1 /*
  2  * Copyright (c) 2018, 2019, Red Hat, Inc. All rights reserved.
  3  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
  4  *
  5  * This code is free software; you can redistribute it and/or modify it
  6  * under the terms of the GNU General Public License version 2 only, as
  7  * published by the Free Software Foundation.
  8  *
  9  * This code is distributed in the hope that it will be useful, but WITHOUT
 10  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
 11  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
 12  * version 2 for more details (a copy is included in the LICENSE file that
 13  * accompanied this code).
 14  *
 15  * You should have received a copy of the GNU General Public License version
 16  * 2 along with this work; if not, write to the Free Software Foundation,
 17  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
 18  *
 19  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
 20  * or visit www.oracle.com if you need additional information or have any
 21  * questions.
 22  *
 23  */
 24 
 25 #include "precompiled.hpp"
 26 
 27 #include "gc/shenandoah/heuristics/shenandoahAdaptiveHeuristics.hpp"
 28 #include "gc/shenandoah/shenandoahCollectionSet.hpp"
 29 #include "gc/shenandoah/shenandoahFreeSet.hpp"
 30 #include "gc/shenandoah/shenandoahGeneration.hpp"
 31 #include "gc/shenandoah/shenandoahHeapRegion.inline.hpp"
 32 #include "logging/log.hpp"
 33 #include "logging/logTag.hpp"
 34 #include "utilities/quickSort.hpp"
 35 
 36 // These constants are used to adjust the margin of error for the moving
 37 // average of the allocation rate and cycle time. The units are standard
 38 // deviations.
 39 const double ShenandoahAdaptiveHeuristics::FULL_PENALTY_SD = 0.2;
 40 const double ShenandoahAdaptiveHeuristics::DEGENERATE_PENALTY_SD = 0.1;
 41 
 42 // These are used to decide if we want to make any adjustments at all
 43 // at the end of a successful concurrent cycle.
 44 const double ShenandoahAdaptiveHeuristics::LOWEST_EXPECTED_AVAILABLE_AT_END = -0.5;
 45 const double ShenandoahAdaptiveHeuristics::HIGHEST_EXPECTED_AVAILABLE_AT_END = 0.5;
 46 
 47 // These values are the confidence interval expressed as standard deviations.
 48 // At the minimum confidence level, there is a 25% chance that the true value of
 49 // the estimate (average cycle time or allocation rate) is not more than
 50 // MINIMUM_CONFIDENCE standard deviations away from our estimate. Similarly, the
 51 // MAXIMUM_CONFIDENCE interval here means there is a one in a thousand chance
 52 // that the true value of our estimate is outside the interval. These are used
 53 // as bounds on the adjustments applied at the outcome of a GC cycle.
 54 const double ShenandoahAdaptiveHeuristics::MINIMUM_CONFIDENCE = 0.319; // 25%
 55 const double ShenandoahAdaptiveHeuristics::MAXIMUM_CONFIDENCE = 3.291; // 99.9%
 56 
 57 ShenandoahAdaptiveHeuristics::ShenandoahAdaptiveHeuristics(ShenandoahGeneration* generation) :
 58   ShenandoahHeuristics(generation),
 59   _margin_of_error_sd(ShenandoahAdaptiveInitialConfidence),
 60   _spike_threshold_sd(ShenandoahAdaptiveInitialSpikeThreshold),
 61   _last_trigger(OTHER) { }
 62 
 63 ShenandoahAdaptiveHeuristics::~ShenandoahAdaptiveHeuristics() {}
 64 
 65 void ShenandoahAdaptiveHeuristics::choose_collection_set_from_regiondata(ShenandoahCollectionSet* cset,
 66                                                                          RegionData* data, size_t size,
 67                                                                          size_t actual_free) {
 68   size_t garbage_threshold = ShenandoahHeapRegion::region_size_bytes() * ShenandoahGarbageThreshold / 100;
 69 
 70   // The logic for cset selection in adaptive is as follows:
 71   //
 72   //   1. We cannot get cset larger than available free space. Otherwise we guarantee OOME
 73   //      during evacuation, and thus guarantee full GC. In practice, we also want to let
 74   //      application to allocate something. This is why we limit CSet to some fraction of
 75   //      available space. In non-overloaded heap, max_cset would contain all plausible candidates
 76   //      over garbage threshold.
