8312116: GenShen: make instantaneous allocation rate triggers more timely by kdnilsen · Pull Request #29039 · openjdk/jdk

kdnilsen · 2026-01-05T15:10:52Z

After studying large numbers of GC logs with degenerated cycles that have resulted from "late" triggers, we propose the following general improvements:

Track trends in GC times rather than always using the average GC time plus standard deviation. In many situations, GC times trend upward due to, for example, increasing amounts of live data that must be marked as a workload builds up its working set of memory.
Sample allocation rates more frequently than once every 100 ms.
Track trends in allocation rates. In some situations, the allocation rate trends upwards due to, for example, the start of a new phase of execution or a spike in client workload.
When we detect acceleration of allocation rate, predict consumption of memory based on accelerated allocation rates rather than assuming constant allocation rate.

Progress

Change must be properly reviewed (1 review required, with at least 1 Reviewer)
Change must not contain extraneous whitespace
Commit message must refer to an issue

Issue

JDK-8312116: GenShen: make instantaneous allocation rate triggers more timely (Sub-task - P3)

Reviewers

William Kemper (@earthling-amzn - Reviewer)

Reviewing

Using git

Checkout this PR locally:
$ git fetch https://git.openjdk.org/jdk.git pull/29039/head:pull/29039
$ git checkout pull/29039

Update a local copy of the PR:
$ git checkout pull/29039
$ git pull https://git.openjdk.org/jdk.git pull/29039/head

Using Skara CLI tools

Checkout this PR locally:
$ git pr checkout 29039

View PR using the GUI difftool:
$ git pr show -t 29039

Using diff file

Download this PR as a diff file:
https://git.openjdk.org/jdk/pull/29039.diff

Using Webrev

Link to Webrev Comment

When allocation rate appears to be accelerating, predict consumption of memory according to an accelerated rate of consumption. This expedites triggers under critical phase changes.

Set both acceleration sample size and momentary spike sample size to 10. Remove the restriction that momentary spike sample size must be strictly less than acceleration sample size. These changes were motivated by experiments with Extremem workloads. More experiments are in progress and further changes may be implemented based on those resuts.

…-triggers

Also refine future predicted gc time to account for possible delay before we start the GC cycle.

Also tidy up the descriptions of new sample size parameters.

As originally implemented, we apply penalties to the triggering heuristic every time we experience a degenerated cycle. This has the effect of forcing GC triggers to spiral out of control. This commit changes the penalty mechanism. When a degen happens through no fault of the heuristic triggering mechanism, we do not pile on additional penalties. Specifically, we consider that heuristic triggering is not responsible for a degenerated cycle that is associated with a GC that began immediately following the end of the previous GC cycle.

Added tag jdk-25+10 for changeset a637ccf

… <= _highest_valid_narrow_klass_id) failed: narrowKlass ID out of range (3131947710) Reviewed-by: shade

(and less expensive monitoring of triggering conditions)

…erated-triggers-gh

kdnilsen · 2026-02-17T22:55:52Z

+    }
+    // else, leave current_rate = y_max, acceleration = 0
+  }
+  // and here also, leave current_rate = y_max, acceleration = 0


y_max is no more. fix these two comments.

kdnilsen · 2026-02-17T23:11:09Z

                            range,                                          \
                            constraint)                                     \
                                                                            \
+  product(double, ShenandoahAccelerationSamplePeriod, 0.0145, EXPERIMENTAL, \


Let's change this option to ms rather than seconds for consistency with existing parameters.

kdnilsen · 2026-02-17T23:33:41Z

-  size_t spike_headroom = capacity / 100 * ShenandoahAllocSpikeFactor;
-  size_t penalties      = capacity / 100 * _gc_time_penalties;
+  avg_cycle_time = _gc_cycle_time_history->davg() + (_margin_of_error_sd * _gc_cycle_time_history->dsd());
+  avg_alloc_rate = _allocation_rate.upper_bound(_margin_of_error_sd);


Before we test any trigger conditions, we should consider whether a certain minimum amount of memory has been allocated. Move the test from accelerated-triggers below to apply to all triggers.

