< prev index next >

src/hotspot/share/gc/shenandoah/heuristics/shenandoahAdaptiveHeuristics.cpp

Print this page

  1 /*
  2  * Copyright (c) 2018, 2019, Red Hat, Inc. All rights reserved.

  3  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
  4  *
  5  * This code is free software; you can redistribute it and/or modify it
  6  * under the terms of the GNU General Public License version 2 only, as
  7  * published by the Free Software Foundation.
  8  *
  9  * This code is distributed in the hope that it will be useful, but WITHOUT
 10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
 11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
 12  * version 2 for more details (a copy is included in the LICENSE file that
 13  * accompanied this code).
 14  *
 15  * You should have received a copy of the GNU General Public License version
 16  * 2 along with this work; if not, write to the Free Software Foundation,
 17  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
 18  *
 19  * Please contact 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 



 27 #include "gc/shenandoah/heuristics/shenandoahAdaptiveHeuristics.hpp"
 28 #include "gc/shenandoah/shenandoahCollectionSet.hpp"

 29 #include "gc/shenandoah/shenandoahFreeSet.hpp"
 30 #include "gc/shenandoah/shenandoahHeap.inline.hpp"
 31 #include "gc/shenandoah/shenandoahHeapRegion.inline.hpp"
 32 #include "logging/log.hpp"
 33 #include "logging/logTag.hpp"

 34 #include "utilities/quickSort.hpp"
 35 
 36 // These constants are used to adjust the margin of error for the moving
 37 // average of the allocation rate and cycle time. The units are standard
 38 // deviations.
 39 const double ShenandoahAdaptiveHeuristics::FULL_PENALTY_SD = 0.2;
 40 const double ShenandoahAdaptiveHeuristics::DEGENERATE_PENALTY_SD = 0.1;
 41 
 42 // These are used to decide if we want to make any adjustments at all
 43 // at the end of a successful concurrent cycle.
 44 const double ShenandoahAdaptiveHeuristics::LOWEST_EXPECTED_AVAILABLE_AT_END = -0.5;
 45 const double ShenandoahAdaptiveHeuristics::HIGHEST_EXPECTED_AVAILABLE_AT_END = 0.5;
 46 
 47 // These values are the confidence interval expressed as standard deviations.
 48 // At the minimum confidence level, there is a 25% chance that the true value of
 49 // the estimate (average cycle time or allocation rate) is not more than
 50 // MINIMUM_CONFIDENCE standard deviations away from our estimate. Similarly, the
 51 // MAXIMUM_CONFIDENCE interval here means there is a one in a thousand chance
 52 // that the true value of our estimate is outside the interval. These are used
 53 // as bounds on the adjustments applied at the outcome of a GC cycle.
 54 const double ShenandoahAdaptiveHeuristics::MINIMUM_CONFIDENCE = 0.319; // 25%
 55 const double ShenandoahAdaptiveHeuristics::MAXIMUM_CONFIDENCE = 3.291; // 99.9%
 56 
 57 ShenandoahAdaptiveHeuristics::ShenandoahAdaptiveHeuristics(ShenandoahSpaceInfo* space_info) :
 58   ShenandoahHeuristics(space_info),
 59   _margin_of_error_sd(ShenandoahAdaptiveInitialConfidence),
 60   _spike_threshold_sd(ShenandoahAdaptiveInitialSpikeThreshold),
 61   _last_trigger(OTHER) { }

