< prev index next >

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

Print this page

 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() :
 58   ShenandoahHeuristics(),
 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    = ShenandoahHeap::heap()->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<RegionData>(data, (int)size, compare_by_garbage, false);
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 = ShenandoahHeap::heap()->free_set()->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

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   ShenandoahHeap* heap = ShenandoahHeap::heap();
200   size_t max_capacity = heap->max_capacity();
201   size_t capacity = heap->soft_max_capacity();
202   size_t available = heap->free_set()->available();
203   size_t allocated = heap->bytes_allocated_since_gc_start();



204 
205   // Make sure the code below treats available without the soft tail.
206   size_t soft_tail = max_capacity - capacity;
207   available = (available > soft_tail) ? (available - soft_tail) : 0;
208 
209   // Track allocation rate even if we decide to start a cycle for other reasons.
210   double rate = _allocation_rate.sample(allocated);
211   _last_trigger = OTHER;
212 
213   size_t min_threshold = capacity / 100 * ShenandoahMinFreeThreshold;

214   if (available < min_threshold) {
215     log_info(gc)("Trigger: Free (" SIZE_FORMAT "%s) is below minimum threshold (" SIZE_FORMAT "%s)",

216                  byte_size_in_proper_unit(available),     proper_unit_for_byte_size(available),
217                  byte_size_in_proper_unit(min_threshold), proper_unit_for_byte_size(min_threshold));
218     return true;
219   }
220 

221   const size_t max_learn = ShenandoahLearningSteps;
222   if (_gc_times_learned < max_learn) {
223     size_t init_threshold = capacity / 100 * ShenandoahInitFreeThreshold;
224     if (available < init_threshold) {
225       log_info(gc)("Trigger: Learning " SIZE_FORMAT " of " SIZE_FORMAT ". Free (" SIZE_FORMAT "%s) is below initial threshold (" SIZE_FORMAT "%s)",
226                    _gc_times_learned + 1, max_learn,
227                    byte_size_in_proper_unit(available),      proper_unit_for_byte_size(available),
228                    byte_size_in_proper_unit(init_threshold), proper_unit_for_byte_size(init_threshold));
229       return true;
230     }
231   }
232 
233   // Check if allocation headroom is still okay. This also factors in:
234   //   1. Some space to absorb allocation spikes
235   //   2. Accumulated penalties from Degenerated and Full GC
236   size_t allocation_headroom = available;
237 













































238   size_t spike_headroom = capacity / 100 * ShenandoahAllocSpikeFactor;
239   size_t penalties      = capacity / 100 * _gc_time_penalties;
240 









241   allocation_headroom -= MIN2(allocation_headroom, spike_headroom);
242   allocation_headroom -= MIN2(allocation_headroom, penalties);
243 
244   double avg_cycle_time = _gc_time_history->davg() + (_margin_of_error_sd * _gc_time_history->dsd());



















245   double avg_alloc_rate = _allocation_rate.upper_bound(_margin_of_error_sd);



246   if (avg_cycle_time > allocation_headroom / avg_alloc_rate) {
247     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)",
248                  avg_cycle_time * 1000,





249                  byte_size_in_proper_unit(avg_alloc_rate), proper_unit_for_byte_size(avg_alloc_rate),
250                  byte_size_in_proper_unit(allocation_headroom), proper_unit_for_byte_size(allocation_headroom),
251                  _margin_of_error_sd);
252 
253     log_info(gc, ergo)("Free headroom: " SIZE_FORMAT "%s (free) - " SIZE_FORMAT "%s (spike) - " SIZE_FORMAT "%s (penalties) = " SIZE_FORMAT "%s",
254                        byte_size_in_proper_unit(available),           proper_unit_for_byte_size(available),
255                        byte_size_in_proper_unit(spike_headroom),      proper_unit_for_byte_size(spike_headroom),
256                        byte_size_in_proper_unit(penalties),           proper_unit_for_byte_size(penalties),
257                        byte_size_in_proper_unit(allocation_headroom), proper_unit_for_byte_size(allocation_headroom));
258 
259     _last_trigger = RATE;
260     return true;
261   }
262 
263   bool is_spiking = _allocation_rate.is_spiking(rate, _spike_threshold_sd);
264   if (is_spiking && avg_cycle_time > allocation_headroom / rate) {
265     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)",
266                  avg_cycle_time * 1000,
267                  byte_size_in_proper_unit(rate), proper_unit_for_byte_size(rate),
268                  byte_size_in_proper_unit(allocation_headroom), proper_unit_for_byte_size(allocation_headroom),

