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 "gc/shenandoah/shenandoahYoungGeneration.hpp"
33 #include "logging/log.hpp"
34 #include "logging/logTag.hpp"
35 #include "utilities/quickSort.hpp"
36
37 // These constants are used to adjust the margin of error for the moving
38 // average of the allocation rate and cycle time. The units are standard
39 // deviations.
40 const double ShenandoahAdaptiveHeuristics::FULL_PENALTY_SD = 0.2;
41 const double ShenandoahAdaptiveHeuristics::DEGENERATE_PENALTY_SD = 0.1;
42
43 // These are used to decide if we want to make any adjustments at all
44 // at the end of a successful concurrent cycle.
45 const double ShenandoahAdaptiveHeuristics::LOWEST_EXPECTED_AVAILABLE_AT_END = -0.5;
46 const double ShenandoahAdaptiveHeuristics::HIGHEST_EXPECTED_AVAILABLE_AT_END = 0.5;
47
48 // These values are the confidence interval expressed as standard deviations.
49 // At the minimum confidence level, there is a 25% chance that the true value of
50 // the estimate (average cycle time or allocation rate) is not more than
51 // MINIMUM_CONFIDENCE standard deviations away from our estimate. Similarly, the
52 // MAXIMUM_CONFIDENCE interval here means there is a one in a thousand chance
53 // that the true value of our estimate is outside the interval. These are used
54 // as bounds on the adjustments applied at the outcome of a GC cycle.
55 const double ShenandoahAdaptiveHeuristics::MINIMUM_CONFIDENCE = 0.319; // 25%
56 const double ShenandoahAdaptiveHeuristics::MAXIMUM_CONFIDENCE = 3.291; // 99.9%
57
58 ShenandoahAdaptiveHeuristics::ShenandoahAdaptiveHeuristics(ShenandoahGeneration* generation) :
59 ShenandoahHeuristics(generation),
60 _margin_of_error_sd(ShenandoahAdaptiveInitialConfidence),
61 _spike_threshold_sd(ShenandoahAdaptiveInitialSpikeThreshold),
62 _last_trigger(OTHER) { }
63
64 ShenandoahAdaptiveHeuristics::~ShenandoahAdaptiveHeuristics() {}
65
66 void ShenandoahAdaptiveHeuristics::choose_collection_set_from_regiondata(ShenandoahCollectionSet* cset,
67 RegionData* data, size_t size,
68 size_t actual_free) {
69 size_t garbage_threshold = ShenandoahHeapRegion::region_size_bytes() * ShenandoahGarbageThreshold / 100;
70 size_t ignore_threshold = ShenandoahHeapRegion::region_size_bytes() * ShenandoahIgnoreGarbageThreshold / 100;
71 ShenandoahHeap* heap = ShenandoahHeap::heap();
72
73 // The logic for cset selection in adaptive is as follows:
74 //
75 // 1. We cannot get cset larger than available free space. Otherwise we guarantee OOME
76 // during evacuation, and thus guarantee full GC. In practice, we also want to let
77 // application to allocate something. This is why we limit CSet to some fraction of
78 // available space. In non-overloaded heap, max_cset would contain all plausible candidates
79 // over garbage threshold.
80 //
81 // 2. We should not get cset too low so that free threshold would not be met right
82 // after the cycle. Otherwise we get back-to-back cycles for no reason if heap is
83 // too fragmented. In non-overloaded non-fragmented heap min_garbage would be around zero.
84 //
85 // Therefore, we start by sorting the regions by garbage. Then we unconditionally add the best candidates
86 // before we meet min_garbage. Then we add all candidates that fit with a garbage threshold before
87 // we hit max_cset. When max_cset is hit, we terminate the cset selection. Note that in this scheme,
88 // ShenandoahGarbageThreshold is the soft threshold which would be ignored until min_garbage is hit.
89
90 // In generational mode, the sort order within the data array is not strictly descending amounts of garbage. In
91 // particular, regions that have reached tenure age will be sorted into this array before younger regions that contain
92 // more garbage. This represents one of the reasons why we keep looking at regions even after we decide, for example,
93 // to exclude one of the regions because it might require evacuation of too much live data.
94 bool is_generational = heap->mode()->is_generational();
95 bool is_global = (_generation->generation_mode() == GLOBAL);
96 size_t capacity = heap->young_generation()->max_capacity();
97
98 // cur_young_garbage represents the amount of memory to be reclaimed from young-gen. In the case that live objects
99 // are known to be promoted out of young-gen, we count this as cur_young_garbage because this memory is reclaimed
100 // from young-gen and becomes available to serve future young-gen allocation requests.