 62 
 63 ShenandoahAdaptiveHeuristics::~ShenandoahAdaptiveHeuristics() {}
 64 
 65 void ShenandoahAdaptiveHeuristics::choose_collection_set_from_regiondata(ShenandoahCollectionSet* cset,
 66                                                                          RegionData* data, size_t size,
 67                                                                          size_t actual_free) {
 68   size_t garbage_threshold = ShenandoahHeapRegion::region_size_bytes() * ShenandoahGarbageThreshold / 100;
 69 
 70   // The logic for cset selection in adaptive is as follows:
 71   //
 72   //   1. We cannot get cset larger than available free space. Otherwise we guarantee OOME
 73   //      during evacuation, and thus guarantee full GC. In practice, we also want to let
 74   //      application to allocate something. This is why we limit CSet to some fraction of
 75   //      available space. In non-overloaded heap, max_cset would contain all plausible candidates
 76   //      over garbage threshold.
 77   //
 78   //   2. We should not get cset too low so that free threshold would not be met right
 79   //      after the cycle. Otherwise we get back-to-back cycles for no reason if heap is
 80   //      too fragmented. In non-overloaded non-fragmented heap min_garbage would be around zero.
 81   //
 82   // Therefore, we start by sorting the regions by garbage. Then we unconditionally add the best candidates
 83   // before we meet min_garbage. Then we add all candidates that fit with a garbage threshold before
 84   // we hit max_cset. When max_cset is hit, we terminate the cset selection. Note that in this scheme,
 85   // ShenandoahGarbageThreshold is the soft threshold which would be ignored until min_garbage is hit.
 86 
 87   size_t capacity    = _space_info->soft_max_capacity();
 88   size_t max_cset    = (size_t)((1.0 * capacity / 100 * ShenandoahEvacReserve) / ShenandoahEvacWaste);
 89   size_t free_target = (capacity / 100 * ShenandoahMinFreeThreshold) + max_cset;
 90   size_t min_garbage = (free_target > actual_free ? (free_target - actual_free) : 0);
 91 
 92   log_info(gc, ergo)("Adaptive CSet Selection. Target Free: " SIZE_FORMAT "%s, Actual Free: "
 93                      SIZE_FORMAT "%s, Max CSet: " SIZE_FORMAT "%s, Min Garbage: " SIZE_FORMAT "%s",
 94                      byte_size_in_proper_unit(free_target), proper_unit_for_byte_size(free_target),
 95                      byte_size_in_proper_unit(actual_free), proper_unit_for_byte_size(actual_free),
 96                      byte_size_in_proper_unit(max_cset),    proper_unit_for_byte_size(max_cset),
 97                      byte_size_in_proper_unit(min_garbage), proper_unit_for_byte_size(min_garbage));
 98 
 99   // Better select garbage-first regions
100   QuickSort::sort(data, size, compare_by_garbage);
101 
102   size_t cur_cset = 0;
103   size_t cur_garbage = 0;
104 
105   for (size_t idx = 0; idx < size; idx++) {
106     ShenandoahHeapRegion* r = data[idx]._region;
107 
108     size_t new_cset    = cur_cset + r->get_live_data_bytes();
109     size_t new_garbage = cur_garbage + r->garbage();
110 
111     if (new_cset > max_cset) {
112       break;
113     }
114 
115     if ((new_garbage < min_garbage) || (r->garbage() > garbage_threshold)) {
116       cset->add_region(r);
117       cur_cset = new_cset;
118       cur_garbage = new_garbage;
119     }
120   }
121 }
122 
123 void ShenandoahAdaptiveHeuristics::record_cycle_start() {
124   ShenandoahHeuristics::record_cycle_start();
125   _allocation_rate.allocation_counter_reset();
126 }
127 
128 void ShenandoahAdaptiveHeuristics::record_success_concurrent() {
129   ShenandoahHeuristics::record_success_concurrent();
130 
131   size_t available = _space_info->available();
132 
133   _available.add(available);
134   double z_score = 0.0;
135   if (_available.sd() > 0) {
136     z_score = (available - _available.avg()) / _available.sd();







137   }
138 
139   log_debug(gc, ergo)("Available: " SIZE_FORMAT " %sB, z-score=%.3f. Average available: %.1f %sB +/- %.1f %sB.",
140                       byte_size_in_proper_unit(available), proper_unit_for_byte_size(available),
141                       z_score,
142                       byte_size_in_proper_unit(_available.avg()), proper_unit_for_byte_size(_available.avg()),
143                       byte_size_in_proper_unit(_available.sd()), proper_unit_for_byte_size(_available.sd()));
144 
145   // In the case when a concurrent GC cycle completes successfully but with an
146   // unusually small amount of available memory we will adjust our trigger
147   // parameters so that they are more likely to initiate a new cycle.
148   // Conversely, when a GC cycle results in an above average amount of available
149   // memory, we will adjust the trigger parameters to be less likely to initiate
150   // a GC cycle.
151   //
152   // The z-score we've computed is in no way statistically related to the
153   // trigger parameters, but it has the nice property that worse z-scores for
154   // available memory indicate making larger adjustments to the trigger
155   // parameters. It also results in fewer adjustments as the application
156   // stabilizes.
157   //
158   // In order to avoid making endless and likely unnecessary adjustments to the
159   // trigger parameters, the change in available memory (with respect to the
160   // average) at the end of a cycle must be beyond these threshold values.
161   if (z_score < LOWEST_EXPECTED_AVAILABLE_AT_END ||
162       z_score > HIGHEST_EXPECTED_AVAILABLE_AT_END) {
163     // The sign is flipped because a negative z-score indicates that the

