1 /*
  2  * Copyright (c) 2018, 2019, Red Hat, Inc. All rights reserved.
  3  * Copyright Amazon.com Inc. or its affiliates. All Rights Reserved.
  4  * DO NOT ALTER OR REMOVE COPYRIGHT NOTICES OR THIS FILE HEADER.
  5  *
  6  * This code is free software; you can redistribute it and/or modify it
  7  * under the terms of the GNU General Public License version 2 only, as
  8  * published by the Free Software Foundation.
  9  *
 10  * This code is distributed in the hope that it will be useful, but WITHOUT
 11  * ANY WARRANTY; without even the implied warranty of MERCHANTABILITY or
 12  * FITNESS FOR A PARTICULAR PURPOSE.  See the GNU General Public License
 13  * version 2 for more details (a copy is included in the LICENSE file that
 14  * accompanied this code).
 15  *
 16  * You should have received a copy of the GNU General Public License version
 17  * 2 along with this work; if not, write to the Free Software Foundation,
 18  * Inc., 51 Franklin St, Fifth Floor, Boston, MA 02110-1301 USA.
 19  *
 20  * Please contact Oracle, 500 Oracle Parkway, Redwood Shores, CA 94065 USA
 21  * or visit www.oracle.com if you need additional information or have any
 22  * questions.
 23  *
 24  */
 25 
 26 #include "precompiled.hpp"
 27 
 28 #include "gc/shared/gcCause.hpp"
 29 #include "gc/shenandoah/heuristics/shenandoahHeuristics.hpp"
 30 #include "gc/shenandoah/heuristics/shenandoahSpaceInfo.hpp"
 31 #include "gc/shenandoah/heuristics/shenandoahAdaptiveHeuristics.hpp"
 32 #include "gc/shenandoah/shenandoahCollectionSet.hpp"
 33 #include "gc/shenandoah/shenandoahCollectorPolicy.hpp"
 34 #include "gc/shenandoah/shenandoahFreeSet.hpp"
 35 #include "gc/shenandoah/shenandoahHeap.inline.hpp"
 36 #include "gc/shenandoah/shenandoahHeapRegion.inline.hpp"
 37 #include "logging/log.hpp"
 38 #include "logging/logTag.hpp"
 39 #include "runtime/globals.hpp"
 40 #include "utilities/quickSort.hpp"
 41 
 42 // These constants are used to adjust the margin of error for the moving
 43 // average of the allocation rate and cycle time. The units are standard
 44 // deviations.
 45 const double ShenandoahAdaptiveHeuristics::FULL_PENALTY_SD = 0.2;
 46 const double ShenandoahAdaptiveHeuristics::DEGENERATE_PENALTY_SD = 0.1;
 47 
 48 // These are used to decide if we want to make any adjustments at all
 49 // at the end of a successful concurrent cycle.
 50 const double ShenandoahAdaptiveHeuristics::LOWEST_EXPECTED_AVAILABLE_AT_END = -0.5;
 51 const double ShenandoahAdaptiveHeuristics::HIGHEST_EXPECTED_AVAILABLE_AT_END = 0.5;
 52 
 53 // These values are the confidence interval expressed as standard deviations.
 54 // At the minimum confidence level, there is a 25% chance that the true value of
 55 // the estimate (average cycle time or allocation rate) is not more than
 56 // MINIMUM_CONFIDENCE standard deviations away from our estimate. Similarly, the
 57 // MAXIMUM_CONFIDENCE interval here means there is a one in a thousand chance
 58 // that the true value of our estimate is outside the interval. These are used
 59 // as bounds on the adjustments applied at the outcome of a GC cycle.
