Change stack size growth strategy to 0.5 rate #8869

lu-pinto · 2025-06-27T10:34:53Z

This is a cherry-pick from the performance branch that we used to test the changes in the latest interop.

original PR: #8814

Fixed Issue(s)

Thanks for sending a pull request! Have you done the following?

Checked out our contribution guidelines?
Considered documentation and added the doc-change-required label to this PR if updates are required.
Considered the changelog and included an update if required.
For database changes (e.g. KeyValueSegmentIdentifier) considered compatibility and performed forwards and backwards compatibility tests

Locally, you can run these tests to catch failures early:

spotless: ./gradlew spotlessApply
unit tests: ./gradlew build
acceptance tests: ./gradlew acceptanceTest
integration tests: ./gradlew integrationTest
reference tests: ./gradlew ethereum:referenceTests:referenceTests

macfarla · 2025-06-30T00:15:03Z

evm/src/main/java/org/hyperledger/besu/evm/internal/FlexStack.java

  }

+  private static int newLength(final int oldCapacity, final int minGrowth, final int prefGrowth) {
+    return oldCapacity + Math.max(minGrowth, prefGrowth);


is there a danger of int overflow

Good catch @macfarla!
There was also another existing problem when someone sets maxSize == Integer.MAX_VALUE which is not a realistic value since the JVM is not able to allocate that many entries. You get something like this:

Exception in thread "main" java.lang.OutOfMemoryError: Requested array size exceeds VM limit at java.base/java.lang.reflect.Array.newArray(Native Method) at java.base/java.lang.reflect.Array.newInstance(Array.java:78) at FlexStack.expandEntries(FlexStack.java:124) at FlexStack.push(FlexStack.java:143) at Test.main(Test.java:6)

Fixed now. Couldn't write tests because we would have to increase the heap size to 16G for such a case so I checked it separately offline.

macfarla · 2025-06-30T23:45:28Z

evm/src/main/java/org/hyperledger/besu/evm/internal/FlexStack.java

    top = nextTop;
  }

+  private int newLength(final int oldCapacity, final int minGrowth, final int prefGrowth) {


Suggested change

private int newLength(final int oldCapacity, final int minGrowth, final int prefGrowth) {

private int newLength(final int currentCapacity, final int minGrowth, final int prefGrowth) {

to make consistent with other changes

evm/src/main/java/org/hyperledger/besu/evm/internal/FlexStack.java

lu-pinto · 2025-07-03T17:37:31Z

After speaking with @ahamlat today, it's best to assess performance on mainnet first and do some benchmarking for CALL* before pulling this in

evm/src/main/java/org/hyperledger/besu/evm/internal/FlexStack.java

macfarla · 2025-07-22T23:17:45Z

moving to draft since we need to do performance testing on this before merging

lu-pinto · 2025-08-06T16:55:38Z

I think the added benchmark might unblock this PR, waiting on feedback.
I got the following results running on a r6a.2xlarge:

Without this PR:

Benchmark                     (stackDepth)  Mode  Cnt     Score    Error  Units
OperandStackBenchmark.fillUp             6  avgt   15    12.804 ±  0.021  ns/op
OperandStackBenchmark.fillUp            15  avgt   15    26.348 ±  0.028  ns/op
OperandStackBenchmark.fillUp            34  avgt   15   145.041 ±  7.118  ns/op
OperandStackBenchmark.fillUp           100  avgt   15   335.065 ±  0.573  ns/op
OperandStackBenchmark.fillUp           234  avgt   15   843.487 ± 16.104  ns/op
OperandStackBenchmark.fillUp           500  avgt   15  2102.937 ±  9.191  ns/op
OperandStackBenchmark.fillUp           800  avgt   15  4502.884 ± 17.090  ns/op
OperandStackBenchmark.fillUp          1024  avgt   15  6943.397 ± 18.922  ns/op

With this PR:

Benchmark                     (stackDepth)  Mode  Cnt     Score    Error  Units
OperandStackBenchmark.fillUp             6  avgt   15    22.623 ±  0.913  ns/op
OperandStackBenchmark.fillUp            15  avgt   15    30.296 ±  0.149  ns/op
OperandStackBenchmark.fillUp            34  avgt   15    58.666 ±  0.090  ns/op
OperandStackBenchmark.fillUp           100  avgt   15   277.010 ±  0.520  ns/op
OperandStackBenchmark.fillUp           234  avgt   15   700.040 ±  1.873  ns/op
OperandStackBenchmark.fillUp           500  avgt   15  1610.791 ±  1.557  ns/op
OperandStackBenchmark.fillUp           800  avgt   15  2632.757 ± 33.295  ns/op
OperandStackBenchmark.fillUp          1024  avgt   15  3150.211 ± 13.124  ns/op

It's noticeable the gains at a higher stack height as expected (x2). On the lower end, up until 32, since there's no resize, there's a slight regression. However, I would like to note that the average used stack height stands at around 75 as we can see below:

This is the maximum stack size expansion in each CALL - green line from a node closer to genesis and yellow line from a node closer to HEAD.

