Improve performance on `IActorRef.Child` API #5242

Aaronontheweb · 2021-09-01T21:02:42Z

Builds upon #5241, but introduces some breaking API changes.

Performance numbers:

`dev`

BenchmarkDotNet=v0.13.1, OS=Windows 10.0.19041.1165 (2004/May2020Update/20H1)
AMD Ryzen 7 1700, 1 CPU, 16 logical and 8 physical cores
.NET SDK=5.0.302
  [Host]     : .NET Core 3.1.17 (CoreCLR 4.700.21.31506, CoreFX 4.700.21.31502), X64 RyuJIT
  DefaultJob : .NET Core 3.1.17 (CoreCLR 4.700.21.31506, CoreFX 4.700.21.31502), X64 RyuJIT

Method	Mean	Error	StdDev	Gen 0	Allocated
ResolveChild	64.69 ns	0.145 ns	0.129 ns	-	-
Resolve3DeepChildRepointableActorRef	652.15 ns	9.237 ns	8.641 ns	0.0839	352 B
Resolve3DeepChildLocalActorRef	311.74 ns	4.924 ns	4.605 ns	0.0629	264 B

This PR

BenchmarkDotNet=v0.13.1, OS=Windows 10.0.19041.1165 (2004/May2020Update/20H1)
AMD Ryzen 7 1700, 1 CPU, 16 logical and 8 physical cores
.NET SDK=5.0.302
  [Host]     : .NET Core 3.1.17 (CoreCLR 4.700.21.31506, CoreFX 4.700.21.31502), X64 RyuJIT
  DefaultJob : .NET Core 3.1.17 (CoreCLR 4.700.21.31506, CoreFX 4.700.21.31502), X64 RyuJIT

Method	Mean	Error	StdDev	Gen 0	Allocated
ResolveChild	47.40 ns	0.018 ns	0.014 ns	-	-
Resolve3DeepChildRepointableActorRef	398.67 ns	6.259 ns	5.855 ns	0.0496	208 B
Resolve3DeepChildLocalActorRef	227.38 ns	2.343 ns	1.830 ns	0.0420	176 B

… of `IEnumerable<T>`

Aaronontheweb · 2021-09-01T21:03:12Z

src/core/Akka.API.Tests/CoreAPISpec.ApproveCore.approved.txt

@@ -1533,7 +1533,7 @@ namespace Akka.Actor
        public override Akka.Actor.ActorPath Path { get; }
        public override Akka.Actor.IActorRefProvider Provider { get; }
        public override Akka.Actor.ICell Underlying { get; }
-        public override Akka.Actor.IActorRef GetChild(System.Collections.Generic.IEnumerable<string> name) { }
+        public override Akka.Actor.IActorRef GetChild(System.Collections.Generic.IReadOnlyList<string> name) { }


Source-compatible, but not necessarily a binary-compatible change.

Aaronontheweb · 2021-09-01T21:11:54Z

Looks like this may have broken cluster formation based on some of the testing I'm doing locally. I'll check it out some once I get reports back from the test suite.

Aaronontheweb · 2021-09-02T01:54:57Z

Looks like there's some work still to be done with the RemoteDaemon and how it resolves remotely deployed actors...

Aaronontheweb · 2021-09-02T02:50:05Z

So once #5243 gets merged, new baseline numbers for dev:

New numbers

BenchmarkDotNet=v0.13.1, OS=Windows 10.0.19041.1165 (2004/May2020Update/20H1)
AMD Ryzen 7 1700, 1 CPU, 16 logical and 8 physical cores
.NET SDK=5.0.302
  [Host]     : .NET Core 3.1.17 (CoreCLR 4.700.21.31506, CoreFX 4.700.21.31502), X64 RyuJIT
  DefaultJob : .NET Core 3.1.17 (CoreCLR 4.700.21.31506, CoreFX 4.700.21.31502), X64 RyuJIT

