1 /*
  2  * Copyright Amazon.com Inc. 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 
 27 #include "gc/shenandoah/heuristics/shenandoahOldHeuristics.hpp"
 28 #include "gc/shenandoah/heuristics/shenandoahYoungHeuristics.hpp"
 29 #include "gc/shenandoah/shenandoahCollectorPolicy.hpp"
 30 #include "gc/shenandoah/shenandoahHeap.inline.hpp"
 31 #include "gc/shenandoah/shenandoahHeapRegion.inline.hpp"
 32 #include "gc/shenandoah/shenandoahYoungGeneration.hpp"
 33 
 34 #include "utilities/quickSort.hpp"
 35 
 36 ShenandoahYoungHeuristics::ShenandoahYoungHeuristics(ShenandoahYoungGeneration* generation)
 37         : ShenandoahGenerationalHeuristics(generation) {
 38 }
 39 
 40 
 41 void ShenandoahYoungHeuristics::choose_collection_set_from_regiondata(ShenandoahCollectionSet* cset,
 42                                                                       RegionData* data, size_t size,
 43                                                                       size_t actual_free) {
 44   // The logic for cset selection in adaptive is as follows:
 45   //
 46   //   1. We cannot get cset larger than available free space. Otherwise we guarantee OOME
 47   //      during evacuation, and thus guarantee full GC. In practice, we also want to let
 48   //      application to allocate something. This is why we limit CSet to some fraction of
 49   //      available space. In non-overloaded heap, max_cset would contain all plausible candidates
 50   //      over garbage threshold.
 51   //
 52   //   2. We should not get cset too low so that free threshold would not be met right
 53   //      after the cycle. Otherwise we get back-to-back cycles for no reason if heap is
 54   //      too fragmented. In non-overloaded non-fragmented heap min_garbage would be around zero.
 55   //
 56   // Therefore, we start by sorting the regions by garbage. Then we unconditionally add the best candidates
 57   // before we meet min_garbage. Then we add all candidates that fit with a garbage threshold before
 58   // we hit max_cset. When max_cset is hit, we terminate the cset selection. Note that in this scheme,
 59   // ShenandoahGarbageThreshold is the soft threshold which would be ignored until min_garbage is hit.
 60 
 61   // In generational mode, the sort order within the data array is not strictly descending amounts of garbage.  In
 62   // particular, regions that have reached tenure age will be sorted into this array before younger regions that contain
 63   // more garbage.  This represents one of the reasons why we keep looking at regions even after we decide, for example,
 64   // to exclude one of the regions because it might require evacuation of too much live data.
 65 
 66   // Better select garbage-first regions
 67   QuickSort::sort<RegionData>(data, (int) size, compare_by_garbage, false);
 68 
 69   size_t cur_young_garbage = add_preselected_regions_to_collection_set(cset, data, size);
 70 
 71   choose_young_collection_set(cset, data, size, actual_free, cur_young_garbage);
 72 
 73   log_cset_composition(cset);
 74 }
 75 
 76 void ShenandoahYoungHeuristics::choose_young_collection_set(ShenandoahCollectionSet* cset,
 77                                                             const RegionData* data,
 78                                                             size_t size, size_t actual_free,
 79                                                             size_t cur_young_garbage) const {
 80 
 81   ShenandoahHeap* heap = ShenandoahHeap::heap();
 82 
 83   size_t capacity = heap->young_generation()->max_capacity();
 84   size_t garbage_threshold = ShenandoahHeapRegion::region_size_bytes() * ShenandoahGarbageThreshold / 100;
 85   size_t ignore_threshold = ShenandoahHeapRegion::region_size_bytes() * ShenandoahIgnoreGarbageThreshold / 100;
 86   const uint tenuring_threshold = heap->age_census()->tenuring_threshold();
 87 
 88   // This is young-gen collection or a mixed evacuation.
 89   // If this is mixed evacuation, the old-gen candidate regions have already been added.
