SIMD vectorization of Array.sum<int>, etc

[Thorium]

Description

Specific overloads (float, float32, int, int64) of Seq.sum, Seq.average, Array.sum and Array.average to take advantage of vectorization in System.Linq.Enumerable module.

This is potentially a naive first try to solve #16230 by the spirit of @T-Gro comment #16230 (comment)

Checklist

• Test cases added
• Performance benchmarks added in case of performance changes
• Release notes entry updated:

  Please make sure to add an entry with short succinct description of the change as well as link to this pull request to the respective release notes file, if applicable.

[github-actions]

❗ Release notes required

---

✅ Found changes and release notes in following paths:

Change path	Release notes path	Description
src/FSharp.Core	docs/release-notes/.FSharp.Core/9.0.300.md

[brianrourkeboll]

I would think that this would be a reasonable place to use static optimization syntax to specify which types should delegate to the LINQ method and which to the existing code, e.g.,

let inline sum (array: ^T array) : ^T =
    existingSumCode array
    when ^T : int   = System.Linq.Enumerable.Sum array
    when ^T : int64 = System.Linq.Enumerable.Sum array
    …

[Thorium]

Sure. I expect "static optimization conditionals" are a compile-time thing and not runtime? Because I can't check easily with sharplab.io, it says "error FS0817: Static optimization conditionals are only for use within the F# library"

[brianrourkeboll]

Yes, the appropriate implementation would be chosen at compile-time once the type parameter was resolved.

You could add some IL tests under https://github.com/dotnet/fsharp/tree/main/tests/FSharp.Compiler.ComponentTests/EmittedIL if you wanted.

[T-Gro]

It looks like the error case is now different for average over an empty collection.
If it is the same exception type, and only a different message, we could justify the break.

(we should not evaluate a seq before calling into Average, since this would be breaking in a different way)

[Thorium]

I did some initial tests to see if this makes sense at all or not. I used the current benchmarks\CompiledCodeBenchmarks\MicroPerf\MicroPerf.fsproj with adding following tests:

    [<Benchmark>]
    member x.ArraySum() = // Array sum
        array
        |> Array.sum 
        |> ignore

    [<Benchmark>]
    member x.ArrayAverage() = // There has an extra map, because average needs float
        array
        |> Array.map float
        |> Array.average 
        |> ignore

    [<Benchmark>]
    member x.ArraySeqSum() = // Seq sum by using array as base data
        array
        |> Seq.sum 
        |> ignore

    [<Benchmark>]
    member x.ListSeqSum() = // Seq sum by using list as base data
        list
        |> Seq.sum 
        |> ignore

And here are the results with current main:

BenchmarkDotNet v0.13.10, Windows 11 (10.0.26100.3915)
13th Gen Intel Core i9-13900H, 1 CPU, 20 logical and 14 physical cores
.NET SDK 9.0.203
  [Host]     : .NET 9.0.4 (9.0.425.16305), X64 RyuJIT AVX2 DEBUG
  Job-ZVNJYC : .NET 9.0.4 (9.0.425.16305), X64 RyuJIT AVX2
  Job-NFLFBL : .NET 9.0.4 (9.0.425.16305), X64 RyuJIT AVX2
  Job-ABSOZP : .NET 9.0.4 (9.0.425.16305), X64 RyuJIT AVX2