 60 const double ShenandoahAdaptiveHeuristics::MINIMUM_CONFIDENCE = 0.319; // 25%
 61 const double ShenandoahAdaptiveHeuristics::MAXIMUM_CONFIDENCE = 3.291; // 99.9%
 62 
 63 ShenandoahAdaptiveHeuristics::ShenandoahAdaptiveHeuristics(ShenandoahSpaceInfo* space_info) :
 64   ShenandoahHeuristics(space_info),
 65   _margin_of_error_sd(ShenandoahAdaptiveInitialConfidence),
 66   _spike_threshold_sd(ShenandoahAdaptiveInitialSpikeThreshold),
 67   _last_trigger(OTHER),
 68   _available(Moving_Average_Samples, ShenandoahAdaptiveDecayFactor) { }
 69 
 70 ShenandoahAdaptiveHeuristics::~ShenandoahAdaptiveHeuristics() {}
 71 
 72 void ShenandoahAdaptiveHeuristics::choose_collection_set_from_regiondata(ShenandoahCollectionSet* cset,
 73                                                                          RegionData* data, size_t size,
 74                                                                          size_t actual_free) {
 75   size_t garbage_threshold = ShenandoahHeapRegion::region_size_bytes() * ShenandoahGarbageThreshold / 100;
 76 
 77   // The logic for cset selection in adaptive is as follows:
 78   //
 79   //   1. We cannot get cset larger than available free space. Otherwise we guarantee OOME
 80   //      during evacuation, and thus guarantee full GC. In practice, we also want to let
 81   //      application to allocate something. This is why we limit CSet to some fraction of
 82   //      available space. In non-overloaded heap, max_cset would contain all plausible candidates
 83   //      over garbage threshold.
 84   //
 85   //   2. We should not get cset too low so that free threshold would not be met right
 86   //      after the cycle. Otherwise we get back-to-back cycles for no reason if heap is
 87   //      too fragmented. In non-overloaded non-fragmented heap min_garbage would be around zero.
 88   //
 89   // Therefore, we start by sorting the regions by garbage. Then we unconditionally add the best candidates
 90   // before we meet min_garbage. Then we add all candidates that fit with a garbage threshold before
 91   // we hit max_cset. When max_cset is hit, we terminate the cset selection. Note that in this scheme,
 92   // ShenandoahGarbageThreshold is the soft threshold which would be ignored until min_garbage is hit.
 93 
 94   size_t capacity    = _space_info->soft_max_capacity();
 95   size_t max_cset    = (size_t)((1.0 * capacity / 100 * ShenandoahEvacReserve) / ShenandoahEvacWaste);
 96   size_t free_target = (capacity * ShenandoahMinFreeThreshold) / 100 + max_cset;
 97   size_t min_garbage = (free_target > actual_free) ? (free_target - actual_free) : 0;
 98 
 99   log_info(gc, ergo)("Adaptive CSet Selection. Target Free: " SIZE_FORMAT "%s, Actual Free: "
100                      SIZE_FORMAT "%s, Max Evacuation: " SIZE_FORMAT "%s, Min Garbage: " SIZE_FORMAT "%s",
101                      byte_size_in_proper_unit(free_target), proper_unit_for_byte_size(free_target),
102                      byte_size_in_proper_unit(actual_free), proper_unit_for_byte_size(actual_free),
103                      byte_size_in_proper_unit(max_cset),    proper_unit_for_byte_size(max_cset),
104                      byte_size_in_proper_unit(min_garbage), proper_unit_for_byte_size(min_garbage));
105 
106   // Better select garbage-first regions
107   QuickSort::sort(data, size, compare_by_garbage);
108 
109   size_t cur_cset = 0;
110   size_t cur_garbage = 0;
111 
112   for (size_t idx = 0; idx < size; idx++) {
113     ShenandoahHeapRegion* r = data[idx].get_region();
114 
115     size_t new_cset    = cur_cset + r->get_live_data_bytes();
116     size_t new_garbage = cur_garbage + r->garbage();
117 
118     if (new_cset > max_cset) {
119       break;
120     }
121 
122     if ((new_garbage < min_garbage) || (r->garbage() > garbage_threshold)) {
123       cset->add_region(r);
124       cur_cset = new_cset;
