Reduce `FSM<TState, TData>` allocations #6145

Aaronontheweb · 2022-10-06T00:58:59Z

Changes

Reduced some unnecessary allocations inside the FSM<TState, TData> class.

Checklist

For significant changes, please ensure that the following have been completed (delete if not relevant):

This change follows the Akka.NET API Compatibility Guidelines.
This change follows the Akka.NET Wire Compatibility Guidelines.
I have reviewed my own pull request.
Changes in public API reviewed, if any.

Latest `v1.4` Benchmarks

BenchmarkDotNet=v0.13.1, OS=Windows 10.0.19044.2006 (21H2)
AMD Ryzen 7 1700, 1 CPU, 16 logical and 8 physical cores
.NET SDK=6.0.201
  [Host]     : .NET 6.0.3 (6.0.322.12309), X64 RyuJIT
  DefaultJob : .NET 6.0.3 (6.0.322.12309), X64 RyuJIT

Method	MsgCount	Mean	Error	StdDev	Ratio	RatioSD	Gen 0	Gen 1	Gen 2	Allocated
BenchmarkFsm	1000000	772.8 ms	13.20 ms	12.35 ms	1.74	0.06	68000.0000	4000.0000	1000.0000	287 MB
BenchmarkUntyped	1000000	444.3 ms	8.88 ms	9.50 ms	1.00	0.00	12000.0000	4000.0000	1000.0000	56 MB

This PR's Benchmarks

BenchmarkDotNet=v0.13.1, OS=Windows 10.0.19044.2006 (21H2)
AMD Ryzen 7 1700, 1 CPU, 16 logical and 8 physical cores
.NET SDK=6.0.201
  [Host]     : .NET 6.0.3 (6.0.322.12309), X64 RyuJIT
  DefaultJob : .NET 6.0.3 (6.0.322.12309), X64 RyuJIT

Method	MsgCount	Mean	Error	StdDev	Ratio	RatioSD	Gen 0	Gen 1	Gen 2	Allocated
BenchmarkFsm	1000000	662.5 ms	12.99 ms	13.34 ms	1.47	0.05	35000.0000	4000.0000	1000.0000	150 MB
BenchmarkUntyped	1000000	453.5 ms	8.97 ms	11.35 ms	1.00	0.00	13000.0000	3000.0000	1000.0000	56 MB

48% memory reduction, 14% throughput improvement

* Made `Event` a `readonly struct` * Eliminated unnecessary `List<object>` allocations * Cleaned up XML-DOC comments

Aaronontheweb

Described API and memory allocation changes.

Aaronontheweb · 2022-10-06T01:11:05Z

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

@@ -736,7 +736,7 @@ namespace Akka.Actor
            public override int GetHashCode() { }
            public override string ToString() { }
        }
-        public sealed class Event<TD> : Akka.Actor.INoSerializationVerificationNeeded
+        public struct Event<TD> : Akka.Actor.INoSerializationVerificationNeeded


Technically a binary breaking change maybe, but it's source compatible.

Aaronontheweb · 2022-10-06T01:11:17Z

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

@@ -773,10 +773,10 @@ namespace Akka.Actor
        {
            public static Akka.Actor.FSMBase.StateTimeout Instance { get; }
        }
-        public class State<TS, TD> : System.IEquatable<Akka.Actor.FSMBase.State<TS, TD>>
+        public sealed class State<TS, TD> : System.IEquatable<Akka.Actor.FSMBase.State<TS, TD>>


sealed the State class for better perf.

Aaronontheweb · 2022-10-06T01:11:37Z

src/core/Akka/Actor/FSM.cs

@@ -435,7 +435,7 @@ public class State<TS, TD> : IEquatable<State<TS, TD>>
            /// <param name="notifies">TBD</param>
            public State(TS stateName, TD stateData, TimeSpan? timeout = null, Reason stopReason = null, IReadOnlyList<object> replies = null, bool notifies = true)
            {
-                Replies = replies ?? new List<object>();
+                Replies = replies ?? Array.Empty<object>();


Eliminated unnecessary List<object> allocation.

