1 /*
2 * Copyright (c) 2018, Red Hat, Inc. and/or its affiliates.
3 *
4 * This code is free software; you can redistribute it and/or modify it
5 * under the terms of the GNU General Public License version 2 only, as
6 * published by the Free Software Foundation.
7 *
8 * This code is distributed in the hope that it will be useful, but WITHOUT
9 * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
10 * FITNESS FOR A PARTICULAR PURPOSE. See the GNU General Public License
11 * version 2 for more details (a copy is included in the LICENSE file that
12 * accompanied this code).
13 *
14 * You should have received a copy of the GNU General Public License version
15 * 2 along with this work; if not, write to the Free Software Foundation,
16 * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
17 *
18 * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
19 * or visit www.oracle.com if you need additional information or have any
20 * questions.
21 *
22 */
23
24 #include "precompiled.hpp"
25
26 #include "gc_implementation/shenandoah/shenandoahPacer.hpp"
27 #include "gc_implementation/shenandoah/shenandoahHeap.hpp"
28 #include "gc_implementation/shenandoah/shenandoahHeap.inline.hpp"
29 #include "gc_implementation/shenandoah/shenandoahFreeSet.hpp"
30
31 /*
32 * In normal concurrent cycle, we have to pace the application to let GC finish.
33 *
34 * Here, we do not know how large would be the collection set, and what are the
35 * relative performances of the each stage in the concurrent cycle, and so we have to
36 * make some assumptions.
37 *
38 * We assume, for pessimistic reasons, that the entire heap is full of alive objects,
39 * and it will be evacuated fully. Therefore, we count live objects visited by all three
40 * stages against the heap used at the beginning of the collection. That means if there
41 * are dead objects, they would not be accounted for in this budget, and that would mean
42 * allocation would be pacified excessively. But that *also* means the collection cycle
43 * would finish earlier than pacer expects.
44 *
45 * The allocatable space when GC is running is "free" at the start of cycle, but the
46 * accounted budget is based on "used". So, we need to adjust the tax knowing that.
47 * Also, since we effectively count the used space three times (mark, evac, update-refs),
48 * we need to multiply the tax by 3. Example: for 10 MB free and 90 MB used, GC would
49 * come back with 3*90 MB budget, and thus for each 1 MB of allocation, we have to pay
50 * 3*90 / 10 MBs. In the end, we would pay back the entire budget.
51 */
52
53 void ShenandoahPacer::setup_for_mark() {
54 assert(ShenandoahPacing, "Only be here when pacing is enabled");
55
56 size_t used = _heap->used();
57 size_t free = _heap->free_set()->available();
58
59 size_t non_taxable = free * ShenandoahPacingCycleSlack / 100;
60 size_t taxable = free - non_taxable;
61
62 double tax = 1.0 * used / taxable; // base tax for available free space
63 tax *= 3; // mark is phase 1 of 3, claim 1/3 of free for it
64 tax = MAX2<double>(1, tax); // never allocate more than GC collects during the cycle
65 tax *= 1.1; // additional surcharge to help unclutter heap
66
67 restart_with(non_taxable, tax);
68
69 log_info(gc, ergo)("Pacer for Mark. Used: " SIZE_FORMAT "M, Free: " SIZE_FORMAT
70 "M, Non-Taxable: " SIZE_FORMAT "M, Alloc Tax Rate: %.1fx",
71 used / M, free / M, non_taxable / M, tax);
72 }
73
74 void ShenandoahPacer::setup_for_evac() {
75 assert(ShenandoahPacing, "Only be here when pacing is enabled");
76
77 size_t cset = _heap->collection_set()->live_data();
78 size_t free = _heap->free_set()->available();
79
80 size_t non_taxable = free * ShenandoahPacingCycleSlack / 100;
81 size_t taxable = free - non_taxable;
82
83 double tax = 1.0 * cset / taxable; // base tax for available free space
84 tax *= 2; // evac is phase 2 of 3, claim 1/2 of remaining free
85 tax = MAX2<double>(1, tax); // never allocate more than GC collects during the cycle
86 tax *= 1.1; // additional surcharge to help unclutter heap
87
88 restart_with(non_taxable, tax);
89
90 log_info(gc, ergo)("Pacer for Evacuation. CSet: " SIZE_FORMAT "M, Free: " SIZE_FORMAT
91 "M, Non-Taxable: " SIZE_FORMAT "M, Alloc Tax Rate: %.1fx",
92 cset / M, free / M, non_taxable / M, tax);
93 }
94
95 void ShenandoahPacer::setup_for_updaterefs() {
96 assert(ShenandoahPacing, "Only be here when pacing is enabled");
97
98 size_t used = _heap->used();
99 size_t free = _heap->free_set()->available();
100
101 size_t non_taxable = free * ShenandoahPacingCycleSlack / 100;
102 size_t taxable = free - non_taxable;
103
104 double tax = 1.0 * used / taxable; // base tax for available free space
105 tax *= 1; // update-refs is phase 3 of 3, claim the remaining free
106 tax = MAX2<double>(1, tax); // never allocate more than GC collects during the cycle
107 tax *= 1.1; // additional surcharge to help unclutter heap
108
109 restart_with(non_taxable, tax);
110
111 log_info(gc, ergo)("Pacer for Update-Refs. Used: " SIZE_FORMAT "M, Free: " SIZE_FORMAT
112 "M, Non-Taxable: " SIZE_FORMAT "M, Alloc Tax Rate: %.1fx",
113 used / M, free / M, non_taxable / M, tax);
114 }
115
116 /*
117 * In idle phase, we have to pace the application to let control thread react with GC start.