Signed-off-by: Luis Pinto <luis.pinto@consensys.net>

Co-authored-by: Sally MacFarlane <macfarla.github@gmail.com> Signed-off-by: Luis Pinto <luis.pinto@consensys.net>

Signed-off-by: Luis Pinto <luis.pinto@consensys.net>

siladu

LGTM.

Regression tested on mainnet and no regressions found.
Also no obvious difference in performance detected.

siladu · 2025-08-12T12:42:27Z

@lu-pinto maybe warrants a changelog entry

Signed-off-by: Luis Pinto <luis.pinto@consensys.net>

This PR changes the stack rate growth strategy of FlexStack to grow at a 50% rate instead of constantly growing by 32 slots as is currently. The former is the approach adopted by the JDK for ArrayList. With a 50% allocation rate and considering a max stack size of 1024, setting the initial size to 91 one can reach the worse case (1024) in just 6 resizes compared to 32 required with the current approach. Also the current usage on mainnet has shown that CALLs never go above 150 deep during execution but this can change in the future. Without this PR: Benchmark (stackDepth) Mode Cnt Score Error Units OperandStackBenchmark.fillUp 6 avgt 15 12.804 ± 0.021 ns/op OperandStackBenchmark.fillUp 15 avgt 15 26.348 ± 0.028 ns/op OperandStackBenchmark.fillUp 34 avgt 15 145.041 ± 7.118 ns/op OperandStackBenchmark.fillUp 100 avgt 15 335.065 ± 0.573 ns/op OperandStackBenchmark.fillUp 234 avgt 15 843.487 ± 16.104 ns/op OperandStackBenchmark.fillUp 500 avgt 15 2102.937 ± 9.191 ns/op OperandStackBenchmark.fillUp 800 avgt 15 4502.884 ± 17.090 ns/op OperandStackBenchmark.fillUp 1024 avgt 15 6943.397 ± 18.922 ns/op With this PR: Benchmark (stackDepth) Mode Cnt Score Error Units OperandStackBenchmark.fillUp 6 avgt 15 22.623 ± 0.913 ns/op OperandStackBenchmark.fillUp 15 avgt 15 30.296 ± 0.149 ns/op OperandStackBenchmark.fillUp 34 avgt 15 58.666 ± 0.090 ns/op OperandStackBenchmark.fillUp 100 avgt 15 277.010 ± 0.520 ns/op OperandStackBenchmark.fillUp 234 avgt 15 700.040 ± 1.873 ns/op OperandStackBenchmark.fillUp 500 avgt 15 1610.791 ± 1.557 ns/op OperandStackBenchmark.fillUp 800 avgt 15 2632.757 ± 33.295 ns/op OperandStackBenchmark.fillUp 1024 avgt 15 3150.211 ± 13.124 ns/op

This PR changes the stack rate growth strategy of FlexStack to grow at a 50% rate instead of constantly growing by 32 slots as is currently. The former is the approach adopted by the JDK for ArrayList. With a 50% allocation rate and considering a max stack size of 1024, setting the initial size to 91 one can reach the worse case (1024) in just 6 resizes compared to 32 required with the current approach. Also the current usage on mainnet has shown that CALLs never go above 150 deep during execution but this can change in the future. Without this PR: Benchmark (stackDepth) Mode Cnt Score Error Units OperandStackBenchmark.fillUp 6 avgt 15 12.804 ± 0.021 ns/op OperandStackBenchmark.fillUp 15 avgt 15 26.348 ± 0.028 ns/op OperandStackBenchmark.fillUp 34 avgt 15 145.041 ± 7.118 ns/op OperandStackBenchmark.fillUp 100 avgt 15 335.065 ± 0.573 ns/op OperandStackBenchmark.fillUp 234 avgt 15 843.487 ± 16.104 ns/op OperandStackBenchmark.fillUp 500 avgt 15 2102.937 ± 9.191 ns/op OperandStackBenchmark.fillUp 800 avgt 15 4502.884 ± 17.090 ns/op OperandStackBenchmark.fillUp 1024 avgt 15 6943.397 ± 18.922 ns/op With this PR: Benchmark (stackDepth) Mode Cnt Score Error Units OperandStackBenchmark.fillUp 6 avgt 15 22.623 ± 0.913 ns/op OperandStackBenchmark.fillUp 15 avgt 15 30.296 ± 0.149 ns/op OperandStackBenchmark.fillUp 34 avgt 15 58.666 ± 0.090 ns/op OperandStackBenchmark.fillUp 100 avgt 15 277.010 ± 0.520 ns/op OperandStackBenchmark.fillUp 234 avgt 15 700.040 ± 1.873 ns/op OperandStackBenchmark.fillUp 500 avgt 15 1610.791 ± 1.557 ns/op OperandStackBenchmark.fillUp 800 avgt 15 2632.757 ± 33.295 ns/op OperandStackBenchmark.fillUp 1024 avgt 15 3150.211 ± 13.124 ns/op Signed-off-by: jflo <justin+github@florentine.us>

