如何使用 actor 重写具有共享依赖项的 for 循环

Question

我们有一些代码需要运行得更快。它已经被分析过，所以我们想使用多个线程。通常我会设置一个内存队列，并让许多线程获取队列的作业并计算结果。对于共享数据，我将使用 ConcurrentHashMap 或类似的。

我真的不想再走那条路了。根据我的阅读，使用 actor 会产生更干净的代码，如果我使用 akka，迁移到 1 个以上的 jvm 应该会更容易。这是真的吗？

然而，我不知道如何思考演员，所以我不知道从哪里开始。

为了更好地了解这个问题，这里有一些示例代码：

case class Trade(price:Double, volume:Int, stock:String) {
  def value(priceCalculator:PriceCalculator) =
    (priceCalculator.priceFor(stock)-> price)*volume
}
class PriceCalculator {
  def priceFor(stock:String) = {
    Thread.sleep(20)//a slow operation which can be cached
    50.0
  }
}
object ValueTrades {

  def valueAll(trades:List[Trade],
      priceCalculator:PriceCalculator):List[(Trade,Double)] = {
    trades.map { trade => (trade,trade.value(priceCalculator)) }
  }

  def main(args:Array[String]) {
    val trades = List(
      Trade(30.5, 10, "Foo"),
      Trade(30.5, 20, "Foo")
      //usually much longer
    )
    val priceCalculator = new PriceCalculator
    val values = valueAll(trades, priceCalculator)
  }

}

如果有使用 Actor 经验的人能够建议这如何映射到 Actor，我将不胜感激。

Answer 1

这是对我对昂贵计算的共享结果的评论的补充。这是：

import scala.actors._
import Actor._
import Futures._

case class PriceFor(stock: String) // Ask for result

// The following could be an "object" as well, if it's supposed to be singleton
class PriceCalculator extends Actor {
  val map = new scala.collection.mutable.HashMap[String, Future[Double]]()
  def act = loop {
    react {
      case PriceFor(stock) => reply(map getOrElseUpdate (stock, future {
        Thread.sleep(2000) // a slow operation
        50.0
      }))
    }
  }
}

这是一个用法示例：

scala> val pc = new PriceCalculator; pc.start
pc: PriceCalculator = PriceCalculator@141fe06

scala> class Test(stock: String) extends Actor {
     |   def act = {
     |     println(System.currentTimeMillis().toString+": Asking for stock "+stock)
     |     val f = (pc !? PriceFor(stock)).asInstanceOf[Future[Double]]
     |     println(System.currentTimeMillis().toString+": Got the future back")
     |     val res = f.apply() // this blocks until the result is ready
     |     println(System.currentTimeMillis().toString+": Value: "+res)
     |   }
     | }
defined class Test

scala> List("abc", "def", "abc").map(new Test(_)).map(_.start)
1269310737461: Asking for stock abc
res37: List[scala.actors.Actor] = List(Test@6d888e, Test@1203c7f, Test@163d118)
1269310737461: Asking for stock abc
1269310737461: Asking for stock def
1269310737464: Got the future back

scala> 1269310737462: Got the future back
1269310737465: Got the future back
1269310739462: Value: 50.0
1269310739462: Value: 50.0
1269310739465: Value: 50.0


scala> new Test("abc").start // Should return instantly
1269310755364: Asking for stock abc
res38: scala.actors.Actor = Test@15b5b68
1269310755365: Got the future back

scala> 1269310755367: Value: 50.0

Answer 2

对于简单的并行化，我会抛出一堆工作来处理，然后等待它们全部返回，我倾向于使用 Futures 模式。

class ActorExample {
  import actors._
  import Actor._
  class Worker(val id: Int) extends Actor {
    def busywork(i0: Int, i1: Int) = {
      var sum,i = i0
      while (i < i1) {
        i += 1
        sum += 42*i
      }
      sum
    }
    def act() { loop { react {
      case (i0:Int,i1:Int) => sender ! busywork(i0,i1)
      case None => exit()
    }}}
  }

  val workforce = (1 to 4).map(i => new Worker(i)).toList

  def parallelFourSums = {
    workforce.foreach(_.start())
    val futures = workforce.map(w => w !! ((w.id,1000000000)) );
    val computed = futures.map(f => f() match {
      case i:Int => i
      case _ => throw new IllegalArgumentException("I wanted an int!")
    })
    workforce.foreach(_ ! None)
    computed
  }

  def serialFourSums = {
    val solo = workforce.head
    workforce.map(w => solo.busywork(w.id,1000000000))
  }

  def timed(f: => List[Int]) = {
    val t0 = System.nanoTime
    val result = f
    val t1 = System.nanoTime
    (result, t1-t0)
  }

  def go {
    val serial = timed( serialFourSums )
    val parallel = timed( parallelFourSums )
    println("Serial result:  " + serial._1)
    println("Parallel result:" + parallel._1)
    printf("Serial took   %.3f seconds\n",serial._2*1e-9)
    printf("Parallel took %.3f seconds\n",parallel._2*1e-9)
  }
}

基本上，这个想法是创建一组工作人员（每个工作负载一个），然后将所有数据扔给他们！这立即回馈了未来。当您尝试读取未来时，发送方会阻塞，直到工作人员实际完成数据处理。

您可以重写上面的代码，以便 PriceCalculator 扩展 Actor，而 valueAll 协调数据的返回。

请注意，传递非不可变数据时必须小心。

不管怎样，在我输入的机器上，如果你运行上面的代码，你会得到：（

scala> (new ActorExample).go
Serial result:  List(-1629056553, -1629056636, -1629056761, -1629056928)
Parallel result:List(-1629056553, -1629056636, -1629056761, -1629056928)
Serial took   1.532 seconds
Parallel took 0.443 seconds

显然我至少有四个核心；并行计时有点不同，具体取决于哪个工作人员获得什么处理器以及其他正在发生的事情机器。）