Scalaz（58）－ scalaz-stream: fs2-并行运算示范，fs2 parallel processing

从表面上来看，Stream代表一连串无穷数据元素。一连串的意思是元素有固定的排列顺序，所以对元素的运算也必须按照顺序来：完成了前面的运算再跟着进行下一个元素的运算。这样来看，Stream应该不是很好的并行运算工具。但是，fs2所支持的并行运算方式不是以数据元素而是以Stream为运算单位的：fs2支持多个Stream同时进行运算，如merge函数。所以fs2使Stream的并行运算成为了可能。

一般来说，我们可能在Stream的几个状态节点要求并行运算：

1、同时运算多个数据源头来产生不排序的数据元素

2、同时对获取的一连串数据元素进行处理，如：map（update）,filter等等

3、同时将一连串数据元素无序存入终点（Sink）

我们可以创建一个例子来示范fs2的并行运算：模拟从3个文件中读取字串，然后统计在这3个文件中母音出现的次数。假设文件读取和母音统计是有任意时间延迟的（latency）,我们看看如何进行并行运算及并行运算能有多少效率上的提升。我们先设定一些跟踪和模拟延迟的帮助函数：

 def log[A](prompt: String): Pipe[Task,A,A] = _.evalMap { a => Task.delay{ println(s"$prompt>"); a }}
                                                   //> log: [A](prompt: String)fs2.Pipe[fs2.Task,A,A]
 def randomDelay[A](max: FiniteDuration): Pipe[Task,A,A] = _.evalMap { a =>
   val delay: Task[Int] = Task.delay { scala.util.Random.nextInt(max.toMillis.toInt) }
   delay.flatMap {d => Task.now(a).schedule(d.millis) }
 }                                                 //> randomDelay: [A](max: scala.concurrent.duration.FiniteDuration)fs2.Pipe[fs2.

log是个跟踪函数，randomDelay是个延迟模拟函数，模拟在max内的任意时间延迟。

与scalaz-stream-0.8不同，fs2重新实现了文件操作功能：不再依赖java的字串（string）处理功能。也不再依赖scodec的二进制数据转换功能。下面是fs2的文件读取方法示范：

 val s1 = io.file.readAll[Task](java.nio.file.Paths.get("/Users/tiger-macpro/basic/BasicBackend.scala"),)
   //> s1  : fs2.Stream[fs2.Task,Byte] = evalScope(Scope(Bind(Eval(Snapshot),<function1>))).flatMap(<function1>)
 val s2 = io.file.readAll[Task](java.nio.file.Paths.get("/Users/tiger-macpro/basic/DatabaseConfig.scala"),)
   //> s2  : fs2.Stream[fs2.Task,Byte] = evalScope(Scope(Bind(Eval(Snapshot),<function1>))).flatMap(<function1>)
 val s3 = io.file.readAll[Task](java.nio.file.Paths.get("/Users/tiger-macpro/basic/BasicProfile.scala"),)
   //> s3  : fs2.Stream[fs2.Task,Byte] = evalScope(Scope(Bind(Eval(Snapshot),<function1>))).flatMap(<function1>)

fs2.io.file.readAll函数的款式如下：

def readAll[F[_]](path: Path, chunkSize: Int)(implicit F: Effect[F]): Stream[F, Byte] ={...}

readAll分批（by chunks）从文件中读取Byte类型数据（当返回数据量小于chunkSize代表完成读取），返回结果类型是Stream[F,Byte]。我们需要进行Byte>>>String转换及分行等处理。fs2在text对象里提供了相关函数：

object text {
  private val utf8Charset = Charset.forName("UTF-8")

  /** Converts UTF-8 encoded byte stream to a stream of `String`. */
  def utf8Decode[F[_]]: Pipe[F, Byte, String] =
    _.chunks.through(utf8DecodeC)

  /** Converts UTF-8 encoded `Chunk[Byte]` inputs to `String`. */
  def utf8DecodeC[F[_]]: Pipe[F, Chunk[Byte], String] = {
    /**
      * Returns the number of continuation bytes if `b` is an ASCII byte or a
      * leading byte of a multi-byte sequence, and -1 otherwise.
      */
    def continuationBytes(b: Byte): Int = {
      if      ((b & 0x80) == 0x00)  // ASCII byte
      else if ((b & 0xE0) == 0xC0)  // leading byte of a 2 byte seq
      else if ((b & 0xF0) == 0xE0)  // leading byte of a 3 byte seq
      else if ((b & 0xF8) == 0xF0)  // leading byte of a 4 byte seq
      else                        - // continuation byte or garbage
    }
...
/** Encodes a stream of `String` in to a stream of bytes using the UTF-8 charset. */
  def utf8Encode[F[_]]: Pipe[F, String, Byte] =
    _.flatMap(s => Stream.chunk(Chunk.bytes(s.getBytes(utf8Charset))))