Changed this code.

kdnilsen · 2026-02-17T23:37:55Z

+}

-ShenandoahAdaptiveHeuristics::~ShenandoahAdaptiveHeuristics() {}
+void ShenandoahAdaptiveHeuristics::compute_headroom_adjustment(size_t mutator_available) {


No need for mutator_available as an argument and no need to compute byte_allocated_at_start_of_idle.

Comment: if someone changes soft_max_capacity(), this should be called to recompute.

I've simplified implementation of compute_headroom_adjustment() and removed unnecessary argument.
I've added a call to compute_headroom_adjustment() when soft_max_capacity is changed.

kdnilsen · 2026-02-17T23:39:36Z

+  // before we need to start the next GC.
+  void start_idle_span() override;
+
+  // If old-generation marking finishes during an idle span and immediate old-generation garbage is identified, we will rebuild


Maybe this is redundant with start_idle_span or not even necessary (since start_idle_span doesn't need bytes_available.

I've removed resume_idle_span().

kdnilsen · 2026-02-17T23:42:23Z

  // source of feedback to adjust trigger parameters.
  TruncatedSeq _available;

+  ShenandoahFreeSet* _free_set;


Can we use TruncatedSeq::predict_next() for this linear prediction?

Also, can we get rid of _regulator_thread, _control_thread, is_generational?

I've removed _regulator_thread, _control_thread, _is_generational from ShenanoahAdaptiveHeuristics.

TruncatedSeq::predict_next() assumes all data samples are equidistant and does not allow a parameter to predict the value at a specific future time, so it does not provide the same functionality as the abstraction introduced in this PR.

…aptive triggers to wait for some garbage to accumulate

…ft_capacity is managed

…andoahAdaptiveHeuristics

kdnilsen · 2026-03-03T01:19:48Z

/integrate

openjdk · 2026-03-03T09:39:49Z

Going to push as commit 0b183bf.
Since your change was applied there have been 192 commits pushed to the master branch:

c0c8bdd: 8378948: Remove unused local variable in RunnerGSInserterThread
7e9e649: 8378083: Mark shenandoah/generational/TestOldGrowthTriggers.java as flagless
f4da2d5: 8378684: Fix -Wdeprecated-declarations warnings from gtest by clang23
... and 189 more: https://git.openjdk.org/jdk/compare/c8338be9ad455445a94972d2d9e483a24adc27cf...master

Your commit was automatically rebased without conflicts.

openjdk · 2026-03-03T09:40:57Z

@kdnilsen Pushed as commit 0b183bf.

💡 You may see a message that your pull request was closed with unmerged commits. This can be safely ignored.

mrserb · 2026-03-03T18:16:01Z

                              _partitions.leftmost(ShenandoahFreeSetPartitionId::OldCollector),
                              _partitions.rightmost(ShenandoahFreeSetPartitionId::OldCollector));
          old_region_count++;
+          assert(ac = ShenandoahHeapRegion::region_size_bytes(), "Cannot move to old unless entire region is in alloc capacity");


Is this assignment expected here?

Thanks for the catch. If the asserted condition is true, this is harmless. But certainly not what was intended.
I will fix in a follow-on patch.

mrserb · 2026-03-04T02:15:28Z

+      if (i > 0) {
+        // first sample not included in weighted average because it has no weight.
+        double sample_weight = x_array[i] - x_array[i-1];
+        weighted_y_sum = y_array[i] * sample_weight;


Should this be changed to
weighted_y_sum += y_array[i] * sample_weight;
?

Will also correct this in a follow-on patch. Thanks.

kdnilsen and others added 30 commits December 6, 2024 23:35

Track mutator allocations

0070f74

Add methods to reveal RegulatorThread wake time and period

7c9097e

respond to acceleration of allocation rate with quicker trigger

908e771

Add accelerated triggers to Shenandoah

6ede1cb

When allocation rate appears to be accelerating, predict consumption of memory according to an accelerated rate of consumption. This expedites triggers under critical phase changes.

