Cross platform weak bag implementation by vasilmkd · Pull Request #2673 · typelevel/cats-effect

vasilmkd · 2021-12-20T12:08:44Z

Completely replaces the WeakHashMap mechanism, on each WorkerThread, the fallback and JS.

A possible remedy for #2634.

Co-authored-by: Arman Bilge <[email protected]>

armanbilge · 2021-12-20T12:21:12Z

build.sbt

+    ("org.scala-js" %%% "scalajs-weakreferences" % JsWeakReferencesVersion)
+      .cross(CrossVersion.for3Use2_13)


Some interesting reading about this scala-js/scala-js-weakreferences#7 (comment)

vasilmkd · 2021-12-20T14:08:26Z

"io.vasilev" %% "cats-effect" % "3.3-87-1d1f0ff"

vasilmkd · 2021-12-20T15:50:07Z

Benchmark results:

Baseline on series/3.3.x with tracing off (this case is unaffected by this PR):

benchmarks/Jmh/run -wi 10 -i 10 -f 2 -t 1 -prof gc --jvmArgs -Dcats.effect.tracing.mode=none --jvmArgs -Dcats.effect.tracing.exceptions.enhanced=false ParallelBenchmark.par
Benchmark                                                       (cpuTokens)  (size)   Mode  Cnt        Score       Error   Units
ParallelBenchmark.parTraverse                                         10000    1000  thrpt   20      316.579 ±     3.402   ops/s
ParallelBenchmark.parTraverse:·gc.alloc.rate                          10000    1000  thrpt   20     1417.315 ±    15.392  MB/sec
ParallelBenchmark.parTraverse:·gc.alloc.rate.norm                     10000    1000  thrpt   20  4929586.675 ±   579.390    B/op
ParallelBenchmark.parTraverse:·gc.churn.G1_Eden_Space                 10000    1000  thrpt   20     1429.037 ±    22.488  MB/sec
ParallelBenchmark.parTraverse:·gc.churn.G1_Eden_Space.norm            10000    1000  thrpt   20  4970221.334 ± 47058.032    B/op
ParallelBenchmark.parTraverse:·gc.churn.G1_Survivor_Space             10000    1000  thrpt   20        0.222 ±     0.022  MB/sec
ParallelBenchmark.parTraverse:·gc.churn.G1_Survivor_Space.norm        10000    1000  thrpt   20      771.322 ±    75.550    B/op
ParallelBenchmark.parTraverse:·gc.count                               10000    1000  thrpt   20      919.000              counts
ParallelBenchmark.parTraverse:·gc.time                                10000    1000  thrpt   20     1325.000                  ms

Baseline on series/3.3.x with tracing on:

benchmarks/Jmh/run -wi 10 -i 10 -f 2 -t 1 -prof gc --jvmArgs -Dcats.effect.tracing.mode=cached --jvmArgs -Dcats.effect.tracing.exceptions.enhanced=true ParallelBenchmark.par
Benchmark                                                       (cpuTokens)  (size)   Mode  Cnt        Score        Error   Units
ParallelBenchmark.parTraverse                                         10000    1000  thrpt   20      204.719 ±     19.250   ops/s
ParallelBenchmark.parTraverse:·gc.alloc.rate                          10000    1000  thrpt   20     1044.091 ±     97.878  MB/sec
ParallelBenchmark.parTraverse:·gc.alloc.rate.norm                     10000    1000  thrpt   20  5616091.869 ±   2648.642    B/op
ParallelBenchmark.parTraverse:·gc.churn.G1_Eden_Space                 10000    1000  thrpt   20     1052.138 ±    110.314  MB/sec
ParallelBenchmark.parTraverse:·gc.churn.G1_Eden_Space.norm            10000    1000  thrpt   20  5665606.945 ± 358765.601    B/op
ParallelBenchmark.parTraverse:·gc.churn.G1_Survivor_Space             10000    1000  thrpt   20        0.814 ±      0.502  MB/sec
ParallelBenchmark.parTraverse:·gc.churn.G1_Survivor_Space.norm        10000    1000  thrpt   20     4325.508 ±   2692.273    B/op
ParallelBenchmark.parTraverse:·gc.count                               10000    1000  thrpt   20       92.000               counts
ParallelBenchmark.parTraverse:·gc.time                                10000    1000  thrpt   20    17783.000                   ms