Server=True  

Method	Job	Arguments	IterationCount	LaunchCount	WarmupCount	Length	Mean	Error	StdDev	Ratio	RatioSD	Gen0	Allocated	Alloc Ratio
ArraySum	Job-ZVNJYC	/p:BUILDING_USING_DOTNET=true	Default	Default	Default	1000	206.4 ns	1.13 ns	1.06 ns	1.00	0.00	-	-	NA
ArraySum	Job-NFLFBL	/p:BUILDING_USING_DOTNET=true	Default	Default	Default	1000	205.9 ns	0.50 ns	0.39 ns	1.00	0.01	-	-	NA
ArraySum	Job-ABSOZP	Default	2	2	1	1000	207.8 ns	9.28 ns	1.44 ns	1.01	0.01	-	-	NA
ArrayAverage	Job-ZVNJYC	/p:BUILDING_USING_DOTNET=true	Default	Default	Default	1000	1,538.6 ns	35.62 ns	105.04 ns	1.00	0.00	0.1736	8024 B	1.00
ArrayAverage	Job-NFLFBL	/p:BUILDING_USING_DOTNET=true	Default	Default	Default	1000	1,471.2 ns	29.23 ns	53.45 ns	0.94	0.07	0.1736	8024 B	1.00
ArrayAverage	Job-ABSOZP	Default	2	2	1	1000	1,495.0 ns	759.33 ns	117.51 ns	0.92	0.07	0.1736	8024 B	1.00
ArraySeqSum	Job-ZVNJYC	/p:BUILDING_USING_DOTNET=true	Default	Default	Default	1000	500.6 ns	3.69 ns	3.46 ns	1.00	0.00	-	32 B	1.00
ArraySeqSum	Job-NFLFBL	/p:BUILDING_USING_DOTNET=true	Default	Default	Default	1000	496.4 ns	1.80 ns	1.40 ns	0.99	0.01	-	32 B	1.00
ArraySeqSum	Job-ABSOZP	Default	2	2	1	1000	498.9 ns	34.99 ns	5.41 ns	1.00	0.01	-	32 B	1.00
ListSeqSum	Job-ZVNJYC	/p:BUILDING_USING_DOTNET=true	Default	Default	Default	1000	1,234.6 ns	3.12 ns	2.91 ns	1.00	0.00	-	40 B	1.00
ListSeqSum	Job-NFLFBL	/p:BUILDING_USING_DOTNET=true	Default	Default	Default	1000	1,243.7 ns	19.42 ns	17.21 ns	1.01	0.01	-	40 B	1.00
ListSeqSum	Job-ABSOZP	Default	2	2	1	1000	1,251.6 ns	55.34 ns	8.56 ns	1.01	0.01	0.0019	40 B	1.00
ArraySum	Job-ZVNJYC	/p:BUILDING_USING_DOTNET=true	Default	Default	Default	10000	1,949.4 ns	8.32 ns	6.95 ns	1.00	0.00	-	-	NA
ArraySum	Job-NFLFBL	/p:BUILDING_USING_DOTNET=true	Default	Default	Default	10000	1,956.5 ns	24.32 ns	22.75 ns	1.00	0.01	-	-	NA
ArraySum	Job-ABSOZP	Default	2	2	1	10000	1,966.3 ns	128.42 ns	19.87 ns	1.01	0.01	-	-	NA
ArrayAverage	Job-ZVNJYC	/p:BUILDING_USING_DOTNET=true	Default	Default	Default	10000	13,892.3 ns	415.41 ns	1,211.78 ns	1.00	0.00	1.8616	80024 B	1.00
ArrayAverage	Job-NFLFBL	/p:BUILDING_USING_DOTNET=true	Default	Default	Default	10000	15,129.6 ns	242.87 ns	227.18 ns	1.09	0.13	1.9379	80024 B	1.00
ArrayAverage	Job-ABSOZP	Default	2	2	1	10000	12,991.8 ns	8,355.23 ns	1,292.98 ns	1.00	0.26	1.8616	80024 B	1.00
ArraySeqSum	Job-ZVNJYC	/p:BUILDING_USING_DOTNET=true	Default	Default	Default	10000	4,889.1 ns	48.55 ns	43.04 ns	1.00	0.00	-	32 B	1.00
ArraySeqSum	Job-NFLFBL	/p:BUILDING_USING_DOTNET=true	Default	Default	Default	10000	4,828.3 ns	45.71 ns	38.17 ns	0.99	0.01	-	32 B	1.00
ArraySeqSum	Job-ABSOZP	Default	2	2	1	10000	4,833.0 ns	122.61 ns	18.97 ns	0.98	0.00	-	32 B	1.00
ListSeqSum	Job-ZVNJYC	/p:BUILDING_USING_DOTNET=true	Default	Default	Default	10000	14,042.3 ns	280.58 ns	483.98 ns	1.00	0.00	-	40 B	1.00
ListSeqSum	Job-NFLFBL	/p:BUILDING_USING_DOTNET=true	Default	Default	Default	10000	13,767.9 ns	240.03 ns	420.40 ns	0.98	0.04	-	40 B	1.00
ListSeqSum	Job-ABSOZP	Default	2	2	1	10000	13,537.7 ns	4,046.83 ns	626.25 ns	0.95	0.07	-	40 B	1.00

Here are the results with this PR:

Method	Job	Arguments	IterationCount	LaunchCount	WarmupCount	Length	Mean	Error	StdDev	Median	Ratio	RatioSD	Gen0	Allocated	Alloc Ratio
ArraySum	Job-WQFBJH	/p:BUILDING_USING_DOTNET=true	Default	Default	Default	1000	60.66 ns	0.489 ns	0.434 ns	60.59 ns	1.00	0.00	-	-	NA
ArraySum	Job-ZZOCDZ	/p:BUILDING_USING_DOTNET=true	Default	Default	Default	1000	207.89 ns	3.431 ns	3.209 ns	206.96 ns	3.43	0.05	-	-	NA
ArraySum	Job-BMQAWT	Default	2	2	1	1000	60.18 ns	3.505 ns	0.542 ns	60.19 ns	0.99	0.02	-	-	NA
ArrayAverage	Job-WQFBJH	/p:BUILDING_USING_DOTNET=true	Default	Default	Default	1000	1,693.99 ns	29.205 ns	28.683 ns	1,694.44 ns	1.00	0.00	0.1736	8024 B	1.00
ArrayAverage	Job-ZZOCDZ	/p:BUILDING_USING_DOTNET=true	Default	Default	Default	1000	1,443.74 ns	47.701 ns	140.649 ns	1,446.84 ns	0.87	0.04	0.1736	8024 B	1.00
ArrayAverage	Job-BMQAWT	Default	2	2	1	1000	1,486.78 ns	226.790 ns	35.096 ns	1,493.10 ns	0.89	0.02	0.1736	8024 B	1.00
ArraySeqSum	Job-WQFBJH	/p:BUILDING_USING_DOTNET=true	Default	Default	Default	1000	60.23 ns	0.426 ns	0.378 ns	60.11 ns	1.00	0.00	-	-	NA
ArraySeqSum	Job-ZZOCDZ	/p:BUILDING_USING_DOTNET=true	Default	Default	Default	1000	497.31 ns	4.195 ns	3.719 ns	496.69 ns	8.26	0.09	-	32 B	NA
ArraySeqSum	Job-BMQAWT	Default	2	2	1	1000	61.30 ns	1.343 ns	0.208 ns	61.31 ns	1.02	0.00	-	-	NA
ListSeqSum	Job-WQFBJH	/p:BUILDING_USING_DOTNET=true	Default	Default	Default	1000	1,154.49 ns	9.687 ns	9.061 ns	1,155.70 ns	1.00	0.00	-	40 B	1.00
ListSeqSum	Job-ZZOCDZ	/p:BUILDING_USING_DOTNET=true	Default	Default	Default	1000	1,228.62 ns	5.630 ns	4.395 ns	1,228.27 ns	1.06	0.01	-	40 B	1.00
ListSeqSum	Job-BMQAWT	Default	2	2	1	1000	1,164.41 ns	66.398 ns	10.275 ns	1,159.92 ns	1.01	0.02	0.0019	40 B	1.00
ArraySum	Job-WQFBJH	/p:BUILDING_USING_DOTNET=true	Default	Default	Default	10000	546.81 ns	1.656 ns	1.549 ns	547.20 ns	1.00	0.00	-	-	NA
ArraySum	Job-ZZOCDZ	/p:BUILDING_USING_DOTNET=true	Default	Default	Default	10000	1,961.44 ns	25.704 ns	24.044 ns	1,961.63 ns	3.59	0.05	-	-	NA
ArraySum	Job-BMQAWT	Default	2	2	1	10000	541.97 ns	33.331 ns	5.158 ns	541.34 ns	0.99	0.01	-	-	NA
ArrayAverage	Job-WQFBJH	/p:BUILDING_USING_DOTNET=true	Default	Default	Default	10000	16,012.49 ns	316.673 ns	721.226 ns	16,238.87 ns	1.00	0.00	1.8463	80024 B	1.00
ArrayAverage	Job-ZZOCDZ	/p:BUILDING_USING_DOTNET=true	Default	Default	Default	10000	13,651.54 ns	310.627 ns	865.903 ns	13,843.05 ns	0.84	0.08	1.8768	80024 B	1.00
ArrayAverage	Job-BMQAWT	Default	2	2	1	10000	17,271.21 ns	3,620.263 ns	560.240 ns	17,365.04 ns	1.10	0.04	1.8921	80024 B	1.00
ArraySeqSum	Job-WQFBJH	/p:BUILDING_USING_DOTNET=true	Default	Default	Default	10000	554.40 ns	3.815 ns	3.569 ns	554.73 ns	1.00	0.00	-	-	NA
ArraySeqSum	Job-ZZOCDZ	/p:BUILDING_USING_DOTNET=true	Default	Default	Default	10000	4,813.90 ns	39.532 ns	33.011 ns	4,801.26 ns	8.69	0.09	-	32 B	NA
ArraySeqSum	Job-BMQAWT	Default	2	2	1	10000	551.93 ns	12.280 ns	1.900 ns	552.36 ns	1.00	0.00	-	-	NA
ListSeqSum	Job-WQFBJH	/p:BUILDING_USING_DOTNET=true	Default	Default	Default	10000	14,523.24 ns	768.122 ns	2,240.648 ns	13,893.55 ns	1.00	0.00	-	40 B	1.00
ListSeqSum	Job-ZZOCDZ	/p:BUILDING_USING_DOTNET=true	Default	Default	Default	10000	15,181.03 ns	458.748 ns	1,338.189 ns	14,985.06 ns	1.07	0.18	-	40 B	1.00
ListSeqSum	Job-BMQAWT	Default	2	2	1	10000	13,985.62 ns	10,680.237 ns	1,652.779 ns	13,530.39 ns	1.00	0.17	-	40 B	1.00