125       cur_garbage = new_garbage;
126     }
127   }
128 }
129 
130 void ShenandoahAdaptiveHeuristics::record_cycle_start() {
131   ShenandoahHeuristics::record_cycle_start();
132   _allocation_rate.allocation_counter_reset();
133 }
134 
135 void ShenandoahAdaptiveHeuristics::record_success_concurrent() {
136   ShenandoahHeuristics::record_success_concurrent();
137 
138   size_t available = _space_info->available();
139 
140   double z_score = 0.0;
141   double available_sd = _available.sd();
142   if (available_sd > 0) {
143     double available_avg = _available.avg();
144     z_score = (double(available) - available_avg) / available_sd;
145     log_debug(gc, ergo)("%s Available: " SIZE_FORMAT " %sB, z-score=%.3f. Average available: %.1f %sB +/- %.1f %sB.",
146                         _space_info->name(),
147                         byte_size_in_proper_unit(available), proper_unit_for_byte_size(available),
148                         z_score,
149                         byte_size_in_proper_unit(available_avg), proper_unit_for_byte_size(available_avg),
150                         byte_size_in_proper_unit(available_sd), proper_unit_for_byte_size(available_sd));
151   }
152 
153   _available.add(double(available));
154 
155   // In the case when a concurrent GC cycle completes successfully but with an
156   // unusually small amount of available memory we will adjust our trigger
157   // parameters so that they are more likely to initiate a new cycle.
158   // Conversely, when a GC cycle results in an above average amount of available
159   // memory, we will adjust the trigger parameters to be less likely to initiate
160   // a GC cycle.
161   //
162   // The z-score we've computed is in no way statistically related to the
163   // trigger parameters, but it has the nice property that worse z-scores for
164   // available memory indicate making larger adjustments to the trigger
165   // parameters. It also results in fewer adjustments as the application
166   // stabilizes.
167   //
168   // In order to avoid making endless and likely unnecessary adjustments to the
169   // trigger parameters, the change in available memory (with respect to the
170   // average) at the end of a cycle must be beyond these threshold values.
171   if (z_score < LOWEST_EXPECTED_AVAILABLE_AT_END ||
172       z_score > HIGHEST_EXPECTED_AVAILABLE_AT_END) {
173     // The sign is flipped because a negative z-score indicates that the
174     // available memory at the end of the cycle is below average. Positive
175     // adjustments make the triggers more sensitive (i.e., more likely to fire).
176     // The z-score also gives us a measure of just how far below normal. This
177     // property allows us to adjust the trigger parameters proportionally.
178     //
179     // The `100` here is used to attenuate the size of our adjustments. This
180     // number was chosen empirically. It also means the adjustments at the end of
181     // a concurrent cycle are an order of magnitude smaller than the adjustments
182     // made for a degenerated or full GC cycle (which themselves were also
183     // chosen empirically).
184     adjust_last_trigger_parameters(z_score / -100);
185   }
186 }
187 
188 void ShenandoahAdaptiveHeuristics::record_success_degenerated() {
189   ShenandoahHeuristics::record_success_degenerated();
190   // Adjust both trigger's parameters in the case of a degenerated GC because
191   // either of them should have triggered earlier to avoid this case.
192   adjust_margin_of_error(DEGENERATE_PENALTY_SD);
193   adjust_spike_threshold(DEGENERATE_PENALTY_SD);
194 }
195 
196 void ShenandoahAdaptiveHeuristics::record_success_full() {
197   ShenandoahHeuristics::record_success_full();
198   // Adjust both trigger's parameters in the case of a full GC because
199   // either of them should have triggered earlier to avoid this case.
