FunDA(15)- 示范:任务并行运算 - user task parallel execution

    FunDA的并行运算施用就是对用户自定义函数的并行运算。原理上就是把一个输入流截分成多个输入流并行地输入到一个自定义函数的多个运行实例。这些函数运行实例同时在各自不同的线程里同步运算直至耗尽所有输入。并行运算的具体函数实例数是用fs2-nondeterminism的算法根据CPU内核数、线程池配置和用户指定的最大运算实例数来决定的。我们在这次示范里可以对比一下同样工作内容的并行运算和串形运算效率。在前面示范里我们获取了一个AQMRPT表。但这个表不够合理化(normalized):state和county还没有实现编码与STATES和COUNTIES表的连接。在这次示范里我们就创建一个新表NORMAQM,把AQMRPT表内数据都搬进来。并在这个过程中把STATENAME和COUNTYNAME字段转换成STATES和COUNTIES表的id字段。下面就是NORMAQM表结构:

  case class NORMAQMModel(rid: Long
                         , mid: Int
                         , state: Int
                         , county: Int
                         , year: Int
                         , value: Int
                         , average: Int
                         ) extends FDAROW

  class NORMAQMTable(tag: Tag) extends Table[NORMAQMModel](tag, "NORMAQM") {
    def rid = column[Long]("ROWID",O.AutoInc,O.PrimaryKey)
    def mid = column[Int]("MEASUREID")
    def state = column[Int]("STATID")
    def county = column[Int]("COUNTYID")
    def year = column[Int]("REPORTYEAR")
    def value = column[Int]("VALUE")
    def average = column[Int]("AVG")

    def * = (rid,mid,state,county,year,value,average) <> (NORMAQMModel.tupled, NORMAQMModel.unapply)
  }


  val NORMAQMQuery = TableQuery[NORMAQMTable]

下面是这个表的初始化铺垫代码: 

  val db = Database.forConfig("h2db")

  //drop original table schema
  val futVectorTables = db.run(MTable.getTables)

  val futDropTable = futVectorTables.flatMap{ tables => {
    val tableNames = tables.map(t => t.name.name)
    if (tableNames.contains(NORMAQMQuery.baseTableRow.tableName))
      db.run(NORMAQMQuery.schema.drop)
    else Future()
  }
  }.andThen {
    case Success(_) => println(s"Table ${NORMAQMQuery.baseTableRow.tableName} dropped successfully! ")
    case Failure(e) => println(s"Failed to drop Table ${NORMAQMQuery.baseTableRow.tableName}, it may not exist! Error: ${e.getMessage}")
  }
  Await.ready(futDropTable,Duration.Inf)

  //create new table to refine AQMRawTable
  val actionCreateTable = Models.NORMAQMQuery.schema.create
  val futCreateTable = db.run(actionCreateTable).andThen {
    case Success(_) => println("Table created successfully!")
    case Failure(e) => println(s"Table may exist already! Error: ${e.getMessage}")
  }
  //would carry on even fail to create table
  Await.ready(futCreateTable,Duration.Inf)


  //truncate data, only available in slick 3.2.1
  val futTruncateTable = futVectorTables.flatMap{ tables => {
    val tableNames = tables.map(t => t.name.name)
    if (tableNames.contains(NORMAQMQuery.baseTableRow.tableName))
      db.run(NORMAQMQuery.schema.truncate)
    else Future()
  }
  }.andThen {
    case Success(_) => println(s"Table ${NORMAQMQuery.baseTableRow.tableName} truncated successfully!")
    case Failure(e) => println(s"Failed to truncate Table ${NORMAQMQuery.baseTableRow.tableName}! Error: ${e.getMessage}")
  }
  Await.ready(futDropTable,Duration.Inf)

我们需要设计一个函数从STATES表里用AQMRPT表的STATENAME查询ID。我故意把这个函数设计成一个完整的FunDA程序。这样可以模拟一个比较消耗io和计算资源的独立过程(不要理会任何合理性,目标是增加io和运算消耗): 

  //a conceived task for the purpose of resource consumption
  //getting id with corresponding name from STATES table
  def getStateID(state: String): Int = {
    //create a stream for state id with state name
    implicit def toState(row:  StateTable#TableElementType) = StateModel(row.id,row.name)
    val stateLoader = FDAViewLoader(slick.jdbc.H2Profile)(toState _)
    val stateSeq = stateLoader.fda_typedRows(StateQuery.result)(db).toSeq
    //constructed a Stream[Task,String]
    val stateStream =  fda_staticSource(stateSeq)()
    var id  = -1
    def getid: FDAUserTask[FDAROW] = row => {
      row match {
        case StateModel(stid,stname) =>   //target row type
          if (stname.contains(state)) {
            id = stid
            fda_break      //exit
          }
          else fda_skip   //take next row
        case _ => fda_skip
      }
    }
    stateStream.appendTask(getid).startRun
    id
  }

可以看到getStateID函数每次运算都重复构建stateStream。这样可以达到增加io操作的目的。

同样,我们也需要设计另一个函数来从COUNTIES表里获取id字段:

