1 /*
2 * Copyright Amazon.com Inc. or its affiliates. All Rights Reserved.
3 * Copyright (c) 2025, Oracle and/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 "gc/shenandoah/heuristics/shenandoahOldHeuristics.hpp"
27 #include "gc/shenandoah/heuristics/shenandoahYoungHeuristics.hpp"
28 #include "gc/shenandoah/shenandoahCollectorPolicy.hpp"
29 #include "gc/shenandoah/shenandoahGenerationalHeap.hpp"
30 #include "gc/shenandoah/shenandoahHeapRegion.inline.hpp"
31 #include "gc/shenandoah/shenandoahOldGeneration.hpp"
32 #include "gc/shenandoah/shenandoahYoungGeneration.hpp"
33 #include "utilities/quickSort.hpp"
34
35 ShenandoahYoungHeuristics::ShenandoahYoungHeuristics(ShenandoahYoungGeneration* generation)
36 : ShenandoahGenerationalHeuristics(generation) {
37 }
38
39
40 void ShenandoahYoungHeuristics::choose_collection_set_from_regiondata(ShenandoahCollectionSet* cset,
41 RegionData* data, size_t size,
42 size_t actual_free) {
43 // See comments in ShenandoahAdaptiveHeuristics::choose_collection_set_from_regiondata():
44 // we do the same here, but with the following adjustments for generational mode:
45 //
46 // In generational mode, the sort order within the data array is not strictly descending amounts
47 // of garbage. In particular, regions that have reached tenure age will be sorted into this
48 // array before younger regions that typically contain more garbage. This is one reason why,
49 // for example, we continue examining regions even after rejecting a region that has
50 // more live data than we can evacuate.
51
52 // Better select garbage-first regions
53 QuickSort::sort<RegionData>(data, (int) size, compare_by_garbage);
54
55 size_t cur_young_garbage = add_preselected_regions_to_collection_set(cset, data, size);
56
57 choose_young_collection_set(cset, data, size, actual_free, cur_young_garbage);
58
59 log_cset_composition(cset);
60 }
61
62 void ShenandoahYoungHeuristics::choose_young_collection_set(ShenandoahCollectionSet* cset,
63 const RegionData* data,
64 size_t size, size_t actual_free,
65 size_t cur_young_garbage) const {
66
67 auto heap = ShenandoahGenerationalHeap::heap();
68
69 size_t capacity = heap->soft_max_capacity();
70 size_t garbage_threshold = ShenandoahHeapRegion::region_size_bytes() * ShenandoahGarbageThreshold / 100;
71 size_t ignore_threshold = ShenandoahHeapRegion::region_size_bytes() * ShenandoahIgnoreGarbageThreshold / 100;
72 const uint tenuring_threshold = heap->age_census()->tenuring_threshold();
73
74 // This is young-gen collection or a mixed evacuation.
75 // If this is mixed evacuation, the old-gen candidate regions have already been added.
76 size_t max_cset = (size_t) (heap->young_generation()->get_evacuation_reserve() / ShenandoahEvacWaste);
77 size_t cur_cset = 0;
78 size_t free_target = (capacity * ShenandoahMinFreeThreshold) / 100 + max_cset;
79 size_t min_garbage = (free_target > actual_free) ? (free_target - actual_free) : 0;
80
81
82 log_info(gc, ergo)(
83 "Adaptive CSet Selection for YOUNG. Max Evacuation: %zu%s, Actual Free: %zu%s.",
84 byte_size_in_proper_unit(max_cset), proper_unit_for_byte_size(max_cset),
85 byte_size_in_proper_unit(actual_free), proper_unit_for_byte_size(actual_free));
86
87 for (size_t idx = 0; idx < size; idx++) {
88 ShenandoahHeapRegion* r = data[idx].get_region();
89 if (cset->is_preselected(r->index())) {
90 continue;
91 }
92 if (r->age() < tenuring_threshold) {
93 size_t new_cset = cur_cset + r->get_live_data_bytes();
94 size_t region_garbage = r->garbage();
95 size_t new_garbage = cur_young_garbage + region_garbage;
96 bool add_regardless = (region_garbage > ignore_threshold) && (new_garbage < min_garbage);
97 assert(r->is_young(), "Only young candidates expected in the data array");
98 if ((new_cset <= max_cset) && (add_regardless || (region_garbage > garbage_threshold))) {
99 cur_cset = new_cset;
100 cur_young_garbage = new_garbage;
101 cset->add_region(r);
102 }
103 }
104 // Note that we do not add aged regions if they were not pre-selected. The reason they were not preselected
105 // is because there is not sufficient room in old-gen to hold their to-be-promoted live objects or because
106 // they are to be promoted in place.
