[英]How to improve a push data pipeline in C# to match F# in performance
對我來說,一個reoccuring寵物項目是在F#中實現基於推送的數據管道。 推送管道比LINQ之類的拉管道更簡單,更快(盡管它們沒有拉管道的所有功能)。
讓我困惑一段時間的事情是,我似乎沒有在C#中實現推送管道,這在F#中作為我的推送管道是有效的。 我正在尋找有關如何讓我的C#實現更接近F#的輸入。
F#中的簡單推送管道可以表示如下:
type Receiver<'T> = 'T -> unit
type Stream<'T> = Receiver<'T> -> unit
在C#中我們可以這樣寫:
public delegate void Receiver<in T>(T v);
public delegate void Stream<out T>(Receiver<T> r);
這里的想法是Stream<>
是一個函數,它給一個值接收器調用接收器,其中包含流中的所有值。
這允許我們在F#中定義這樣的map
#Select`:
let inline map (m : 'T -> 'U) (s : Stream<'T>) : Stream<'U> =
fun r -> s (fun v -> r (m v))
C#:
public static Stream<U> Map<T, U>(this Stream<T> t, Func<T, U> m) =>
r => t(v => r(m(v)));
我們可以實現其他功能,直到我們可以定義測試開銷的數據管道。
let trivialTest n =
TrivialStream.range 0 1 n
|> TrivialStream.map int64
|> TrivialStream.filter (fun v -> v &&& 1L = 0L)
|> TrivialStream.map ((+) 1L)
|> TrivialStream.sum
let trivialTestCs n =
Stream
.Range(0,1,n)
.Map(fun v -> int64 v)
.Filter(fun v -> v &&& 1L = 0L)
.Map(fun v -> v + 1L)
.Sum()
在此管道中,每個操作都非常便宜,因此在我們測量它時,底層實現的任何開銷都會顯示出來。
當比較4個不同的數據流水線時,命令式(不是真正的流水線,但有理由檢查實現),trivialpush,trivialpush(C#)和linq這些是.NET 4.7.1 / x64上的數字:
Running imperative with total=100000000, outer=1000000, inner=100 ...
... 87 ms, cc0=0, cc1=0, cc2=0, result=2601L
Running trivialpush with total=100000000, outer=1000000, inner=100 ...
... 414 ms, cc0=53, cc1=0, cc2=0, result=2601L
Running trivialpush(C#) with total=100000000, outer=1000000, inner=100 ...
... 1184 ms, cc0=322, cc1=0, cc2=0, result=2601L
Running linq with total=100000000, outer=1000000, inner=100 ...
... 2080 ms, cc0=157, cc1=0, cc2=0, result=2601L
必要的解決方案是更快和LINQ開始拉數據管道是最慢的。 這是預料之中的。
不期望的是,盡管具有非常相似的實現並以類似的方式使用,但F#push管道的開銷似乎比C#管道少3倍。
如何更改C#數據管道以匹配或取代F#數據管道? 我希望數據管道的API大致相同。
更新2018-06-18
@scrwtp詢問如果我在F#中刪除inline
聯會發生什么。 現在我添加inline
以使sum
按預期工作(在F# inline
允許更高級的泛型)
Running imperative with total=100000000, outer=1000000, inner=100 ...
... 85 ms, cc0=0, cc1=0, cc2=0, result=2601L
Running trivialpush with total=100000000, outer=1000000, inner=100 ...
... 773 ms, cc0=106, cc1=0, cc2=0, result=2601L
Running trivialpush(C#) with total=100000000, outer=1000000, inner=100 ...
... 1181 ms, cc0=322, cc1=0, cc2=0, result=2601L
Running linq with total=100000000, outer=1000000, inner=100 ...
... 2124 ms, cc0=157, cc1=0, cc2=0, result=2601L
這顯着減慢了F#版本,但它仍然比我的C#流庫好50%。
有趣的是,當內聯的唯一內容是構建回調管道時, inline
對性能有如此深遠的影響。 構建完成后,回調管道應該看起來完全一樣。
更新2018-06-24
我決定詳細研究F#和C#數據管道之間的區別。
以下是Filter(fun v -> v &&& 1L = 0L)
的jitted代碼如何查找F#:
; TrivialPush, F#, filter operation
00007ffc`b7d01160 488bc2 mov rax,rdx
; F# inlines the filter function: (fun v -> v &&& 1 = 0L)
; Is even?