101 size_t cur_young_garbage = 0;
102
103 // Better select garbage-first regions
104 QuickSort::sort<RegionData>(data, (int)size, compare_by_garbage, false);
105
106 if (is_generational) {
107 if (is_global) {
108 size_t max_young_cset = (size_t) (heap->get_young_evac_reserve() / ShenandoahEvacWaste);
109 size_t young_cur_cset = 0;
110 size_t max_old_cset = (size_t) (heap->get_old_evac_reserve() / ShenandoahEvacWaste);
111 size_t old_cur_cset = 0;
112 size_t free_target = (capacity * ShenandoahMinFreeThreshold) / 100 + max_young_cset;
113 size_t min_garbage = (free_target > actual_free) ? (free_target - actual_free) : 0;
114
115 log_info(gc, ergo)("Adaptive CSet Selection for GLOBAL. Max Young Cset: " SIZE_FORMAT
116 "%s, Max Old CSet: " SIZE_FORMAT "%s, Actual Free: " SIZE_FORMAT "%s.",
117 byte_size_in_proper_unit(max_young_cset), proper_unit_for_byte_size(max_young_cset),
118 byte_size_in_proper_unit(max_old_cset), proper_unit_for_byte_size(max_old_cset),
119 byte_size_in_proper_unit(actual_free), proper_unit_for_byte_size(actual_free));
120
121 for (size_t idx = 0; idx < size; idx++) {
122 ShenandoahHeapRegion* r = data[idx]._region;
123 bool add_region = false;
124 if (r->is_old()) {
125 size_t new_cset = old_cur_cset + r->get_live_data_bytes();
126 if ((new_cset <= max_old_cset) && (r->garbage() > garbage_threshold)) {
127 add_region = true;
128 old_cur_cset = new_cset;
129 }
130 } else if (cset->is_preselected(r->index())) {
131 assert(r->age() >= InitialTenuringThreshold, "Preselected regions must have tenure age");
132 // Entire region will be promoted, This region does not impact young-gen or old-gen evacuation reserve.
133 // This region has been pre-selected and its impact on promotion reserve is already accounted for.
134 add_region = true;
135 // r->used() is r->garbage() + r->get_live_data_bytes()
136 // Since all live data in this region is being evacuated from young-gen, it is as if this memory
137 // is garbage insofar as young-gen is concerned. Counting this as garbage reduces the need to
138 // reclaim highly utilized young-gen regions just for the sake of finding min_garbage to reclaim
139 // within youn-gen memory.
140 cur_young_garbage += r->used();
141 } else if (r->age() < InitialTenuringThreshold) {
142 size_t new_cset = young_cur_cset + r->get_live_data_bytes();
143 size_t region_garbage = r->garbage();
144 size_t new_garbage = cur_young_garbage + region_garbage;
145 bool add_regardless = (region_garbage > ignore_threshold) && (new_garbage < min_garbage);
146 if ((new_cset <= max_young_cset) && (add_regardless || (region_garbage > garbage_threshold))) {
147 add_region = true;
148 young_cur_cset = new_cset;
149 cur_young_garbage = new_garbage;
150 }
151 }
152 // Note that we do not add aged regions if they were not pre-selected. The reason they were not preselected
153 // is because there is not sufficient room in old-gen to hold their to-be-promoted live objects.
154
155 if (add_region) {
156 cset->add_region(r);
157 }
158 }
159 } else {
160 // This is young-gen collection or a mixed evacuation. If this is mixed evacuation, the old-gen candidate regions
161 // have already been added.
