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_implementation/shenandoah/shenandoahFreeSet.hpp" 27 #include "gc_implementation/shenandoah/shenandoahHeap.inline.hpp" 28 #include "gc_implementation/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 "%s, Free: " SIZE_FORMAT "%s, " 74 "Non-Taxable: " SIZE_FORMAT "%s, Alloc Tax Rate: %.1fx", 75 byte_size_in_proper_unit(live), proper_unit_for_byte_size(live), 76 byte_size_in_proper_unit(free), proper_unit_for_byte_size(free), 77 byte_size_in_proper_unit(non_taxable), proper_unit_for_byte_size(non_taxable), 78 tax); 79 } 80 81 void ShenandoahPacer::setup_for_evac() { 82 assert(ShenandoahPacing, "Only be here when pacing is enabled"); 83 84 size_t used = _heap->collection_set()->used(); 85 size_t free = _heap->free_set()->available(); 86 87 size_t non_taxable = free * ShenandoahPacingCycleSlack / 100; 88 size_t taxable = free - non_taxable; 89 90 double tax = 1.0 * used / taxable; // base tax for available free space 91 tax *= 2; // evac is phase 2 of 3, claim 1/2 of remaining free 92 tax = MAX2<double>(1, tax); // never allocate more than GC processes during the phase 93 tax *= ShenandoahPacingSurcharge; // additional surcharge to help unclutter heap 94 95 restart_with(non_taxable, tax); 96 97 log_info(gc, ergo)("Pacer for Evacuation. Used CSet: " SIZE_FORMAT "%s, Free: " SIZE_FORMAT "%s, " 98 "Non-Taxable: " SIZE_FORMAT "%s, Alloc Tax Rate: %.1fx", 99 byte_size_in_proper_unit(used), proper_unit_for_byte_size(used), 100 byte_size_in_proper_unit(free), proper_unit_for_byte_size(free), 101 byte_size_in_proper_unit(non_taxable), proper_unit_for_byte_size(non_taxable), 102 tax); 103 } 104 105 void ShenandoahPacer::setup_for_updaterefs() { 106 assert(ShenandoahPacing, "Only be here when pacing is enabled"); 107 108 size_t used = _heap->used(); 109 size_t free = _heap->free_set()->available(); 110 111 size_t non_taxable = free * ShenandoahPacingCycleSlack / 100; 112 size_t taxable = free - non_taxable; 113 114 double tax = 1.0 * used / taxable; // base tax for available free space 115 tax *= 1; // update-refs is phase 3 of 3, claim the remaining free 116 tax = MAX2<double>(1, tax); // never allocate more than GC processes during the phase 117 tax *= ShenandoahPacingSurcharge; // additional surcharge to help unclutter heap 118 119 restart_with(non_taxable, tax); 120 121 log_info(gc, ergo)("Pacer for Update Refs. Used: " SIZE_FORMAT "%s, Free: " SIZE_FORMAT "%s, " 122 "Non-Taxable: " SIZE_FORMAT "%s, Alloc Tax Rate: %.1fx", 123 byte_size_in_proper_unit(used), proper_unit_for_byte_size(used), 124 byte_size_in_proper_unit(free), proper_unit_for_byte_size(free), 125 byte_size_in_proper_unit(non_taxable), proper_unit_for_byte_size(non_taxable), 126 tax); 127 } 128 129 /* 130 * Traversal walks the entire heap once, and therefore we have to make assumptions about its 131 * liveness, like concurrent mark does. 132 */ 133 134 void ShenandoahPacer::setup_for_traversal() { 135 assert(ShenandoahPacing, "Only be here when pacing is enabled"); 136 137 size_t live = update_and_get_progress_history(); 138 size_t free = _heap->free_set()->available(); 139 140 size_t non_taxable = free * ShenandoahPacingCycleSlack / 100; 141 size_t taxable = free - non_taxable; 142 143 double tax = 1.0 * live / taxable; // base tax for available free space 144 tax *= ShenandoahPacingSurcharge; // additional surcharge to help unclutter heap 145 146 restart_with(non_taxable, tax); 147 148 log_info(gc, ergo)("Pacer for Traversal. Expected Live: " SIZE_FORMAT "%s, Free: " SIZE_FORMAT "%s, " 149 "Non-Taxable: " SIZE_FORMAT "%s, Alloc Tax Rate: %.1fx", 150 byte_size_in_proper_unit(live), proper_unit_for_byte_size(live), 151 byte_size_in_proper_unit(free), proper_unit_for_byte_size(free), 152 byte_size_in_proper_unit(non_taxable), proper_unit_for_byte_size(non_taxable), 153 tax); 154 } 155 156 /* 157 * In idle phase, we have to pace the application to let control thread react with GC start. 158 * 159 * Here, we have rendezvous with concurrent thread that adds up the budget as it acknowledges 160 * it had seen recent allocations. It will naturally pace the allocations if control thread is 161 * not catching up. To bootstrap this feedback cycle, we need to start with some initial budget 162 * for applications to allocate at. 163 */ 164 165 void ShenandoahPacer::setup_for_idle() { 166 assert(ShenandoahPacing, "Only be here when pacing is enabled"); 167 168 size_t initial = _heap->max_capacity() / 100 * ShenandoahPacingIdleSlack; 169 double tax = 1; 170 171 restart_with(initial, tax); 172 173 log_info(gc, ergo)("Pacer for Idle. Initial: " SIZE_FORMAT "%s, Alloc Tax Rate: %.1fx", 174 byte_size_in_proper_unit(initial), proper_unit_for_byte_size(initial), 175 tax); 176 } 177 178 size_t ShenandoahPacer::update_and_get_progress_history() { 179 if (_progress == -1) { 180 // First initialization, report some prior 181 Atomic::store((intptr_t)PACING_PROGRESS_ZERO, &_progress); 182 return (size_t) (_heap->max_capacity() * 0.1); 183 } else { 184 // Record history, and reply historical data 185 _progress_history->add(_progress); 186 Atomic::store((intptr_t)PACING_PROGRESS_ZERO, &_progress); 187 return (size_t) (_progress_history->avg() * HeapWordSize); 188 } 189 } 190 191 void ShenandoahPacer::restart_with(jlong non_taxable_bytes, jdouble tax_rate) { 192 STATIC_ASSERT(sizeof(size_t) <= sizeof(intptr_t)); 193 { 194 intptr_t initial = (size_t) (non_taxable_bytes * tax_rate) >> LogHeapWordSize; 195 intptr_t cur; 196 do { 197 cur = OrderAccess::load_acquire(&_budget); 198 } while (Atomic::cmpxchg(initial, &_budget, cur) != cur); 199 } 200 201 OrderAccess::release_store(&_tax_rate, tax_rate); 202 203 { 204 intptr_t cur, val; 205 do { 206 cur = OrderAccess::load_acquire(&_epoch); 207 val = cur + 1; 208 } while (Atomic::cmpxchg(val, &_epoch, cur) != cur); 209 } 210 } 211 212 bool ShenandoahPacer::claim_for_alloc(size_t words, bool force) { 213 assert(ShenandoahPacing, "Only be here when pacing is enabled"); 214 215 intptr_t tax = MAX2<intptr_t>(1, (intptr_t)(words * OrderAccess::load_acquire(&_tax_rate))); 216 217 intptr_t cur = 0; 218 intptr_t new_val = 0; 219 do { 220 cur = OrderAccess::load_acquire(&_budget); 221 if (cur < tax) { 222 // Progress depleted, alas. 223 return false; 224 } 225 new_val = cur - tax; 226 } while (Atomic::cmpxchg(new_val, &_budget, cur) != cur); 227 return true; 228 } 229 230 void ShenandoahPacer::unpace_for_alloc(intptr_t epoch, size_t words) { 231 assert(ShenandoahPacing, "Only be here when pacing is enabled"); 232 233 if (_epoch != epoch) { 234 // Stale ticket, no need to unpace. 