107 }
108 }
109
110
111 bool ShenandoahYoungHeuristics::should_start_gc() {
112 auto heap = ShenandoahGenerationalHeap::heap();
113 ShenandoahOldGeneration* old_generation = heap->old_generation();
114 ShenandoahOldHeuristics* old_heuristics = old_generation->heuristics();
115
116 // Checks that an old cycle has run for at least ShenandoahMinimumOldTimeMs before allowing a young cycle.
117 if (ShenandoahMinimumOldTimeMs > 0) {
118 if (old_generation->is_preparing_for_mark() || old_generation->is_concurrent_mark_in_progress()) {
119 size_t old_time_elapsed = size_t(old_heuristics->elapsed_cycle_time() * 1000);
120 if (old_time_elapsed < ShenandoahMinimumOldTimeMs) {
121 // Do not decline_trigger() when waiting for minimum quantum of Old-gen marking. It is not at our discretion
122 // to trigger at this time.
123 log_debug(gc)("Young heuristics declines to trigger because old_time_elapsed < ShenandoahMinimumOldTimeMs");
124 return false;
125 }
126 }
127 }
128
129 // inherited triggers have already decided to start a cycle, so no further evaluation is required
130 if (ShenandoahAdaptiveHeuristics::should_start_gc()) {
131 return true;
132 }
133
134 // Get through promotions and mixed evacuations as quickly as possible. These cycles sometimes require significantly
135 // more time than traditional young-generation cycles so start them up as soon as possible. This is a "mitigation"
136 // for the reality that old-gen and young-gen activities are not truly "concurrent". If there is old-gen work to
137 // be done, we start up the young-gen GC threads so they can do some of this old-gen work. As implemented, promotion
138 // gets priority over old-gen marking.
139 size_t promo_expedite_threshold = percent_of(heap->young_generation()->max_capacity(), ShenandoahExpeditePromotionsThreshold);
140 size_t promo_potential = old_generation->get_promotion_potential();
141 if (promo_potential > promo_expedite_threshold) {
142 // Detect unsigned arithmetic underflow
143 assert(promo_potential < heap->capacity(), "Sanity");
144 log_trigger("Expedite promotion of " PROPERFMT, PROPERFMTARGS(promo_potential));
145 accept_trigger();
146 return true;
147 }
148
149 size_t mixed_candidates = old_heuristics->unprocessed_old_collection_candidates();
150 if (mixed_candidates > ShenandoahExpediteMixedThreshold && !heap->is_concurrent_weak_root_in_progress()) {
151 // We need to run young GC in order to open up some free heap regions so we can finish mixed evacuations.
152 // If concurrent weak root processing is in progress, it means the old cycle has chosen mixed collection
153 // candidates, but has not completed. There is no point in trying to start the young cycle before the old
154 // cycle completes.