269                  _spike_threshold_sd);
270     _last_trigger = SPIKE;
271     return true;
272   }
273 
274   return ShenandoahHeuristics::should_start_gc();
275 }
276 
277 void ShenandoahAdaptiveHeuristics::adjust_last_trigger_parameters(double amount) {
278   switch (_last_trigger) {
279     case RATE:
280       adjust_margin_of_error(amount);
281       break;
282     case SPIKE:
283       adjust_spike_threshold(amount);
284       break;
285     case OTHER:
286       // nothing to adjust here.
287       break;
288     default:

 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/shenandoahGeneration.hpp"
 31 #include "gc/shenandoah/shenandoahHeapRegion.inline.hpp"
 32 #include "gc/shenandoah/shenandoahYoungGeneration.hpp"
 33 #include "logging/log.hpp"
 34 #include "logging/logTag.hpp"
 35 #include "utilities/quickSort.hpp"
 36 
 37 // These constants are used to adjust the margin of error for the moving
 38 // average of the allocation rate and cycle time. The units are standard
 39 // deviations.
 40 const double ShenandoahAdaptiveHeuristics::FULL_PENALTY_SD = 0.2;
 41 const double ShenandoahAdaptiveHeuristics::DEGENERATE_PENALTY_SD = 0.1;
 42 
 43 // These are used to decide if we want to make any adjustments at all
 44 // at the end of a successful concurrent cycle.
 45 const double ShenandoahAdaptiveHeuristics::LOWEST_EXPECTED_AVAILABLE_AT_END = -0.5;
 46 const double ShenandoahAdaptiveHeuristics::HIGHEST_EXPECTED_AVAILABLE_AT_END = 0.5;
 47 
 48 // These values are the confidence interval expressed as standard deviations.
 49 // At the minimum confidence level, there is a 25% chance that the true value of
 50 // the estimate (average cycle time or allocation rate) is not more than
 51 // MINIMUM_CONFIDENCE standard deviations away from our estimate. Similarly, the
 52 // MAXIMUM_CONFIDENCE interval here means there is a one in a thousand chance
 53 // that the true value of our estimate is outside the interval. These are used
 54 // as bounds on the adjustments applied at the outcome of a GC cycle.
 55 const double ShenandoahAdaptiveHeuristics::MINIMUM_CONFIDENCE = 0.319; // 25%
 56 const double ShenandoahAdaptiveHeuristics::MAXIMUM_CONFIDENCE = 3.291; // 99.9%
 57 
 58 ShenandoahAdaptiveHeuristics::ShenandoahAdaptiveHeuristics(ShenandoahGeneration* generation) :
 59   ShenandoahHeuristics(generation),
 60   _margin_of_error_sd(ShenandoahAdaptiveInitialConfidence),
 61   _spike_threshold_sd(ShenandoahAdaptiveInitialSpikeThreshold),
 62   _last_trigger(OTHER) { }
 63 
 64 ShenandoahAdaptiveHeuristics::~ShenandoahAdaptiveHeuristics() {}
 65 
 66 void ShenandoahAdaptiveHeuristics::choose_collection_set_from_regiondata(ShenandoahCollectionSet* cset,
 67                                                                          RegionData* data, size_t size,
 68                                                                          size_t actual_free) {
 69   size_t garbage_threshold = ShenandoahHeapRegion::region_size_bytes() * ShenandoahGarbageThreshold / 100;
 70 
 71   // The logic for cset selection in adaptive is as follows:
 72   //
 73   //   1. We cannot get cset larger than available free space. Otherwise we guarantee OOME
 74   //      during evacuation, and thus guarantee full GC. In practice, we also want to let
 75   //      application to allocate something. This is why we limit CSet to some fraction of
 76   //      available space. In non-overloaded heap, max_cset would contain all plausible candidates
 77   //      over garbage threshold.
 78   //
 79   //   2. We should not get cset too low so that free threshold would not be met right
 80   //      after the cycle. Otherwise we get back-to-back cycles for no reason if heap is
 81   //      too fragmented. In non-overloaded non-fragmented heap min_garbage would be around zero.
 82   //
 83   // Therefore, we start by sorting the regions by garbage. Then we unconditionally add the best candidates
 84   // before we meet min_garbage. Then we add all candidates that fit with a garbage threshold before
 85   // we hit max_cset. When max_cset is hit, we terminate the cset selection. Note that in this scheme,
 86   // ShenandoahGarbageThreshold is the soft threshold which would be ignored until min_garbage is hit.
 87 
 88   size_t max_cset    = (ShenandoahHeap::heap()->get_young_evac_reserve() / ShenandoahEvacWaste);
 89   size_t capacity    = ShenandoahHeap::heap()->young_generation()->soft_max_capacity();