Edit: I used Net 9 but the FSharp.Core.dll netstandard2.0 version in both, which was probably a mistake because Spans are truely efficient on netstandard2.1 only (?)

[Thorium]

The IL for Enumerable is very complicated, and is heavily relying on the JIT to simplify it. In Release mode, I observe Enumerable to take half the time; in Debug mode Enumerable takes about 4x the time.

Hmm, I'm thinking if default FSI is compiled with debug mode and people use F# as a scripting language, then performance optimizations like this could be marked as #if !DEBUG. However, it would be even weirder if an error message were different between debug and release.

I did run the same performance tests with FSharp.Core.dll netstandard2.1 version

Main branch:

Method	Job	Arguments	IterationCount	LaunchCount	WarmupCount	Length	Mean	Error	StdDev	Ratio	RatioSD	Gen0	Allocated	Alloc Ratio
ArraySum	Job-WCJSKA	/p:BUILDING_USING_DOTNET=true	Default	Default	Default	1000	207.2 ns	2.34 ns	2.19 ns	1.00	0.00	-	-	NA
ArraySum	Job-ZZTGPQ	/p:BUILDING_USING_DOTNET=true	Default	Default	Default	1000	204.2 ns	3.47 ns	2.90 ns	0.99	0.02	-	-	NA
ArraySum	Job-EYKYAG	Default	2	2	1	1000	206.8 ns	22.17 ns	3.43 ns	1.01	0.02	-	-	NA
ArrayAverage	Job-WCJSKA	/p:BUILDING_USING_DOTNET=true	Default	Default	Default	1000	1,503.2 ns	19.56 ns	16.33 ns	1.00	0.00	0.1736	8024 B	1.00
ArrayAverage	Job-ZZTGPQ	/p:BUILDING_USING_DOTNET=true	Default	Default	Default	1000	1,498.6 ns	20.87 ns	19.52 ns	1.00	0.02	0.1736	8024 B	1.00
ArrayAverage	Job-EYKYAG	Default	2	2	1	1000	1,514.8 ns	164.72 ns	25.49 ns	1.01	0.01	0.1736	8024 B	1.00
ArraySeqSum	Job-WCJSKA	/p:BUILDING_USING_DOTNET=true	Default	Default	Default	1000	490.6 ns	2.15 ns	1.80 ns	1.00	0.00	-	32 B	1.00
ArraySeqSum	Job-ZZTGPQ	/p:BUILDING_USING_DOTNET=true	Default	Default	Default	1000	486.8 ns	2.41 ns	2.13 ns	0.99	0.01	-	32 B	1.00