178 void ShenandoahAdaptiveHeuristics::record_success_degenerated() {
179   ShenandoahHeuristics::record_success_degenerated();
180   // Adjust both trigger's parameters in the case of a degenerated GC because
181   // either of them should have triggered earlier to avoid this case.
182   adjust_margin_of_error(DEGENERATE_PENALTY_SD);
183   adjust_spike_threshold(DEGENERATE_PENALTY_SD);
184 }
185 
186 void ShenandoahAdaptiveHeuristics::record_success_full() {
187   ShenandoahHeuristics::record_success_full();
188   // Adjust both trigger's parameters in the case of a full GC because
189   // either of them should have triggered earlier to avoid this case.
190   adjust_margin_of_error(FULL_PENALTY_SD);
191   adjust_spike_threshold(FULL_PENALTY_SD);
192 }
193 
194 static double saturate(double value, double min, double max) {
195   return MAX2(MIN2(value, max), min);
196 }
197 




























198 bool ShenandoahAdaptiveHeuristics::should_start_gc() {
199   size_t max_capacity = _space_info->max_capacity();
200   size_t capacity = _space_info->soft_max_capacity();
201   size_t available = _space_info->available();
202   size_t allocated = _space_info->bytes_allocated_since_gc_start();
203 
204   // Make sure the code below treats available without the soft tail.
205   size_t soft_tail = max_capacity - capacity;
206   available = (available > soft_tail) ? (available - soft_tail) : 0;
207 
208   // Track allocation rate even if we decide to start a cycle for other reasons.
209   double rate = _allocation_rate.sample(allocated);
210   _last_trigger = OTHER;
211 
212   size_t min_threshold = capacity / 100 * ShenandoahMinFreeThreshold;
213   if (available < min_threshold) {
214     log_info(gc)("Trigger: Free (" SIZE_FORMAT "%s) is below minimum threshold (" SIZE_FORMAT "%s)",
215                  byte_size_in_proper_unit(available),     proper_unit_for_byte_size(available),
216                  byte_size_in_proper_unit(min_threshold), proper_unit_for_byte_size(min_threshold));
217     return true;
218   }
219 

220   const size_t max_learn = ShenandoahLearningSteps;
221   if (_gc_times_learned < max_learn) {
222     size_t init_threshold = capacity / 100 * ShenandoahInitFreeThreshold;
223     if (available < init_threshold) {
224       log_info(gc)("Trigger: Learning " SIZE_FORMAT " of " SIZE_FORMAT ". Free (" SIZE_FORMAT "%s) is below initial threshold (" SIZE_FORMAT "%s)",
225                    _gc_times_learned + 1, max_learn,
226                    byte_size_in_proper_unit(available),      proper_unit_for_byte_size(available),
227                    byte_size_in_proper_unit(init_threshold), proper_unit_for_byte_size(init_threshold));
228       return true;
229     }
230   }
231 
232   // Check if allocation headroom is still okay. This also factors in:
233   //   1. Some space to absorb allocation spikes
234   //   2. Accumulated penalties from Degenerated and Full GC
235   size_t allocation_headroom = available;
236 
237   size_t spike_headroom = capacity / 100 * ShenandoahAllocSpikeFactor;
238   size_t penalties      = capacity / 100 * _gc_time_penalties;
239 
240   allocation_headroom -= MIN2(allocation_headroom, spike_headroom);
241   allocation_headroom -= MIN2(allocation_headroom, penalties);
242 
243   double avg_cycle_time = _gc_time_history->davg() + (_margin_of_error_sd * _gc_time_history->dsd());
244   double avg_alloc_rate = _allocation_rate.upper_bound(_margin_of_error_sd);



245   if (avg_cycle_time * avg_alloc_rate > allocation_headroom) {
246     log_info(gc)("Trigger: Average GC time (%.2f ms) is above the time for average allocation rate (%.0f %sB/s) to deplete free headroom (" SIZE_FORMAT "%s) (margin of error = %.2f)",