 90 
 91   // As currently implemented, we are not enforcing that new_garbage > min_garbage
 92   // size_t free_target = (capacity / 100) * ShenandoahMinFreeThreshold + max_cset;
 93   // size_t min_garbage = (free_target > actual_free ? (free_target - actual_free) : 0);
 94 
 95   log_info(gc, ergo)("Adaptive CSet Selection. Max CSet: " SIZE_FORMAT "%s, Actual Free: " SIZE_FORMAT "%s.",
 96                      byte_size_in_proper_unit(max_cset),    proper_unit_for_byte_size(max_cset),
 97                      byte_size_in_proper_unit(actual_free), proper_unit_for_byte_size(actual_free));
 98 
 99   // Better select garbage-first regions
100   QuickSort::sort<RegionData>(data, (int)size, compare_by_garbage, false);
101 
102   size_t cur_cset = 0;
103   // size_t cur_garbage = 0;
104 
105   // In generational mode, the sort order within the data array is not strictly descending amounts of garbage.  In
106   // particular, regions that have reached tenure age will be sorted into this array before younger regions that contain
107   // more garbage.  This represents one of the reasons why we keep looking at regions even after we decide, for example,
108   // to exclude one of the regions because it might require evacuation of too much live data.
109 
110   for (size_t idx = 0; idx < size; idx++) {
111     ShenandoahHeapRegion* r = data[idx]._region;
112     size_t biased_garbage = data[idx]._garbage;
113 
114     size_t new_cset    = cur_cset + r->get_live_data_bytes();

115 
116     // As currently implemented, we are not enforcing that new_garbage > min_garbage
117     // size_t new_garbage = cur_garbage + r->garbage();
118 
119     // Note that live data bytes within a region is not the same as heap_region_size - garbage.  This is because
120     // each region contains a combination of used memory (which is garbage plus live) and unused memory, which has not
121     // yet been allocated.  It may be the case that the region on this iteration has too much live data to be added to
122     // the collection set while one or more regions seen on subsequent iterations of this loop can be added to the collection
123     // set because they have smaller live memory, even though they also have smaller garbage (and necessarily a larger
124     // amount of unallocated memory).
125 
126     // BANDAID: In an earlier version of this code, this was written:
127     //   if ((new_cset <= max_cset) && ((new_garbage < min_garbage) || (r->garbage() > garbage_threshold)))
128     // The problem with the original code is that in some cases the collection set would include hundreds of regions,
129     // each with less than 100 bytes of garbage.  Evacuating these regions is counterproductive.
130 
131     // TODO: Think about changing the description and defaults for ShenandoahGarbageThreshold and ShenandoahMinFreeThreshold.
132     // If "customers" want to evacuate regions with smaller amounts of garbage contained therein, they should specify a lower
133     // value of ShenandoahGarbageThreshold.  As implemented currently, we may experience back-to-back collections if there is
134     // not enough memory to be reclaimed.  Let's not let pursuit of min_garbage drive us to make poor decisions.  Maybe we
135     // want yet another global parameter to allow a region to be placed into the collection set if
136     // (((new_garbage < min_garbage) && (r->garbage() > ShenandoahSmallerGarbageThreshold)) || (r->garbage() > garbage_threshold))
137 
138     if ((new_cset <= max_cset) && ((r->garbage() > garbage_threshold) || (r->age() >= InitialTenuringThreshold))) {
139       cset->add_region(r);
140       cur_cset = new_cset;
141       // cur_garbage = new_garbage;
142     } else if (biased_garbage == 0) {
143       break;
144     }
145   }
146 }
147 
148 void ShenandoahAdaptiveHeuristics::record_cycle_start() {
149   ShenandoahHeuristics::record_cycle_start();
150   _allocation_rate.allocation_counter_reset();
151 }
152 
153 void ShenandoahAdaptiveHeuristics::record_success_concurrent(bool abbreviated) {
154   ShenandoahHeuristics::record_success_concurrent(abbreviated);
155 
156   size_t available = ShenandoahHeap::heap()->free_set()->available();
157 
158   _available.add(available);
159   double z_score = 0.0;
160   if (_available.sd() > 0) {
161     z_score = (available - _available.avg()) / _available.sd();
162   }
163 
164   log_debug(gc, ergo)("Available: " SIZE_FORMAT " %sB, z-score=%.3f. Average available: %.1f %sB +/- %.1f %sB.",
165                       byte_size_in_proper_unit(available), proper_unit_for_byte_size(available),
166                       z_score,
167                       byte_size_in_proper_unit(_available.avg()), proper_unit_for_byte_size(_available.avg()),
168                       byte_size_in_proper_unit(_available.sd()), proper_unit_for_byte_size(_available.sd()));
169 
170   // In the case when a concurrent GC cycle completes successfully but with an
171   // unusually small amount of available memory we will adjust our trigger
172   // parameters so that they are more likely to initiate a new cycle.
173   // Conversely, when a GC cycle results in an above average amount of available
174   // memory, we will adjust the trigger parameters to be less likely to initiate