 90   size_t max_cset = (size_t) (heap->get_young_evac_reserve() / ShenandoahEvacWaste);
 91   size_t cur_cset = 0;
 92   size_t free_target = (capacity * ShenandoahMinFreeThreshold) / 100 + max_cset;
 93   size_t min_garbage = (free_target > actual_free) ? (free_target - actual_free) : 0;
 94 
 95 
 96   log_info(gc, ergo)(
 97           "Adaptive CSet Selection for YOUNG. Max Evacuation: " SIZE_FORMAT "%s, Actual Free: " SIZE_FORMAT "%s.",
 98           byte_size_in_proper_unit(max_cset), proper_unit_for_byte_size(max_cset),
 99           byte_size_in_proper_unit(actual_free), proper_unit_for_byte_size(actual_free));
100 
101   for (size_t idx = 0; idx < size; idx++) {
102     ShenandoahHeapRegion* r = data[idx]._region;
103     if (cset->is_preselected(r->index())) {
104       continue;
105     }
106     if (r->age() < tenuring_threshold) {
107       size_t new_cset = cur_cset + r->get_live_data_bytes();
108       size_t region_garbage = r->garbage();
109       size_t new_garbage = cur_young_garbage + region_garbage;
110       bool add_regardless = (region_garbage > ignore_threshold) && (new_garbage < min_garbage);
111       assert(r->is_young(), "Only young candidates expected in the data array");
112       if ((new_cset <= max_cset) && (add_regardless || (region_garbage > garbage_threshold))) {
113         cur_cset = new_cset;
114         cur_young_garbage = new_garbage;
115         cset->add_region(r);
116       }
117     }
118     // Note that we do not add aged regions if they were not pre-selected.  The reason they were not preselected
119     // is because there is not sufficient room in old-gen to hold their to-be-promoted live objects or because
120     // they are to be promoted in place.
121   }
122 }
123 
124 
125 bool ShenandoahYoungHeuristics::should_start_gc() {
126   ShenandoahHeap* heap = ShenandoahHeap::heap();
127   ShenandoahOldHeuristics* old_heuristics = heap->old_heuristics();
128 
129   // Checks that an old cycle has run for at least ShenandoahMinimumOldMarkTimeMs before allowing a young cycle.
130   if (ShenandoahMinimumOldMarkTimeMs > 0 && ShenandoahHeap::heap()->is_concurrent_old_mark_in_progress()) {
131     size_t old_mark_elapsed = size_t(old_heuristics->elapsed_cycle_time() * 1000);
132     if (old_mark_elapsed < ShenandoahMinimumOldMarkTimeMs) {
133       return false;
134     }
135   }
136 
137   // inherited triggers have already decided to start a cycle, so no further evaluation is required
138   if (ShenandoahAdaptiveHeuristics::should_start_gc()) {
139     return true;
140   }
141 
142   // Get through promotions and mixed evacuations as quickly as possible.  These cycles sometimes require significantly
143   // more time than traditional young-generation cycles so start them up as soon as possible.  This is a "mitigation"
144   // for the reality that old-gen and young-gen activities are not truly "concurrent".  If there is old-gen work to
145   // be done, we start up the young-gen GC threads so they can do some of this old-gen work.  As implemented, promotion
146   // gets priority over old-gen marking.
147   size_t promo_expedite_threshold = percent_of(heap->young_generation()->max_capacity(), ShenandoahExpeditePromotionsThreshold);
148   size_t promo_potential = heap->get_promotion_potential();
149   if (promo_potential > promo_expedite_threshold) {
150     // Detect unsigned arithmetic underflow
151     assert(promo_potential < heap->capacity(), "Sanity");
152     log_info(gc)("Trigger (%s): expedite promotion of " SIZE_FORMAT "%s",
153                  _space_info->name(),
154                  byte_size_in_proper_unit(promo_potential),
155                  proper_unit_for_byte_size(promo_potential));
156     return true;
157   }
158 
159   size_t mixed_candidates = old_heuristics->unprocessed_old_collection_candidates();
160   if (mixed_candidates > ShenandoahExpediteMixedThreshold && !heap->is_concurrent_weak_root_in_progress()) {
161     // We need to run young GC in order to open up some free heap regions so we can finish mixed evacuations.
162     // If concurrent weak root processing is in progress, it means the old cycle has chosen mixed collection
163     // candidates, but has not completed. There is no point in trying to start the young cycle before the old
164     // cycle completes.
165     log_info(gc)("Trigger (%s): expedite mixed evacuation of " SIZE_FORMAT " regions",
166                  _space_info->name(), mixed_candidates);
167     return true;
168   }
169 
170   return false;
171 }
172 
173 // Return a conservative estimate of how much memory can be allocated before we need to start GC. The estimate is based
174 // on memory that is currently available within young generation plus all of the memory that will be added to the young
175 // generation at the end of the current cycle (as represented by young_regions_to_be_reclaimed) and on the anticipated
176 // amount of time required to perform a GC.