00007ffc`b7d01163 a801 test al,1
00007ffc`b7d01165 7512 jne 00007ffc`b7d01179
; Yes, call next chain in pipeline
; Load pointer next step in pipeline
00007ffc`b7d01167 488b4908 mov rcx,qword ptr [rcx+8]
; Load Object Method Table
00007ffc`b7d0116b 488b01 mov rax,qword ptr [rcx]
; Load Table of methods
00007ffc`b7d0116e 488b4040 mov rax,qword ptr [rax+40h]
; Load address of Invoke
00007ffc`b7d01172 488b4020 mov rax,qword ptr [rax+20h]
; Jump to Invoke (tail call)
00007ffc`b7d01176 48ffe0 jmp rax
; No, the number was odd, bail out
00007ffc`b7d01179 33c0 xor eax,eax
00007ffc`b7d0117b c3 ret
關於這段代碼唯一真正的大抱怨是抖動無法內聯尾部調用,我們最終得到一個虛擬尾調用。
讓我們看看C#中的相同數據管道
; TrivialPush, C#, filter operation
; Method prelude
00007ffc`b75c1a10 57 push rdi
00007ffc`b75c1a11 56 push rsi
; Allocate space on stack
00007ffc`b75c1a12 4883ec28 sub rsp,28h
00007ffc`b75c1a16 488bf1 mov rsi,rcx
00007ffc`b75c1a19 488bfa mov rdi,rdx
; Load pointer test delegate (fun v -> v &&& 1 = 0L)
00007ffc`b75c1a1c 488b4e10 mov rcx,qword ptr [rsi+10h]
; Load Method Table
00007ffc`b75c1a20 488b4110 mov rax,qword ptr [rcx+10h]
; Setup this pointer for delegate
00007ffc`b75c1a24 488d4808 lea rcx,[rax+8]
00007ffc`b75c1a28 488b09 mov rcx,qword ptr [rcx]
00007ffc`b75c1a2b 488bd7 mov rdx,rdi
; Load address to Invoke and call
00007ffc`b75c1a2e ff5018 call qword ptr [rax+18h]
; Did filter return true?
00007ffc`b75c1a31 84c0 test al,al
00007ffc`b75c1a33 7411 je 00007ffc`b75c1a46
; Yes, call next step in data pipeline
; Load Method Table
00007ffc`b75c1a35 488b4608 mov rax,qword ptr [rsi+8]
00007ffc`b75c1a39 488d4808 lea rcx,[rax+8]
; Setup this pointer for delegate
00007ffc`b75c1a3d 488b09 mov rcx,qword ptr [rcx]
00007ffc`b75c1a40 488bd7 mov rdx,rdi
; Load address to Invoke and call
00007ffc`b75c1a43 ff5018 call qword ptr [rax+18h]
; Method prelude epilogue
00007ffc`b75c1a46 90 nop
00007ffc`b75c1a47 4883c428 add rsp,28h
00007ffc`b75c1a4b 5e pop rsi
00007ffc`b75c1a4c 5f pop rdi
00007ffc`b75c1a4d c3 ret
; (fun v -> v &&& 1 = 0L) redirect
00007ffc`b75c0408 e963160000 jmp 00007ffc`b75c1a70
; (fun v -> v &&& 1 = 0L)
00007ffc`b75c1a70 488bc2 mov rax,rdx
; Is even?
00007ffc`b75c1a73 a801 test al,1
00007ffc`b75c1a75 0f94c0 sete al
; return result
00007ffc`b75c1a78 0fb6c0 movzx eax,al
; We are done!
00007ffc`b75c1a7b c3 ret
比較F#數據管道,很容易看出上面的代碼更昂貴:
在64位模式下,似乎主要的性能優勢來自於
我們看到F#數據管道步驟沒有內聯,它是內聯的數據管道構建代碼。 然而,這似乎帶來了一些性能上的好處。 也許是因為抖動更容易獲得信息?