204   ShenandoahHeuristics::record_success_degenerated();
205   // Adjust both trigger's parameters in the case of a degenerated GC because
206   // either of them should have triggered earlier to avoid this case.
207   adjust_margin_of_error(DEGENERATE_PENALTY_SD);
208   adjust_spike_threshold(DEGENERATE_PENALTY_SD);
209 }
210 
211 void ShenandoahAdaptiveHeuristics::record_success_full() {
212   ShenandoahHeuristics::record_success_full();
213   // Adjust both trigger's parameters in the case of a full GC because
214   // either of them should have triggered earlier to avoid this case.
215   adjust_margin_of_error(FULL_PENALTY_SD);
216   adjust_spike_threshold(FULL_PENALTY_SD);
217 }
218 
219 static double saturate(double value, double min, double max) {
220   return MAX2(MIN2(value, max), min);
221 }
222 
223 bool ShenandoahAdaptiveHeuristics::should_start_gc() {
224   size_t max_capacity = _generation->max_capacity();
225   size_t capacity = _generation->soft_max_capacity();
226   size_t available = _generation->available();
227   size_t allocated = _generation->bytes_allocated_since_gc_start();
228 
229   log_debug(gc)("should_start_gc (%s)? available: " SIZE_FORMAT ", soft_max_capacity: " SIZE_FORMAT
230                 ", max_capacity: " SIZE_FORMAT ", allocated: " SIZE_FORMAT,
231                 _generation->name(), available, capacity, max_capacity, allocated);
232 
233   // Make sure the code below treats available without the soft tail.
234   size_t soft_tail = max_capacity - capacity;
235   available = (available > soft_tail) ? (available - soft_tail) : 0;
236 
237   // Track allocation rate even if we decide to start a cycle for other reasons.
238   double rate = _allocation_rate.sample(allocated);
239   _last_trigger = OTHER;
240 
241   size_t min_threshold = capacity / 100 * ShenandoahMinFreeThreshold;
242 
243   if (available < min_threshold) {
244     log_info(gc)("Trigger (%s): Free (" SIZE_FORMAT "%s) is below minimum threshold (" SIZE_FORMAT "%s)",
245                  _generation->name(),
246                  byte_size_in_proper_unit(available),     proper_unit_for_byte_size(available),
247                  byte_size_in_proper_unit(min_threshold), proper_unit_for_byte_size(min_threshold));
248     return true;
249   }
250 
251   // Check if we need to learn a bit about the application
252   const size_t max_learn = ShenandoahLearningSteps;
253   if (_gc_times_learned < max_learn) {
254     size_t init_threshold = capacity / 100 * ShenandoahInitFreeThreshold;
255     if (available < init_threshold) {
256       log_info(gc)("Trigger (%s): Learning " SIZE_FORMAT " of " SIZE_FORMAT ". Free (" SIZE_FORMAT "%s) is below initial threshold (" SIZE_FORMAT "%s)",
257                    _generation->name(), _gc_times_learned + 1, max_learn,
258                    byte_size_in_proper_unit(available),      proper_unit_for_byte_size(available),
259                    byte_size_in_proper_unit(init_threshold), proper_unit_for_byte_size(init_threshold));
260       return true;
261     }
262   }
263 
264   // Check if allocation headroom is still okay. This also factors in:
265   //   1. Some space to absorb allocation spikes
266   //   2. Accumulated penalties from Degenerated and Full GC
267   size_t allocation_headroom = available;
268 
269   // ShenandoahAllocSpikeFactor is the percentage of capacity that we endeavor to assure to be free at the end of the GC
270   // cycle.
271   // TODO: Correct the representation of this quantity
272   //       (and dive deeper into _gc_time_penalties as this may also need to be corrected)
273   //
274   //       Allocation spikes are a characteristic of both the application ahd the JVM configuration.  On the JVM command line,
275   //       the application developer may want to supply a hint of the nature of spikes that are inherent in the application
276   //       workload, and this information would normally be independent of heap size (not a percentage thereof).  On the
277   //       other hand, some allocation spikes are correlated with JVM configuration.  For example, there are allocation
278   //       spikes at the starts of concurrent marking and evacuation to refresh all local allocation buffers.  The nature