Fix bugs and make tuning adjustments

edfcddd

Change default to be more sensitive to acceleration, less to spikes

7801380

Fix bugs and change defaults

b6f3b85

Improve debug messages

fbafa11

Fix whitespace

4015eb0

Merge branch 'master' of https://git.openjdk.org/jdk into accelerated…

6b5f17e

…-triggers

Remove deprecated line of code

6ea2439

Enhance log message and remove dead code

8701275

Also refine future predicted gc time to account for possible delay before we start the GC cycle.

Change default for momentary spike sample size

929a533

Also tidy up the descriptions of new sample size parameters.

Fix idle span invocations and add some debug instrumentation

14d82ac

Fiddle with debug instrumentation

f072099

Merge tag 'jdk-25+10' into accelerated-triggers

1a99c04

Added tag jdk-25+10 for changeset a637ccf

8348092: Shenandoah: assert(nk >= _lowest_valid_narrow_klass_id && nk…

10c5dfc

… <= _highest_valid_narrow_klass_id) failed: narrowKlass ID out of range (3131947710) Reviewed-by: shade

Merge remote-tracking branch 'jdk/master' into accelerated-triggers

7f71fa0

Improve comments and SIZE_FORMAT encodings

b6a2fed

Change defaults to make trigger less aggressive

b245cf7

Change defaults for less aggressive triggering

4fe0c2d

(and less expensive monitoring of triggering conditions)

make triggers even more conservative

9e05aef

Merge remote-tracking branch 'jdk/master' into accelerated-triggers

d1a0949

Merge remote-tracking branch 'origin/accelerated-triggers' into accel…

ca77c49

…erated-triggers-gh

Fixup conflicts introduced by upstream merge

23cc728

Merge remote-tracking branch 'origin/accelerated-triggers' into accel…

95e7105

…erated-triggers-gh

sample allocation rate half as frequently

60afa5d

Merge remote-tracking branch 'origin/accelerated-triggers' into accel…

13503e0

…erated-triggers-gh

kdnilsen added 2 commits February 16, 2026 23:27

Remove dead code and unused variables and rename one function

d85193f

Improve comments

6bd4c9e

kdnilsen commented Feb 17, 2026

View reviewed changes

kdnilsen added 6 commits February 18, 2026 19:23

Improve formatting and comments

9c93198

Represent ShenandoahAccelerationSamplePeriod in ms and require all ad…

88f214b

…aptive triggers to wait for some garbage to accumulate

ShenandoahAccelertaionSamplePeriod is measured in ms

7b04174

Remove arg to compute_headroom_adjustment() and update headroom if so…

2f1c593

…ft_capacity is managed

Remove resume_idle_span()

a4890a2

remove _is_generational, _regulator_thread, _control_thread from Shen…

0b70897

…andoahAdaptiveHeuristics

openjdk Bot added the merge-conflict Pull request has merge conflict with target branch label Feb 18, 2026

Merge remote-tracking branch 'jdk/master' into accelerated-triggers

a20e69f

openjdk Bot removed the merge-conflict Pull request has merge conflict with target branch label Feb 18, 2026

Allow generic triggers even when we have not allocated minimal threshold

ea9141c

earthling-amzn approved these changes Mar 2, 2026

View reviewed changes

openjdk Bot added the ready Pull request is ready to be integrated label Mar 3, 2026

openjdk Bot added the integrated Pull request has been integrated label Mar 3, 2026

openjdk Bot closed this Mar 3, 2026

openjdk Bot removed ready Pull request is ready to be integrated rfr Pull request is ready for review labels Mar 3, 2026

mrserb reviewed Mar 3, 2026

View reviewed changes

mrserb reviewed Mar 4, 2026

View reviewed changes

earthling-amzn mentioned this pull request May 5, 2026

8383892: Shenandoah: Decouple allocation rate sampling from GC cycle #31047

Open

5 tasks

Conversation

kdnilsen commented Jan 5, 2026 • edited by openjdk Bot Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Progress

Issue

Reviewers

Reviewing

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

kdnilsen commented Mar 3, 2026

Uh oh!

openjdk Bot commented Mar 3, 2026

Uh oh!

openjdk Bot commented Mar 3, 2026

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Reviewers

Assignees

Labels

Milestone

Development

Uh oh!

3 participants

kdnilsen commented Jan 5, 2026 •

edited by openjdk Bot

Loading