我正在寻找在Scala 2.9/Akka 2.0 RC2代码中提高并发性和性能的机会。给定以下代码:
import akka.actor._
case class DataDelivery(data:Double)
class ComputeActor extends Actor {
var buffer = scala.collection.mutable.ArrayBuffer[Double]()
val functionsToCompute = List("f1","f2","f3","f4","f5")
var functionMap = scala.collection.mutable.LinkedHashMap[String,(Map[String,Any]) => Double]()
functionMap += {"f1" -> f1}
functionMap += {"f2" -> f2}
functionMap += {"f3" -> f3}
functionMap += {"f4" -> f4}
functionMap += {"f5" -> f5}
def updateData(data:Double):scala.collection.mutable.ArrayBuffer[Double] = {
buffer += data
buffer
}
def f1(map:Map[String,Any]):Double = {
// println("hello from f1")
0.0
}
def f2(map:Map[String,Any]):Double = {
// println("hello from f2")
0.0
}
def f3(map:Map[String,Any]):Double = {
// println("hello from f3")
0.0
}
def f4(map:Map[String,Any]):Double = {
// println("hello from f4")
0.0
}
def f5(map:Map[String,Any]):Double = {
// println("hello from f5")
0.0
}
def computeValues(immutableBuffer:IndexedSeq[Double]):Map[String,Double] = {
var map = Map[String,Double]()
try {
functionsToCompute.foreach(function => {
val value = functionMap(function)
function match {
case "f1" =>
var v = value(Map("lookback"->10,"buffer"->immutableBuffer,"parm1"->0.0))
map += {function -> v}
case "f2" =>
var v = value(Map("lookback"->20,"buffer"->immutableBuffer))
map += {function -> v}
case "f3" =>
var v = value(Map("lookback"->30,"buffer"->immutableBuffer,"parm1"->1.0,"parm2"->false))
map += {function -> v}
case "f4" =>
var v = value(Map("lookback"->40,"buffer"->immutableBuffer))
map += {function -> v}
case "f5" =>
var v = value(Map("buffer"->immutableBuffer))
map += {function -> v}
case _ =>
println(this.unhandled())
}
})
} catch {
case ex: Exception =>
ex.printStackTrace()
}
map
}
def receive = {
case DataDelivery(data) =>
val startTime = System.nanoTime()/1000
val answers = computeValues(updateData(data))
val endTime = System.nanoTime()/1000
val elapsedTime = endTime - startTime
println("elapsed time is " + elapsedTime)
// reply or forward
case msg =>
println("msg is " + msg)
}
}
object Test {
def main(args:Array[String]) {
val system = ActorSystem("actorSystem")
val computeActor = system.actorOf(Props(new ComputeActor),"computeActor")
var i = 0
while (i < 1000) {
computeActor ! DataDelivery(i.toDouble)
i += 1
}
}
}
当我运行这个时,输出(转换为微秒)是
elapsed time is 4898
elapsed time is 184
elapsed time is 144
.
.
.
elapsed time is 109
elapsed time is 103
您可以看到JVM的增量编译器正在启动。
我认为一个快速的胜利可能是改变
functionsToCompute.foreach(function => {
至
functionsToCompute.par.foreach(function => {
但这会导致以下经过时间
elapsed time is 31689
elapsed time is 4874
elapsed time is 622
.
.
.
elapsed time is 698
elapsed time is 2171
一些信息:
1) 我在2核的Macbook Pro上运行这个。
2) 在完整版本中,函数是在可变共享缓冲区的部分上循环的长时间运行的操作。这似乎不是问题,因为从参与者的邮箱中检索消息控制着流,但我怀疑这可能是并发性增加的问题。这就是我转换为IndexedSeq的原因。
3) 在完整版本中,函数ToCompute列表可能会有所不同,因此并非函数Map中的所有项目都必须调用(即)functionMap.size可能比函数ToCompute.size 大得多
4) 函数可以并行计算,但在返回之前,生成的映射必须完整
一些问题:
1) 我能做些什么来使并行版本运行得更快?
2) 增加非阻塞和阻塞期货在哪里有意义?
3) 将计算转发给另一个参与者在哪里有意义?
4) 提高不变性/安全性的机会有哪些?
谢谢,Bruce
根据请求提供一个示例(很抱歉延迟…我没有SO的通知)。
Akka文档中有一个关于"构建未来"的好例子,但我会为您提供一些更适合您情况的内容。
现在,看完这篇文章后,请花点时间阅读Akka网站上的教程和文档。您缺少了这些文档将为您提供的许多关键信息。
import akka.dispatch.{Await, Future, ExecutionContext}
import akka.util.duration._
import java.util.concurrent.Executors
object Main {
// This just makes the example work. You probably have enough context
// set up already to not need these next two lines
val pool = Executors.newCachedThreadPool()
implicit val ec = ExecutionContext.fromExecutorService(pool)
// I'm simulating your function. It just has to return a tuple, I believe
// with a String and a Double
def theFunction(s: String, d: Double) = (s, d)
def main(args: Array[String]) {
// Here we run your functions - I'm just doing a thousand of them
// for fun. You do what yo need to do
val listOfFutures = (1 to 1000) map { i =>
// Run them in parallel in the future
Future {
theFunction(i.toString, i.toDouble)
}
}
// These lines can be composed better, but breaking them up should
// be more illustrative.
//
// Turn the list of Futures (i.e. Seq[Future[(String, Double)]]) into a
// Future with a sequence of results (i.e. Future[Seq[(String, Double)]])
val futureOfResults = Future.sequence(listOfFutures)
// Convert that future into another future that contains a map instead
// instead of a sequence
val intermediate = futureOfResults map { _.toList.toMap }
// Wait for it complete. Ideally you don't do this. Continue to
// transform the future into other forms or use pipeTo() to get it to go
// as a result to some other Actor. "Await" is really just evil... the
// only place you should really use it is in silly programs like this or
// some other special purpose app.
val resultingMap = Await.result(intermediate, 1 second)
println(resultingMap)
// Again, just to make the example work
pool.shutdown()
}
}
在类路径中运行这个程序所需要的只是akka-actor
jar。阿卡网站会告诉你如何设置你需要的东西,但它真的非常简单。