162 size_t max_cset = (size_t) (heap->get_young_evac_reserve() / ShenandoahEvacWaste);
163 size_t cur_cset = 0;
164 size_t free_target = (capacity * ShenandoahMinFreeThreshold) / 100 + max_cset;
165 size_t min_garbage = (free_target > actual_free) ? (free_target - actual_free) : 0;
166
167 log_info(gc, ergo)("Adaptive CSet Selection for YOUNG. Max CSet: " SIZE_FORMAT "%s, Actual Free: " SIZE_FORMAT "%s.",
168 byte_size_in_proper_unit(max_cset), proper_unit_for_byte_size(max_cset),
169 byte_size_in_proper_unit(actual_free), proper_unit_for_byte_size(actual_free));
170
171 for (size_t idx = 0; idx < size; idx++) {
172 ShenandoahHeapRegion* r = data[idx]._region;
173 size_t new_cset;
174 size_t region_garbage = r->garbage();
175 size_t new_garbage = cur_young_garbage + region_garbage;
176 bool add_region = false;
177
178 if (!r->is_old()) {
179 if (cset->is_preselected(r->index())) {
180 assert(r->age() >= InitialTenuringThreshold, "Preselected regions must have tenure age");
181 // Entire region will be promoted, This region does not impact young-gen evacuation reserve. Memory has already
182 // been set aside to hold evacuation results as advance_promotion_reserve.
183 add_region = true;
184 new_cset = cur_cset;
185 // Since all live data in this region is being evacuated from young-gen, it is as if this memory
186 // is garbage insofar as young-gen is concerned. Counting this as garbage reduces the need to
187 // reclaim highly utilized young-gen regions just for the sake of finding min_garbage to reclaim
188 // within youn-gen memory
189 cur_young_garbage += r->get_live_data_bytes();
190 } else if (r->age() < InitialTenuringThreshold) {
191 new_cset = cur_cset + r->get_live_data_bytes();
192 size_t region_garbage = r->garbage();
193 size_t new_garbage = cur_young_garbage + region_garbage;
194 bool add_regardless = (region_garbage > ignore_threshold) && (new_garbage < min_garbage);
195 if ((new_cset <= max_cset) && (add_regardless || (region_garbage > garbage_threshold))) {
196 add_region = true;
197 cur_cset = new_cset;
198 cur_young_garbage = new_garbage;
199 }
200 }
201 // Note that we do not add aged regions if they were not pre-selected. The reason they were not preselected
202 // is because there is not sufficient room in old-gen to hold their to-be-promoted live objects.
203
204 if (add_region) {
205 cset->add_region(r);
206 }
207 }
208 }
209 }
210 } else {
211 // Traditional Shenandoah (non-generational)
212 size_t max_cset = (size_t) (heap->get_young_evac_reserve() / ShenandoahEvacWaste);
213 size_t cur_cset = 0;
214 size_t free_target = (capacity * ShenandoahMinFreeThreshold) / 100 + max_cset;
215 size_t min_garbage = (free_target > actual_free) ? (free_target - actual_free) : 0;
216
217 log_info(gc, ergo)("Adaptive CSet Selection. Max CSet: " SIZE_FORMAT "%s, Actual Free: " SIZE_FORMAT "%s.",
218 byte_size_in_proper_unit(max_cset), proper_unit_for_byte_size(max_cset),
219 byte_size_in_proper_unit(actual_free), proper_unit_for_byte_size(actual_free));
220
221 for (size_t idx = 0; idx < size; idx++) {
222 ShenandoahHeapRegion* r = data[idx]._region;
223 size_t new_cset = cur_cset + r->get_live_data_bytes();
224 size_t region_garbage = r->garbage();
225 size_t new_garbage = cur_young_garbage + region_garbage;
226 bool add_regardless = (region_garbage > ignore_threshold) && (new_garbage < min_garbage);
227 if ((new_cset <= max_cset) && (add_regardless || (region_garbage > garbage_threshold))) {
228 cset->add_region(r);
229 cur_cset = new_cset;
230 cur_young_garbage = new_garbage;
231 }
232 }
233 }
234 }
235
236 void ShenandoahAdaptiveHeuristics::record_cycle_start() {
237 ShenandoahHeuristics::record_cycle_start();
238 _allocation_rate.allocation_counter_reset();
239 }
240
241 void ShenandoahAdaptiveHeuristics::record_success_concurrent(bool abbreviated) {
242 ShenandoahHeuristics::record_success_concurrent(abbreviated);
243
244 size_t available = ShenandoahHeap::heap()->free_set()->available();
245
246 _available.add(available);
247 double z_score = 0.0;
248 if (_available.sd() > 0) {
249 z_score = (available - _available.avg()) / _available.sd();
250 }
251
252 log_debug(gc, ergo)("Available: " SIZE_FORMAT " %sB, z-score=%.3f. Average available: %.1f %sB +/- %.1f %sB.",
253 byte_size_in_proper_unit(available), proper_unit_for_byte_size(available),
254 z_score,
255 byte_size_in_proper_unit(_available.avg()), proper_unit_for_byte_size(_available.avg()),
256 byte_size_in_proper_unit(_available.sd()), proper_unit_for_byte_size(_available.sd()));
257
258 // In the case when a concurrent GC cycle completes successfully but with an
259 // unusually small amount of available memory we will adjust our trigger
260 // parameters so that they are more likely to initiate a new cycle.