200   adjust_margin_of_error(FULL_PENALTY_SD);
201   adjust_spike_threshold(FULL_PENALTY_SD);
202 }
203 
204 static double saturate(double value, double min, double max) {
205   return MAX2(MIN2(value, max), min);
206 }
207 
208 //  Rationale:
209 //    The idea is that there is an average allocation rate and there are occasional abnormal bursts (or spikes) of
210 //    allocations that exceed the average allocation rate.  What do these spikes look like?
211 //
212 //    1. At certain phase changes, we may discard large amounts of data and replace it with large numbers of newly
213 //       allocated objects.  This "spike" looks more like a phase change.  We were in steady state at M bytes/sec
214 //       allocation rate and now we're in a "reinitialization phase" that looks like N bytes/sec.  We need the "spike"
215 //       accommodation to give us enough runway to recalibrate our "average allocation rate".
216 //
217 //   2. The typical workload changes.  "Suddenly", our typical workload of N TPS increases to N+delta TPS.  This means
218 //       our average allocation rate needs to be adjusted.  Once again, we need the "spike" accomodation to give us
219 //       enough runway to recalibrate our "average allocation rate".
220 //
221 //    3. Though there is an "average" allocation rate, a given workload's demand for allocation may be very bursty.  We
222 //       allocate a bunch of LABs during the 5 ms that follow completion of a GC, then we perform no more allocations for
223 //       the next 150 ms.  It seems we want the "spike" to represent the maximum divergence from average within the
224 //       period of time between consecutive evaluation of the should_start_gc() service.  Here's the thinking:
225 //
226 //       a) Between now and the next time I ask whether should_start_gc(), we might experience a spike representing
227 //          the anticipated burst of allocations.  If that would put us over budget, then we should start GC immediately.
228 //       b) Between now and the anticipated depletion of allocation pool, there may be two or more bursts of allocations.
229 //          If there are more than one of these bursts, we can "approximate" that these will be separated by spans of
230 //          time with very little or no allocations so the "average" allocation rate should be a suitable approximation
231 //          of how this will behave.
232 //
233 //    For cases 1 and 2, we need to "quickly" recalibrate the average allocation rate whenever we detect a change
234 //    in operation mode.  We want some way to decide that the average rate has changed, while keeping average
235 //    allocation rate computation independent.
236 bool ShenandoahAdaptiveHeuristics::should_start_gc() {
237   size_t capacity = _space_info->soft_max_capacity();
238   size_t available = _space_info->soft_available();
239   size_t allocated = _space_info->bytes_allocated_since_gc_start();
240 
241   log_debug(gc)("should_start_gc (%s)? available: " SIZE_FORMAT ", soft_max_capacity: " SIZE_FORMAT
242                 ", allocated: " SIZE_FORMAT,
243                 _space_info->name(), available, capacity, allocated);
244 
245   // Track allocation rate even if we decide to start a cycle for other reasons.