This PR with tracing on:

benchmarks/Jmh/run -wi 10 -i 10 -f 2 -t 1 -prof gc --jvmArgs -Dcats.effect.tracing.mode=cached --jvmArgs -Dcats.effect.tracing.exceptions.enhanced=true ParallelBenchmark.par
Benchmark                                                       (cpuTokens)  (size)   Mode  Cnt        Score       Error   Units
ParallelBenchmark.parTraverse                                         10000    1000  thrpt   20      278.251 ±     1.785   ops/s
ParallelBenchmark.parTraverse:·gc.alloc.rate                          10000    1000  thrpt   20     1391.663 ±     9.091  MB/sec
ParallelBenchmark.parTraverse:·gc.alloc.rate.norm                     10000    1000  thrpt   20  5507256.392 ±   575.319    B/op
ParallelBenchmark.parTraverse:·gc.churn.G1_Eden_Space                 10000    1000  thrpt   20     1407.085 ±    15.138  MB/sec
ParallelBenchmark.parTraverse:·gc.churn.G1_Eden_Space.norm            10000    1000  thrpt   20  5568406.214 ± 57756.867    B/op
ParallelBenchmark.parTraverse:·gc.churn.G1_Survivor_Space             10000    1000  thrpt   20        0.246 ±     0.020  MB/sec
ParallelBenchmark.parTraverse:·gc.churn.G1_Survivor_Space.norm        10000    1000  thrpt   20      972.881 ±    79.171    B/op
ParallelBenchmark.parTraverse:·gc.count                               10000    1000  thrpt   20      812.000              counts
ParallelBenchmark.parTraverse:·gc.time                                10000    1000  thrpt   20     1386.000                  ms

This PR improves performance significantly and reduces GC time to the levels of the case without tracing.

vasilmkd · 2021-12-20T17:00:46Z

Under the advice of @armanbilge, this PR shades https://github.com/scala-js/scala-js-weakreferences by copying the source code (~100 LOC, not including license headers). This is due to the existence of https://github.com/scala-js/scala-js-fake-weakreferences which are implemented in terms of strong references and would wreak havoc if selected over the weak reference implementations.

vasilmkd · 2021-12-20T17:25:55Z

For reference these results are from series/3.2.x and the 3.2.9 release:

With tracing:

benchmarks/Jmh/run -wi 10 -i 10 -f 2 -t 1 -prof gc --jvmArgs -Dcats.effect.tracing.mode=cached --jvmArgs -Dcats.effect.tracing.exceptions.enhanced=true ParallelBenchmark.par
Benchmark                                                       (cpuTokens)  (size)   Mode  Cnt        Score       Error   Units
ParallelBenchmark.parTraverse                                         10000    1000  thrpt   20      290.942 ±     2.351   ops/s
ParallelBenchmark.parTraverse:·gc.alloc.rate                          10000    1000  thrpt   20     1521.536 ±    12.438  MB/sec
ParallelBenchmark.parTraverse:·gc.alloc.rate.norm                     10000    1000  thrpt   20  5758396.046 ±   580.724    B/op
ParallelBenchmark.parTraverse:·gc.churn.G1_Eden_Space                 10000    1000  thrpt   20     1536.991 ±    20.210  MB/sec
ParallelBenchmark.parTraverse:·gc.churn.G1_Eden_Space.norm            10000    1000  thrpt   20  5816806.929 ± 53534.035    B/op
ParallelBenchmark.parTraverse:·gc.churn.G1_Survivor_Space             10000    1000  thrpt   20        0.303 ±     0.024  MB/sec
ParallelBenchmark.parTraverse:·gc.churn.G1_Survivor_Space.norm        10000    1000  thrpt   20     1147.693 ±    89.317    B/op
ParallelBenchmark.parTraverse:·gc.count                               10000    1000  thrpt   20      887.000              counts
ParallelBenchmark.parTraverse:·gc.time                                10000    1000  thrpt   20     1316.000                  ms

Without tracing:

Benchmark                                                       (cpuTokens)  (size)   Mode  Cnt        Score       Error   Units
ParallelBenchmark.parTraverse                                         10000    1000  thrpt   20      295.814 ±     2.993   ops/s
ParallelBenchmark.parTraverse:·gc.alloc.rate                          10000    1000  thrpt   20     1559.392 ±    10.000  MB/sec
ParallelBenchmark.parTraverse:·gc.alloc.rate.norm                     10000    1000  thrpt   20  5804947.416 ± 40176.019    B/op
ParallelBenchmark.parTraverse:·gc.churn.G1_Eden_Space                 10000    1000  thrpt   20     1575.089 ±    15.339  MB/sec
ParallelBenchmark.parTraverse:·gc.churn.G1_Eden_Space.norm            10000    1000  thrpt   20  5863488.584 ± 67068.118    B/op
ParallelBenchmark.parTraverse:·gc.churn.G1_Survivor_Space             10000    1000  thrpt   20        0.318 ±     0.025  MB/sec
ParallelBenchmark.parTraverse:·gc.churn.G1_Survivor_Space.norm        10000    1000  thrpt   20     1181.857 ±    90.029    B/op
ParallelBenchmark.parTraverse:·gc.count                               10000    1000  thrpt   20      909.000              counts
ParallelBenchmark.parTraverse:·gc.time                                10000    1000  thrpt   20     1338.000                  ms

vasilmkd and others added 8 commits December 20, 2021 12:12

Implement a GC-aware WeakBag data structure

aee9b65

Implement a simple synchronized WeakBag as a fallback

d8486cf

Replace the fallback mechanism with the new weak bag

ee490d8

Replace the main suspension monitoring mechanism

895a10b

Fix compilation on JS

3cbcf52

Add ref classes for JS

7202365

Co-authored-by: Arman Bilge <[email protected]>

Remove unused weak hash map iteration code

6cbdd56

Remove weak map JS code

71e1d2e

armanbilge reviewed Dec 20, 2021

View reviewed changes

vasilmkd added 2 commits December 20, 2021 13:32

Add mima exclusion

9cc802e

Group imports together

e521303

vasilmkd mentioned this pull request Dec 20, 2021

Create a custom weak data structure #2666

Closed

vasilmkd added 4 commits December 20, 2021 17:37

Copy the weak references source code

cf627c5

Remove the new dependency

8e71057

Add type aliases on JVM

4df25d8

Switch to the internal APIs

8c08080

vasilmkd marked this pull request as ready for review December 20, 2021 16:49

djspiewak approved these changes Dec 20, 2021

View reviewed changes

djspiewak merged commit 77bb7b8 into typelevel:series/3.3.x Dec 20, 2021

armanbilge mentioned this pull request Dec 23, 2021

Create benchmarking infrastructure for JS (and Native?) #2684

Open

vasilmkd deleted the weak-bag branch January 29, 2022 10:37

armanbilge mentioned this pull request Apr 5, 2022

Securely implement UUIDGen for Scala.js #2945

Merged

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Uh oh!

Cross platform weak bag implementation#2673

Cross platform weak bag implementation#2673
djspiewak merged 14 commits intotypelevel:series/3.3.xfrom
vasilmkd:weak-bag

vasilmkd commented Dec 20, 2021 •

edited

Loading

Uh oh!

armanbilge Dec 20, 2021

Uh oh!

vasilmkd commented Dec 20, 2021

Uh oh!

vasilmkd commented Dec 20, 2021 •

edited

Loading

Uh oh!

vasilmkd commented Dec 20, 2021

Uh oh!

vasilmkd commented Dec 20, 2021 •

edited

Loading

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

		("org.scala-js" %%% "scalajs-weakreferences" % JsWeakReferencesVersion)
		.cross(CrossVersion.for3Use2_13)

Uh oh!

Conversation

vasilmkd commented Dec 20, 2021 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

armanbilge Dec 20, 2021

Choose a reason for hiding this comment

Uh oh!

vasilmkd commented Dec 20, 2021

Uh oh!

vasilmkd commented Dec 20, 2021 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

vasilmkd commented Dec 20, 2021

Uh oh!

vasilmkd commented Dec 20, 2021 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Uh oh!

Reviewers

Assignees

Labels

Projects

Milestone

Development

Uh oh!

3 participants

vasilmkd commented Dec 20, 2021 •

edited

Loading

vasilmkd commented Dec 20, 2021 •

edited

Loading

vasilmkd commented Dec 20, 2021 •

edited

Loading