261 // Conversely, when a GC cycle results in an above average amount of available
262 // memory, we will adjust the trigger parameters to be less likely to initiate
263 // a GC cycle.
264 //
265 // The z-score we've computed is in no way statistically related to the
266 // trigger parameters, but it has the nice property that worse z-scores for
267 // available memory indicate making larger adjustments to the trigger
268 // parameters. It also results in fewer adjustments as the application
269 // stabilizes.
270 //
271 // In order to avoid making endless and likely unnecessary adjustments to the
272 // trigger parameters, the change in available memory (with respect to the
273 // average) at the end of a cycle must be beyond these threshold values.
274 if (z_score < LOWEST_EXPECTED_AVAILABLE_AT_END ||
275 z_score > HIGHEST_EXPECTED_AVAILABLE_AT_END) {
276 // The sign is flipped because a negative z-score indicates that the
277 // available memory at the end of the cycle is below average. Positive
278 // adjustments make the triggers more sensitive (i.e., more likely to fire).
279 // The z-score also gives us a measure of just how far below normal. This
280 // property allows us to adjust the trigger parameters proportionally.
281 //
282 // The `100` here is used to attenuate the size of our adjustments. This
283 // number was chosen empirically. It also means the adjustments at the end of
284 // a concurrent cycle are an order of magnitude smaller than the adjustments
285 // made for a degenerated or full GC cycle (which themselves were also
286 // chosen empirically).
287 adjust_last_trigger_parameters(z_score / -100);
288 }
289 }
290
291 void ShenandoahAdaptiveHeuristics::record_success_degenerated() {
292 ShenandoahHeuristics::record_success_degenerated();
293 // Adjust both trigger's parameters in the case of a degenerated GC because
294 // either of them should have triggered earlier to avoid this case.
295 adjust_margin_of_error(DEGENERATE_PENALTY_SD);
296 adjust_spike_threshold(DEGENERATE_PENALTY_SD);
297 }
298
299 void ShenandoahAdaptiveHeuristics::record_success_full() {
300 ShenandoahHeuristics::record_success_full();
301 // Adjust both trigger's parameters in the case of a full GC because
302 // either of them should have triggered earlier to avoid this case.
303 adjust_margin_of_error(FULL_PENALTY_SD);
304 adjust_spike_threshold(FULL_PENALTY_SD);
305 }
306
307 static double saturate(double value, double min, double max) {
308 return MAX2(MIN2(value, max), min);
309 }
310
311 bool ShenandoahAdaptiveHeuristics::should_start_gc() {
312 size_t max_capacity = _generation->max_capacity();
313 size_t capacity = _generation->soft_max_capacity();
314 size_t available = _generation->available();
315 size_t allocated = _generation->bytes_allocated_since_gc_start();
316
317 log_debug(gc)("should_start_gc (%s)? available: " SIZE_FORMAT ", soft_max_capacity: " SIZE_FORMAT
318 ", max_capacity: " SIZE_FORMAT ", allocated: " SIZE_FORMAT,
319 _generation->name(), available, capacity, max_capacity, allocated);
320
321 // Make sure the code below treats available without the soft tail.
322 size_t soft_tail = max_capacity - capacity;
323 available = (available > soft_tail) ? (available - soft_tail) : 0;
324
325 // The collector reserve may eat into what the mutator is allowed to use. Make sure we are looking
326 // at what is available to the mutator when deciding whether to start a GC.