lu-pinto force-pushed the operand-stack-growth-strategy branch from 86da553 to da60750 Compare June 29, 2025 08:01

macfarla reviewed Jun 30, 2025

View reviewed changes

lu-pinto force-pushed the operand-stack-growth-strategy branch from 4ffbfe5 to 8104be6 Compare July 3, 2025 13:36

macfarla reviewed Jul 3, 2025

View reviewed changes

evm/src/main/java/org/hyperledger/besu/evm/internal/FlexStack.java Outdated Show resolved Hide resolved

lu-pinto force-pushed the operand-stack-growth-strategy branch 2 times, most recently from 396ed88 to 92ed95e Compare July 8, 2025 18:23

macfarla marked this pull request as draft July 22, 2025 23:17

macfarla added the performance label Jul 31, 2025

macfarla assigned lu-pinto Aug 5, 2025

lu-pinto force-pushed the operand-stack-growth-strategy branch 3 times, most recently from fbd2f92 to 3ecd27a Compare August 6, 2025 16:45

lu-pinto and others added 7 commits August 7, 2025 12:07

Change stack size growth strategy to 0.5 rate

fbb2306

Signed-off-by: Luis Pinto <luis.pinto@consensys.net>

fix potential overflow for large enough max stack size

75fa923

Signed-off-by: Luis Pinto <luis.pinto@consensys.net>

fix growing rate on stack

05c37de

Signed-off-by: Luis Pinto <luis.pinto@consensys.net>

Rename SOFT_MAX_ARRAY_LENGTH to MAX_ARRAY_LENGTH

f6d04e0

Co-authored-by: Sally MacFarlane <macfarla.github@gmail.com> Signed-off-by: Luis Pinto <luis.pinto@consensys.net>

add operand stack benchmark and optimize flex stack init

88bb522

Signed-off-by: Luis Pinto <luis.pinto@consensys.net>

lower error rate in benchmark

1eeb943

Signed-off-by: Luis Pinto <luis.pinto@consensys.net>

spotless

ac0e781

Signed-off-by: Luis Pinto <luis.pinto@consensys.net>

lu-pinto force-pushed the operand-stack-growth-strategy branch from 3ecd27a to ac0e781 Compare August 7, 2025 11:07

lu-pinto marked this pull request as ready for review August 7, 2025 11:07

Merge branch 'main' into operand-stack-growth-strategy

c84e6c5

siladu approved these changes Aug 12, 2025

View reviewed changes

lu-pinto added 2 commits August 18, 2025 11:45

Merge branch 'main' into operand-stack-growth-strategy

dacebe5

add Changelog entry

c8b6730

Signed-off-by: Luis Pinto <luis.pinto@consensys.net>

lu-pinto enabled auto-merge (squash) August 18, 2025 11:10

lu-pinto merged commit f21630e into hyperledger:main Aug 18, 2025
46 checks passed

macfarla mentioned this pull request Aug 25, 2025

RC3 changelog - 8869 included 25.8.0 #9106

Merged

9 tasks

	private int newLength(final int oldCapacity, final int minGrowth, final int prefGrowth) {
	private int newLength(final int currentCapacity, final int minGrowth, final int prefGrowth) {

Change stack size growth strategy to 0.5 rate #8869

Change stack size growth strategy to 0.5 rate #8869

Uh oh!

Conversation

lu-pinto commented Jun 27, 2025

Fixed Issue(s)

Thanks for sending a pull request! Have you done the following?

Locally, you can run these tests to catch failures early:

Uh oh!

macfarla Jun 30, 2025

Choose a reason for hiding this comment

Uh oh!

lu-pinto Jun 30, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

lu-pinto Jun 30, 2025

Choose a reason for hiding this comment

Uh oh!

macfarla Jun 30, 2025

Choose a reason for hiding this comment

Uh oh!

macfarla Jun 30, 2025

Choose a reason for hiding this comment

Uh oh!

Uh oh!

Uh oh!

lu-pinto commented Jul 3, 2025 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Uh oh!

macfarla commented Jul 22, 2025

Uh oh!

lu-pinto commented Aug 6, 2025

Uh oh!

siladu left a comment

Choose a reason for hiding this comment

Uh oh!

siladu commented Aug 12, 2025

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

lu-pinto Jun 30, 2025 •

edited

Loading

lu-pinto commented Jul 3, 2025 •

edited

Loading