Register type definitions lazily #11938

4e6 · 2024-12-24T10:51:57Z

Pull Request Description

close #10923

Changelog:

update: registerTypeDefinitions registers constructors lazily

Important Notes

The improvement of time spent in the registerTypeDefinitions method (highlighted):

before

after

Checklist

Please ensure that the following checklist has been satisfied before submitting the PR:

The documentation has been updated, if necessary.
Screenshots/screencasts have been attached, if there are any visual changes. For interactive or animated visual changes, a screencast is preferred.
All code follows the
Scala,
Java,
TypeScript,
and
Rust
style guides. In case you are using a language not listed above, follow the Rust style guide.
Unit tests have been written where possible.
No regressions in benchmarks
Speedup in Startup benchmarks

JaroslavTulach

Few style-like comments inline. Treat as opinions.
Overall this seems like the right change.
what's the impact on benchmarks?
explicitly: what's the impact on test/Benchmarks/src/Startup/*.enso startups?

JaroslavTulach · 2024-12-27T04:10:38Z

engine/runtime/src/main/java/org/enso/interpreter/runtime/data/atom/AtomConstructor.java

+   * @param annotations the list of attached annotations
+   * @param args the list of argument definitions
+   */
+  public record InitializationParts(


Does this new record have to be public? Because the record defines not just public constructor, but also public getters. I do not think we want people around the code base to call the getters.

enso/engine/runtime/src/main/java/org/enso/interpreter/runtime/callable/function/FunctionSchema.java

Line 235 in 0c41e8d

public static final class Builder {

Could we rather follow this example by @Akirathan where Pavel introduced FunctionSchema.Builder?

JaroslavTulach · 2024-12-27T04:15:06Z

engine/runtime/src/main/java/org/enso/interpreter/runtime/util/CachingSupplier.java

+   * @return the supplier providing the value with the mapping function applied
+   * @param <R> the result type
+   */
+  public <R> CachingSupplier<R> map(Function<T, R> f) {


The "tendency towards functional style" is clear...

I, on the other hand, really like it :)

There are benchmark regressions and the only change that doesn't seem straightforward to me is this map method. But of course, I may be biased...

JaroslavTulach · 2024-12-27T04:29:49Z

engine/runtime/src/main/java/org/enso/interpreter/runtime/data/atom/AtomConstructor.java

    this.constructorFunction =
-        buildConstructorFunction(
-            language, section, localScope, scopeBuilder, assignments, varReads, annotations, args);
+        initializationResultSupplier.map(


Interesting use of the new map function. The alternative would be to have a single field Supplier<InitializationResult> stored in the AtomConstructor. I seem to like it a bit more - shows the "fixed" vs. the lazily computed part of the structure.

JaroslavTulach · 2024-12-27T04:35:58Z

engine/runtime/src/main/scala/org/enso/interpreter/runtime/IrToTruffle.scala

+          new RuntimeAnnotation(annotation.name, closureRootNode)
+        }
+
+        new InitializationParts(


With the builder style, this would become:

AtomConstructor.newBuilder(). section(...). scope(...). assignments(...). reads(...). args(defs);

Akirathan

LGTM. Agree with @JaroslavTulach on the commenteds about style. You don't have to schedule the benchmarks and compare them on this PR, but remember to please at least add a comment on this PR about the results.

…erhead

4e6 · 2024-12-28T21:11:03Z

Here is the benchmarks run, although I'm not sure how to compare it to the reference values https://github.com/enso-org/enso/actions/runs/12528446812

GitHub
Benchmark Engine · enso-org/enso@ca0fabc
Enso Analytics is a self-service data prep and analysis platform designed for data teams. - Benchmark Engine · ca0fabc

JaroslavTulach · 2024-12-30T06:26:15Z

although I'm not sure how to compare it to the reference values

@Akirathan has provided the information here.

bench_download -h for documentation and usage

I've just tried:

enso/tools/performance/engine-benchmarks$ ./bench_download.py -s engine -b wip/db/10923-register-type-definitions-overhead develop

there are many slowdowns. Looks like allocations of atoms are slower than then used to be.

There is another set of benchmarks including startup benchmarks. Run following CI action and then generate comparition with:

enso/tools/performance/engine-benchmarks$ ./bench_download.py -s stdlib -b wip/db/10923-register-type-definitions-overhead develop

At least the startup benchmarks should show some improvement.

JaroslavTulach

We need to eliminate the performance regressions before integrating.