327 size_t usable = ShenandoahHeap::heap()->free_set()->available();
328 if (usable < available) {
329 log_debug(gc)("Usable (" SIZE_FORMAT "%s) is less than available (" SIZE_FORMAT "%s)",
330 byte_size_in_proper_unit(usable), proper_unit_for_byte_size(usable),
331 byte_size_in_proper_unit(available), proper_unit_for_byte_size(available));
332 available = usable;
333 }
334
335 // Allocation spikes are a characteristic of both the application ahd the JVM configuration. On the JVM command line,
336 // the application developer may want to supply a hint of the nature of spikes that are inherent in the application
337 // workload, and this information would normally be independent of heap size (not a percentage thereof). On the
338 // other hand, some allocation spikes are correlated with JVM configuration. For example, there are allocation
339 // spikes at the starts of concurrent marking and evacuation to refresh all local allocation buffers. The nature
340 // of these spikes is determined by LAB min and max sizes and numbers of threads, but also on frequency of GC passes,
341 // and on "periodic" behavior of these threads If GC frequency is much higher than the periodic trigger for mutator
342 // threads, then many of the mutator threads may be able to "sit out" of most GC passes. Though the thread's stack
343 // must be scanned, the thread does not need to refresh its LABs if it sits idle throughout the duration of the GC
344 // pass. The best prediction for this aspect of spikes in allocation patterns is probably recent past history.
345 // TODO: and dive deeper into _gc_time_penalties as this may also need to be corrected
346
347 // Check if allocation headroom is still okay. This also factors in:
348 // 1. Some space to absorb allocation spikes (ShenandoahAllocSpikeFactor)
349 // 2. Accumulated penalties from Degenerated and Full GC
350 size_t allocation_headroom = available;
351 size_t spike_headroom = capacity / 100 * ShenandoahAllocSpikeFactor;
352 size_t penalties = capacity / 100 * _gc_time_penalties;
353
354 allocation_headroom -= MIN2(allocation_headroom, penalties);
355 allocation_headroom -= MIN2(allocation_headroom, spike_headroom);
356
357 // Track allocation rate even if we decide to start a cycle for other reasons.
358 double rate = _allocation_rate.sample(allocated);
359 _last_trigger = OTHER;
360
361 size_t min_threshold = min_free_threshold();
362
363 if (allocation_headroom < min_threshold) {
364 log_info(gc)("Trigger (%s): Free (" SIZE_FORMAT "%s) is below minimum threshold (" SIZE_FORMAT "%s)",
365 _generation->name(),
366 byte_size_in_proper_unit(available), proper_unit_for_byte_size(available),
367 byte_size_in_proper_unit(min_threshold), proper_unit_for_byte_size(min_threshold));
368 return true;
369 }
370
371 // Check if we need to learn a bit about the application
372 const size_t max_learn = ShenandoahLearningSteps;
373 if (_gc_times_learned < max_learn) {
374 size_t init_threshold = capacity / 100 * ShenandoahInitFreeThreshold;
375 if (allocation_headroom < init_threshold) {
376 log_info(gc)("Trigger (%s): Learning " SIZE_FORMAT " of " SIZE_FORMAT ". Free (" SIZE_FORMAT "%s) is below initial threshold (" SIZE_FORMAT "%s)",
377 _generation->name(), _gc_times_learned + 1, max_learn,
378 byte_size_in_proper_unit(available), proper_unit_for_byte_size(available),
379 byte_size_in_proper_unit(init_threshold), proper_unit_for_byte_size(init_threshold));
380 return true;
381 }
382 }
383
384 // Rationale:
385 // The idea is that there is an average allocation rate and there are occasional abnormal bursts (or spikes) of
386 // allocations that exceed the average allocation rate. What do these spikes look like?
387 //
388 // 1. At certain phase changes, we may discard large amounts of data and replace it with large numbers of newly
389 // allocated objects. This "spike" looks more like a phase change. We were in steady state at M bytes/sec
390 // allocation rate and now we're in a "reinitialization phase" that looks like N bytes/sec. We need the "spike"
391 // accomodation to give us enough runway to recalibrate our "average allocation rate".
392 //
393 // 2. The typical workload changes. "Suddenly", our typical workload of N TPS increases to N+delta TPS. This means
394 // our average allocation rate needs to be adjusted. Once again, we need the "spike" accomodation to give us
395 // enough runway to recalibrate our "average allocation rate".
396 //
397 // 3. Though there is an "average" allocation rate, a given workload's demand for allocation may be very bursty. We
398 // allocate a bunch of LABs during the 5 ms that follow completion of a GC, then we perform no more allocations for
399 // the next 150 ms. It seems we want the "spike" to represent the maximum divergence from average within the
400 // period of time between consecutive evaluation of the should_start_gc() service. Here's the thinking:
401 //
402 // a) Between now and the next time I ask whether should_start_gc(), we might experience a spike representing
403 // the anticipated burst of allocations. If that would put us over budget, then we should start GC immediately.