  /** Encodes a stream of `String` in to a stream of `Chunk[Byte]` using the UTF-8 charset. */
  def utf8EncodeC[F[_]]: Pipe[F, String, Chunk[Byte]] =
    _.map(s => Chunk.bytes(s.getBytes(utf8Charset)))

  /** Transforms a stream of `String` such that each emitted `String` is a line from the input. */
  def lines[F[_]]: Pipe[F, String, String] = {
...

utf8Encode,utf8Decode,lines这几个函数正是我们需要的，它们都是Pipe类型。我们可以把这几个Pipe直接用through接到Stream上：

 val startTime = System.currentTimeMillis         //> startTime  : Long = 1472444756321
  val s1lines = s1.through(text.utf8Decode).through(text.lines)
      .through(randomDelay( millis)).runFold()((b,_) => b + ).unsafeRun
                                                   //> s1lines  : Int = 479
  println(s"reading s1 $s1lines lines in ${System.currentTimeMillis - startTime}ms")
                                                   //> reading s1 479 lines in 5370ms

  val startTime2 = System.currentTimeMillis        //> startTime2  : Long = 1472444761691
  val s2lines = s2.through(text.utf8Decode).through(text.lines)
    .through(randomDelay( millis)).runFold()((b,_) => b + ).unsafeRun
                                                   //> s2lines  : Int = 174
  println(s"reading s2 $s2lines lines in ${System.currentTimeMillis - startTime2}ms")
                                                   //> reading s2 174 lines in 1923ms
  val startTime3 = System.currentTimeMillis        //> startTime3  : Long = 1472444763614
  val s3lines = s3.through(text.utf8Decode).through(text.lines)
    .through(randomDelay( millis)).runFold()((b,_) => b + ).unsafeRun
                                                   //> s3lines  : Int = 174
 println(s"reading s3 $s3lines lines in ${System.currentTimeMillis - startTime3}ms")
                                                   //> reading s3 174 lines in 1928ms
 println(s"reading all three files ${s1lines+s2lines+s3lines} total lines in ${System.currentTimeMillis - startTime}ms")
                                                   //> reading all three files 827 total lines in 9221ms

在以上的例子里我们用runFold函数统计文件的文字行数并在读取过程中用randomDelay来制造了随意长度的拖延。上面3个文件的字串读取和转换处理一共877行、9221ms。

我们知道fs2的并行运算函数concurrent.join函数类型款式是这样的：

def join[F[_],O](maxOpen: Int)(outer: Stream[F,Stream[F,O]])(implicit F: Async[F]): Stream[F,O] = {...}

join运算的对象outer是个两层Stream（Streams of Stream）：Stream[F,Stream[F,P]]，我们需要先进行类型款式调整：

 val lines1 = s1.through(text.utf8Decode).through(text.lines).through(randomDelay( millis))
   //> lines1  : fs2.Stream[fs2.Task,String] = evalScope(Scope(Bind(Eval(Snapshot),<function1>))).flatMap(<function1>).flatMap(<function1>)
 val lines2 = s2.through(text.utf8Decode).through(text.lines).through(randomDelay( millis))
   //> lines2  : fs2.Stream[fs2.Task,String] = evalScope(Scope(Bind(Eval(Snapshot),<function1>))).flatMap(<function1>).flatMap(<function1>)
 val lines3 = s3.through(text.utf8Decode).through(text.lines).through(randomDelay( millis))
   //> lines3  : fs2.Stream[fs2.Task,String] = evalScope(Scope(Bind(Eval(Snapshot),<function1>))).flatMap(<function1>).flatMap(<function1>)
 val ss: Stream[Task,Stream[Task,String]] = Stream(lines1,lines2,lines3)
   //> ss  : fs2.Stream[fs2.Task,fs2.Stream[fs2.Task,String]] = Segment(Emit(Chunk(evalScope(Scope(Bind(Eval(Snapshot),<function1>))).flatMap(<function1>).flatMap(<function1>), evalScope(Scope(Bind(Eval(Snapshot),<function1>))).flatMap(<function1>).flatMap(<function1>), evalScope(Scope(Bind(Eval(Snapshot),<function1>))).flatMap(<function1>).flatMap(<function1>))))

现在这个ss的类型复合我们的要求。我们可以测试一下并行运算的效率：

 val ss_start = System.currentTimeMillis           //> ss_start  : Long = 1472449962698
 val ss_lines = fs2.concurrent.join()(ss).runFold()((b,_) => b + ).unsafeRun
                                                   //> ss_lines  : Int = 827
 println(s"parallel reading all files ${ss_lines} total lines in ${System.currentTimeMillis - ss_start}ms")
                                                   //> parallel reading all files 827 total lines in 5173ms

