Microsoft MVP성태의 닷넷 이야기
개발 환경 구성: 383. BenchmarkDotNet 사용 시 주의 사항 [링크 복사], [링크+제목 복사]
조회: 12388
글쓴 사람
정성태 (techsharer at outlook.com)
홈페이지
첨부 파일
 

BenchmarkDotNet 사용 시 주의 사항

BenchmarkDotNet으로,

BenchmarkDotNet 라이브러리 소개
; https://www.sysnet.pe.kr/2/0/11547

struct 타입을 좀 더 테스트해봤습니다.

C# 7.2의 특징 - GC 및 메모리 복사 방지를 위한 struct 타입 개선
; https://www.sysnet.pe.kr/2/0/11546

성능 비교를 위해 다음의 타입들을 생성하고,

ClassVector                 : class로 구현
StructVector                : struct로 구현
InStructVector              : StructVector이고, op_Addition 메서드의 인자를 in 처리
ReadonlyInStructVector      : InStructVector이고, readonly struct로 정의
ReadonlyInlineStructVector  : ReadonlyInStructVector이고, op_Addition에 AggressiveInlining 특성 부여

각각의 타입에 대해 op_Addition을 테스트하는 코드만을 Benchmark에 추가했습니다.

public class VectorBenchmark
{
    [Benchmark]
    public void ClassVectorTest()
    {
        var player = new ClassVector(10.0, 20.0, 30.0);
        var speed = new ClassVector(10.0, 20.0, 30.0);
        var result = player + speed;
    }

    [Benchmark]
    public void StructVectorTest()
    {
        var player = new StructVector(10.0, 20.0, 30.0);
        var speed = new StructVector(10.0, 20.0, 30.0);
        var result = player + speed;
    }

    [Benchmark]
    public void InStructVectorTest()
    {
        var player = new InStructVector(10.0, 20.0, 30.0);
        var speed = new InStructVector(10.0, 20.0, 30.0);
        var result = player + speed;
    }

    [Benchmark]
    public void ReadonlyInStructVectorTest()
    {
        var player = new ReadonlyInStructVector(10.0, 20.0, 30.0);
        var speed = new ReadonlyInStructVector(10.0, 20.0, 30.0);
        var result = player + speed;
    }

    [Benchmark]
    public void ReadonlyInlineStructVectorTest()
    {
        var player = new ReadonlyInlineStructVector(10.0, 20.0, 30.0);
        var speed = new ReadonlyInlineStructVector(10.0, 20.0, 30.0);
        var result = player + speed;
    }
}

결과는 이렇습니다.

// * Summary *

BenchmarkDotNet=v0.10.14, OS=Windows 10.0.17134
Intel Core i5-4670 CPU 3.40GHz (Haswell), 1 CPU, 4 logical and 4 physical cores
  [Host]     : .NET Framework 4.7.1 (CLR 4.0.30319.42000), 64bit RyuJIT-v4.7.3101.0
  DefaultJob : .NET Framework 4.7.1 (CLR 4.0.30319.42000), 64bit RyuJIT-v4.7.3101.0


                             Method |       Mean |     Error |    StdDev |
----------------------------------- |-----------:|----------:|----------:|
                    ClassVectorTest | 13.7698 ns | 0.1722 ns | 0.1611 ns |
                   StructVectorTest |  2.5874 ns | 0.0320 ns | 0.0299 ns |
                 InStructVectorTest | 10.5442 ns | 0.0892 ns | 0.0834 ns |
         ReadonlyInStructVectorTest | 10.2489 ns | 0.0843 ns | 0.0788 ns |
     ReadonlyInlineStructVectorTest |  0.5599 ns | 0.0106 ns | 0.0099 ns |

메서드가 inline 처리된 StructVectorTest, ReadonlyInlineStructVectorTest의 성능이 우세한 것은 당연해 보이고, 또한 그 둘 중에서도 복사가 발생하지 않는 ReadonlyInlineStructVectorTest의 성능이 좀 더 높게 나옵니다.

