1 | /* |
2 | * Copyright (c) 2018, 2019, Red Hat, Inc. All rights reserved. |
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/shenandoah/shenandoahFreeSet.hpp" |
27 | #include "gc/shenandoah/shenandoahHeap.inline.hpp" |
28 | #include "gc/shenandoah/shenandoahPacer.hpp" |
29 | |
30 | /* |
31 | * In normal concurrent cycle, we have to pace the application to let GC finish. |
32 | * |
33 | * Here, we do not know how large would be the collection set, and what are the |
34 | * relative performances of the each stage in the concurrent cycle, and so we have to |
35 | * make some assumptions. |
36 | * |
37 | * For concurrent mark, there is no clear notion of progress. The moderately accurate |
38 | * and easy to get metric is the amount of live objects the mark had encountered. But, |
39 | * that does directly correlate with the used heap, because the heap might be fully |
40 | * dead or fully alive. We cannot assume either of the extremes: we would either allow |
41 | * application to run out of memory if we assume heap is fully dead but it is not, and, |
42 | * conversely, we would pacify application excessively if we assume heap is fully alive |
43 | * but it is not. So we need to guesstimate the particular expected value for heap liveness. |
44 | * The best way to do this is apparently recording the past history. |
45 | * |
46 | * For concurrent evac and update-refs, we are walking the heap per-region, and so the |
47 | * notion of progress is clear: we get reported the "used" size from the processed regions |
48 | * and use the global heap-used as the baseline. |
49 | * |
50 | * The allocatable space when GC is running is "free" at the start of cycle, but the |
51 | * accounted budget is based on "used". So, we need to adjust the tax knowing that. |
52 | * Also, since we effectively count the used space three times (mark, evac, update-refs), |
53 | * we need to multiply the tax by 3. Example: for 10 MB free and 90 MB used, GC would |
54 | * come back with 3*90 MB budget, and thus for each 1 MB of allocation, we have to pay |
55 | * 3*90 / 10 MBs. In the end, we would pay back the entire budget. |
56 | */ |
57 | |
58 | void ShenandoahPacer::setup_for_mark() { |
59 | assert(ShenandoahPacing, "Only be here when pacing is enabled" ); |
60 | |
61 | size_t live = update_and_get_progress_history(); |
62 | size_t free = _heap->free_set()->available(); |
63 | |
64 | size_t non_taxable = free * ShenandoahPacingCycleSlack / 100; |
65 | size_t taxable = free - non_taxable; |
66 | |
67 | double tax = 1.0 * live / taxable; // base tax for available free space |
68 | tax *= 3; // mark is phase 1 of 3, claim 1/3 of free for it |
69 | tax *= ShenandoahPacingSurcharge; // additional surcharge to help unclutter heap |
70 | |
71 | restart_with(non_taxable, tax); |
72 | |
73 | log_info(gc, ergo)("Pacer for Mark. Expected Live: " SIZE_FORMAT "M, Free: " SIZE_FORMAT |
74 | "M, Non-Taxable: " SIZE_FORMAT "M, Alloc Tax Rate: %.1fx" , |
75 | live / M, free / M, non_taxable / M, tax); |
76 | } |
77 | |
78 | void ShenandoahPacer::setup_for_evac() { |
79 | assert(ShenandoahPacing, "Only be here when pacing is enabled" ); |
80 | |
81 | size_t used = _heap->collection_set()->used(); |
82 | size_t free = _heap->free_set()->available(); |
83 | |
84 | size_t non_taxable = free * ShenandoahPacingCycleSlack / 100; |
85 | size_t taxable = free - non_taxable; |
86 | |
87 | double tax = 1.0 * used / taxable; // base tax for available free space |
88 | tax *= 2; // evac is phase 2 of 3, claim 1/2 of remaining free |
89 | tax = MAX2<double>(1, tax); // never allocate more than GC processes during the phase |
90 | tax *= ShenandoahPacingSurcharge; // additional surcharge to help unclutter heap |
91 | |
92 | restart_with(non_taxable, tax); |
93 | |
94 | log_info(gc, ergo)("Pacer for Evacuation. Used CSet: " SIZE_FORMAT "M, Free: " SIZE_FORMAT |
95 | "M, Non-Taxable: " SIZE_FORMAT "M, Alloc Tax Rate: %.1fx" , |
96 | used / M, free / M, non_taxable / M, tax); |
97 | } |
98 | |
99 | void ShenandoahPacer::setup_for_updaterefs() { |
100 | assert(ShenandoahPacing, "Only be here when pacing is enabled" ); |
101 | |
102 | size_t used = _heap->used(); |