Aaronontheweb · 2022-10-06T01:12:05Z

src/core/Akka/Actor/FSM.cs

@@ -539,15 +540,16 @@ public State(TS stateName, TD stateData, TimeSpan? timeout = null, Reason stopRe
            /// </summary>
            internal State<TS, TD> WithNotification(bool notifies)
            {
+                // don't bother allocating even a stack type if the notifies value is identical.
+                if (Notifies == notifies)


Don't return a new copy of this object if the current Notifies value matches the new one (it usually does.)

Aaronontheweb · 2022-10-06T01:12:32Z

src/core/Akka/Actor/FSM.cs

@@ -585,26 +587,26 @@ public override int GetHashCode()
        /// which allows pattern matching to extract both state and data.
        /// </summary>
        /// <typeparam name="TD">The state data for this event</typeparam>
-        public sealed class Event<TD> : INoSerializationVerificationNeeded
+        public readonly struct Event<TD> : INoSerializationVerificationNeeded


We allocate one of these for every message we process, so changing to a readonly struct here greatly reduces Gen 0 / 1 GC.

Aaronontheweb · 2022-10-06T01:12:58Z

src/core/Akka/Actor/FSM.cs

@@ -750,7 +752,7 @@ public void StartWith(TState stateName, TData stateData, TimeSpan? timeout = nul
        /// <returns>Descriptor for staying in the current state.</returns>
        public State<TState, TData> Stay()
        {
-            return GoTo(_currentState.StateName).WithNotification(false);
+            return _currentState.WithNotification(false);


Biggest memory savings - don't allocate a new State object each time we call Stay. Re-use the same value.

FWIW, there's no benefit to copying the same state over and over again here. The Reason, Replies properties et al don't have any effects when used here in combination with WithNotification(false).

Aaronontheweb · 2022-10-06T01:43:32Z

Looked at making State<TData, TState> into a struct as well but that's a significant breaking API change (we have lots of FSM behaviors that return null in order to trigger Unhandled behavior) that doesn't yield much benefit once we stop returning new states during Stay() operations.

Aaronontheweb · 2022-10-06T01:54:31Z

Looks like this change definitely caused an issue with the FSMTimingSpec. Going to need to look at that before I ask for review.

Aaronontheweb · 2022-10-06T01:54:48Z

Looks like this change definitely caused an issue with the FSMTimingSpec. Going to need to look at that before I ask for review.

Spec could also be racy - worked on my machine.

Aaronontheweb · 2022-10-06T16:07:39Z

Looks like this change definitely caused an issue with the FSMTimingSpec. Going to need to look at that before I ask for review.

Spec could also be racy - worked on my machine.

Nope, definitely an error related to my changes. Will investigate.

Aaronontheweb · 2022-10-06T23:28:35Z

Benchmarks based on latest changes.

BenchmarkDotNet=v0.13.1, OS=Windows 10.0.19044.2006 (21H2)
AMD Ryzen 7 1700, 1 CPU, 16 logical and 8 physical cores
.NET SDK=6.0.201
  [Host]     : .NET 6.0.3 (6.0.322.12309), X64 RyuJIT
  DefaultJob : .NET 6.0.3 (6.0.322.12309), X64 RyuJIT

Method	MsgCount	Mean	Error	StdDev	Ratio	RatioSD	Gen 0	Gen 1	Gen 2	Allocated
BenchmarkFsm	1000000	641.1 ms	12.73 ms	14.15 ms	1.35	0.05	35000.0000	4000.0000	1000.0000	150 MB
BenchmarkUntyped	1000000	474.1 ms	9.38 ms	10.04 ms	1.00	0.00	12000.0000	4000.0000	1000.0000	56 MB

Arkatufus

LGTM

* close akkadotnet#2560 - added performance benchmarks for FSM * Improved FSM memory consumption * Made `Event` a `readonly struct` * Eliminated unnecessary `List<object>` allocations * Cleaned up XML-DOC comments * don't return new `State<TState, TData>` during `Stay()` * API approvals * fixed `State<TS.,TD>` errors