 77   //
 78   //   2. We should not get cset too low so that free threshold would not be met right
 79   //      after the cycle. Otherwise we get back-to-back cycles for no reason if heap is
 80   //      too fragmented. In non-overloaded non-fragmented heap min_garbage would be around zero.
 81   //
 82   // Therefore, we start by sorting the regions by garbage. Then we unconditionally add the best candidates
 83   // before we meet min_garbage. Then we add all candidates that fit with a garbage threshold before
 84   // we hit max_cset. When max_cset is hit, we terminate the cset selection. Note that in this scheme,
 85   // ShenandoahGarbageThreshold is the soft threshold which would be ignored until min_garbage is hit.
 86 
 87   size_t capacity    = _generation->soft_max_capacity();
 88   size_t max_cset    = (size_t)((1.0 * capacity / 100 * ShenandoahEvacReserve) / ShenandoahEvacWaste);
 89   size_t free_target = (capacity / 100 * ShenandoahMinFreeThreshold) + max_cset;
 90   size_t min_garbage = (free_target > actual_free ? (free_target - actual_free) : 0);
 91 
 92   log_info(gc, ergo)("Adaptive CSet Selection. Target Free: " SIZE_FORMAT "%s, Actual Free: "
 93                      SIZE_FORMAT "%s, Max CSet: " SIZE_FORMAT "%s, Min Garbage: " SIZE_FORMAT "%s",
 94                      byte_size_in_proper_unit(free_target), proper_unit_for_byte_size(free_target),
 95                      byte_size_in_proper_unit(actual_free), proper_unit_for_byte_size(actual_free),
 96                      byte_size_in_proper_unit(max_cset),    proper_unit_for_byte_size(max_cset),
 97                      byte_size_in_proper_unit(min_garbage), proper_unit_for_byte_size(min_garbage));
 98 
 99   // Better select garbage-first regions
100   QuickSort::sort<RegionData>(data, (int)size, compare_by_garbage, false);
101 
102   size_t cur_cset = 0;
103   size_t cur_garbage = 0;
104 
105   for (size_t idx = 0; idx < size; idx++) {
106     ShenandoahHeapRegion* r = data[idx]._region;
107 
108     size_t new_cset    = cur_cset + r->get_live_data_bytes();
109     size_t new_garbage = cur_garbage + r->garbage();
110 
111     if (new_cset > max_cset) {
112       break;
113     }
114 
115     if ((new_garbage < min_garbage) || (r->garbage() > garbage_threshold)) {
116       cset->add_region(r);
117       cur_cset = new_cset;
118       cur_garbage = new_garbage;
119     }
120   }
121 }
122 
123 void ShenandoahAdaptiveHeuristics::record_cycle_start() {
124   ShenandoahHeuristics::record_cycle_start();
125   _allocation_rate.allocation_counter_reset();
126 }
127 
128 void ShenandoahAdaptiveHeuristics::record_success_concurrent() {
129   ShenandoahHeuristics::record_success_concurrent();
130 
131   size_t available = ShenandoahHeap::heap()->free_set()->available();
132 
133   _available.add(available);
134   double z_score = 0.0;
135   if (_available.sd() > 0) {
136     z_score = (available - _available.avg()) / _available.sd();
137   }
138 
139   log_debug(gc, ergo)("Available: " SIZE_FORMAT " %sB, z-score=%.3f. Average available: %.1f %sB +/- %.1f %sB.",
140                       byte_size_in_proper_unit(available), proper_unit_for_byte_size(available),
141                       z_score,
142                       byte_size_in_proper_unit(_available.avg()), proper_unit_for_byte_size(_available.avg()),
143                       byte_size_in_proper_unit(_available.sd()), proper_unit_for_byte_size(_available.sd()));
144 
145   // In the case when a concurrent GC cycle completes successfully but with an
146   // unusually small amount of available memory we will adjust our trigger
147   // parameters so that they are more likely to initiate a new cycle.