103 | size_t free = _heap->free_set()->available(); |
104 | |
105 | size_t non_taxable = free * ShenandoahPacingCycleSlack / 100; |
106 | size_t taxable = free - non_taxable; |
107 | |
108 | double tax = 1.0 * used / taxable; // base tax for available free space |
109 | tax *= 1; // update-refs is phase 3 of 3, claim the remaining free |
110 | tax = MAX2<double>(1, tax); // never allocate more than GC processes during the phase |
111 | tax *= ShenandoahPacingSurcharge; // additional surcharge to help unclutter heap |
112 | |
113 | restart_with(non_taxable, tax); |
114 | |
115 | log_info(gc, ergo)("Pacer for Update Refs. Used: " SIZE_FORMAT "M, Free: " SIZE_FORMAT |
116 | "M, Non-Taxable: " SIZE_FORMAT "M, Alloc Tax Rate: %.1fx" , |
117 | used / M, free / M, non_taxable / M, tax); |
118 | } |
119 | |
120 | /* |
121 | * Traversal walks the entire heap once, and therefore we have to make assumptions about its |
122 | * liveness, like concurrent mark does. |
123 | */ |
124 | |
125 | void ShenandoahPacer::setup_for_traversal() { |
126 | assert(ShenandoahPacing, "Only be here when pacing is enabled" ); |
127 | |
128 | size_t live = update_and_get_progress_history(); |
129 | size_t free = _heap->free_set()->available(); |
130 | |
131 | size_t non_taxable = free * ShenandoahPacingCycleSlack / 100; |
132 | size_t taxable = free - non_taxable; |
133 | |
134 | double tax = 1.0 * live / taxable; // base tax for available free space |
135 | tax *= ShenandoahPacingSurcharge; // additional surcharge to help unclutter heap |
136 | |
137 | restart_with(non_taxable, tax); |
138 | |
139 | log_info(gc, ergo)("Pacer for Traversal. Expected Live: " SIZE_FORMAT "M, Free: " SIZE_FORMAT |
140 | "M, Non-Taxable: " SIZE_FORMAT "M, Alloc Tax Rate: %.1fx" , |
141 | live / M, free / M, non_taxable / M, tax); |
142 | } |
143 | |
144 | /* |
145 | * In idle phase, we have to pace the application to let control thread react with GC start. |
146 | * |
147 | * Here, we have rendezvous with concurrent thread that adds up the budget as it acknowledges |
148 | * it had seen recent allocations. It will naturally pace the allocations if control thread is |
149 | * not catching up. To bootstrap this feedback cycle, we need to start with some initial budget |
150 | * for applications to allocate at. |
151 | */ |
152 | |
153 | void ShenandoahPacer::setup_for_idle() { |
154 | assert(ShenandoahPacing, "Only be here when pacing is enabled" ); |
155 | |
156 | size_t initial = _heap->max_capacity() / 100 * ShenandoahPacingIdleSlack; |
157 | double tax = 1; |
158 | |
159 | restart_with(initial, tax); |
160 | |
161 | log_info(gc, ergo)("Pacer for Idle. Initial: " SIZE_FORMAT "M, Alloc Tax Rate: %.1fx" , |
162 | initial / M, tax); |
163 | } |
164 | |
165 | size_t ShenandoahPacer::update_and_get_progress_history() { |
166 | if (_progress == -1) { |
167 | // First initialization, report some prior |
168 | Atomic::store((intptr_t)PACING_PROGRESS_ZERO, &_progress); |
169 | return (size_t) (_heap->max_capacity() * 0.1); |
170 | } else { |
171 | // Record history, and reply historical data |
172 | _progress_history->add(_progress); |
173 | Atomic::store((intptr_t)PACING_PROGRESS_ZERO, &_progress); |
174 | return (size_t) (_progress_history->avg() * HeapWordSize); |
175 | } |
176 | } |
177 | |
178 | void ShenandoahPacer::restart_with(size_t non_taxable_bytes, double tax_rate) { |
179 | size_t initial = (size_t)(non_taxable_bytes * tax_rate) >> LogHeapWordSize; |
180 | STATIC_ASSERT(sizeof(size_t) <= sizeof(intptr_t)); |
181 | Atomic::xchg((intptr_t)initial, &_budget); |
182 | Atomic::store(tax_rate, &_tax_rate); |
183 | Atomic::inc(&_epoch); |
184 | } |
185 | |
186 | bool ShenandoahPacer::claim_for_alloc(size_t words, bool force) { |
187 | assert(ShenandoahPacing, "Only be here when pacing is enabled" ); |
188 | |
189 | intptr_t tax = MAX2<intptr_t>(1, words * Atomic::load(&_tax_rate)); |
190 | |
191 | intptr_t cur = 0; |
192 | intptr_t new_val = 0; |
193 | do { |
194 | cur = Atomic::load(&_budget); |
195 | if (cur < tax && !force) { |
196 | // Progress depleted, alas. |
197 | return false; |
198 | } |
199 | new_val = cur - tax; |