이 상태에서 테스트 코드마다 GetHashCode 호출을 추가했더니,

public void ...VectorTest()
{
    var player = new ...Vector(10.0, 20.0, 30.0);
    var speed = new ...Vector(10.0, 20.0, 30.0);
    var result = player + speed;

    result.GetHashCode();
}

단지 하나의 코드를 추가했을 뿐인데 이제 결과가 완전히 달라집니다.

                             Method |       Mean |     Error |    StdDev |
----------------------------------- |-----------:|----------:|----------:|
                    ClassVectorTest | 34.8412 ns | 0.2300 ns | 0.2151 ns |
                   StructVectorTest | 47.8031 ns | 0.3171 ns | 0.2966 ns |
                 InStructVectorTest | 53.4127 ns | 1.2505 ns | 1.4401 ns |
         ReadonlyInStructVectorTest | 52.1917 ns | 0.3648 ns | 0.3412 ns |
     ReadonlyInlineStructVectorTest | 39.5150 ns | 0.2622 ns | 0.2452 ns |

확인은 해보지 않았지만, GetHashCode 시 값 형식에서 object.GetHashCode 호출로 인한 내부 박싱이 발생해 힙을 사용하도록 변경된 것이 아닌가 예상됩니다. 일단 그건 그렇다 치고, 문제는 모든 Vector 코드에 GetHashCode를 다음과 같이 직접 구현했을 때 발생합니다.

public override int GetHashCode()
{
    return (int)(_x + _y + _z);
}

그럼, 각각의 결과물들이 다음과 같이 바뀝니다.

// GetHashCode() 추가
                             Method |       Mean |     Error |    StdDev |
----------------------------------- |-----------:|----------:|----------:|
                    ClassVectorTest | 15.7360 ns | 0.1863 ns | 0.1651 ns |
                   StructVectorTest | 10.0317 ns | 0.0746 ns | 0.0623 ns |
                 InStructVectorTest | 12.0021 ns | 0.0687 ns | 0.0643 ns |
         ReadonlyInStructVectorTest | 11.8757 ns | 0.1092 ns | 0.1022 ns |
     ReadonlyInlineStructVectorTest |  0.0000 ns | 0.0000 ns | 0.0000 ns |

ReadonlyInlineStructVectorTest의 결과가 놀라운데요, 어떻게 저럴 수 있을까요? 저 때의 JIT 코드를 windbg로 살펴보면 다음과 같이 나옵니다.

0:000> !DumpMD /d 00007ffa11f47790
Method Name:  ConsoleApp1.VectorBenchmark.ReadonlyInlineStructVectorTest()
Class:        00007ffa11f54858
MethodTable:  00007ffa11f477a8
mdToken:      0000000006000007
Module:       00007ffa11f470d0
IsJitted:     yes
CodeAddr:     00007ffa11ea6450
Transparency: Critical

0:000> !U /d 00007ffa11ea6450
Normal JIT generated code
ConsoleApp1.VectorBenchmark.ReadonlyInlineStructVectorTest()
Begin 00007ffa11ea6450, size 1
>>> 00007ffa`11ea6450 c3              ret

보는 바와 같이 정상적인 테스트 코드를 생성하지 못하고 ret만 포함하는 메서드로 테스트하고 있는 것입니다. 따라서 BenchmarkDotNet으로 성능 측정을 할 때는 그 자체의 버그나 Release 모드에서의 최적화에 따른 의도치 않은 코드 삭제를 주의해야 합니다.

게다가 전통적인 Stopwatch를 이용할 때는 Visual Studio 내에서 disassembly 코드를 보며 곧바로 확인할 수 있는 여지가 있는데, BenchmarkDotNet으로 하게 되면 ConsoleApp1.exe의 자식 프로세스로 임시 exe 프로세스를 생성시켜 성능 측정을 하기 때문에 그런 부분을 파헤치기가 불편한 점도 있습니다.




[이 글에 대해서 여러분들과 의견을 공유하고 싶습니다. 틀리거나 미흡한 부분 또는 의문 사항이 있으시면 언제든 댓글 남겨주십시오.]