读取同等行数但只用了5173ms，与之前的9221ms相比，大约有成倍的提速。

join(3)(ss)返回了一个合并的Stream，类型是Stream[Task,String]。我们可以运算这个Stream里母音出现的频率。我们先设计这个统计函数：

 //c 是个vowl
 def vowls(c: Char): Boolean = List('A','E','I','O','U').contains(c)
                                                   //> vowls: (c: Char)Boolean
 //直接用scala标准库实现
 def pipeVowlsCount: Pipe[Task,String,Map[Char,Int]] =
   _.evalMap (text => Task.delay{
      text.toUpperCase.toList.filter(vowls).groupBy(s => s).mapValues(_.size)
      }.schedule((text.length / ).millis))       //> pipeVowlsCount: => fs2.Pipe[fs2.Task,String,Map[Char,Int]]

注意我们使用了text => Task.delay{...}.schedule(d)，实际上我们完全可以用 text => Thread.sleep(d)，但是这样会造成了不纯代码，所以我们用evalMap来实现纯代码运算。试试统计全部字串内母音出现的总数：

 import scalaz.{Monoid}
 //为runFold提供一个Map[Char,Int]Monoid实例
 implicit object mapMonoid extends Monoid[Map[Char,Int]]  {
    def zero: Map[Char,Int] = Map()
    def append(m1: Map[Char,Int], m2: => Map[Char,Int]): Map[Char,Int] = {
      (m1.keySet ++ m2.keySet).map { k =>
        (k, m1.getOrElse(k,) + m2.getOrElse(k,))
      }.toMap
    }
 }
 val vc_start = System.currentTimeMillis           //> vc_start  : Long = 1472464772465
 val vowlsLine = fs2.concurrent.join()(ss).through(pipeVowlsCount)
     .runFold(Map[Char,Int]())(mapMonoid.append(_,_)).unsafeRun
   //> vowlsLine  : scala.collection.immutable.Map[Char,Int] = Map(E -> 3381, U - 838, A -> 2361, I -> 2031, O -> 1824)
 println(s"parallel reading all files and counted vowls sequencially in ${System.currentTimeMillis - vc_start}ms")
   //> parallel reading all files and counted vowls sequencially in 10466ms

我们必须为runFold提供一个Monoid[Map[Char,Int]]实例mapMonoid。

那我们又如何实现统计功能的并行运算呢? fs2.concurrent.join(maxOpen)(...)函数能把一个Stream截成maxOpen数的子Stream，然后对这些子Stream进行并行运算。那么我们又如何转换Stream[F,Stream[F,O]]类型呢？我们必须把Stream[F,O]的O升格成Stream[F,O]。我们先用一个函数来把O转换成Map[Char,Int]，然后把这个函数升格成Stream[Task,Map[Char,Int]，这个可以用Stream.eval实现：

 def fVowlsCount(text: String): Map[Char,Int] =
   text.toUpperCase.toList.filter(vowls).groupBy(s => s).mapValues(_.size)
                                                   //> fVowlsCount: (text: String)Map[Char,Int]
 val parVowlsLine: Stream[Task,Stream[Task,Map[Char,Int]]] = fs2.concurrent.join()(ss)
     .map {text => Stream.eval(Task {fVowlsCount(text)}.schedule((text.length / ).millis))}
     //> parVowlsLine  : fs2.Stream[fs2.Task,fs2.Stream[fs2.Task,Map[Char,Int]]] = attemptEval(Task).flatMap(<function1>).flatMap(<function1>).mapChunks(<function1>)

我们来检查一下运行效率：

 val parvc_start = System.currentTimeMillis        //> parvc_start  : Long = 1472465844694
 fs2.concurrent.join()(parVowlsLine)
   .runFold(Map[Char,Int]())(mapMonoid.append(_,_)).unsafeRun
   //> res0: scala.collection.immutable.Map[Char,Int] = Map(E -> 3381, U -> 838, A-> 2361, I -> 2031, O -> 1824)
 println(s"parallel reading all files and counted vowls in ${System.currentTimeMillis - parvc_start}ms")
   //> parallel reading all files and counted vowls in 4984ms