200 | } while (Atomic::cmpxchg(new_val, &_budget, cur) != cur); |
201 | return true; |
202 | } |
203 | |
204 | void ShenandoahPacer::unpace_for_alloc(intptr_t epoch, size_t words) { |
205 | assert(ShenandoahPacing, "Only be here when pacing is enabled" ); |
206 | |
207 | if (_epoch != epoch) { |
208 | // Stale ticket, no need to unpace. |
209 | return; |
210 | } |
211 | |
212 | intptr_t tax = MAX2<intptr_t>(1, words * Atomic::load(&_tax_rate)); |
213 | Atomic::add(tax, &_budget); |
214 | } |
215 | |
216 | intptr_t ShenandoahPacer::epoch() { |
217 | return Atomic::load(&_epoch); |
218 | } |
219 | |
220 | void ShenandoahPacer::pace_for_alloc(size_t words) { |
221 | assert(ShenandoahPacing, "Only be here when pacing is enabled" ); |
222 | |
223 | // Fast path: try to allocate right away |
224 | if (claim_for_alloc(words, false)) { |
225 | return; |
226 | } |
227 | |
228 | size_t max = ShenandoahPacingMaxDelay; |
229 | double start = os::elapsedTime(); |
230 | |
231 | size_t total = 0; |
232 | size_t cur = 0; |
233 | |
234 | while (true) { |
235 | // We could instead assist GC, but this would suffice for now. |
236 | // This code should also participate in safepointing. |
237 | // Perform the exponential backoff, limited by max. |
238 | |
239 | cur = cur * 2; |
240 | if (total + cur > max) { |
241 | cur = (max > total) ? (max - total) : 0; |
242 | } |
243 | cur = MAX2<size_t>(1, cur); |
244 | |
245 | os::sleep(Thread::current(), cur, true); |
246 | |
247 | double end = os::elapsedTime(); |
248 | total = (size_t)((end - start) * 1000); |
249 | |
250 | if (total > max) { |
251 | // Spent local time budget to wait for enough GC progress. |
252 | // Breaking out and allocating anyway, which may mean we outpace GC, |
253 | // and start Degenerated GC cycle. |
254 | _delays.add(total); |
255 | |
256 | // Forcefully claim the budget: it may go negative at this point, and |
257 | // GC should replenish for this and subsequent allocations |
258 | claim_for_alloc(words, true); |
259 | break; |
260 | } |
261 | |
262 | if (claim_for_alloc(words, false)) { |
263 | // Acquired enough permit, nice. Can allocate now. |
264 | _delays.add(total); |
265 | break; |
266 | } |
267 | } |
268 | } |
269 | |
270 | void ShenandoahPacer::print_on(outputStream* out) const { |
271 | out->print_cr("ALLOCATION PACING:" ); |
272 | out->cr(); |
273 | |
274 | out->print_cr("Max pacing delay is set for " UINTX_FORMAT " ms." , ShenandoahPacingMaxDelay); |
275 | out->cr(); |
276 | |
277 | out->print_cr("Higher delay would prevent application outpacing the GC, but it will hide the GC latencies" ); |
278 | out->print_cr("from the STW pause times. Pacing affects the individual threads, and so it would also be" ); |
279 | out->print_cr("invisible to the usual profiling tools, but would add up to end-to-end application latency." ); |
280 | out->print_cr("Raise max pacing delay with care." ); |
281 | out->cr(); |
282 | |
283 | out->print_cr("Actual pacing delays histogram:" ); |
284 | out->cr(); |
285 | |
286 | out->print_cr("%10s - %10s %12s%12s" , "From" , "To" , "Count" , "Sum" ); |
287 | |
288 | size_t total_count = 0; |
289 | size_t total_sum = 0; |
290 | for (int c = _delays.min_level(); c <= _delays.max_level(); c++) { |
291 | int l = (c == 0) ? 0 : 1 << (c - 1); |
292 | int r = 1 << c; |
293 | size_t count = _delays.level(c); |
294 | size_t sum = count * (r - l) / 2; |
295 | total_count += count; |
296 | total_sum += sum; |
297 | |
298 | out->print_cr("%7d ms - %7d ms: " SIZE_FORMAT_W(12) SIZE_FORMAT_W(12) " ms" , l, r, count, sum); |
299 | } |
300 | out->print_cr("%23s: " SIZE_FORMAT_W(12) SIZE_FORMAT_W(12) " ms" , "Total" , total_count, total_sum); |
301 | out->cr(); |
302 | out->print_cr("Pacing delays are measured from entering the pacing code till exiting it. Therefore," ); |
303 | out->print_cr("observed pacing delays may be higher than the threshold when paced thread spent more" ); |
304 | out->print_cr("time in the pacing code. It usually happens when thread is de-scheduled while paced," ); |
305 | out->print_cr("OS takes longer to unblock the thread, or JVM experiences an STW pause." ); |
306 | out->cr(); |
307 | } |
308 | |