為了提高C#管道的性能,似乎我需要構建我的C#代碼,以便抖動虛擬化並內聯調用。 抖動具有這些功能,但為什么它們不適用?
有沒有我可以構建我的F#代碼,以便尾部調用可以被內聯的虛擬化?
完整的F#控制台程序:
module TrivialStream =
// A very simple push stream
type Receiver<'T> = 'T -> unit
type Stream<'T> = Receiver<'T> -> unit
module Details =
module Loop =
let rec range s e r i = if i <= e then r i; range s e r (i + s)
open Details
let inline range b s e : Stream<int> =
fun r -> Loop.range s e r b
let inline filter (f : 'T -> bool) (s : Stream<'T>) : Stream<'T> =
fun r -> s (fun v -> if f v then r v)
let inline map (m : 'T -> 'U) (s : Stream<'T>) : Stream<'U> =
fun r -> s (fun v -> r (m v))
let inline sum (s : Stream<'T>) : 'T =
let mutable ss = LanguagePrimitives.GenericZero
s (fun v -> ss <- ss + v)
ss
module PerformanceTests =
open System
open System.Diagnostics
open System.IO
open System.Linq
open TrivialStreams
let now =
let sw = Stopwatch ()
sw.Start ()
fun () -> sw.ElapsedMilliseconds
let time n a =
let inline cc i = GC.CollectionCount i
let v = a ()
GC.Collect (2, GCCollectionMode.Forced, true)
let bcc0, bcc1, bcc2 = cc 0, cc 1, cc 2
let b = now ()
for i in 1..n do
a () |> ignore
let e = now ()
let ecc0, ecc1, ecc2 = cc 0, cc 1, cc 2
v, (e - b), ecc0 - bcc0, ecc1 - bcc1, ecc2 - bcc2
let trivialTest n =
TrivialStream.range 0 1 n
|> TrivialStream.map int64
|> TrivialStream.filter (fun v -> v &&& 1L = 0L)
|> TrivialStream.map ((+) 1L)
|> TrivialStream.sum
let trivialTestCs n =
Stream
.Range(0,1,n)
.Map(fun v -> int64 v)
.Filter(fun v -> v &&& 1L = 0L)
.Map(fun v -> v + 1L)
.Sum()
let linqTest n =
Enumerable
.Range(0, n + 1)
.Select(fun v -> int64 v)
.Where(fun v -> v &&& 1L = 0L)
.Select(fun v -> v + 1L)
.Sum()
let imperativeTest n =
let rec loop s i =
if i >= 0L then
if i &&& 1L = 0L then
loop (s + i + 1L) (i - 1L)
else
loop s (i - 1L)
else
s
loop 0L (int64 n)
let test () =
printfn "Running performance tests..."
let testCases =
[|
"imperative" , imperativeTest
"trivialpush" , trivialTest
"trivialpush(C#)" , trivialTestCs
"linq" , linqTest
|]
do
// Just in case tiered compilation is activated on dotnet core 2.1+
let warmups = 100
printfn "Warming up..."