JaroslavTulach · 2024-12-30T06:29:40Z

engine/runtime/src/main/java/org/enso/interpreter/runtime/data/atom/AtomConstructor.java

+  /** Builder required for initialization of the atom constructor. */
+  public static final class InitializationBuilder {
+
+    private final SourceSection section;


private fields are good.

JaroslavTulach · 2024-12-30T06:30:13Z

engine/runtime/src/main/java/org/enso/interpreter/runtime/data/atom/AtomConstructor.java

+      this.args = args;
+    }
+
+    private SourceSection getSection() {


Maybe the getters aren't really necessary and one could access directly the private fields.

JaroslavTulach · 2024-12-30T06:33:52Z

engine/runtime/src/main/java/org/enso/interpreter/runtime/util/CachingSupplier.java

+   * @return the supplier providing the value with the mapping function applied
+   * @param <R> the result type
+   */
+  public <R> CachingSupplier<R> map(Function<T, R> f) {


There are benchmark regressions and the only change that doesn't seem straightforward to me is this map method. But of course, I may be biased...

JaroslavTulach · 2024-12-30T07:44:41Z

engine/runtime/src/main/java/org/enso/interpreter/runtime/data/atom/AtomConstructor.java

@@ -278,8 +384,9 @@ public String toString() {
   *
   * @return the constructor function of this constructor.
   */
+  @TruffleBoundary


This is slowing execution down.

This method needs to be inlined, it cannot be @TruffleBoundary. The IGV graph has been generated by:

sbt:runtime-benchmarks> withDebug --dumpGraphs benchOnly -- AtomBenchmarks.benchGenerateList

JaroslavTulach · 2024-12-30T07:45:51Z

engine/runtime/src/main/java/org/enso/interpreter/runtime/data/atom/AtomConstructor.java

-   * @param varReads the expressions that read field values from local vars
+   * @param initializationBuilderSupplier the function supplying the parts required for
+   *     initialization
+   * @param fieldNames the argument names
   * @return {@code this}, for convenience
   */
  public AtomConstructor initializeFields(


Can initializeFields be private now, when there is the builder?

4e6 · 2024-12-30T10:39:10Z

On my machine, after removing truffle boundaries b94f81f it is:

[info] Benchmark                                    Mode  Cnt   Score   Error  Units
[info] AtomBenchmarks.benchGenerateList             avgt    5   4.502 ± 1.034  ms/op
[info] AtomBenchmarks.benchGenerateListAutoscoping  avgt    5  10.528 ± 0.146  ms/op
[info] AtomBenchmarks.benchGenerateListQualified    avgt    5   4.560 ± 0.379  ms/op

Before:

[info] Benchmark                                    Mode  Cnt   Score    Error  Units
[info] AtomBenchmarks.benchGenerateList             avgt    5  20.476 ±  1.154  ms/op
[info] AtomBenchmarks.benchGenerateListAutoscoping  avgt    5  32.763 ± 19.428  ms/op
[info] AtomBenchmarks.benchGenerateListQualified    avgt    5  19.932 ±  2.811  ms/op

engine/runtime/src/main/java/org/enso/interpreter/runtime/data/atom/AtomConstructor.java

JaroslavTulach

If there is no regression in engine benchmarks and there is some speedup in startup then feel free to integrate in current state.

4e6 · 2025-01-01T12:55:46Z

Startup benchmark results are about the same:

`wip/db/10923-register-type-definitions-overhead`

$ ./built-distribution/enso-engine*/enso*/bin/enso --run test/Benchmarks/src/Startup/Startup.enso
Benchmarking 'Startup.empty_startup' with configuration: [warmup={2 iterations, 5 seconds each}, measurement={3 iterations, 5 seconds each}]
Measurement avg time:    2136.23 ms (+-191.653)
Benchmarking 'Startup.hello_world_startup' with configuration: [warmup={2 iterations, 5 seconds each}, measurement={3 iterations, 5 seconds each}]
Measurement avg time:    3559.529 ms (+-47.995)
Benchmarking 'Startup.import_world_startup' with configuration: [warmup={2 iterations, 5 seconds each}, measurement={3 iterations, 5 seconds each}]
Measurement avg time:    4100.263 ms (+-48.712)

`develop`

$ ./built-distribution/enso-engine*/enso*/bin/enso --run test/Benchmarks/src/Startup/Startup.enso
Benchmarking 'Startup.empty_startup' with configuration: [warmup={2 iterations, 5 seconds each}, measurement={3 iterations, 5 seconds each}]
Measurement avg time:    2150.757 ms (+-168.048)
Benchmarking 'Startup.hello_world_startup' with configuration: [warmup={2 iterations, 5 seconds each}, measurement={3 iterations, 5 seconds each}]
Measurement avg time:    3621.465 ms (+-54.777)
Benchmarking 'Startup.import_world_startup' with configuration: [warmup={2 iterations, 5 seconds each}, measurement={3 iterations, 5 seconds each}]
Measurement avg time:    4185.714 ms (+-49.514)

