出于学习目的,我尝试实现MemoryManager和MemoryPool,并尝试与标准实现竞争。但是特别是我的MemoryManager真慢。有人可以指出我的方向吗?
我的内存池:
internal abstract class ByteMemoryPool : MemoryPool<byte>
{
private const int POOL_USAGE_BORDER_BYTES = 85000;
public override int MaxBufferSize => Int32.MaxValue;
public new static ByteMemoryPool.Impl Shared { get; } = new ByteMemoryPool.Impl();
public override IMemoryOwner<byte> Rent(int minBufferSize = -1)
{
return RentCore(minBufferSize);
}
protected override void Dispose(bool disposing)
{
}
private Rental RentCore(int minBufferSize)
{
return new Rental(minBufferSize);
}
public sealed class Impl : ByteMemoryPool
{
public new Rental Rent(int minBufferSize) => RentCore(minBufferSize);
}
public struct Rental : IMemoryOwner<byte>
{
private byte[]? _array;
private readonly bool _notRented;
public Rental(int minBufferSize)
{
if (minBufferSize < POOL_USAGE_BORDER_BYTES)
{
_array = new byte[minBufferSize];
_notRented = true;
}
else
{
_array = ArrayPool<byte>.Shared.Rent(minBufferSize);
_notRented = false;
}
}
public Memory<byte> Memory
{
get
{
if (_array == null)
throw new ObjectDisposedException(nameof(_array));
return new Memory<byte>(_array);
}
}
public void Dispose()
{
if (_array != null && !_notRented)
{
ArrayPool<byte>.Shared.Return(_array, true);
_array = null;
}
else
{
_array = null;
}
}
}
}
我的内存管理器:
internal sealed class NativeByteMemoryManager : MemoryManager<byte>
{
private IntPtr _memoryPtr;
private readonly int _length;
public unsafe NativeByteMemoryManager(int length)
{
_length = length;
_memoryPtr = Marshal.AllocHGlobal(length);
Unsafe.InitBlock((void*)_memoryPtr, 0, (uint)_length);
}
public override Memory<byte> Memory => CreateMemory(_length);
public override unsafe Span<byte> GetSpan()
{
return new Span<byte>(_memoryPtr.ToPointer(), _length);
}
public override unsafe MemoryHandle Pin(int elementIndex = 0)
{
void* pointer = (void*) ((byte*) _memoryPtr + elementIndex);
return new MemoryHandle(pointer, default, this);
}
public override void Unpin()
{
Marshal.FreeHGlobal(_memoryPtr);
_memoryPtr = IntPtr.Zero;
}
protected override void Dispose(bool disposing)
{
if (_memoryPtr != IntPtr.Zero)
{
Marshal.FreeHGlobal(_memoryPtr);
_memoryPtr = IntPtr.Zero;
}
}
}
基准:
public class MemoryManagerBenchmark
{
[Params(1000, 8000, 64000, 4000000)]
[System.Diagnostics.CodeAnalysis.SuppressMessage("Design", "CA1051:Do not declare visible instance fields", Justification = "<Pending>")]
public int ArraySize;
[Benchmark(Baseline = true)]
public int MemoryPoolDefault()
{
var x = ArrayPool<byte>.Shared.Rent(ArraySize);
var l = x.Length;
ArrayPool<byte>.Shared.Return(x, true);
return l;
}
[Benchmark(Baseline = false)]
public int MemoryPoolByte()
{
using var x = ByteMemoryPool.Shared.Rent(ArraySize);
var l = x.Memory.Length;
return l;
}
[Benchmark(Baseline = false)]
public int MemoryManager()
{
using var x = new NativeByteMemoryManager(ArraySize);
var l = x.Memory.Length;
return l;
}
}
和结果:
BenchmarkDotNet=v0.11.5, OS=Windows 10.0.17763.107 (1809/October2018Update/Redstone5)
Intel Core i7-2600 CPU 3.40GHz (Sandy Bridge), 1 CPU, 8 logical and 4 physical cores
.NET Core SDK=3.0.100
[Host] : .NET Core 3.0.0 (CoreCLR 4.700.19.46205, CoreFX 4.700.19.46214), 64bit RyuJIT
Job-MXYBLG : .NET Core 3.0.0 (CoreCLR 4.700.19.46205, CoreFX 4.700.19.46214), 64bit RyuJIT
