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 * In idle phase, we have to pace the application to let control thread react with GC start. 131 * 132 * Here, we have rendezvous with concurrent thread that adds up the budget as it acknowledges 133 * it had seen recent allocations. It will naturally pace the allocations if control thread is 134 * not catching up. To bootstrap this feedback cycle, we need to start with some initial budget 135 * for applications to allocate at. 136 */ 137 138 void ShenandoahPacer::setup_for_idle() { 139 assert(ShenandoahPacing, "Only be here when pacing is enabled"); 140 141 size_t initial = _heap->max_capacity() / 100 * ShenandoahPacingIdleSlack; 142 double tax = 1; 143 144 restart_with(initial, tax); 145 146 log_info(gc, ergo)("Pacer for Idle. Initial: " SIZE_FORMAT "%s, Alloc Tax Rate: %.1fx", 147 byte_size_in_proper_unit(initial), proper_unit_for_byte_size(initial), 148 tax); 149 } 150 151 size_t ShenandoahPacer::update_and_get_progress_history() { 152 if (_progress == -1) { 153 // First initialization, report some prior 154 Atomic::store((intptr_t)PACING_PROGRESS_ZERO, &_progress); 155 return (size_t) (_heap->max_capacity() * 0.1); 156 } else { 157 // Record history, and reply historical data 158 _progress_history->add(_progress); 159 Atomic::store((intptr_t)PACING_PROGRESS_ZERO, &_progress); 160 return (size_t) (_progress_history->avg() * HeapWordSize); 161 } 162 } 163 164 void ShenandoahPacer::restart_with(jlong non_taxable_bytes, jdouble tax_rate) { 165 STATIC_ASSERT(sizeof(size_t) <= sizeof(intptr_t)); 166 { 167 intptr_t initial = (size_t) (non_taxable_bytes * tax_rate) >> LogHeapWordSize; 168 intptr_t cur; 169 do { 170 cur = OrderAccess::load_acquire(&_budget); 171 } while (Atomic::cmpxchg(initial, &_budget, cur) != cur); 172 } 173 174 OrderAccess::release_store(&_tax_rate, tax_rate); 175 176 { 177 intptr_t cur, val; 178 do { 179 cur = OrderAccess::load_acquire(&_epoch); 180 val = cur + 1; 181 } while (Atomic::cmpxchg(val, &_epoch, cur) != cur); 182 } 183 } 184 185 bool ShenandoahPacer::claim_for_alloc(size_t words, bool force) { 186 assert(ShenandoahPacing, "Only be here when pacing is enabled"); 187 188 intptr_t tax = MAX2<intptr_t>(1, (intptr_t)(words * OrderAccess::load_acquire(&_tax_rate))); 189 190 intptr_t cur = 0; 191 intptr_t new_val = 0; 192 do { 193 cur = OrderAccess::load_acquire(&_budget); 194 if (cur < tax) { 195 // Progress depleted, alas. 196 return false; 197 } 198 new_val = cur - tax; 199 } while (Atomic::cmpxchg(new_val, &_budget, cur) != cur); 200 return true; 201 } 202 203 void ShenandoahPacer::unpace_for_alloc(intptr_t epoch, size_t words) { 204 assert(ShenandoahPacing, "Only be here when pacing is enabled"); 205 206 if (_epoch != epoch) { 207 // Stale ticket, no need to unpace. 208 return; 209 } 210 211 intptr_t tax = MAX2<intptr_t>(1, (intptr_t)(words * OrderAccess::load_acquire(&_tax_rate))); 212 Atomic::add(tax, &_budget); 213 } 214 215 intptr_t ShenandoahPacer::epoch() { 216 return OrderAccess::load_acquire(&_epoch); 217 } 218 219 void ShenandoahPacer::pace_for_alloc(size_t words) { 220 assert(ShenandoahPacing, "Only be here when pacing is enabled"); 221 222 // Fast path: try to allocate right away 223 if (claim_for_alloc(words, false)) { 224 return; 225 } 226 227 size_t max = ShenandoahPacingMaxDelay; 228 double start = os::elapsedTime(); 229 230 size_t total = 0; 231 size_t cur = 0; 232 233 while (true) { 234 // We could instead assist GC, but this would suffice for now. 235 // This code should also participate in safepointing. 236 // Perform the exponential backoff, limited by max. 237 238 cur = cur * 2; 239 if (total + cur > max) { 240 cur = (max > total) ? (max - total) : 0; 241 } 242 cur = MAX2<size_t>(1, cur); 243 244 os::sleep(Thread::current(), cur, true); 245 246 double end = os::elapsedTime(); 247 total = (size_t)((end - start) * 1000); 248 249 if (total > max) { 250 // Spent local time budget to wait for enough GC progress. 251 // Breaking out and allocating anyway, which may mean we outpace GC, 252 // and start Degenerated GC cycle. 253 _delays.add(total); 254 255 // Forcefully claim the budget: it may go negative at this point, and 256 // GC should replenish for this and subsequent allocations 257 claim_for_alloc(words, true); 258 break; 259 } 260 261 if (claim_for_alloc(words, false)) { 262 // Acquired enough permit, nice. Can allocate now. 263 _delays.add(total); 264 break; 265 } 266 } 267 } 268 269 void ShenandoahPacer::print_on(outputStream* out) const { 270 out->print_cr("ALLOCATION PACING:"); 271 out->cr(); 272 273 out->print_cr("Max pacing delay is set for " UINTX_FORMAT " ms.", ShenandoahPacingMaxDelay); 274 out->cr(); 275 276 out->print_cr("Higher delay would prevent application outpacing the GC, but it will hide the GC latencies"); 277 out->print_cr("from the STW pause times. Pacing affects the individual threads, and so it would also be"); 278 out->print_cr("invisible to the usual profiling tools, but would add up to end-to-end application latency."); 279 out->print_cr("Raise max pacing delay with care."); 280 out->cr(); 281 282 out->print_cr("Actual pacing delays histogram:"); 283 out->cr(); 284 285 out->print_cr("%10s - %10s %12s%12s", "From", "To", "Count", "Sum"); 286 287 size_t total_count = 0; 288 size_t total_sum = 0; 289 for (int c = _delays.min_level(); c <= _delays.max_level(); c++) { 290 int l = (c == 0) ? 0 : 1 << (c - 1); 291 int r = 1 << c; 292 size_t count = _delays.level(c); 293 size_t sum = count * (r - l) / 2; 294 total_count += count; 295 total_sum += sum; 296 297 out->print_cr("%7d ms - %7d ms: " SIZE_FORMAT_W(12) SIZE_FORMAT_W(12) " ms", l, r, count, sum); 298 } 299 out->print_cr("%23s: " SIZE_FORMAT_W(12) SIZE_FORMAT_W(12) " ms", "Total", total_count, total_sum); 300 out->cr(); 301 out->print_cr("Pacing delays are measured from entering the pacing code till exiting it. Therefore,"); 302 out->print_cr("observed pacing delays may be higher than the threshold when paced thread spent more"); 303 out->print_cr("time in the pacing code. It usually happens when thread is de-scheduled while paced,"); 304 out->print_cr("OS takes longer to unblock the thread, or JVM experiences an STW pause."); 305 out->cr(); 306 }