148   // Conversely, when a GC cycle results in an above average amount of available
149   // memory, we will adjust the trigger parameters to be less likely to initiate
150   // a GC cycle.
151   //
152   // The z-score we've computed is in no way statistically related to the
153   // trigger parameters, but it has the nice property that worse z-scores for
154   // available memory indicate making larger adjustments to the trigger
155   // parameters. It also results in fewer adjustments as the application
156   // stabilizes.
157   //
158   // In order to avoid making endless and likely unnecessary adjustments to the
159   // trigger parameters, the change in available memory (with respect to the
160   // average) at the end of a cycle must be beyond these threshold values.
161   if (z_score < LOWEST_EXPECTED_AVAILABLE_AT_END ||
162       z_score > HIGHEST_EXPECTED_AVAILABLE_AT_END) {
163     // The sign is flipped because a negative z-score indicates that the
164     // available memory at the end of the cycle is below average. Positive
165     // adjustments make the triggers more sensitive (i.e., more likely to fire).
166     // The z-score also gives us a measure of just how far below normal. This
167     // property allows us to adjust the trigger parameters proportionally.
168     //
169     // The `100` here is used to attenuate the size of our adjustments. This
170     // number was chosen empirically. It also means the adjustments at the end of
171     // a concurrent cycle are an order of magnitude smaller than the adjustments
172     // made for a degenerated or full GC cycle (which themselves were also
173     // chosen empirically).
174     adjust_last_trigger_parameters(z_score / -100);
175   }
176 }
177 
178 void ShenandoahAdaptiveHeuristics::record_success_degenerated() {
179   ShenandoahHeuristics::record_success_degenerated();
180   // Adjust both trigger's parameters in the case of a degenerated GC because
181   // either of them should have triggered earlier to avoid this case.
182   adjust_margin_of_error(DEGENERATE_PENALTY_SD);
183   adjust_spike_threshold(DEGENERATE_PENALTY_SD);
184 }
185 
186 void ShenandoahAdaptiveHeuristics::record_success_full() {
187   ShenandoahHeuristics::record_success_full();
188   // Adjust both trigger's parameters in the case of a full GC because
189   // either of them should have triggered earlier to avoid this case.
190   adjust_margin_of_error(FULL_PENALTY_SD);
191   adjust_spike_threshold(FULL_PENALTY_SD);
192 }
193 
194 static double saturate(double value, double min, double max) {
195   return MAX2(MIN2(value, max), min);
196 }
197 
198 bool ShenandoahAdaptiveHeuristics::should_start_gc() {
199   size_t max_capacity = _generation->max_capacity();
200   size_t capacity = _generation->soft_max_capacity();
201   size_t available = _generation->available();
202   size_t allocated = _generation->bytes_allocated_since_gc_start();
203 
204   log_debug(gc)("should_start_gc (%s)? available: " SIZE_FORMAT ", soft_max_capacity: " SIZE_FORMAT
205                 ", max_capacity: " SIZE_FORMAT ", allocated: " SIZE_FORMAT,
206                 _generation->name(), available, capacity, max_capacity, allocated);
207 
208   // Make sure the code below treats available without the soft tail.
209   size_t soft_tail = max_capacity - capacity;
210   available = (available > soft_tail) ? (available - soft_tail) : 0;
211 
212   // Track allocation rate even if we decide to start a cycle for other reasons.