ArraySeqSum	Job-EYKYAG	Default	2	2	1	1000	488.8 ns	74.19 ns	11.48 ns	1.00	0.02	-	32 B	1.00
ListSeqSum	Job-WCJSKA	/p:BUILDING_USING_DOTNET=true	Default	Default	Default	1000	1,214.5 ns	2.83 ns	2.51 ns	1.00	0.00	-	40 B	1.00
ListSeqSum	Job-ZZTGPQ	/p:BUILDING_USING_DOTNET=true	Default	Default	Default	1000	1,241.1 ns	19.15 ns	17.91 ns	1.02	0.01	-	40 B	1.00
ListSeqSum	Job-EYKYAG	Default	2	2	1	1000	1,216.6 ns	34.26 ns	5.30 ns	1.00	0.00	0.0019	40 B	1.00
ArraySum	Job-WCJSKA	/p:BUILDING_USING_DOTNET=true	Default	Default	Default	10000	1,905.3 ns	8.41 ns	7.02 ns	1.00	0.00	-	-	NA
ArraySum	Job-ZZTGPQ	/p:BUILDING_USING_DOTNET=true	Default	Default	Default	10000	1,924.9 ns	12.14 ns	11.36 ns	1.01	0.01	-	-	NA
ArraySum	Job-EYKYAG	Default	2	2	1	10000	1,933.3 ns	95.32 ns	14.75 ns	1.01	0.01	-	-	NA
ArrayAverage	Job-WCJSKA	/p:BUILDING_USING_DOTNET=true	Default	Default	Default	10000	13,970.7 ns	474.45 ns	1,398.93 ns	1.00	0.00	1.8463	80024 B	1.00
ArrayAverage	Job-ZZTGPQ	/p:BUILDING_USING_DOTNET=true	Default	Default	Default	10000	13,967.8 ns	334.64 ns	986.68 ns	1.01	0.12	1.8311	80024 B	1.00
ArrayAverage	Job-EYKYAG	Default	2	2	1	10000	12,963.1 ns	7,565.88 ns	1,170.83 ns	0.85	0.07	1.8768	80024 B	1.00
ArraySeqSum	Job-WCJSKA	/p:BUILDING_USING_DOTNET=true	Default	Default	Default	10000	4,710.8 ns	18.93 ns	15.80 ns	1.00	0.00	-	32 B	1.00
ArraySeqSum	Job-ZZTGPQ	/p:BUILDING_USING_DOTNET=true	Default	Default	Default	10000	4,760.8 ns	17.52 ns	15.53 ns	1.01	0.00	-	32 B	1.00
ArraySeqSum	Job-EYKYAG	Default	2	2	1	10000	4,776.5 ns	345.62 ns	53.49 ns	1.01	0.01	-	32 B	1.00
ListSeqSum	Job-WCJSKA	/p:BUILDING_USING_DOTNET=true	Default	Default	Default	10000	14,616.1 ns	355.10 ns	1,030.22 ns	1.00	0.00	-	40 B	1.00
ListSeqSum	Job-ZZTGPQ	/p:BUILDING_USING_DOTNET=true	Default	Default	Default	10000	13,809.1 ns	270.69 ns	300.87 ns	0.94	0.08	-	40 B	1.00
ListSeqSum	Job-EYKYAG	Default	2	2	1	10000	15,396.3 ns	3,747.35 ns	579.91 ns	1.05	0.14	-	40 B	1.00