247                  avg_cycle_time * 1000,
248                  byte_size_in_proper_unit(avg_alloc_rate), proper_unit_for_byte_size(avg_alloc_rate),
249                  byte_size_in_proper_unit(allocation_headroom), proper_unit_for_byte_size(allocation_headroom),
250                  _margin_of_error_sd);
251 
252     log_info(gc, ergo)("Free headroom: " SIZE_FORMAT "%s (free) - " SIZE_FORMAT "%s (spike) - " SIZE_FORMAT "%s (penalties) = " SIZE_FORMAT "%s",
253                        byte_size_in_proper_unit(available),           proper_unit_for_byte_size(available),
254                        byte_size_in_proper_unit(spike_headroom),      proper_unit_for_byte_size(spike_headroom),
255                        byte_size_in_proper_unit(penalties),           proper_unit_for_byte_size(penalties),
256                        byte_size_in_proper_unit(allocation_headroom), proper_unit_for_byte_size(allocation_headroom));
257 
258     _last_trigger = RATE;
259     return true;
260   }
261 
262   bool is_spiking = _allocation_rate.is_spiking(rate, _spike_threshold_sd);
263   if (is_spiking && avg_cycle_time > allocation_headroom / rate) {
264     log_info(gc)("Trigger: Average GC time (%.2f ms) is above the time for instantaneous allocation rate (%.0f %sB/s) to deplete free headroom (" SIZE_FORMAT "%s) (spike threshold = %.2f)",
265                  avg_cycle_time * 1000,
266                  byte_size_in_proper_unit(rate), proper_unit_for_byte_size(rate),
267                  byte_size_in_proper_unit(allocation_headroom), proper_unit_for_byte_size(allocation_headroom),
268                  _spike_threshold_sd);
269     _last_trigger = SPIKE;
270     return true;
271   }
272 
273   return ShenandoahHeuristics::should_start_gc();
274 }
275 
276 void ShenandoahAdaptiveHeuristics::adjust_last_trigger_parameters(double amount) {
277   switch (_last_trigger) {
278     case RATE:
279       adjust_margin_of_error(amount);
280       break;
281     case SPIKE:
282       adjust_spike_threshold(amount);
283       break;
284     case OTHER:
285       // nothing to adjust here.
286       break;
287     default:
288       ShouldNotReachHere();
289   }
290 }
291 
292 void ShenandoahAdaptiveHeuristics::adjust_margin_of_error(double amount) {
293   _margin_of_error_sd = saturate(_margin_of_error_sd + amount, MINIMUM_CONFIDENCE, MAXIMUM_CONFIDENCE);
294   log_debug(gc, ergo)("Margin of error now %.2f", _margin_of_error_sd);
295 }
296 
297 void ShenandoahAdaptiveHeuristics::adjust_spike_threshold(double amount) {
298   _spike_threshold_sd = saturate(_spike_threshold_sd - amount, MINIMUM_CONFIDENCE, MAXIMUM_CONFIDENCE);
299   log_debug(gc, ergo)("Spike threshold now: %.2f", _spike_threshold_sd);
300 }
301 







302 ShenandoahAllocationRate::ShenandoahAllocationRate() :
303   _last_sample_time(os::elapsedTime()),
304   _last_sample_value(0),
305   _interval_sec(1.0 / ShenandoahAdaptiveSampleFrequencyHz),
306   _rate(int(ShenandoahAdaptiveSampleSizeSeconds * ShenandoahAdaptiveSampleFrequencyHz), ShenandoahAdaptiveDecayFactor),
307   _rate_avg(int(ShenandoahAdaptiveSampleSizeSeconds * ShenandoahAdaptiveSampleFrequencyHz), ShenandoahAdaptiveDecayFactor) {
308 }
309 
310 double ShenandoahAllocationRate::sample(size_t allocated) {
311   double now = os::elapsedTime();
312   double rate = 0.0;
313   if (now - _last_sample_time > _interval_sec) {
314     if (allocated >= _last_sample_value) {
315       rate = instantaneous_rate(now, allocated);
316       _rate.add(rate);
317       _rate_avg.add(_rate.avg());
318     }
319 
320     _last_sample_time = now;
321     _last_sample_value = allocated;