Method	Mean	Error	StdDev	Gen 0	Allocated
ResolveChild	67.12 ns	0.041 ns	0.035 ns	-	-
Resolve3DeepChildRepointableActorRef	534.29 ns	2.166 ns	1.809 ns	0.0324	136 B
Resolve3DeepChildLocalActorRef	367.67 ns	0.650 ns	0.543 ns	0.0095	40 B

Numbers for this PR:

BenchmarkDotNet=v0.13.1, OS=Windows 10.0.19041.1165 (2004/May2020Update/20H1)
AMD Ryzen 7 1700, 1 CPU, 16 logical and 8 physical cores
.NET SDK=5.0.302
  [Host]     : .NET Core 3.1.17 (CoreCLR 4.700.21.31506, CoreFX 4.700.21.31502), X64 RyuJIT
  DefaultJob : .NET Core 3.1.17 (CoreCLR 4.700.21.31506, CoreFX 4.700.21.31502), X64 RyuJIT

Method	Mean	Error	StdDev	Gen 0	Allocated
ResolveChild	49.00 ns	0.150 ns	0.141 ns	-	-
Resolve3DeepChildRepointableActorRef	382.39 ns	2.349 ns	2.197 ns	0.0172	72 B
Resolve3DeepChildLocalActorRef	362.73 ns	1.708 ns	1.514 ns	0.0095	40 B

Top level actor resolves (i.e. ones that go through Akka.Cluster.Sharding) got a lot faster, but ones beneath the top-level actors stayed pretty much the same.

…heweb/akka.net into perf/improve-ResolveAPI

This reverts commit 8cce9b7.

Aaronontheweb · 2021-09-02T03:13:16Z

VirtualPathContainer resolve performance baseline

BenchmarkDotNet=v0.13.1, OS=Windows 10.0.19041.1165 (2004/May2020Update/20H1)
AMD Ryzen 7 1700, 1 CPU, 16 logical and 8 physical cores
.NET SDK=5.0.302
  [Host]     : .NET Core 3.1.17 (CoreCLR 4.700.21.31506, CoreFX 4.700.21.31502), X64 RyuJIT
  DefaultJob : .NET Core 3.1.17 (CoreCLR 4.700.21.31506, CoreFX 4.700.21.31502), X64 RyuJIT

Method	Mean	Error	StdDev	Gen 0	Allocated
ResolveChild	66.71 ns	0.048 ns	0.040 ns	-	-
Resolve3DeepChildRepointableActorRef	540.70 ns	1.754 ns	1.555 ns	0.0324	136 B
Resolve3DeepChildLocalActorRef	368.26 ns	1.574 ns	1.472 ns	0.0095	40 B
ResolveVirtualPathContainer	93.03 ns	0.532 ns	0.471 ns	0.0305	128 B

On this PR:

Forgot to add baseline numbers for this:

BenchmarkDotNet=v0.13.1, OS=Windows 10.0.19041.1165 (2004/May2020Update/20H1)
AMD Ryzen 7 1700, 1 CPU, 16 logical and 8 physical cores
.NET SDK=5.0.302
  [Host]     : .NET Core 3.1.17 (CoreCLR 4.700.21.31506, CoreFX 4.700.21.31502), X64 RyuJIT
  DefaultJob : .NET Core 3.1.17 (CoreCLR 4.700.21.31506, CoreFX 4.700.21.31502), X64 RyuJIT

Method	Mean	Error	StdDev	Gen 0	Allocated
ResolveChild	49.81 ns	0.171 ns	0.160 ns	-	-
Resolve3DeepChildRepointableActorRef	372.47 ns	2.948 ns	2.758 ns	0.0172	72 B
Resolve3DeepChildLocalActorRef	398.80 ns	2.803 ns	2.485 ns	0.0095	40 B
ResolveVirtualPathContainer	94.21 ns	0.885 ns	0.828 ns	0.0172	72 B

to11mtm · 2021-09-02T11:59:17Z

src/core/Akka/Util/Internal/Collections/ListSlice.cs

+
+        public T this[int index]
+        {
+            get => _array.ElementAt(Offset + index);


Since _array is an IReadOnlyList<T> here we may be better off using the Indexer on that interface rather than ElementAt.