  //another conceived task for the purpose of resource consumption
  //getting id with corresponding names from COUNTIES table
  def getCountyID(state: String, county: String): Int = {
    //create a stream for county id with state name and county name
    implicit def toCounty(row:  CountyTable#TableElementType) = CountyModel(row.id,row.name)
    val countyLoader = FDAViewLoader(slick.jdbc.H2Profile)(toCounty _)
    val countySeq = countyLoader.fda_typedRows(CountyQuery.result)(db).toSeq
    //constructed a Stream[Task,String]
    val countyStream =  fda_staticSource(countySeq)()
    var id  = -1
    def getid: FDAUserTask[FDAROW] = row => {
      row match {
        case CountyModel(cid,cname) =>   //target row type
          if (cname.contains(state) && cname.contains(county)) {
            id = cid
            fda_break      //exit
          }
          else fda_skip   //take next row
        case _ => fda_skip
      }
    }
    countyStream.appendTask(getid).startRun
    id
  }

我们可以如下这样获取这个程序的数据源:

  //original table listing
  implicit def toAQMRPT(row: AQMRPTTable#TableElementType) =
    AQMRPTModel(row.rid,row.mid,row.state,row.county,row.year,row.value,row.total,row.valid)
  val AQMRPTLoader = FDAStreamLoader(slick.jdbc.H2Profile)(toAQMRPT _)
  val AQMRPTStream = AQMRPTLoader.fda_typedStream(AQMRPTQuery.result)(db)(256,256)()

按照正常的FunDA流程我们设计了两个用户自定义函数:一个根据数据行内的state和county字段调用函数getStateID和getCountyID获取相应id后构建一条新的NORMAQM表插入指令行,然后传给下个自定义函数。下个自定义函数就直接运算收到的动作行:

  def getIdsThenInsertAction: FDAUserTask[FDAROW] = row => {
    row match {
      case aqm: AQMRPTModel =>
        if (aqm.valid) {
          val stateId = getStateID(aqm.state)
          val countyId = getCountyID(aqm.state,aqm.county)
          val action = NORMAQMQuery += NORMAQMModel(0,aqm.mid, stateId, countyId, aqm.year,aqm.value,aqm.total)
          fda_next(FDAActionRow(action))
        }
        else fda_skip
      case _ => fda_skip
    }
  }
  val runner = FDAActionRunner(slick.jdbc.H2Profile)
  def runInsertAction: FDAUserTask[FDAROW] = row =>
   row match {
    case FDAActionRow(action) =>
      runner.fda_execAction(action)(db)
      fda_skip
    case _ => fda_skip
  }

像前面几篇示范那样我们把这两个用户自定义函数与数据源组合起来成为完整的FunDA程序后startRun就可以得到实际效果了:

    AQMRPTStream.take(10000)
      .appendTask(getIdsThenInsertAction)
      .appendTask(runInsertAction)
      .startRun

这个程序运算了579秒,不过这是个单一线程运算。我们想知道并行运算结果。那么我们首先要把这个getIdsThenInsertAction转成一个并行运算函数FDAParTask:

AQMRPTStream.toPar(getIdsThenInsertAction)

FunDA提供了并行运算器fda_runPar:

      implicit val strategy = Strategy.fromCachedDaemonPool("cachedPool")
      fda_runPar(AQMRPTStream.take(100000).toPar(getIdsThenInsertAction))(8)  //max 8 open computations
        .appendTask(runInsertAction)
        .startRun

我们可以自定义线程池。fda_runPar返回标准的FunDA FDAPipeLine,所以我们可以在后面挂上runInsertAction函数。下面是不同行数的运算时间对比结果:

    //processing 10000 rows in a single thread in 570 seconds
          // processing 10000 rows parallelly  in 316 seconds

    //processing 20000 rows in a single thread in 1090 seconds
            //processing 20000 rows parallelly  in 614 seconds

    //processing 100000 rows in a single thread in 2+ hrs
      //processing 100000 rows parallelly  in 3885 seconds

可以得出,并行运算对越大数据集有更大的效率提高。下面就是这次示范的源代码:

import slick.jdbc.meta._
import com.bayakala.funda._
import api._
import scala.language.implicitConversions
import scala.concurrent.ExecutionContext.Implicits.global
import scala.concurrent.duration._
import scala.concurrent.{Await, Future}
import scala.util.{Failure, Success}
import slick.jdbc.H2Profile.api._
import Models._
import fs2.Strategy

object ParallelTasks extends App {

  val db = Database.forConfig("h2db")

  //drop original table schema
  val futVectorTables = db.run(MTable.getTables)

  val futDropTable = futVectorTables.flatMap{ tables => {
    val tableNames = tables.map(t => t.name.name)
    if (tableNames.contains(NORMAQMQuery.baseTableRow.tableName))
      db.run(NORMAQMQuery.schema.drop)
    else Future()
  }
  }.andThen {
    case Success(_) => println(s"Table ${NORMAQMQuery.baseTableRow.tableName} dropped successfully! ")
    case Failure(e) => println(s"Failed to drop Table ${NORMAQMQuery.baseTableRow.tableName}, it may not exist! Error: ${e.getMessage}")
  }
  Await.ready(futDropTable,Duration.Inf)