279   //       of these spikes is determined by LAB min and max sizes and numbers of threads, but also on frequency of GC passes,
280   //       and on "periodic" behavior of these threads  If GC frequency is much higher than the periodic trigger for mutator
281   //       threads, then many of the mutator threads may be able to "sit out" of most GC passes.  Though the thread's stack
282   //       must be scanned, the thread does not need to refresh its LABs if it sits idle throughout the duration of the GC
283   //       pass.  The best prediction for this aspect of spikes in allocation patterns is probably recent past history.
284   //
285   //  Rationale:
286   //    The idea is that there is an average allocation rate and there are occasional abnormal bursts (or spikes) of
287   //    allocations that exceed the average allocation rate.  What do these spikes look like?
288   //
289   //    1. At certain phase changes, we may discard large amounts of data and replace it with large numbers of newly
290   //       allocated objects.  This "spike" looks more like a phase change.  We were in steady state at M bytes/sec
291   //       allocation rate and now we're in a "reinitialization phase" that looks like N bytes/sec.  We need the "spike"
292   //       accomodation to give us enough runway to recalibrate our "average allocation rate".
293   //
294   //   2. The typical workload changes.  "Suddenly", our typical workload of N TPS increases to N+delta TPS.  This means
295   //       our average allocation rate needs to be adjusted.  Once again, we need the "spike" accomodation to give us
296   //       enough runway to recalibrate our "average allocation rate".
297   //
298   //    3. Though there is an "average" allocation rate, a given workload's demand for allocation may be very bursty.  We
299   //       allocate a bunch of LABs during the 5 ms that follow completion of a GC, then we perform no more allocations for
300   //       the next 150 ms.  It seems we want the "spike" to represent the maximum divergence from average within the
301   //       period of time between consecutive evaluation of the should_start_gc() service.  Here's the thinking:
302   //
303   //       a) Between now and the next time I ask whether should_start_gc(), we might experience a spike representing
304   //          the anticipated burst of allocations.  If that would put us over budget, then we should start GC immediately.
305   //       b) Between now and the anticipated depletion of allocation pool, there may be two or more bursts of allocations.
306   //          If there are more than one of these bursts, we can "approximate" that these will be separated by spans of
307   //          time with very little or no allocations so the "average" allocation rate should be a suitable approximation
308   //          of how this will behave.
309   //
310   //    For cases 1 and 2, we need to "quickly" recalibrate the average allocation rate whenever we detect a change
311   //    in operation mode.  We want some way to decide that the average rate has changed.  Make average allocation rate
312   //    computations an independent effort.
313 
314   size_t spike_headroom = capacity / 100 * ShenandoahAllocSpikeFactor;
315   size_t penalties      = capacity / 100 * _gc_time_penalties;
316 
317   // TODO: Account for inherent delays in responding to GC triggers
318   //  1. It has been observed that delays of 200 ms or greater are common between the moment we return true from should_start_gc()
319   //     and the moment at which we begin execution of the concurrent reset phase.  Add this time into the calculation of
320   //     avg_cycle_time below.  (What is "this time"?  Perhaps we should remember recent history of this delay for the
321   //     running workload and use the maximum delay recently seen for "this time".)
322   //  2. The frequency of inquiries to should_start_gc() is adaptive, ranging between ShenandoahControlIntervalMin and
323   //     ShenandoahControlIntervalMax.  The current control interval (or the max control interval) should also be added into
324   //     the calculation of avg_cycle_time below.
325 