If _array does not also implement IList<T> we will force an enumeration

.ElementAt adds additional null and type checks even in best case.

Ah yes, that's a bit of leftovers from when I was using IReadOnlyCollection instead of IReadOnlyList - the former doesn't have an indexer. I'll update that.

Arkatufus

LGTM

Aaronontheweb · 2021-09-02T15:03:48Z

Even better numbers after fixing the indexer operation

BenchmarkDotNet=v0.13.1, OS=Windows 10.0.19041.1165 (2004/May2020Update/20H1)
AMD Ryzen 7 1700, 1 CPU, 16 logical and 8 physical cores
.NET SDK=5.0.302
  [Host]     : .NET Core 3.1.17 (CoreCLR 4.700.21.31506, CoreFX 4.700.21.31502), X64 RyuJIT
  DefaultJob : .NET Core 3.1.17 (CoreCLR 4.700.21.31506, CoreFX 4.700.21.31502), X64 RyuJIT

Method	Mean	Error	StdDev	Gen 0	Allocated
ResolveChild	49.69 ns	0.224 ns	0.187 ns	-	-
Resolve3DeepChildRepointableActorRef	371.25 ns	2.587 ns	2.419 ns	0.0172	72 B
Resolve3DeepChildLocalActorRef	387.38 ns	2.858 ns	2.534 ns	0.0095	40 B
ResolveVirtualPathContainer	89.60 ns	0.842 ns	0.788 ns	0.0172	72 B

Aaronontheweb · 2021-09-02T15:04:37Z

Although IMHO those differences are probably in the margin of error

Aaronontheweb · 2021-09-02T15:04:59Z

I wonder if getting rid of these deeply nested TryGet(... out) calls would also help....

Aaronontheweb · 2021-09-02T16:23:41Z

Perf numbers from simplified branching:

BenchmarkDotNet=v0.13.1, OS=Windows 10.0.19041.1165 (2004/May2020Update/20H1)
AMD Ryzen 7 1700, 1 CPU, 16 logical and 8 physical cores
.NET SDK=5.0.302
  [Host]     : .NET Core 3.1.17 (CoreCLR 4.700.21.31506, CoreFX 4.700.21.31502), X64 RyuJIT
  DefaultJob : .NET Core 3.1.17 (CoreCLR 4.700.21.31506, CoreFX 4.700.21.31502), X64 RyuJIT

Method	Mean	Error	StdDev	Gen 0	Allocated
ResolveChild	50.69 ns	0.028 ns	0.022 ns	-	-
Resolve3DeepChildRepointableActorRef	354.04 ns	2.459 ns	2.300 ns	0.0172	72 B
Resolve3DeepChildLocalActorRef	362.21 ns	3.093 ns	2.893 ns	0.0095	40 B
ResolveVirtualPathContainer	89.11 ns	1.173 ns	1.097 ns	0.0172	72 B

Aaronontheweb · 2021-09-02T16:33:47Z

Perf numbers from removing redundant Try(... out) methods

BenchmarkDotNet=v0.13.1, OS=Windows 10.0.19041.1165 (2004/May2020Update/20H1)
AMD Ryzen 7 1700, 1 CPU, 16 logical and 8 physical cores
.NET SDK=5.0.302
  [Host]     : .NET Core 3.1.17 (CoreCLR 4.700.21.31506, CoreFX 4.700.21.31502), X64 RyuJIT
  DefaultJob : .NET Core 3.1.17 (CoreCLR 4.700.21.31506, CoreFX 4.700.21.31502), X64 RyuJIT