213   double rate = _allocation_rate.sample(allocated);
214   _last_trigger = OTHER;
215 
216   size_t min_threshold = capacity / 100 * ShenandoahMinFreeThreshold;
217 
218   log_debug(gc)("  available adjusted to: " SIZE_FORMAT ", min_threshold: " SIZE_FORMAT ", ShenandoahMinFreeThreshold: " SIZE_FORMAT,
219                 available, min_threshold, ShenandoahMinFreeThreshold);
220 
221   if (available < min_threshold) {
222     log_info(gc)("Trigger (%s): Free (" SIZE_FORMAT "%s) is below minimum threshold (" SIZE_FORMAT "%s)",
223                  _generation->name(),
224                  byte_size_in_proper_unit(available),     proper_unit_for_byte_size(available),
225                  byte_size_in_proper_unit(min_threshold), proper_unit_for_byte_size(min_threshold));
226     return true;
227   }
228 
229   // Check if we need to learn a bit about the application
230   const size_t max_learn = ShenandoahLearningSteps;
231   if (_gc_times_learned < max_learn) {
232     size_t init_threshold = capacity / 100 * ShenandoahInitFreeThreshold;
233     if (available < init_threshold) {
234       log_info(gc)("Trigger (%s): Learning " SIZE_FORMAT " of " SIZE_FORMAT ". Free (" SIZE_FORMAT "%s) is below initial threshold (" SIZE_FORMAT "%s)",
235                    _generation->name(), _gc_times_learned + 1, max_learn,
236                    byte_size_in_proper_unit(available),      proper_unit_for_byte_size(available),
237                    byte_size_in_proper_unit(init_threshold), proper_unit_for_byte_size(init_threshold));
238       return true;
239     }
240   }
241 
242   // Check if allocation headroom is still okay. This also factors in:
243   //   1. Some space to absorb allocation spikes
244   //   2. Accumulated penalties from Degenerated and Full GC
245   size_t allocation_headroom = available;
246 
247   size_t spike_headroom = capacity / 100 * ShenandoahAllocSpikeFactor;
248   size_t penalties      = capacity / 100 * _gc_time_penalties;
249 
250   allocation_headroom -= MIN2(allocation_headroom, spike_headroom);
251   allocation_headroom -= MIN2(allocation_headroom, penalties);
252 
253   double avg_cycle_time = _gc_time_history->davg() + (_margin_of_error_sd * _gc_time_history->dsd());
254   double avg_alloc_rate = _allocation_rate.upper_bound(_margin_of_error_sd);
255   log_debug(gc)("%s: average GC time: %.2f ms, allocation rate: %.0f %s/s",
256     _generation->name(), avg_cycle_time * 1000, byte_size_in_proper_unit(avg_alloc_rate), proper_unit_for_byte_size(avg_alloc_rate));
257 
258   if (avg_cycle_time > allocation_headroom / avg_alloc_rate) {
259     log_info(gc)("Trigger (%s): Average GC time (%.2f ms) is above the time for average allocation rate (%.0f %sB/s) to deplete free headroom (" SIZE_FORMAT "%s) (margin of error = %.2f)",
260                  _generation->name(), avg_cycle_time * 1000,
261                  byte_size_in_proper_unit(avg_alloc_rate), proper_unit_for_byte_size(avg_alloc_rate),
262                  byte_size_in_proper_unit(allocation_headroom), proper_unit_for_byte_size(allocation_headroom),
263                  _margin_of_error_sd);
264 
265     log_info(gc, ergo)("Free headroom: " SIZE_FORMAT "%s (free) - " SIZE_FORMAT "%s (spike) - " SIZE_FORMAT "%s (penalties) = " SIZE_FORMAT "%s",
266                        byte_size_in_proper_unit(available),           proper_unit_for_byte_size(available),
267                        byte_size_in_proper_unit(spike_headroom),      proper_unit_for_byte_size(spike_headroom),
268                        byte_size_in_proper_unit(penalties),           proper_unit_for_byte_size(penalties),
269                        byte_size_in_proper_unit(allocation_headroom), proper_unit_for_byte_size(allocation_headroom));
270 
271     _last_trigger = RATE;
272     return true;
273   }
274 
275   bool is_spiking = _allocation_rate.is_spiking(rate, _spike_threshold_sd);
276   if (is_spiking && avg_cycle_time > allocation_headroom / rate) {
277     log_info(gc)("Trigger (%s): Average GC time (%.2f ms) is above the time for instantaneous allocation rate (%.0f %sB/s) to deplete free headroom (" SIZE_FORMAT "%s) (spike threshold = %.2f)",
278                  _generation->name(), avg_cycle_time * 1000,
279                  byte_size_in_proper_unit(rate), proper_unit_for_byte_size(rate),
280                  byte_size_in_proper_unit(allocation_headroom), proper_unit_for_byte_size(allocation_headroom),
281                  _spike_threshold_sd);
282     _last_trigger = SPIKE;
283     return true;
284   }
285 
286   return ShenandoahHeuristics::should_start_gc();
287 }
288 
289 void ShenandoahAdaptiveHeuristics::adjust_last_trigger_parameters(double amount) {
290   switch (_last_trigger) {
291     case RATE:
292       adjust_margin_of_error(amount);
293       break;
294     case SPIKE:
295       adjust_spike_threshold(amount);
296       break;
297     case OTHER:
298       // nothing to adjust here.