[최초 등록일: ]
[최종 수정일: 6/14/2018]

Creative Commons License
이 저작물은 크리에이티브 커먼즈 코리아 저작자표시-비영리-변경금지 2.0 대한민국 라이센스에 따라 이용하실 수 있습니다.
by SeongTae Jeong, mailto:techsharer at outlook.com

비밀번호

댓글 작성자
 




1  2  3  4  [5]  6  7  8  9  10  11  12  13  14  15  ...
NoWriterDateCnt.TitleFile(s)
13500정성태12/24/20232176디버깅 기술: 194. Windbg - x64 가상 주소를 물리 주소로 변환
13498정성태12/23/20232812닷넷: 2186. 한국투자증권 KIS Developers OpenAPI의 C# 래퍼 버전 - eFriendOpenAPI NuGet 패키지
13497정성태12/22/20232293오류 유형: 885. Visual Studiio - error : Could not connect to the remote system. Please verify your connection settings, and that your machine is on the network and reachable.
13496정성태12/21/20232306Linux: 66. 리눅스 - 실행 중인 프로세스 내부의 환경변수 설정을 구하는 방법 (gdb)
13495정성태12/20/20232316Linux: 65. clang++로 공유 라이브러리의 -static 옵션 빌드가 가능할까요?
13494정성태12/20/20232499Linux: 64. Linux 응용 프로그램의 (C++) so 의존성 줄이기(ReleaseMinDependency) - 두 번째 이야기
13493정성태12/19/20232541닷넷: 2185. C# - object를 QueryString으로 직렬화하는 방법
13492정성태12/19/20232261개발 환경 구성: 699. WSL에 nopCommerce 예제 구성
13491정성태12/19/20232231Linux: 63. 리눅스 - 다중 그룹 또는 사용자를 리소스에 권한 부여
13490정성태12/19/20232344개발 환경 구성: 698. Golang - GLIBC 의존을 없애는 정적 빌드 방법
13489정성태12/19/20232132개발 환경 구성: 697. GoLand에서 ldflags 지정 방법
13488정성태12/18/20232066오류 유형: 884. HTTP 500.0 - 명령행에서 실행한 ASP.NET Core 응용 프로그램을 실행하는 방법
13487정성태12/16/20232381개발 환경 구성: 696. C# - 리눅스용 AOT 빌드를 docker에서 수행 [1]
13486정성태12/15/20232194개발 환경 구성: 695. Nuget config 파일에 값 설정/삭제 방법
13485정성태12/15/20232088오류 유형: 883. dotnet build/restore - error : Root element is missing
13484정성태12/14/20232161개발 환경 구성: 694. Windows 디렉터리 경로를 WSL의 /mnt 포맷으로 구하는 방법
13483정성태12/14/20232295닷넷: 2184. C# - 하나의 resource 파일을 여러 프로그램에서 (AOT 시에도) 사용하는 방법파일 다운로드1
13482정성태12/13/20232812닷넷: 2183. C# - eFriend Expert OCX 예제를 .NET Core/5+ Console App에서 사용하는 방법 [2]파일 다운로드1
13481정성태12/13/20232265개발 환경 구성: 693. msbuild - .NET Core/5+ 프로젝트에서 resgen을 이용한 리소스 파일 생성 방법파일 다운로드1
13480정성태12/12/20232597개발 환경 구성: 692. Windows WSL 2 + Chrome 웹 브라우저 설치
13479정성태12/11/20232284개발 환경 구성: 691. WSL 2 (Ubuntu) + nginx 환경 설정
13477정성태12/8/20232465닷넷: 2182. C# - .NET 7부터 추가된 Int128, UInt128 [1]파일 다운로드1
13476정성태12/8/20232202닷넷: 2181. C# - .NET 8 JsonStringEnumConverter의 AOT를 위한 개선파일 다운로드1
13475정성태12/7/20232261닷넷: 2180. .NET 8 - 함수 포인터에 대한 Reflection 정보 조회파일 다운로드1
13474정성태12/6/20232133개발 환경 구성: 690. 닷넷 코어/5+ 버전의 ilasm/ildasm 실행 파일 구하는 방법 - 두 번째 이야기
13473정성태12/5/20232309닷넷: 2179. C# - 값 형식(Blittable)을 메모리 복사를 이용해 바이트 배열로 직렬화/역직렬화파일 다운로드1
1  2  3  4  [5]  6  7  8  9  10  11  12  13  14  15  ...