118 *
119 * Here, we have rendezvous with concurrent thread that adds up the budget as it acknowledges
120 * it had seen recent allocations. It will naturally pace the allocations if control thread is
121 * not catching up. To bootstrap this feedback cycle, we need to start with some initial budget
122 * for applications to allocate at.
123 */
124
125 void ShenandoahPacer::setup_for_idle() {
126 assert(ShenandoahPacing, "Only be here when pacing is enabled");
127
128 size_t initial = _heap->capacity() * ShenandoahPacingIdleSlack / 100;
129 double tax = 1;
130
131 restart_with(initial, tax);
132
133 log_info(gc, ergo)("Pacer for Idle. Initial: " SIZE_FORMAT "M, Alloc Tax Rate: %.1fx",
134 initial / M, tax);
135 }
136
137 void ShenandoahPacer::restart_with(jlong non_taxable_bytes, jdouble tax_rate) {
138 STATIC_ASSERT(sizeof(size_t) <= sizeof(intptr_t));
139 intptr_t initial = (size_t)(non_taxable_bytes * tax_rate) >> LogHeapWordSize;
140 intptr_t cur;
141 do {
142 cur = OrderAccess::load_acquire(&_budget);
143 } while (Atomic::cmpxchg(initial, &_budget, cur) != cur);
144 OrderAccess::release_store(&_tax_rate, tax_rate);
145 }
146
147 bool ShenandoahPacer::claim_for_alloc(size_t words, bool force) {
148 assert(ShenandoahPacing, "Only be here when pacing is enabled");
149
150 intptr_t tax = MAX2<intptr_t>(1, words * OrderAccess::load_acquire(&_tax_rate));
151
152 intptr_t cur = 0;
153 intptr_t new_val = 0;
154 do {
155 cur = OrderAccess::load_acquire(&_budget);
156 if (cur < tax) {
157 // Progress depleted, alas.
158 return false;
159 }
160 new_val = cur - tax;
161 } while (Atomic::cmpxchg(new_val, &_budget, cur) != cur);
162 return true;
163 }
164
165 void ShenandoahPacer::pace_for_alloc(size_t words) {
166 assert(ShenandoahPacing, "Only be here when pacing is enabled");
167
168 // Fast path: try to allocate right away
169 if (claim_for_alloc(words, false)) {
170 return;
171 }
172
173 size_t max_wait_ms = ShenandoahPacingMaxDelay;
174 double start = os::elapsedTime();
175
176 while (true) {
177 // We could instead assist GC, but this would suffice for now.
178 // This code should also participate in safepointing.
179 os::sleep(Thread::current(), 1, true);
180
181 double end = os::elapsedTime();
182 size_t ms = (size_t)((end - start) * 1000);
183 if (ms > max_wait_ms) {
184 // Spent local time budget to wait for enough GC progress.
185 // Breaking out and allocating anyway, which may mean we outpace GC,
186 // and start Degenerated GC cycle.
187 _delays.add(ms);
188
189 // Forcefully claim the budget: it may go negative at this point, and
190 // GC should replenish for this and subsequent allocations
191 claim_for_alloc(words, true);
192 break;
193 }
194
195 if (claim_for_alloc(words, false)) {
196 // Acquired enough permit, nice. Can allocate now.
197 _delays.add(ms);
198 break;
199 }
200 }
201 }
202
203 void ShenandoahPacer::print_on(outputStream* out) const {
204 out->print_cr("ALLOCATION PACING:");
205 out->cr();
206
207 out->print_cr("Max pacing delay is set for " UINTX_FORMAT " ms.", ShenandoahPacingMaxDelay);
208 out->cr();
209
210 out->print_cr("Higher delay would prevent application outpacing the GC, but it will hide the GC latencies");
211 out->print_cr("from the STW pause times. Pacing affects the individual threads, and so it would also be");
212 out->print_cr("invisible to the usual profiling tools, but would add up to end-to-end application latency.");
213 out->print_cr("Raise max pacing delay with care.");
214 out->cr();
215
216 out->print_cr("Actual pacing delays histogram:");
217 out->cr();
218
219 out->print_cr("%10s - %10s %12s", "From", "To", "Count");
220 for (int c = _delays.min_level(); c <= _delays.max_level(); c++) {
221 out->print("%7d ms - %7d ms:", (c == 0) ? 0 : 1 << (c - 1), 1 << c);
222 out->print_cr(SIZE_FORMAT_W(12), _delays.level(c));
223 }
224 out->cr();
225 }