155 log_trigger("Expedite mixed evacuation of %zu regions", mixed_candidates);
156 accept_trigger();
157 return true;
158 }
159
160 // Don't decline_trigger() here That was done in ShenandoahAdaptiveHeuristics::should_start_gc()
161 return false;
162 }
163
164 // Return a conservative estimate of how much memory can be allocated before we need to start GC. The estimate is based
165 // on memory that is currently available within young generation plus all of the memory that will be added to the young
166 // generation at the end of the current cycle (as represented by young_regions_to_be_reclaimed) and on the anticipated
167 // amount of time required to perform a GC.
168 size_t ShenandoahYoungHeuristics::bytes_of_allocation_runway_before_gc_trigger(size_t young_regions_to_be_reclaimed) {
169 size_t capacity = _space_info->max_capacity();
170 size_t usage = _space_info->used();
171 size_t available = (capacity > usage)? capacity - usage: 0;
172 size_t allocated = _space_info->bytes_allocated_since_gc_start();
173
174 size_t available_young_collected = ShenandoahHeap::heap()->collection_set()->get_young_available_bytes_collected();
175 size_t anticipated_available =
176 available + young_regions_to_be_reclaimed * ShenandoahHeapRegion::region_size_bytes() - available_young_collected;
177 size_t spike_headroom = capacity * ShenandoahAllocSpikeFactor / 100;
178 size_t penalties = capacity * _gc_time_penalties / 100;
179
180 double rate = _allocation_rate.sample(allocated);
181
182 // At what value of available, would avg and spike triggers occur?
183 // if allocation_headroom < avg_cycle_time * avg_alloc_rate, then we experience avg trigger
184 // if allocation_headroom < avg_cycle_time * rate, then we experience spike trigger if is_spiking
185 //
186 // allocation_headroom =
187 // 0, if penalties > available or if penalties + spike_headroom > available
188 // available - penalties - spike_headroom, otherwise
189 //
190 // so we trigger if available - penalties - spike_headroom < avg_cycle_time * avg_alloc_rate, which is to say
191 // available < avg_cycle_time * avg_alloc_rate + penalties + spike_headroom
192 // or if available < penalties + spike_headroom
193 //
194 // since avg_cycle_time * avg_alloc_rate > 0, the first test is sufficient to test both conditions
195 //
196 // thus, evac_slack_avg is MIN2(0, available - avg_cycle_time * avg_alloc_rate + penalties + spike_headroom)
197 //
198 // similarly, evac_slack_spiking is MIN2(0, available - avg_cycle_time * rate + penalties + spike_headroom)
199 // but evac_slack_spiking is only relevant if is_spiking, as defined below.
200
201 double avg_cycle_time = _gc_cycle_time_history->davg() + (_margin_of_error_sd * _gc_cycle_time_history->dsd());
202 double avg_alloc_rate = _allocation_rate.upper_bound(_margin_of_error_sd);
203 size_t evac_slack_avg;
204 if (anticipated_available > avg_cycle_time * avg_alloc_rate + penalties + spike_headroom) {
205 evac_slack_avg = anticipated_available - (avg_cycle_time * avg_alloc_rate + penalties + spike_headroom);
206 } else {
207 // we have no slack because it's already time to trigger
208 evac_slack_avg = 0;
209 }
210
211 bool is_spiking = _allocation_rate.is_spiking(rate, _spike_threshold_sd);
212 size_t evac_slack_spiking;
213 if (is_spiking) {
214 if (anticipated_available > avg_cycle_time * rate + penalties + spike_headroom) {
215 evac_slack_spiking = anticipated_available - (avg_cycle_time * rate + penalties + spike_headroom);
216 } else {
217 // we have no slack because it's already time to trigger
218 evac_slack_spiking = 0;
219 }
220 } else {
221 evac_slack_spiking = evac_slack_avg;
222 }
223
224 size_t threshold = min_free_threshold();
225 size_t evac_min_threshold = (anticipated_available > threshold)? anticipated_available - threshold: 0;
226 return MIN3(evac_slack_spiking, evac_slack_avg, evac_min_threshold);
227 }