404 // b) Between now and the anticipated depletion of allocation pool, there may be two or more bursts of allocations.
405 // If there are more than one of these bursts, we can "approximate" that these will be separated by spans of
406 // time with very little or no allocations so the "average" allocation rate should be a suitable approximation
407 // of how this will behave.
408 //
409 // For cases 1 and 2, we need to "quickly" recalibrate the average allocation rate whenever we detect a change
410 // in operation mode. We want some way to decide that the average rate has changed. Make average allocation rate
411 // computations an independent effort.
412
413
414 // TODO: Account for inherent delays in responding to GC triggers
415 // 1. It has been observed that delays of 200 ms or greater are common between the moment we return true from should_start_gc()
416 // and the moment at which we begin execution of the concurrent reset phase. Add this time into the calculation of
417 // avg_cycle_time below. (What is "this time"? Perhaps we should remember recent history of this delay for the
418 // running workload and use the maximum delay recently seen for "this time".)
419 // 2. The frequency of inquiries to should_start_gc() is adaptive, ranging between ShenandoahControlIntervalMin and
420 // ShenandoahControlIntervalMax. The current control interval (or the max control interval) should also be added into
421 // the calculation of avg_cycle_time below.
422
423 double avg_cycle_time = _gc_time_history->davg() + (_margin_of_error_sd * _gc_time_history->dsd());
424
425 size_t last_live_memory = get_last_live_memory();
426 size_t penultimate_live_memory = get_penultimate_live_memory();
427 double original_cycle_time = avg_cycle_time;
428 if ((penultimate_live_memory < last_live_memory) && (penultimate_live_memory != 0)) {
429 // If the live-memory size is growing, our estimates of cycle time are based on lighter workload, so adjust.
430 // TODO: Be more precise about how to scale when live memory is growing. Existing code is a very rough approximation
431 // tuned with very limited workload observations.
432 avg_cycle_time = (avg_cycle_time * 2 * last_live_memory) / penultimate_live_memory;
433 } else {
434 int degen_cycles = degenerated_cycles_in_a_row();
435 if (degen_cycles > 0) {
436 // If we've degenerated recently, we might be waiting too long between triggers so adjust trigger forward.
437 // TODO: Be more precise about how to scale when we've experienced recent degenerated GC. Existing code is a very
438 // rough approximation tuned with very limited workload observations.
439 avg_cycle_time += degen_cycles * avg_cycle_time;
440 }
441 }
442
443 double avg_alloc_rate = _allocation_rate.upper_bound(_margin_of_error_sd);
444 log_debug(gc)("%s: average GC time: %.2f ms, allocation rate: %.0f %s/s",
445 _generation->name(), avg_cycle_time * 1000, byte_size_in_proper_unit(avg_alloc_rate), proper_unit_for_byte_size(avg_alloc_rate));
446
447 if (avg_cycle_time > allocation_headroom / avg_alloc_rate) {
448 if (avg_cycle_time > original_cycle_time) {
449 log_debug(gc)("%s: average GC time adjusted from: %.2f ms to %.2f ms because upward trend in live memory retention",
450 _generation->name(), original_cycle_time, avg_cycle_time);
451 }
452
453 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)",
454 _generation->name(), avg_cycle_time * 1000,
455 byte_size_in_proper_unit(avg_alloc_rate), proper_unit_for_byte_size(avg_alloc_rate),
456 byte_size_in_proper_unit(allocation_headroom), proper_unit_for_byte_size(allocation_headroom),
457 _margin_of_error_sd);
458
459 log_info(gc, ergo)("Free headroom: " SIZE_FORMAT "%s (free) - " SIZE_FORMAT "%s (spike) - " SIZE_FORMAT "%s (penalties) = " SIZE_FORMAT "%s",
460 byte_size_in_proper_unit(available), proper_unit_for_byte_size(available),
461 byte_size_in_proper_unit(spike_headroom), proper_unit_for_byte_size(spike_headroom),
462 byte_size_in_proper_unit(penalties), proper_unit_for_byte_size(penalties),
463 byte_size_in_proper_unit(allocation_headroom), proper_unit_for_byte_size(allocation_headroom));
464
465 _last_trigger = RATE;
466 return true;
467 }
468
469 bool is_spiking = _allocation_rate.is_spiking(rate, _spike_threshold_sd);
470 if (is_spiking && avg_cycle_time > allocation_headroom / rate) {
471 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)",
472 _generation->name(), avg_cycle_time * 1000,
473 byte_size_in_proper_unit(rate), proper_unit_for_byte_size(rate),
474 byte_size_in_proper_unit(allocation_headroom), proper_unit_for_byte_size(allocation_headroom),
475
476 _spike_threshold_sd);
477 _last_trigger = SPIKE;
478 return true;
479 }
480
481 return ShenandoahHeuristics::should_start_gc();
482 }
483
484 void ShenandoahAdaptiveHeuristics::adjust_last_trigger_parameters(double amount) {
485 switch (_last_trigger) {
486 case RATE:
487 adjust_margin_of_error(amount);
488 break;
489 case SPIKE:
490 adjust_spike_threshold(amount);
491 break;
492 case OTHER:
493 // nothing to adjust here.