326   allocation_headroom -= MIN2(allocation_headroom, spike_headroom);
327   allocation_headroom -= MIN2(allocation_headroom, penalties);
328 
329   double avg_cycle_time = _gc_time_history->davg() + (_margin_of_error_sd * _gc_time_history->dsd());
330 
331   size_t last_live_memory = get_last_live_memory();
332   size_t penultimate_live_memory = get_penultimate_live_memory();
333   double original_cycle_time = avg_cycle_time;
334   if ((penultimate_live_memory < last_live_memory) && (penultimate_live_memory != 0)) {
335     // If the live-memory size is growing, our estimates of cycle time are based on lighter workload, so adjust.
336     // TODO: Be more precise about how to scale when live memory is growing.  Existing code is a very rough approximation
337     // tuned with very limited workload observations.
338     avg_cycle_time = (avg_cycle_time * 2 * last_live_memory) / penultimate_live_memory;
339   } else {
340     int degen_cycles = degenerated_cycles_in_a_row();
341     if (degen_cycles > 0) {
342       // If we've degenerated recently, we might be waiting too long between triggers so adjust trigger forward.
343       // TODO: Be more precise about how to scale when we've experienced recent degenerated GC.  Existing code is a very
344       // rough approximation tuned with very limited workload observations.
345       avg_cycle_time += degen_cycles * avg_cycle_time;
346     }
347   }
348 
349   double avg_alloc_rate = _allocation_rate.upper_bound(_margin_of_error_sd);
350   log_debug(gc)("%s: average GC time: %.2f ms, allocation rate: %.0f %s/s",
351     _generation->name(), avg_cycle_time * 1000, byte_size_in_proper_unit(avg_alloc_rate), proper_unit_for_byte_size(avg_alloc_rate));
352 
353   if (avg_cycle_time > allocation_headroom / avg_alloc_rate) {
354     if (avg_cycle_time > original_cycle_time) {
355       log_debug(gc)("%s: average GC time adjusted from: %.2f ms to %.2f ms because upward trend in live memory retention",
356                     _generation->name(), original_cycle_time, avg_cycle_time);
357     }
358 
359     log_info(gc)("Trigger (%s): 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)",
360                  _generation->name(), avg_cycle_time * 1000,
361                  byte_size_in_proper_unit(avg_alloc_rate), proper_unit_for_byte_size(avg_alloc_rate),
362                  byte_size_in_proper_unit(allocation_headroom), proper_unit_for_byte_size(allocation_headroom),
363                  _margin_of_error_sd);
364 
365     log_info(gc, ergo)("Free headroom: " SIZE_FORMAT "%s (free) - " SIZE_FORMAT "%s (spike) - " SIZE_FORMAT "%s (penalties) = " SIZE_FORMAT "%s",
366                        byte_size_in_proper_unit(available),           proper_unit_for_byte_size(available),
367                        byte_size_in_proper_unit(spike_headroom),      proper_unit_for_byte_size(spike_headroom),
368                        byte_size_in_proper_unit(penalties),           proper_unit_for_byte_size(penalties),
369                        byte_size_in_proper_unit(allocation_headroom), proper_unit_for_byte_size(allocation_headroom));
370 
371     _last_trigger = RATE;
372     return true;
373   }
374 
375   bool is_spiking = _allocation_rate.is_spiking(rate, _spike_threshold_sd);
376   if (is_spiking && avg_cycle_time > allocation_headroom / rate) {
377     log_info(gc)("Trigger (%s): 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)",
378                  _generation->name(), avg_cycle_time * 1000,
379                  byte_size_in_proper_unit(rate), proper_unit_for_byte_size(rate),
380                  byte_size_in_proper_unit(allocation_headroom), proper_unit_for_byte_size(allocation_headroom),
381 
382                  _spike_threshold_sd);
383     _last_trigger = SPIKE;
384     return true;
385   }
386 
387   return ShenandoahHeuristics::should_start_gc();
388 }
389 
390 void ShenandoahAdaptiveHeuristics::adjust_last_trigger_parameters(double amount) {
391   switch (_last_trigger) {
392     case RATE:
393       adjust_margin_of_error(amount);
394       break;
395     case SPIKE:
396       adjust_spike_threshold(amount);
397       break;
398     case OTHER:
399       // nothing to adjust here.
400       break;
401     default:
< prev index next >