Method	Mean	Error	StdDev	Gen 0	Allocated
ResolveChild	50.04 ns	0.137 ns	0.122 ns	-	-
Resolve3DeepChildRepointableActorRef	345.53 ns	1.424 ns	1.112 ns	0.0172	72 B
Resolve3DeepChildLocalActorRef	328.63 ns	1.803 ns	1.598 ns	0.0095	40 B
ResolveVirtualPathContainer	88.01 ns	1.020 ns	0.905 ns	0.0172	72 B

Aaronontheweb · 2021-09-02T16:35:18Z

src/core/Akka.API.Tests/CoreAPISpec.ApproveCore.approved.txt

@@ -93,7 +93,6 @@ namespace Akka.Actor
        public System.Collections.Generic.IEnumerable<Akka.Actor.IInternalActorRef> GetChildren() { }
        public static Akka.Actor.IActorRef GetCurrentSelfOrNoSender() { }
        public static Akka.Actor.IActorRef GetCurrentSenderOrNoSender() { }
-        [System.ObsoleteAttribute("Use TryGetSingleChild [0.7.1]")]


Removed TryGetSingleChild since no one was actually using its boolean functionality at all

Aaronontheweb · 2021-09-02T16:35:48Z

src/core/Akka.Remote/RemoteSystemDaemon.cs


-                var child = GetChild(concatenatedChildNames);
+                var child = GetChild(concatenatedChildNames.ToList());


Extra allocation in remote deployments here.... might see if I can go about fixing that still....

Aaronontheweb · 2021-09-02T16:35:59Z

src/core/Akka.Remote/RemoteSystemDaemon.cs

@@ -313,7 +311,7 @@ public override IActorRef GetChild(IEnumerable<string> name)
                {
                    if (uid != ActorCell.UndefinedUid && uid != child.Path.Uid)
                        return Nobody.Instance;
-                    return n == 0 ? child : child.GetChild(name.TakeRight(n));
+                    return n == 0 ? child : child.GetChild(name.TakeRight(n).ToList());


Another allocation in remote deployments

Aaronontheweb · 2021-09-02T16:36:50Z

src/core/Akka/Actor/ActorCell.Children.cs

@@ -353,58 +353,37 @@ private bool TryGetChildRestartStatsByName(string name, out ChildRestartStats ch
        /// </summary>
        /// <param name="name">N/A</param>
        /// <returns>N/A</returns>
-        [Obsolete("Use TryGetSingleChild [0.7.1]")]
        public IInternalActorRef GetSingleChild(string name)


Half the perf optimizations happened here:

Simplified branching

Remove Try... out stuff since it was all functionally being converted back into ActorRefs.Nobody anyway.

Aaronontheweb · 2021-09-02T16:37:33Z

src/core/Akka/Util/Internal/Collections/ListSlice.cs

+    /// <see cref="ArraySegment{T}"/> but for <see cref="IReadOnlyList{T}"/>
+    /// </summary>
+    /// <typeparam name="T"></typeparam>
+    internal struct ListSlice<T> : IList<T>, IReadOnlyList<T>


Basically an ArraySegment that takes a type of IReadOnlyList<T>

Arkatufus

LGTM

Aaronontheweb added 2 commits September 1, 2021 14:53

improveme child lookup performance

9a50416

modified IActorRef.Child API to accept a IReadOnlyList<T> instead…

0aa7122

… of `IEnumerable<T>`

Aaronontheweb added akka-actor api-change perf labels Sep 1, 2021

Aaronontheweb commented Sep 1, 2021

View reviewed changes

Aaronontheweb added 7 commits September 1, 2021 16:50

fixed a bug with low-level resolution

bf6e162

approved core API changes

fdac3b9

approved again....

bc91ad8

fixed - was missing one approval

d754014

added extension method to simplify copying

3910d11

cleaned up RepointableActorRef

f8f5342

cleaned up branching

3197d0c

Aaronontheweb added 3 commits September 1, 2021 21:28

implemented CopyTo method for debugging

6b590d7

added child-resolve program

8cce9b7

removed expensive allocation from MinimalActorRef.GetChild

16e33c8

Aaronontheweb added 4 commits September 1, 2021 21:50

Merge branch 'dev' into perf/improve-ResolveAPI

702d7b8

Merge branch 'perf/improve-ResolveAPI' of https://github.com/Aaronont…

26db634

…heweb/akka.net into perf/improve-ResolveAPI

Revert "added child-resolve program"

bc66127

This reverts commit 8cce9b7.