Force=False IterationCount=15 LaunchCount=2
WarmupCount=10
Method | ArraySize | Mean | Error | StdDev | Median | Kurtosis | Skewness | Ratio | RatioSD | Rank | Baseline | Gen 0 | Gen 1 | Gen 2 | Allocated |
------------------ |---------- |---------------:|--------------:|--------------:|----------------:|---------:|---------:|------:|--------:|-----:|--------- |---------:|---------:|---------:|----------:|
**MemoryManager** | **1000** | **187.3 ns** | **28.466 ns** | **41.726 ns** | **168.24 ns** | **1.210** | **0.1370** | **1.88** | **0.45** | **3** | **No** | **0.0076** | **-** | **-** | **32 B** |
**MemoryPoolByte** | **1000** | **123.1 ns** | **6.342 ns** | **9.492 ns** | **121.44 ns** | **1.730** | **0.4267** | **1.24** | **0.14** | **2** | **No** | **0.2447** | **-** | **-** | **1024 B** |
**MemoryPoolDefault** | **1000** | **100.2 ns** | **5.226 ns** | **7.821 ns** | **97.50 ns** | **2.284** | **0.6929** | **1.00** | **0.00** | **1** | **Yes** | **-** | **-** | **-** | **-** |
| | | | | | | | | | | | | | | |
**MemoryManager** | **8000** | **374.1 ns** | **25.279 ns** | **37.054 ns** | **349.88 ns** | **1.264** | **0.2485** | **1.54** | **0.22** | **2** | **No** | **0.0076** | **-** | **-** | **32 B** |
**MemoryPoolByte** | **8000** | **842.4 ns** | **12.637 ns** | **18.523 ns** | **839.46 ns** | **2.485** | **0.7287** | **3.46** | **0.26** | **3** | **No** | **1.9150** | **-** | **-** | **8024 B** |
**MemoryPoolDefault** | **8000** | **245.1 ns** | **12.542 ns** | **17.988 ns** | **236.31 ns** | **5.935** | **1.9246** | **1.00** | **0.00** | **1** | **Yes** | **-** | **-** | **-** | **-** |
| | | | | | | | | | | | | | | |
**MemoryManager** | **64000** | **2,311.8 ns** | **87.763 ns** | **131.359 ns** | **2,266.83 ns** | **2.146** | **0.6641** | **1.06** | **0.06** | **2** | **No** | **0.0076** | **-** | **-** | **32 B** |
**MemoryPoolByte** | **64000** | **5,351.5 ns** | **82.720 ns** | **118.634 ns** | **5,298.23 ns** | **4.749** | **1.5884** | **2.46** | **0.14** | **3** | **No** | **15.1443** | **-** | **-** | **64024 B** |
**MemoryPoolDefault** | **64000** | **2,187.6 ns** | **83.603 ns** | **125.133 ns** | **2,102.50 ns** | **2.154** | **0.9189** | **1.00** | **0.00** | **1** | **Yes** | **-** | **-** | **-** | **-** |
| | | | | | | | | | | | | | | |
**MemoryManager** | **4000000** | **2,188,789.3 ns** | **65,843.021 ns** | **98,550.733 ns** | **2,165,661.52 ns** | **4.130** | **1.3955** | **10.78** | **0.72** | **2** | **No** | **-** | **-** | **-** | **32 B** |
**MemoryPoolByte** | **4000000** | **199,434.5 ns** | **2,634.057 ns** | **3,777.686 ns** | **198,360.50 ns** | **3.567** | **0.9854** | **0.98** | **0.04** | **1** | **No** | **999.7559** | **999.7559** | **999.7559** | **-** |
**MemoryPoolDefault** | **4000000** | **203,299.8 ns** | **3,986.979 ns** | **5,967.523 ns** | **201,993.74 ns** | **3.340** | **0.7295** | **1.00** | **0.00** | **1** | **Yes** | **999.7559** | **999.7559** | **999.7559** | **-** |
答案 0 :(得分:1)
基准测试工具只能告诉您运行给定代码所需的时间。 探查器可以告诉您为什么要花这么长时间。因此,您需要分析代码以找出答案。
我已经在我的博客上介绍了使用BenchmarkDotNet和PerfView进行的示例性能调查,这可能是您的一个很好的起点:https://adamsitnik.com/Sample-Perf-Investigation/
如果PerfView不符合您的要求,您可以尝试其他分析器:https://github.com/dotnet/performance/blob/master/docs/profiling-workflow-dotnet-runtime.md