494 break;
495 default:
496 ShouldNotReachHere();
497 }
498 }
499
500 void ShenandoahAdaptiveHeuristics::adjust_margin_of_error(double amount) {
501 _margin_of_error_sd = saturate(_margin_of_error_sd + amount, MINIMUM_CONFIDENCE, MAXIMUM_CONFIDENCE);
502 log_debug(gc, ergo)("Margin of error now %.2f", _margin_of_error_sd);
503 }
504
505 void ShenandoahAdaptiveHeuristics::adjust_spike_threshold(double amount) {
506 _spike_threshold_sd = saturate(_spike_threshold_sd - amount, MINIMUM_CONFIDENCE, MAXIMUM_CONFIDENCE);
507 log_debug(gc, ergo)("Spike threshold now: %.2f", _spike_threshold_sd);
508 }
509
510 ShenandoahAllocationRate::ShenandoahAllocationRate() :
511 _last_sample_time(os::elapsedTime()),
512 _last_sample_value(0),
513 _interval_sec(1.0 / ShenandoahAdaptiveSampleFrequencyHz),
514 _rate(int(ShenandoahAdaptiveSampleSizeSeconds * ShenandoahAdaptiveSampleFrequencyHz), ShenandoahAdaptiveDecayFactor),
515 _rate_avg(int(ShenandoahAdaptiveSampleSizeSeconds * ShenandoahAdaptiveSampleFrequencyHz), ShenandoahAdaptiveDecayFactor) {
516 }
517
518 double ShenandoahAllocationRate::sample(size_t allocated) {
519 double now = os::elapsedTime();
520 double rate = 0.0;
521 if (now - _last_sample_time > _interval_sec) {
522 if (allocated >= _last_sample_value) {
523 rate = instantaneous_rate(now, allocated);
524 _rate.add(rate);
525 _rate_avg.add(_rate.avg());
526 }
527
528 _last_sample_time = now;
529 _last_sample_value = allocated;
530 }
531 return rate;
532 }
533
534 double ShenandoahAllocationRate::upper_bound(double sds) const {
535 // Here we are using the standard deviation of the computed running
536 // average, rather than the standard deviation of the samples that went
537 // into the moving average. This is a much more stable value and is tied
538 // to the actual statistic in use (moving average over samples of averages).
539 return _rate.davg() + (sds * _rate_avg.dsd());
540 }
541
542 void ShenandoahAllocationRate::allocation_counter_reset() {
543 _last_sample_time = os::elapsedTime();
544 _last_sample_value = 0;
545 }
546
547 bool ShenandoahAllocationRate::is_spiking(double rate, double threshold) const {
548 if (rate <= 0.0) {
549 return false;
550 }
551
552 double sd = _rate.sd();
553 if (sd > 0) {
554 // There is a small chance that that rate has already been sampled, but it
555 // seems not to matter in practice.
556 double z_score = (rate - _rate.avg()) / sd;
557 if (z_score > threshold) {
558 return true;
559 }
560 }
561 return false;
562 }
563
564 double ShenandoahAllocationRate::instantaneous_rate(size_t allocated) const {
565 return instantaneous_rate(os::elapsedTime(), allocated);
566 }
567
568 double ShenandoahAllocationRate::instantaneous_rate(double time, size_t allocated) const {
569 size_t last_value = _last_sample_value;
570 double last_time = _last_sample_time;
571 size_t allocation_delta = (allocated > last_value) ? (allocated - last_value) : 0;
572 double time_delta_sec = time - last_time;
573 return (time_delta_sec > 0) ? (allocation_delta / time_delta_sec) : 0;
574 }