  1 /*
  2  * Copyright (c) 2018, 2019, Red Hat, Inc. All rights reserved.
  3  * Copyright Amazon.com Inc. or its affiliates. All Rights Reserved.
  4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
  5  *
  6  * This code is free software; you can redistribute it and/or modify it
  7  * under the terms of the GNU General Public License version 2 only, as
  8  * published by the Free Software Foundation.
  9  *
 10  * This code is distributed in the hope that it will be useful, but WITHOUT
 11  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
 12  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
 13  * version 2 for more details (a copy is included in the LICENSE file that
 14  * accompanied this code).
 15  *
 16  * You should have received a copy of the GNU General Public License version
 17  * 2 along with this work; if not, write to the Free Software Foundation,
 18  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
 19  *
 20  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
 21  * or visit www.oracle.com if you need additional information or have any
 22  * questions.
 23  *
 24  */
 25 
 26 #include "precompiled.hpp"
 27 
 28 #include "gc/shared/gcCause.hpp"
 29 #include "gc/shenandoah/heuristics/shenandoahHeuristics.hpp"
 30 #include "gc/shenandoah/heuristics/shenandoahSpaceInfo.hpp"
 31 #include "gc/shenandoah/heuristics/shenandoahAdaptiveHeuristics.hpp"
 32 #include "gc/shenandoah/shenandoahCollectionSet.hpp"
 33 #include "gc/shenandoah/shenandoahCollectorPolicy.hpp"
 34 #include "gc/shenandoah/shenandoahFreeSet.hpp"
 35 #include "gc/shenandoah/shenandoahHeap.inline.hpp"
 36 #include "gc/shenandoah/shenandoahHeapRegion.inline.hpp"
 37 #include "logging/log.hpp"
 38 #include "logging/logTag.hpp"
 39 #include "runtime/globals.hpp"
 40 #include "utilities/quickSort.hpp"
 41 
 42 // These constants are used to adjust the margin of error for the moving
 43 // average of the allocation rate and cycle time. The units are standard
 44 // deviations.
 45 const double ShenandoahAdaptiveHeuristics::FULL_PENALTY_SD = 0.2;
 46 const double ShenandoahAdaptiveHeuristics::DEGENERATE_PENALTY_SD = 0.1;
 47 
 48 // These are used to decide if we want to make any adjustments at all
 49 // at the end of a successful concurrent cycle.
 50 const double ShenandoahAdaptiveHeuristics::LOWEST_EXPECTED_AVAILABLE_AT_END = -0.5;
 51 const double ShenandoahAdaptiveHeuristics::HIGHEST_EXPECTED_AVAILABLE_AT_END = 0.5;
 52 
 53 // These values are the confidence interval expressed as standard deviations.
 54 // At the minimum confidence level, there is a 25% chance that the true value of
 55 // the estimate (average cycle time or allocation rate) is not more than
 56 // MINIMUM_CONFIDENCE standard deviations away from our estimate. Similarly, the
 57 // MAXIMUM_CONFIDENCE interval here means there is a one in a thousand chance
 58 // that the true value of our estimate is outside the interval. These are used
 59 // as bounds on the adjustments applied at the outcome of a GC cycle.
 60 const double ShenandoahAdaptiveHeuristics::MINIMUM_CONFIDENCE = 0.319; // 25%
 61 const double ShenandoahAdaptiveHeuristics::MAXIMUM_CONFIDENCE = 3.291; // 99.9%
 62 
 63 ShenandoahAdaptiveHeuristics::ShenandoahAdaptiveHeuristics(ShenandoahSpaceInfo* space_info) :
 64   ShenandoahHeuristics(space_info),
 65   _margin_of_error_sd(ShenandoahAdaptiveInitialConfidence),
 66   _spike_threshold_sd(ShenandoahAdaptiveInitialSpikeThreshold),
 67   _last_trigger(OTHER),
 68   _available(Moving_Average_Samples, ShenandoahAdaptiveDecayFactor) { }
 69 
 70 ShenandoahAdaptiveHeuristics::~ShenandoahAdaptiveHeuristics() {}
 71 
 72 void ShenandoahAdaptiveHeuristics::choose_collection_set_from_regiondata(ShenandoahCollectionSet* cset,
 73                                                                          RegionData* data, size_t size,
 74                                                                          size_t actual_free) {
 75   size_t garbage_threshold = ShenandoahHeapRegion::region_size_bytes() * ShenandoahGarbageThreshold / 100;
 76 
 77   // The logic for cset selection in adaptive is as follows:
 78   //
 79   //   1. We cannot get cset larger than available free space. Otherwise we guarantee OOME
 80   //      during evacuation, and thus guarantee full GC. In practice, we also want to let
 81   //      application to allocate something. This is why we limit CSet to some fraction of
 82   //      available space. In non-overloaded heap, max_cset would contain all plausible candidates
 83   //      over garbage threshold.
 84   //
 85   //   2. We should not get cset too low so that free threshold would not be met right
 86   //      after the cycle. Otherwise we get back-to-back cycles for no reason if heap is
 87   //      too fragmented. In non-overloaded non-fragmented heap min_garbage would be around zero.
 88   //
 89   // Therefore, we start by sorting the regions by garbage. Then we unconditionally add the best candidates
 90   // before we meet min_garbage. Then we add all candidates that fit with a garbage threshold before
 91   // we hit max_cset. When max_cset is hit, we terminate the cset selection. Note that in this scheme,
 92   // ShenandoahGarbageThreshold is the soft threshold which would be ignored until min_garbage is hit.
 93 
 94   size_t capacity    = _space_info->soft_max_capacity();
 95   size_t max_cset    = (size_t)((1.0 * capacity / 100 * ShenandoahEvacReserve) / ShenandoahEvacWaste);
 96   size_t free_target = (capacity / 100 * ShenandoahMinFreeThreshold) + max_cset;
 97   size_t min_garbage = (free_target > actual_free ? (free_target - actual_free) : 0);
 98 
 99   log_info(gc, ergo)("Adaptive CSet Selection. Target Free: " SIZE_FORMAT "%s, Actual Free: "
100                      SIZE_FORMAT "%s, Max Evacuation: " SIZE_FORMAT "%s, Min Garbage: " SIZE_FORMAT "%s",
101                      byte_size_in_proper_unit(free_target), proper_unit_for_byte_size(free_target),
102                      byte_size_in_proper_unit(actual_free), proper_unit_for_byte_size(actual_free),
103                      byte_size_in_proper_unit(max_cset),    proper_unit_for_byte_size(max_cset),
104                      byte_size_in_proper_unit(min_garbage), proper_unit_for_byte_size(min_garbage));
105 
106   // Better select garbage-first regions
107   QuickSort::sort(data, size, compare_by_garbage);
108 
109   size_t cur_cset = 0;
110   size_t cur_garbage = 0;
111 
112   for (size_t idx = 0; idx < size; idx++) {
113     ShenandoahHeapRegion* r = data[idx].get_region();
114 
115     size_t new_cset    = cur_cset + r->get_live_data_bytes();
116     size_t new_garbage = cur_garbage + r->garbage();
117 
118     if (new_cset > max_cset) {
119       break;
120     }
121 
122     if ((new_garbage < min_garbage) || (r->garbage() > garbage_threshold)) {
123       cset->add_region(r);
124       cur_cset = new_cset;
125       cur_garbage = new_garbage;
126     }
127   }
128 }
129 
130 void ShenandoahAdaptiveHeuristics::record_cycle_start() {
131   ShenandoahHeuristics::record_cycle_start();
132   _allocation_rate.allocation_counter_reset();
133 }
134 
135 void ShenandoahAdaptiveHeuristics::record_success_concurrent() {
136   ShenandoahHeuristics::record_success_concurrent();
137 
138   size_t available = _space_info->available();
139 

