", new_tenuring_threshold_arg);
 502   } else if (increment_tenuring_threshold_for_gc_cost()) {
 503     log_debug(gc, ergo)("Tenuring threshold: (attempted to increase to balance GC costs) = %u", new_tenuring_threshold_arg);
 504   } else {
 505     assert(!tenuring_threshold_change(), "(no change was attempted)");
 506   }
 507 }