299       break;
300     default:
301       ShouldNotReachHere();
302   }
303 }
304 
305 void ShenandoahAdaptiveHeuristics::adjust_margin_of_error(double amount) {
306   _margin_of_error_sd = saturate(_margin_of_error_sd + amount, MINIMUM_CONFIDENCE, MAXIMUM_CONFIDENCE);
307   log_debug(gc, ergo)("Margin of error now %.2f", _margin_of_error_sd);
308 }
309 
310 void ShenandoahAdaptiveHeuristics::adjust_spike_threshold(double amount) {
311   _spike_threshold_sd = saturate(_spike_threshold_sd - amount, MINIMUM_CONFIDENCE, MAXIMUM_CONFIDENCE);
312   log_debug(gc, ergo)("Spike threshold now: %.2f", _spike_threshold_sd);
313 }
314 
315 ShenandoahAllocationRate::ShenandoahAllocationRate() :
316   _last_sample_time(os::elapsedTime()),
317   _last_sample_value(0),
318   _interval_sec(1.0 / ShenandoahAdaptiveSampleFrequencyHz),
319   _rate(int(ShenandoahAdaptiveSampleSizeSeconds * ShenandoahAdaptiveSampleFrequencyHz), ShenandoahAdaptiveDecayFactor),
320   _rate_avg(int(ShenandoahAdaptiveSampleSizeSeconds * ShenandoahAdaptiveSampleFrequencyHz), ShenandoahAdaptiveDecayFactor) {
321 }
322 
323 double ShenandoahAllocationRate::sample(size_t allocated) {
324   double now = os::elapsedTime();
325   double rate = 0.0;
326   if (now - _last_sample_time > _interval_sec) {
327     if (allocated >= _last_sample_value) {
328       rate = instantaneous_rate(now, allocated);
329       _rate.add(rate);
330       _rate_avg.add(_rate.avg());
331     }
332 
333     _last_sample_time = now;
334     _last_sample_value = allocated;
335   }
336   return rate;
337 }
338 
339 double ShenandoahAllocationRate::upper_bound(double sds) const {
340   // Here we are using the standard deviation of the computed running
341   // average, rather than the standard deviation of the samples that went
342   // into the moving average. This is a much more stable value and is tied
343   // to the actual statistic in use (moving average over samples of averages).
344   return _rate.davg() + (sds * _rate_avg.dsd());
345 }
346 
347 void ShenandoahAllocationRate::allocation_counter_reset() {
348   _last_sample_time = os::elapsedTime();
349   _last_sample_value = 0;
350 }
351 
352 bool ShenandoahAllocationRate::is_spiking(double rate, double threshold) const {
353   if (rate <= 0.0) {
354     return false;
355   }
356 
357   double sd = _rate.sd();
358   if (sd > 0) {
359     // There is a small chance that that rate has already been sampled, but it
360     // seems not to matter in practice.
361     double z_score = (rate - _rate.avg()) / sd;
362     if (z_score > threshold) {
363       return true;
364     }
365   }
366   return false;
367 }
368 
369 double ShenandoahAllocationRate::instantaneous_rate(size_t allocated) const {
370   return instantaneous_rate(os::elapsedTime(), allocated);
371 }
372 
373 double ShenandoahAllocationRate::instantaneous_rate(double time, size_t allocated) const {
374   size_t last_value = _last_sample_value;
375   double last_time = _last_sample_time;
376   size_t allocation_delta = (allocated > last_value) ? (allocated - last_value) : 0;
377   double time_delta_sec = time - last_time;
378   return (time_delta_sec > 0)  ? (allocation_delta / time_delta_sec) : 0;
379 }