140   double z_score = 0.0;
141   double available_sd = _available.sd();
142   if (available_sd > 0) {
143     double available_avg = _available.avg();
144     z_score = (double(available) - available_avg) / available_sd;
145     log_debug(gc, ergo)("Available: " SIZE_FORMAT " %sB, z-score=%.3f. Average available: %.1f %sB +/- %.1f %sB.",
146                         byte_size_in_proper_unit(available), proper_unit_for_byte_size(available),
147                         z_score,
148                         byte_size_in_proper_unit(available_avg), proper_unit_for_byte_size(available_avg),
149                         byte_size_in_proper_unit(available_sd), proper_unit_for_byte_size(available_sd));
150   }
151 
152   _available.add(double(available));




153 
154   // In the case when a concurrent GC cycle completes successfully but with an
155   // unusually small amount of available memory we will adjust our trigger
156   // parameters so that they are more likely to initiate a new cycle.
157   // Conversely, when a GC cycle results in an above average amount of available
158   // memory, we will adjust the trigger parameters to be less likely to initiate
159   // a GC cycle.
160   //
161   // The z-score we've computed is in no way statistically related to the
162   // trigger parameters, but it has the nice property that worse z-scores for
163   // available memory indicate making larger adjustments to the trigger
164   // parameters. It also results in fewer adjustments as the application
165   // stabilizes.
166   //
167   // In order to avoid making endless and likely unnecessary adjustments to the
168   // trigger parameters, the change in available memory (with respect to the
169   // average) at the end of a cycle must be beyond these threshold values.
170   if (z_score < LOWEST_EXPECTED_AVAILABLE_AT_END ||
171       z_score > HIGHEST_EXPECTED_AVAILABLE_AT_END) {
172     // The sign is flipped because a negative z-score indicates that the