Aaronontheweb added 4 commits October 5, 2022 15:03

close akkadotnet#2560 - added performance benchmarks for FSM

b5d7d98

Improved FSM memory consumption

b06a0d7

* Made `Event` a `readonly struct` * Eliminated unnecessary `List<object>` allocations * Cleaned up XML-DOC comments

don't return new State<TState, TData> during Stay()

3227a6a

Merge branch 'v1.4' into fsm-actor-allocations

1b72186

Aaronontheweb added akka-actor perf akka.net v1.4 Issues affecting Akka.NET v1.4 labels Oct 6, 2022

Aaronontheweb added this to the 1.4.44 milestone Oct 6, 2022

API approvals

9aad12b

Aaronontheweb commented Oct 6, 2022

View reviewed changes

Aaronontheweb added 2 commits October 6, 2022 17:09

Merge branch 'v1.4' into fsm-actor-allocations

cb2b110

fixed State<TS.,TD> errors

33cb3cb

Arkatufus approved these changes Oct 7, 2022

View reviewed changes

Aaronontheweb enabled auto-merge (squash) October 7, 2022 15:28

Aaronontheweb disabled auto-merge October 7, 2022 15:59

Aaronontheweb merged commit 234bb8f into akkadotnet:v1.4 Oct 7, 2022

Aaronontheweb deleted the fsm-actor-allocations branch October 7, 2022 15:59

Aaronontheweb mentioned this pull request Oct 17, 2022

added v1.4.44 release notes #6197

Merged

Provide feedback

Saved searches

Use saved searches to filter your results more quickly

Reduce `FSM<TState, TData>` allocations #6145

Reduce `FSM<TState, TData>` allocations #6145

Aaronontheweb commented Oct 6, 2022 •

edited

Loading

Aaronontheweb left a comment

Aaronontheweb Oct 6, 2022

Aaronontheweb Oct 6, 2022

Aaronontheweb Oct 6, 2022

Aaronontheweb Oct 6, 2022

Aaronontheweb Oct 6, 2022

Aaronontheweb Oct 6, 2022

Aaronontheweb Oct 6, 2022

Aaronontheweb commented Oct 6, 2022

Aaronontheweb commented Oct 6, 2022

Aaronontheweb commented Oct 6, 2022

Aaronontheweb commented Oct 6, 2022

Aaronontheweb commented Oct 6, 2022

Arkatufus left a comment

Reduce FSM<TState, TData> allocations #6145

Reduce FSM<TState, TData> allocations #6145

Conversation

Aaronontheweb commented Oct 6, 2022 • edited Loading

Changes

Checklist

Latest v1.4 Benchmarks

This PR's Benchmarks

Aaronontheweb left a comment

Choose a reason for hiding this comment

Aaronontheweb Oct 6, 2022

Choose a reason for hiding this comment

Aaronontheweb Oct 6, 2022

Choose a reason for hiding this comment

Aaronontheweb Oct 6, 2022

Choose a reason for hiding this comment

Aaronontheweb Oct 6, 2022

Choose a reason for hiding this comment

Aaronontheweb Oct 6, 2022

Choose a reason for hiding this comment

Aaronontheweb Oct 6, 2022

Choose a reason for hiding this comment

Aaronontheweb Oct 6, 2022

Choose a reason for hiding this comment

Aaronontheweb commented Oct 6, 2022

Aaronontheweb commented Oct 6, 2022

Aaronontheweb commented Oct 6, 2022

Aaronontheweb commented Oct 6, 2022

Aaronontheweb commented Oct 6, 2022

Arkatufus left a comment

Choose a reason for hiding this comment

Reduce `FSM<TState, TData>` allocations #6145

Reduce `FSM<TState, TData>` allocations #6145

Aaronontheweb commented Oct 6, 2022 •

edited

Loading

Latest `v1.4` Benchmarks