246   double rate = _allocation_rate.sample(allocated);
247   _last_trigger = OTHER;
248 
249   size_t min_threshold = min_free_threshold();
250   if (available < min_threshold) {
251     log_info(gc)("Trigger (%s): Free (" SIZE_FORMAT "%s) is below minimum threshold (" SIZE_FORMAT "%s)", _space_info->name(),
252                  byte_size_in_proper_unit(available), proper_unit_for_byte_size(available),
253                  byte_size_in_proper_unit(min_threshold), proper_unit_for_byte_size(min_threshold));
254     return true;
255   }
256 
257   // Check if we need to learn a bit about the application
258   const size_t max_learn = ShenandoahLearningSteps;
259   if (_gc_times_learned < max_learn) {
260     size_t init_threshold = capacity / 100 * ShenandoahInitFreeThreshold;
261     if (available < init_threshold) {
262       log_info(gc)("Trigger (%s): Learning " SIZE_FORMAT " of " SIZE_FORMAT ". Free (" SIZE_FORMAT "%s) is below initial threshold (" SIZE_FORMAT "%s)",
263                    _space_info->name(), _gc_times_learned + 1, max_learn,
264                    byte_size_in_proper_unit(available), proper_unit_for_byte_size(available),
265                    byte_size_in_proper_unit(init_threshold), proper_unit_for_byte_size(init_threshold));
266       return true;
267     }
268   }
269   // Check if allocation headroom is still okay. This also factors in:
270   //   1. Some space to absorb allocation spikes (ShenandoahAllocSpikeFactor)
271   //   2. Accumulated penalties from Degenerated and Full GC
272   size_t allocation_headroom = available;
273 
274   size_t spike_headroom = capacity / 100 * ShenandoahAllocSpikeFactor;
275   size_t penalties      = capacity / 100 * _gc_time_penalties;
276 
277   allocation_headroom -= MIN2(allocation_headroom, spike_headroom);
278   allocation_headroom -= MIN2(allocation_headroom, penalties);
279 
280   double avg_cycle_time = _gc_cycle_time_history->davg() + (_margin_of_error_sd * _gc_cycle_time_history->dsd());
281   double avg_alloc_rate = _allocation_rate.upper_bound(_margin_of_error_sd);
282 
283   log_debug(gc)("%s: average GC time: %.2f ms, allocation rate: %.0f %s/s",
284                 _space_info->name(),
285           avg_cycle_time * 1000, byte_size_in_proper_unit(avg_alloc_rate), proper_unit_for_byte_size(avg_alloc_rate));
286   if (avg_cycle_time * avg_alloc_rate > allocation_headroom) {
287     log_info(gc)("Trigger (%s): Average GC time (%.2f ms) is above the time for average allocation rate (%.0f %sB/s)"
288                  " to deplete free headroom (" SIZE_FORMAT "%s) (margin of error = %.2f)",
289                  _space_info->name(), avg_cycle_time * 1000,
290                  byte_size_in_proper_unit(avg_alloc_rate), proper_unit_for_byte_size(avg_alloc_rate),
291                  byte_size_in_proper_unit(allocation_headroom), proper_unit_for_byte_size(allocation_headroom),
292                  _margin_of_error_sd);
293     log_info(gc, ergo)("Free headroom: " SIZE_FORMAT "%s (free) - " SIZE_FORMAT "%s (spike) - " SIZE_FORMAT "%s (penalties) = " SIZE_FORMAT "%s",
294                        byte_size_in_proper_unit(available),           proper_unit_for_byte_size(available),
295                        byte_size_in_proper_unit(spike_headroom),      proper_unit_for_byte_size(spike_headroom),
296                        byte_size_in_proper_unit(penalties),           proper_unit_for_byte_size(penalties),
297                        byte_size_in_proper_unit(allocation_headroom), proper_unit_for_byte_size(allocation_headroom));
298     _last_trigger = RATE;
299     return true;
300   }
301 
302   bool is_spiking = _allocation_rate.is_spiking(rate, _spike_threshold_sd);
303   if (is_spiking && avg_cycle_time > allocation_headroom / rate) {
304     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)",
305                  _space_info->name(), avg_cycle_time * 1000,
306                  byte_size_in_proper_unit(rate), proper_unit_for_byte_size(rate),
307                  byte_size_in_proper_unit(allocation_headroom), proper_unit_for_byte_size(allocation_headroom),
308                  _spike_threshold_sd);
309     _last_trigger = SPIKE;
310     return true;
311   }
312 
313   return ShenandoahHeuristics::should_start_gc();
314 }
315 
316 void ShenandoahAdaptiveHeuristics::adjust_last_trigger_parameters(double amount) {
317   switch (_last_trigger) {
318     case RATE:
319       adjust_margin_of_error(amount);
320       break;
321     case SPIKE:
322       adjust_spike_threshold(amount);
323       break;
324     case OTHER:
325       // nothing to adjust here.