187 void ShenandoahAdaptiveHeuristics::record_success_degenerated() {
188   ShenandoahHeuristics::record_success_degenerated();
189   // Adjust both trigger's parameters in the case of a degenerated GC because
190   // either of them should have triggered earlier to avoid this case.
191   adjust_margin_of_error(DEGENERATE_PENALTY_SD);
192   adjust_spike_threshold(DEGENERATE_PENALTY_SD);
193 }
194 
195 void ShenandoahAdaptiveHeuristics::record_success_full() {
196   ShenandoahHeuristics::record_success_full();
197   // Adjust both trigger's parameters in the case of a full GC because
198   // either of them should have triggered earlier to avoid this case.
199   adjust_margin_of_error(FULL_PENALTY_SD);
200   adjust_spike_threshold(FULL_PENALTY_SD);
201 }
202 
203 static double saturate(double value, double min, double max) {
204   return MAX2(MIN2(value, max), min);
205 }
206 
207 //  Rationale:
208 //    The idea is that there is an average allocation rate and there are occasional abnormal bursts (or spikes) of
209 //    allocations that exceed the average allocation rate.  What do these spikes look like?
210 //
211 //    1. At certain phase changes, we may discard large amounts of data and replace it with large numbers of newly
212 //       allocated objects.  This "spike" looks more like a phase change.  We were in steady state at M bytes/sec
213 //       allocation rate and now we're in a "reinitialization phase" that looks like N bytes/sec.  We need the "spike"
214 //       accommodation to give us enough runway to recalibrate our "average allocation rate".
215 //
216 //   2. The typical workload changes.  "Suddenly", our typical workload of N TPS increases to N+delta TPS.  This means
217 //       our average allocation rate needs to be adjusted.  Once again, we need the "spike" accomodation to give us
218 //       enough runway to recalibrate our "average allocation rate".
219 //
220 //    3. Though there is an "average" allocation rate, a given workload's demand for allocation may be very bursty.  We
221 //       allocate a bunch of LABs during the 5 ms that follow completion of a GC, then we perform no more allocations for
222 //       the next 150 ms.  It seems we want the "spike" to represent the maximum divergence from average within the
223 //       period of time between consecutive evaluation of the should_start_gc() service.  Here's the thinking:
224 //
225 //       a) Between now and the next time I ask whether should_start_gc(), we might experience a spike representing
226 //          the anticipated burst of allocations.  If that would put us over budget, then we should start GC immediately.
227 //       b) Between now and the anticipated depletion of allocation pool, there may be two or more bursts of allocations.
228 //          If there are more than one of these bursts, we can "approximate" that these will be separated by spans of
229 //          time with very little or no allocations so the "average" allocation rate should be a suitable approximation
230 //          of how this will behave.
231 //
232 //    For cases 1 and 2, we need to "quickly" recalibrate the average allocation rate whenever we detect a change
233 //    in operation mode.  We want some way to decide that the average rate has changed, while keeping average
234 //    allocation rate computation independent.
235 bool ShenandoahAdaptiveHeuristics::should_start_gc() {

236   size_t capacity = _space_info->soft_max_capacity();
237   size_t available = _space_info->soft_available();
238   size_t allocated = _space_info->bytes_allocated_since_gc_start();
239 
240   log_debug(gc)("should_start_gc? available: " SIZE_FORMAT ", soft_max_capacity: " SIZE_FORMAT
241                 ", allocated: " SIZE_FORMAT, available, capacity, allocated);

242 
243   // Track allocation rate even if we decide to start a cycle for other reasons.
244   double rate = _allocation_rate.sample(allocated);
245   _last_trigger = OTHER;
246 
247   size_t min_threshold = min_free_threshold();
248   if (available < min_threshold) {
249     log_trigger("Free (" SIZE_FORMAT "%s) is below minimum threshold (" SIZE_FORMAT "%s)",
250                  byte_size_in_proper_unit(available), proper_unit_for_byte_size(available),
251                  byte_size_in_proper_unit(min_threshold), proper_unit_for_byte_size(min_threshold));
252     return true;
253   }
254 
255   // Check if we need to learn a bit about the application
256   const size_t max_learn = ShenandoahLearningSteps;
257   if (_gc_times_learned < max_learn) {
258     size_t init_threshold = capacity / 100 * ShenandoahInitFreeThreshold;
259     if (available < init_threshold) {
260       log_trigger("Learning " SIZE_FORMAT " of " SIZE_FORMAT ". Free (" SIZE_FORMAT "%s) is below initial threshold (" SIZE_FORMAT "%s)",
261                    _gc_times_learned + 1, max_learn,
262                    byte_size_in_proper_unit(available), proper_unit_for_byte_size(available),
263                    byte_size_in_proper_unit(init_threshold), proper_unit_for_byte_size(init_threshold));
264       return true;
265     }
266   }

267   // Check if allocation headroom is still okay. This also factors in:
268   //   1. Some space to absorb allocation spikes (ShenandoahAllocSpikeFactor)
269   //   2. Accumulated penalties from Degenerated and Full GC
270   size_t allocation_headroom = available;
271 
272   size_t spike_headroom = capacity / 100 * ShenandoahAllocSpikeFactor;
273   size_t penalties      = capacity / 100 * _gc_time_penalties;
274 
275   allocation_headroom -= MIN2(allocation_headroom, spike_headroom);
276   allocation_headroom -= MIN2(allocation_headroom, penalties);
277 
278   double avg_cycle_time = _gc_cycle_time_history->davg() + (_margin_of_error_sd * _gc_cycle_time_history->dsd());
279   double avg_alloc_rate = _allocation_rate.upper_bound(_margin_of_error_sd);
280 
281   log_debug(gc)("average GC time: %.2f ms, allocation rate: %.0f %s/s",
282           avg_cycle_time * 1000, byte_size_in_proper_unit(avg_alloc_rate), proper_unit_for_byte_size(avg_alloc_rate));
283   if (avg_cycle_time * avg_alloc_rate > allocation_headroom) {
284     log_trigger("Average GC time (%.2f ms) is above the time for average allocation rate (%.0f %sB/s)"
285                  " to deplete free headroom (" SIZE_FORMAT "%s) (margin of error = %.2f)",
286                  avg_cycle_time * 1000,
287                  byte_size_in_proper_unit(avg_alloc_rate), proper_unit_for_byte_size(avg_alloc_rate),
288                  byte_size_in_proper_unit(allocation_headroom), proper_unit_for_byte_size(allocation_headroom),
289                  _margin_of_error_sd);