Merge branch 'dev' into perf/improve-ResolveAPI

075159b

to11mtm reviewed Sep 2, 2021

View reviewed changes

Arkatufus approved these changes Sep 2, 2021

View reviewed changes

Aaronontheweb added 2 commits September 2, 2021 09:57

Merge branch 'dev' into perf/improve-ResolveAPI

d04e5a7

fixed indexer operation

9f8b852

simplified branching inside TryGetSingleChild

83224be

removed redundant TryOut methods

60b3537

Aaronontheweb marked this pull request as ready for review September 2, 2021 16:34

Aaronontheweb commented Sep 2, 2021

View reviewed changes

Aaronontheweb requested a review from Arkatufus September 2, 2021 16:58

Arkatufus approved these changes Sep 2, 2021

View reviewed changes

Aaronontheweb merged commit d25db5d into akkadotnet:dev Sep 2, 2021

Aaronontheweb deleted the perf/improve-ResolveAPI branch September 2, 2021 17:22

Aaronontheweb added this to the 1.4.25 milestone Sep 8, 2021

This was referenced Sep 8, 2021

Akka.NET v1.4.25 Release Notes #5274

Merged

Akka.NET v1.4.25 Release #5275

Merged

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Improve performance on `IActorRef.Child` API #5242

Improve performance on `IActorRef.Child` API #5242

Aaronontheweb commented Sep 1, 2021

Aaronontheweb Sep 1, 2021

Aaronontheweb commented Sep 1, 2021

Aaronontheweb commented Sep 2, 2021

Aaronontheweb commented Sep 2, 2021

Aaronontheweb commented Sep 2, 2021 •

edited

Loading

to11mtm Sep 2, 2021

Aaronontheweb Sep 2, 2021

Arkatufus left a comment

Aaronontheweb commented Sep 2, 2021

Aaronontheweb commented Sep 2, 2021

Aaronontheweb commented Sep 2, 2021

Aaronontheweb commented Sep 2, 2021

Aaronontheweb commented Sep 2, 2021

Aaronontheweb Sep 2, 2021

Aaronontheweb Sep 2, 2021

Aaronontheweb Sep 2, 2021

Aaronontheweb Sep 2, 2021

Aaronontheweb Sep 2, 2021

Arkatufus left a comment


		var child = GetChild(concatenatedChildNames);
		var child = GetChild(concatenatedChildNames.ToList());

Improve performance on IActorRef.Child API #5242

Improve performance on IActorRef.Child API #5242

Conversation

Aaronontheweb commented Sep 1, 2021

dev

This PR

Choose a reason for hiding this comment

Aaronontheweb commented Sep 1, 2021

Aaronontheweb commented Sep 2, 2021

Aaronontheweb commented Sep 2, 2021

Aaronontheweb commented Sep 2, 2021 • edited Loading

Choose a reason for hiding this comment

Choose a reason for hiding this comment

Arkatufus left a comment

Choose a reason for hiding this comment

Aaronontheweb commented Sep 2, 2021

Aaronontheweb commented Sep 2, 2021

Aaronontheweb commented Sep 2, 2021

Aaronontheweb commented Sep 2, 2021

Aaronontheweb commented Sep 2, 2021

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

Arkatufus left a comment

Choose a reason for hiding this comment

Improve performance on `IActorRef.Child` API #5242

Improve performance on `IActorRef.Child` API #5242

`dev`

Aaronontheweb commented Sep 2, 2021 •

edited

Loading