326       break;
327     default:
328       ShouldNotReachHere();
329   }
330 }
331 
332 void ShenandoahAdaptiveHeuristics::adjust_margin_of_error(double amount) {
333   _margin_of_error_sd = saturate(_margin_of_error_sd + amount, MINIMUM_CONFIDENCE, MAXIMUM_CONFIDENCE);
334   log_debug(gc, ergo)("Margin of error now %.2f", _margin_of_error_sd);
335 }
336 
337 void ShenandoahAdaptiveHeuristics::adjust_spike_threshold(double amount) {
338   _spike_threshold_sd = saturate(_spike_threshold_sd - amount, MINIMUM_CONFIDENCE, MAXIMUM_CONFIDENCE);
339   log_debug(gc, ergo)("Spike threshold now: %.2f", _spike_threshold_sd);
340 }
341 
342 size_t ShenandoahAdaptiveHeuristics::min_free_threshold() {
343   // Note that soft_max_capacity() / 100 * min_free_threshold is smaller than max_capacity() / 100 * min_free_threshold.
344   // We want to behave conservatively here, so use max_capacity().  By returning a larger value, we cause the GC to
345   // trigger when the remaining amount of free shrinks below the larger threshold.
346   return _space_info->max_capacity() / 100 * ShenandoahMinFreeThreshold;
347 }
348 
349 ShenandoahAllocationRate::ShenandoahAllocationRate() :
350   _last_sample_time(os::elapsedTime()),
351   _last_sample_value(0),
352   _interval_sec(1.0 / ShenandoahAdaptiveSampleFrequencyHz),
353   _rate(int(ShenandoahAdaptiveSampleSizeSeconds * ShenandoahAdaptiveSampleFrequencyHz), ShenandoahAdaptiveDecayFactor),
354   _rate_avg(int(ShenandoahAdaptiveSampleSizeSeconds * ShenandoahAdaptiveSampleFrequencyHz), ShenandoahAdaptiveDecayFactor) {
355 }
356 
357 double ShenandoahAllocationRate::sample(size_t allocated) {
358   double now = os::elapsedTime();
359   double rate = 0.0;
360   if (now - _last_sample_time > _interval_sec) {
361     if (allocated >= _last_sample_value) {
362       rate = instantaneous_rate(now, allocated);
363       _rate.add(rate);
364       _rate_avg.add(_rate.avg());
365     }
366 
367     _last_sample_time = now;
368     _last_sample_value = allocated;
369   }
370   return rate;
371 }
372 
373 double ShenandoahAllocationRate::upper_bound(double sds) const {
374   // Here we are using the standard deviation of the computed running
375   // average, rather than the standard deviation of the samples that went
376   // into the moving average. This is a much more stable value and is tied
377   // to the actual statistic in use (moving average over samples of averages).
378   return _rate.davg() + (sds * _rate_avg.dsd());
379 }
380 
381 void ShenandoahAllocationRate::allocation_counter_reset() {
382   _last_sample_time = os::elapsedTime();
383   _last_sample_value = 0;
384 }
385 
386 bool ShenandoahAllocationRate::is_spiking(double rate, double threshold) const {
387   if (rate <= 0.0) {
388     return false;
389   }
390 
391   double sd = _rate.sd();
392   if (sd > 0) {
393     // There is a small chance that that rate has already been sampled, but it
394     // seems not to matter in practice.
395     double z_score = (rate - _rate.avg()) / sd;
396     if (z_score > threshold) {
397       return true;
398     }
399   }
400   return false;
401 }
402 
403 double ShenandoahAllocationRate::instantaneous_rate(double time, size_t allocated) const {
404   size_t last_value = _last_sample_value;
405   double last_time = _last_sample_time;
406   size_t allocation_delta = (allocated > last_value) ? (allocated - last_value) : 0;
407   double time_delta_sec = time - last_time;
408   return (time_delta_sec > 0)  ? (allocation_delta / time_delta_sec) : 0;
409 }