for name, a in testCases do
time warmups (fun () -> a warmups) |> ignore
let total = 100000000
let outers =
[|
10
1000
1000000
|]
for outer in outers do
let inner = total / outer
for name, a in testCases do
printfn "Running %s with total=%d, outer=%d, inner=%d ..." name total outer inner
let v, ms, cc0, cc1, cc2 = time outer (fun () -> a inner)
printfn " ... %d ms, cc0=%d, cc1=%d, cc2=%d, result=%A" ms cc0 cc1 cc2 v
printfn "Performance tests completed"
[<EntryPoint>]
let main argv =
PerformanceTests.test ()
0
完整的C#庫:
namespace TrivialStreams
{
using System;
public delegate void Receiver<in T>(T v);
public delegate void Stream<out T>(Receiver<T> r);
public static class Stream
{
public static Stream<int> Range(int b, int s, int e) =>
r =>
{
for(var i = 0; i <= e; i += s)
{
r(i);
}
};
public static Stream<T> Filter<T>(this Stream<T> t, Func<T, bool> f) =>
r => t(v =>
{
if (f(v)) r(v);
});
public static Stream<U> Map<T, U>(this Stream<T> t, Func<T, U> m) =>
r => t(v => r(m(v)));
public static long Sum(this Stream<long> t)
{
var sum = 0L;
t(v => sum += v);
return sum;
}
}
}
F#編譯器有時會在沒有明確指示的情況下內聯函數。 您可以使用[<MethodImpl(MethodImplOptions.NoInlining)>]
注釋函數以防止這種情況。
像這樣更新你的TrivialStream
:
open System.Runtime.CompilerServices
[<MethodImpl(MethodImplOptions.NoInlining)>]
let range b s e : Stream<int> =
fun r -> Loop.range s e r b
[<MethodImpl(MethodImplOptions.NoInlining)>]
let filter (f : 'T -> bool) (s : Stream<'T>) : Stream<'T> =
fun r -> s (fun v -> if f v then r v)
[<MethodImpl(MethodImplOptions.NoInlining)>]
let map (m : 'T -> 'U) (s : Stream<'T>) : Stream<'U> =
fun r -> s (fun v -> r (m v))
[<MethodImpl(MethodImplOptions.NoInlining)>]
let sum (s : Stream<'T>) : 'T =
let mutable ss = LanguagePrimitives.GenericZero
s (fun v -> ss <- ss + v)
ss
然后像這樣更新測試本身:
open System.Runtime.CompilerServices
[<MethodImpl(MethodImplOptions.NoInlining)>]
let parseToInt64 = int64
[<MethodImpl(MethodImplOptions.NoInlining)>]
let filterImpl = fun v -> v &&& 1L = 0L
[<MethodImpl(MethodImplOptions.NoInlining)>]
let mapImpl = ((+) 1L)
let trivialTest n =
TrivialStream.range 0 1 n
|> TrivialStream.map parseToInt64
|> TrivialStream.filter filterImpl
|> TrivialStream.map mapImpl
|> TrivialStream.sum
當作為32位應用程序運行時,這會導致F#運行實際上比C#版本慢 。 對於32位版本的尾調用還有一些額外的奇怪行為。
對於64位版本,這些更改使F#和C#版本相互之間的比例達到15%。
如果你為C# Receiver
交換F# Receiver
和Stream
(或者只是Action<'t>
和Action<Action<'t>>
),那么兩者的性能大致相當,所以我懷疑還有其他優化使用FSharpFunc
正在發揮作用。
open TrivialStreams
// A very simple push stream
//type Receiver<'T> = 'T -> unit
//type Stream<'T> = Receiver<'T> -> unit
module Details =
module Loop =
let rec range s e (r:Receiver<'t> ) i = if i <= e then r.Invoke i; range s e r (i + s)
open Details
open System.Runtime.CompilerServices
[<MethodImpl(MethodImplOptions.NoInlining)>]
let range b s e =
Stream<'t>(fun r -> (Loop.range s e r b))
[<MethodImpl(MethodImplOptions.NoInlining)>]
let filter f (s : Stream<'T>) =
Stream<'T>(fun r -> s.Invoke (fun v -> if f v then r.Invoke v))
[<MethodImpl(MethodImplOptions.NoInlining)>]
let map m (s : Stream<'T>) =
Stream<'U>(fun r -> s.Invoke (fun v -> r.Invoke (m v)))
[<MethodImpl(MethodImplOptions.NoInlining)>]
let sum (s : Stream<'T>) : 'T =
let mutable ss = LanguagePrimitives.GenericZero
s.Invoke (fun v -> ss <- ss + v)
ss
您可以通過使用[MethodImpl(MethodImplOptions.AggressiveInlining)]
屬性注釋方法,將一小部分F#編譯器優化應用於C#,但這只是一個微小的改進。
聲明:本站的技術帖子網頁,遵循CC BY-SA 4.0協議,如果您需要轉載,請注明本站網址或者原文地址。任何問題請咨詢:yoyou2525@163.com.