290     log_info(gc, ergo)("Free headroom: " SIZE_FORMAT "%s (free) - " SIZE_FORMAT "%s (spike) - " SIZE_FORMAT "%s (penalties) = " SIZE_FORMAT "%s",
291                        byte_size_in_proper_unit(available),           proper_unit_for_byte_size(available),
292                        byte_size_in_proper_unit(spike_headroom),      proper_unit_for_byte_size(spike_headroom),
293                        byte_size_in_proper_unit(penalties),           proper_unit_for_byte_size(penalties),
294                        byte_size_in_proper_unit(allocation_headroom), proper_unit_for_byte_size(allocation_headroom));

295     _last_trigger = RATE;
296     return true;
297   }
298 
299   bool is_spiking = _allocation_rate.is_spiking(rate, _spike_threshold_sd);
300   if (is_spiking && avg_cycle_time > allocation_headroom / rate) {
301     log_trigger("Average GC time (%.2f ms) is above the time for instantaneous allocation rate (%.0f %sB/s) to deplete free headroom (" SIZE_FORMAT "%s) (spike threshold = %.2f)",
302                  avg_cycle_time * 1000,
303                  byte_size_in_proper_unit(rate), proper_unit_for_byte_size(rate),
304                  byte_size_in_proper_unit(allocation_headroom), proper_unit_for_byte_size(allocation_headroom),
305                  _spike_threshold_sd);
306     _last_trigger = SPIKE;
307     return true;
308   }
309 
310   return ShenandoahHeuristics::should_start_gc();
311 }
312 
313 void ShenandoahAdaptiveHeuristics::adjust_last_trigger_parameters(double amount) {
314   switch (_last_trigger) {
315     case RATE:
316       adjust_margin_of_error(amount);
317       break;
318     case SPIKE:
319       adjust_spike_threshold(amount);
320       break;
321     case OTHER:
322       // nothing to adjust here.
323       break;
324     default:
325       ShouldNotReachHere();
326   }
327 }
328 
329 void ShenandoahAdaptiveHeuristics::adjust_margin_of_error(double amount) {
330   _margin_of_error_sd = saturate(_margin_of_error_sd + amount, MINIMUM_CONFIDENCE, MAXIMUM_CONFIDENCE);
331   log_debug(gc, ergo)("Margin of error now %.2f", _margin_of_error_sd);
332 }
333 
334 void ShenandoahAdaptiveHeuristics::adjust_spike_threshold(double amount) {
335   _spike_threshold_sd = saturate(_spike_threshold_sd - amount, MINIMUM_CONFIDENCE, MAXIMUM_CONFIDENCE);
336   log_debug(gc, ergo)("Spike threshold now: %.2f", _spike_threshold_sd);
337 }
338 
339 size_t ShenandoahAdaptiveHeuristics::min_free_threshold() {
340   // Note that soft_max_capacity() / 100 * min_free_threshold is smaller than max_capacity() / 100 * min_free_threshold.
341   // We want to behave conservatively here, so use max_capacity().  By returning a larger value, we cause the GC to
342   // trigger when the remaining amount of free shrinks below the larger threshold.
343   return _space_info->max_capacity() / 100 * ShenandoahMinFreeThreshold;
344 }
345 
346 ShenandoahAllocationRate::ShenandoahAllocationRate() :
347   _last_sample_time(os::elapsedTime()),
348   _last_sample_value(0),
349   _interval_sec(1.0 / ShenandoahAdaptiveSampleFrequencyHz),
350   _rate(int(ShenandoahAdaptiveSampleSizeSeconds * ShenandoahAdaptiveSampleFrequencyHz), ShenandoahAdaptiveDecayFactor),
351   _rate_avg(int(ShenandoahAdaptiveSampleSizeSeconds * ShenandoahAdaptiveSampleFrequencyHz), ShenandoahAdaptiveDecayFactor) {
352 }
353 
354 double ShenandoahAllocationRate::sample(size_t allocated) {
355   double now = os::elapsedTime();
356   double rate = 0.0;
357   if (now - _last_sample_time > _interval_sec) {
358     if (allocated >= _last_sample_value) {
359       rate = instantaneous_rate(now, allocated);
360       _rate.add(rate);
361       _rate_avg.add(_rate.avg());
362     }
363 
364     _last_sample_time = now;
365     _last_sample_value = allocated;
< prev index next >