177 size_t ShenandoahYoungHeuristics::bytes_of_allocation_runway_before_gc_trigger(size_t young_regions_to_be_reclaimed) {
178   size_t capacity = _space_info->soft_max_capacity();
179   size_t usage = _space_info->used();
180   size_t available = (capacity > usage)? capacity - usage: 0;
181   size_t allocated = _space_info->bytes_allocated_since_gc_start();
182 
183   size_t available_young_collected = ShenandoahHeap::heap()->collection_set()->get_young_available_bytes_collected();
184   size_t anticipated_available =
185           available + young_regions_to_be_reclaimed * ShenandoahHeapRegion::region_size_bytes() - available_young_collected;
186   size_t spike_headroom = capacity * ShenandoahAllocSpikeFactor / 100;
187   size_t penalties      = capacity * _gc_time_penalties / 100;
188 
189   double rate = _allocation_rate.sample(allocated);
190 
191   // At what value of available, would avg and spike triggers occur?
192   //  if allocation_headroom < avg_cycle_time * avg_alloc_rate, then we experience avg trigger
193   //  if allocation_headroom < avg_cycle_time * rate, then we experience spike trigger if is_spiking
194   //
195   // allocation_headroom =
196   //     0, if penalties > available or if penalties + spike_headroom > available
197   //     available - penalties - spike_headroom, otherwise
198   //
199   // so we trigger if available - penalties - spike_headroom < avg_cycle_time * avg_alloc_rate, which is to say
200   //                  available < avg_cycle_time * avg_alloc_rate + penalties + spike_headroom
201   //            or if available < penalties + spike_headroom
202   //
203   // since avg_cycle_time * avg_alloc_rate > 0, the first test is sufficient to test both conditions
204   //
205   // thus, evac_slack_avg is MIN2(0,  available - avg_cycle_time * avg_alloc_rate + penalties + spike_headroom)
206   //
207   // similarly, evac_slack_spiking is MIN2(0, available - avg_cycle_time * rate + penalties + spike_headroom)
208   // but evac_slack_spiking is only relevant if is_spiking, as defined below.
209 
210   double avg_cycle_time = _gc_cycle_time_history->davg() + (_margin_of_error_sd * _gc_cycle_time_history->dsd());
211 
212   // TODO: Consider making conservative adjustments to avg_cycle_time, such as: (avg_cycle_time *= 2) in cases where
213   // we expect a longer-than-normal GC duration.  This includes mixed evacuations, evacuation that perform promotion
214   // including promotion in place, and OLD GC bootstrap cycles.  It has been observed that these cycles sometimes
215   // require twice or more the duration of "normal" GC cycles.  We have experimented with this approach.  While it
216   // does appear to reduce the frequency of degenerated cycles due to late triggers, it also has the effect of reducing
217   // evacuation slack so that there is less memory available to be transferred to OLD.  The result is that we
218   // throttle promotion and it takes too long to move old objects out of the young generation.
219 
220   double avg_alloc_rate = _allocation_rate.upper_bound(_margin_of_error_sd);
221   size_t evac_slack_avg;
222   if (anticipated_available > avg_cycle_time * avg_alloc_rate + penalties + spike_headroom) {
223     evac_slack_avg = anticipated_available - (avg_cycle_time * avg_alloc_rate + penalties + spike_headroom);
224   } else {
225     // we have no slack because it's already time to trigger
226     evac_slack_avg = 0;
227   }
228 
229   bool is_spiking = _allocation_rate.is_spiking(rate, _spike_threshold_sd);
230   size_t evac_slack_spiking;
231   if (is_spiking) {
232     if (anticipated_available > avg_cycle_time * rate + penalties + spike_headroom) {
233       evac_slack_spiking = anticipated_available - (avg_cycle_time * rate + penalties + spike_headroom);
234     } else {
235       // we have no slack because it's already time to trigger
236       evac_slack_spiking = 0;
237     }
238   } else {
239     evac_slack_spiking = evac_slack_avg;
240   }
241 
242   size_t threshold = min_free_threshold();
243   size_t evac_min_threshold = (anticipated_available > threshold)? anticipated_available - threshold: 0;
244   return MIN3(evac_slack_spiking, evac_slack_avg, evac_min_threshold);
245 }
246