This PR

Method	Job	Arguments	IterationCount	LaunchCount	WarmupCount	Length	Mean	Error	StdDev	Median	Ratio	RatioSD	Gen0	Allocated	Alloc Ratio
ArraySum	Job-YRWBVV	/p:BUILDING_USING_DOTNET=true	Default	Default	Default	1000	59.19 ns	0.550 ns	0.514 ns	59.01 ns	1.00	0.00	-	-	NA
ArraySum	Job-YQEOCP	/p:BUILDING_USING_DOTNET=true	Default	Default	Default	1000	206.10 ns	0.720 ns	0.674 ns	206.08 ns	3.48	0.03	-	-	NA
ArraySum	Job-SEOIAS	Default	2	2	1	1000	59.65 ns	4.451 ns	0.689 ns	59.64 ns	1.01	0.02	-	-	NA
ArrayAverage	Job-YRWBVV	/p:BUILDING_USING_DOTNET=true	Default	Default	Default	1000	1,729.86 ns	34.440 ns	72.645 ns	1,750.76 ns	1.00	0.00	0.1736	8024 B	1.00
ArrayAverage	Job-YQEOCP	/p:BUILDING_USING_DOTNET=true	Default	Default	Default	1000	1,521.22 ns	27.082 ns	22.615 ns	1,522.65 ns	0.94	0.05	0.1736	8024 B	1.00
ArrayAverage	Job-SEOIAS	Default	2	2	1	1000	1,658.32 ns	130.915 ns	20.259 ns	1,666.57 ns	0.98	0.02	0.1736	8024 B	1.00
ArraySeqSum	Job-YRWBVV	/p:BUILDING_USING_DOTNET=true	Default	Default	Default	1000	59.88 ns	0.326 ns	0.305 ns	59.84 ns	1.00	0.00	-	-	NA
ArraySeqSum	Job-YQEOCP	/p:BUILDING_USING_DOTNET=true	Default	Default	Default	1000	496.87 ns	6.676 ns	6.245 ns	493.92 ns	8.30	0.10	-	32 B	NA
ArraySeqSum	Job-SEOIAS	Default	2	2	1	1000	60.61 ns	3.319 ns	0.514 ns	60.63 ns	1.01	0.01	-	-	NA
ListSeqSum	Job-YRWBVV	/p:BUILDING_USING_DOTNET=true	Default	Default	Default	1000	1,163.60 ns	18.969 ns	17.743 ns	1,162.85 ns	1.00	0.00	-	40 B	1.00
ListSeqSum	Job-YQEOCP	/p:BUILDING_USING_DOTNET=true	Default	Default	Default	1000	1,230.86 ns	6.546 ns	6.123 ns	1,233.38 ns	1.06	0.02	-	40 B	1.00
ListSeqSum	Job-SEOIAS	Default	2	2	1	1000	1,156.84 ns	66.957 ns	10.362 ns	1,153.00 ns	0.99	0.02	0.0019	40 B	1.00
ArraySum	Job-YRWBVV	/p:BUILDING_USING_DOTNET=true	Default	Default	Default	10000	553.52 ns	6.542 ns	6.119 ns	554.30 ns	1.00	0.00	-	-	NA
ArraySum	Job-YQEOCP	/p:BUILDING_USING_DOTNET=true	Default	Default	Default	10000	1,944.70 ns	16.488 ns	15.423 ns	1,944.27 ns	3.51	0.05	-	-	NA
ArraySum	Job-SEOIAS	Default	2	2	1	10000	547.22 ns	53.119 ns	8.220 ns	543.16 ns	1.00	0.02	-	-	NA
ArrayAverage	Job-YRWBVV	/p:BUILDING_USING_DOTNET=true	Default	Default	Default	10000	16,386.69 ns	261.673 ns	244.769 ns	16,295.85 ns	1.00	0.00	1.8768	80024 B	1.00
ArrayAverage	Job-YQEOCP	/p:BUILDING_USING_DOTNET=true	Default	Default	Default	10000	13,693.86 ns	512.901 ns	1,512.301 ns	14,049.32 ns	0.87	0.11	1.8921	80024 B	1.00
ArrayAverage	Job-SEOIAS	Default	2	2	1	10000	16,549.34 ns	5,509.571 ns	852.612 ns	16,654.92 ns	1.01	0.05	1.8311	80024 B	1.00
ArraySeqSum	Job-YRWBVV	/p:BUILDING_USING_DOTNET=true	Default	Default	Default	10000	549.69 ns	2.934 ns	2.600 ns	549.51 ns	1.00	0.00	-	-	NA
ArraySeqSum	Job-YQEOCP	/p:BUILDING_USING_DOTNET=true	Default	Default	Default	10000	4,843.34 ns	64.337 ns	60.181 ns	4,826.96 ns	8.82	0.12	-	32 B	NA
ArraySeqSum	Job-SEOIAS	Default	2	2	1	10000	551.21 ns	42.723 ns	6.611 ns	550.96 ns	1.00	0.02	-	-	NA
ListSeqSum	Job-YRWBVV	/p:BUILDING_USING_DOTNET=true	Default	Default	Default	10000	13,349.78 ns	349.754 ns	992.195 ns	12,981.18 ns	1.00	0.00	-	40 B	1.00
ListSeqSum	Job-YQEOCP	/p:BUILDING_USING_DOTNET=true	Default	Default	Default	10000	14,501.57 ns	311.609 ns	894.065 ns	14,248.72 ns	1.09	0.10	-	40 B	1.00
ListSeqSum	Job-SEOIAS	Default	2	2	1	10000	12,715.10 ns	817.355 ns	126.487 ns	12,726.11 ns	0.99	0.03	-	40 B	1.00

By quick look of this, the Sum seems to be improved but the Average not really. Should I revert the Average change?