235 return; 236 } 237 238 intptr_t tax = MAX2<intptr_t>(1, (intptr_t)(words * OrderAccess::load_acquire(&_tax_rate))); 239 Atomic::add(tax, &_budget); 240 } 241 242 intptr_t ShenandoahPacer::epoch() { 243 return OrderAccess::load_acquire(&_epoch); 244 } 245 246 void ShenandoahPacer::pace_for_alloc(size_t words) { 247 assert(ShenandoahPacing, "Only be here when pacing is enabled"); 248 249 // Fast path: try to allocate right away 250 if (claim_for_alloc(words, false)) { 251 return; 252 } 253 254 size_t max = ShenandoahPacingMaxDelay; 255 double start = os::elapsedTime(); 256 257 size_t total = 0; 258 size_t cur = 0; 259 260 while (true) { 261 // We could instead assist GC, but this would suffice for now. 262 // This code should also participate in safepointing. 263 // Perform the exponential backoff, limited by max. 264 265 cur = cur * 2; 266 if (total + cur > max) { 267 cur = (max > total) ? (max - total) : 0; 268 } 269 cur = MAX2<size_t>(1, cur); 270 271 os::sleep(Thread::current(), cur, true); 272 273 double end = os::elapsedTime(); 274 total = (size_t)((end - start) * 1000); 275 276 if (total > max) { 277 // Spent local time budget to wait for enough GC progress. 278 // Breaking out and allocating anyway, which may mean we outpace GC, 279 // and start Degenerated GC cycle. 280 _delays.add(total); 281 282 // Forcefully claim the budget: it may go negative at this point, and 283 // GC should replenish for this and subsequent allocations 284 claim_for_alloc(words, true); 285 break; 286 } 287 288 if (claim_for_alloc(words, false)) { 289 // Acquired enough permit, nice. Can allocate now. 290 _delays.add(total); 291 break; 292 } 293 } 294 } 295 296 void ShenandoahPacer::print_on(outputStream* out) const { 297 out->print_cr("ALLOCATION PACING:"); 298 out->cr(); 299 300 out->print_cr("Max pacing delay is set for " UINTX_FORMAT " ms.", ShenandoahPacingMaxDelay); 301 out->cr(); 302 303 out->print_cr("Higher delay would prevent application outpacing the GC, but it will hide the GC latencies"); 304 out->print_cr("from the STW pause times. Pacing affects the individual threads, and so it would also be"); 305 out->print_cr("invisible to the usual profiling tools, but would add up to end-to-end application latency."); 306 out->print_cr("Raise max pacing delay with care."); 307 out->cr(); 308 309 out->print_cr("Actual pacing delays histogram:"); 310 out->cr(); 311 312 out->print_cr("%10s - %10s %12s%12s", "From", "To", "Count", "Sum"); 313 314 size_t total_count = 0; 315 size_t total_sum = 0; 316 for (int c = _delays.min_level(); c <= _delays.max_level(); c++) { 317 int l = (c == 0) ? 0 : 1 << (c - 1); 318 int r = 1 << c; 319 size_t count = _delays.level(c); 320 size_t sum = count * (r - l) / 2; 321 total_count += count; 322 total_sum += sum; 323 324 out->print_cr("%7d ms - %7d ms: " SIZE_FORMAT_W(12) SIZE_FORMAT_W(12) " ms", l, r, count, sum); 325 } 326 out->print_cr("%23s: " SIZE_FORMAT_W(12) SIZE_FORMAT_W(12) " ms", "Total", total_count, total_sum); 327 out->cr(); 328 out->print_cr("Pacing delays are measured from entering the pacing code till exiting it. Therefore,"); 329 out->print_cr("observed pacing delays may be higher than the threshold when paced thread spent more"); 330 out->print_cr("time in the pacing code. It usually happens when thread is de-scheduled while paced,"); 331 out->print_cr("OS takes longer to unblock the thread, or JVM experiences an STW pause."); 332 out->cr(); 333 }