4e6 added 5 commits December 24, 2024 10:38

feat: atom constructor supplier

948e330

feat: initialization parts supplier

cf95eb8

feat: register method supplier

ab7bb45

fix: for constructor size 1

34cee96

misc: cleanup

0663c2a

4e6 added the CI: No changelog needed Do not require a changelog entry for this PR. label Dec 24, 2024

4e6 self-assigned this Dec 24, 2024

4e6 requested review from JaroslavTulach, hubertp and Akirathan as code owners December 24, 2024 10:51

fix: native image

d15c59f

JaroslavTulach approved these changes Dec 27, 2024

View reviewed changes

Akirathan approved these changes Dec 27, 2024

View reviewed changes

hubertp approved these changes Dec 27, 2024

View reviewed changes

enso-bot bot mentioned this pull request Dec 28, 2024

Text.find slows down with the number of invocations #11859

Closed

4e6 added 3 commits December 28, 2024 17:21

Merge branch 'develop' into wip/db/10923-register-type-definitions-ov…

ca0fabc

…erhead

feat: atom constructor initialization builder

06d3fb5

misc: fmt

a2c78e7

JaroslavTulach self-requested a review December 30, 2024 06:26

JaroslavTulach requested changes Dec 30, 2024

View reviewed changes

JaroslavTulach reviewed Dec 30, 2024

View reviewed changes

update: remove truffle boundaries

b94f81f

4e6 commented Dec 30, 2024

View reviewed changes

engine/runtime/src/main/java/org/enso/interpreter/runtime/data/atom/AtomConstructor.java Show resolved Hide resolved

JaroslavTulach approved these changes Dec 31, 2024

View reviewed changes

4e6 added the CI: Ready to merge This PR is eligible for automatic merge label Jan 1, 2025

mergify bot merged commit bc6a8aa into develop Jan 1, 2025
49 checks passed

mergify bot deleted the wip/db/10923-register-type-definitions-overhead branch January 1, 2025 12:57

JaroslavTulach mentioned this pull request Jan 4, 2025

Remove 519ms startup overhead caused by registerTypeDefinitions #10923

Closed

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Register type definitions lazily #11938

Register type definitions lazily #11938

4e6 commented Dec 24, 2024 •

edited by JaroslavTulach

Loading

JaroslavTulach left a comment

JaroslavTulach Dec 27, 2024

JaroslavTulach Dec 27, 2024

hubertp Dec 27, 2024

JaroslavTulach Dec 30, 2024

JaroslavTulach Dec 27, 2024

JaroslavTulach Dec 27, 2024

Akirathan left a comment

4e6 commented Dec 28, 2024 •

edited by unfurl-links bot

Loading

JaroslavTulach commented Dec 30, 2024

JaroslavTulach left a comment

JaroslavTulach Dec 30, 2024

JaroslavTulach Dec 30, 2024

JaroslavTulach Dec 30, 2024

JaroslavTulach Dec 30, 2024 •

edited

Loading

JaroslavTulach Dec 30, 2024

4e6 commented Dec 30, 2024

JaroslavTulach left a comment

4e6 commented Jan 1, 2025

Register type definitions lazily #11938

Register type definitions lazily #11938

Conversation

4e6 commented Dec 24, 2024 • edited by JaroslavTulach Loading

Pull Request Description

Important Notes

before

after

Checklist

JaroslavTulach left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Akirathan left a comment

Choose a reason for hiding this comment

4e6 commented Dec 28, 2024 • edited by unfurl-links bot Loading

JaroslavTulach commented Dec 30, 2024

JaroslavTulach left a comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Choose a reason for hiding this comment

JaroslavTulach Dec 30, 2024 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

4e6 commented Dec 30, 2024

JaroslavTulach left a comment

Choose a reason for hiding this comment

4e6 commented Jan 1, 2025

wip/db/10923-register-type-definitions-overhead

develop

4e6 commented Dec 24, 2024 •

edited by JaroslavTulach

Loading

4e6 commented Dec 28, 2024 •

edited by unfurl-links bot

Loading

JaroslavTulach Dec 30, 2024 •

edited

Loading

`wip/db/10923-register-type-definitions-overhead`

`develop`