并行运算只需要4985ms，而流程运算需要10466+(9221-5173)=14xxx，这里有3，4倍的速度提升。

下面是这次讨论的示范源代码：

 import fs2._
 import scala.language.{higherKinds,implicitConversions,postfixOps}
 import scala.concurrent.duration._
 object fs2Merge {
 implicit val strategy = Strategy.fromFixedDaemonPool()
 implicit val scheduler = Scheduler.fromFixedDaemonPool()
 def log[A](prompt: String): Pipe[Task,A,A] = _.evalMap { a => Task.delay{ println(s"$prompt>"); a }}
 def randomDelay[A](max: FiniteDuration): Pipe[Task,A,A] = _.evalMap { a =>
   val delay: Task[Int] = Task.delay { scala.util.Random.nextInt(max.toMillis.toInt) }
   delay.flatMap {d => Task.now(a).schedule(d.millis) }
 }

  val s1 = io.file.readAll[Task](java.nio.file.Paths.get("/Users/tiger-macpro/basic/BasicBackend.scala"),)
  val s2 = io.file.readAll[Task](java.nio.file.Paths.get("/Users/tiger-macpro/basic/DatabaseConfig.scala"),)
  val s3 = io.file.readAll[Task](java.nio.file.Paths.get("/Users/tiger-macpro/basic/BasicProfile.scala"),)

  val startTime = System.currentTimeMillis
  val s1lines = s1.through(text.utf8Decode).through(text.lines)
      .through(randomDelay( millis)).runFold()((b,_) => b + ).unsafeRun
  println(s"reading s1 $s1lines lines in ${System.currentTimeMillis - startTime}ms")

  val startTime2 = System.currentTimeMillis
  val s2lines = s2.through(text.utf8Decode).through(text.lines)
    .through(randomDelay( millis)).runFold()((b,_) => b + ).unsafeRun
  println(s"reading s2 $s2lines lines in ${System.currentTimeMillis - startTime2}ms")
  val startTime3 = System.currentTimeMillis
  val s3lines = s3.through(text.utf8Decode).through(text.lines)
    .through(randomDelay( millis)).runFold()((b,_) => b + ).unsafeRun
 println(s"reading s3 $s3lines lines in ${System.currentTimeMillis - startTime3}ms")
 println(s"reading all three files ${s1lines+s2lines+s3lines} total lines in ${System.currentTimeMillis - startTime}ms")
 val lines1 = s1.through(text.utf8Decode).through(text.lines).through(randomDelay( millis))
 val lines2 = s2.through(text.utf8Decode).through(text.lines).through(randomDelay( millis))
 val lines3 = s3.through(text.utf8Decode).through(text.lines).through(randomDelay( millis))
 val ss: Stream[Task,Stream[Task,String]] = Stream(lines1,lines2,lines3)
 val ss_start = System.currentTimeMillis
 val ss_lines = fs2.concurrent.join()(ss).runFold()((b,_) => b + ).unsafeRun
 println(s"parallel reading all files ${ss_lines} total lines in ${System.currentTimeMillis - ss_start}ms")

 //c 是个vowl
 def vowls(c: Char): Boolean = List('A','E','I','O','U').contains(c)
 //直接用scala标准库实现
 def pipeVowlsCount: Pipe[Task,String,Map[Char,Int]] =
   _.evalMap (text => Task.delay{
      text.toUpperCase.toList.filter(vowls).groupBy(s => s).mapValues(_.size)
      }.schedule((text.length / ).millis))

 import scalaz.{Monoid}
 //为runFold提供一个Map[Char,Int]Monoid实例
 implicit object mapMonoid extends Monoid[Map[Char,Int]]  {
    def zero: Map[Char,Int] = Map()
    def append(m1: Map[Char,Int], m2: => Map[Char,Int]): Map[Char,Int] = {
      (m1.keySet ++ m2.keySet).map { k =>
        (k, m1.getOrElse(k,) + m2.getOrElse(k,))
      }.toMap
    }
 }
 val vc_start = System.currentTimeMillis
 val vowlsLine = fs2.concurrent.join()(ss).through(pipeVowlsCount)
     .runFold(Map[Char,Int]())(mapMonoid.append(_,_)).unsafeRun
 println(s"parallel reading all files and counted vowls sequencially in ${System.currentTimeMillis - vc_start}ms")
 def fVowlsCount(text: String): Map[Char,Int] =
   text.toUpperCase.toList.filter(vowls).groupBy(s => s).mapValues(_.size)
 val parVowlsLine: Stream[Task,Stream[Task,Map[Char,Int]]] = fs2.concurrent.join()(ss)
     .map {text => Stream.eval(Task {fVowlsCount(text)}.schedule((text.length / ).millis))}
 val parvc_start = System.currentTimeMillis
 fs2.concurrent.join()(parVowlsLine)
   .runFold(Map[Char,Int]())(mapMonoid.append(_,_)).unsafeRun
 println(s"parallel reading all files and counted vowls in ${System.currentTimeMillis - parvc_start}ms")
 }