  //create new table to refine AQMRawTable
  val actionCreateTable = Models.NORMAQMQuery.schema.create
  val futCreateTable = db.run(actionCreateTable).andThen {
    case Success(_) => println("Table created successfully!")
    case Failure(e) => println(s"Table may exist already! Error: ${e.getMessage}")
  }
  //would carry on even fail to create table
  Await.ready(futCreateTable,Duration.Inf)


  //truncate data, only available in slick 3.2.1
  val futTruncateTable = futVectorTables.flatMap{ tables => {
    val tableNames = tables.map(t => t.name.name)
    if (tableNames.contains(NORMAQMQuery.baseTableRow.tableName))
      db.run(NORMAQMQuery.schema.truncate)
    else Future()
  }
  }.andThen {
    case Success(_) => println(s"Table ${NORMAQMQuery.baseTableRow.tableName} truncated successfully!")
    case Failure(e) => println(s"Failed to truncate Table ${NORMAQMQuery.baseTableRow.tableName}! Error: ${e.getMessage}")
  }
  Await.ready(futDropTable,Duration.Inf)

  //a conceived task for the purpose of resource consumption
  //getting id with corresponding name from STATES table
  def getStateID(state: String): Int = {
    //create a stream for state id with state name
    implicit def toState(row:  StateTable#TableElementType) = StateModel(row.id,row.name)
    val stateLoader = FDAViewLoader(slick.jdbc.H2Profile)(toState _)
    val stateSeq = stateLoader.fda_typedRows(StateQuery.result)(db).toSeq
    //constructed a Stream[Task,String]
    val stateStream =  fda_staticSource(stateSeq)()
    var id  = -1
    def getid: FDAUserTask[FDAROW] = row => {
      row match {
        case StateModel(stid,stname) =>   //target row type
          if (stname.contains(state)) {
            id = stid
            fda_break      //exit
          }
          else fda_skip   //take next row
        case _ => fda_skip
      }
    }
    stateStream.appendTask(getid).startRun
    id
  }
  //another conceived task for the purpose of resource consumption
  //getting id with corresponding names from COUNTIES table
  def getCountyID(state: String, county: String): Int = {
    //create a stream for county id with state name and county name
    implicit def toCounty(row:  CountyTable#TableElementType) = CountyModel(row.id,row.name)
    val countyLoader = FDAViewLoader(slick.jdbc.H2Profile)(toCounty _)
    val countySeq = countyLoader.fda_typedRows(CountyQuery.result)(db).toSeq
    //constructed a Stream[Task,String]
    val countyStream =  fda_staticSource(countySeq)()
    var id  = -1
    def getid: FDAUserTask[FDAROW] = row => {
      row match {
        case CountyModel(cid,cname) =>   //target row type
          if (cname.contains(state) && cname.contains(county)) {
            id = cid
            fda_break      //exit
          }
          else fda_skip   //take next row
        case _ => fda_skip
      }
    }
    countyStream.appendTask(getid).startRun
    id
  }

  //original table listing
  implicit def toAQMRPT(row: AQMRPTTable#TableElementType) =
    AQMRPTModel(row.rid,row.mid,row.state,row.county,row.year,row.value,row.total,row.valid)
  val AQMRPTLoader = FDAStreamLoader(slick.jdbc.H2Profile)(toAQMRPT _)
  val AQMRPTStream = AQMRPTLoader.fda_typedStream(AQMRPTQuery.result)(db)(256,256)()

  def getIdsThenInsertAction: FDAUserTask[FDAROW] = row => {
    row match {
      case aqm: AQMRPTModel =>
        if (aqm.valid) {
          val stateId = getStateID(aqm.state)
          val countyId = getCountyID(aqm.state,aqm.county)
          val action = NORMAQMQuery += NORMAQMModel(0,aqm.mid, stateId, countyId, aqm.year,aqm.value,aqm.total)
          fda_next(FDAActionRow(action))
        }
        else fda_skip
      case _ => fda_skip
    }
  }
  val runner = FDAActionRunner(slick.jdbc.H2Profile)
  def runInsertAction: FDAUserTask[FDAROW] = row =>
   row match {
    case FDAActionRow(action) =>
      runner.fda_execAction(action)(db)
      fda_skip
    case _ => fda_skip
  }

  val cnt_start = System.currentTimeMillis()
/*
    AQMRPTStream.take(100000)
      .appendTask(getIdsThenInsertAction)
      .appendTask(runInsertAction)
      .startRun
    //println(s"processing 10000 rows in a single thread in ${(System.currentTimeMillis - cnt_start)/1000} seconds")
    //processing 10000 rows in a single thread in 570 seconds
    //println(s"processing 20000 rows in a single thread in ${(System.currentTimeMillis - cnt_start)/1000} seconds")
    //processing 20000 rows in a single thread in 1090 seconds
    //println(s"processing 100000 rows in a single thread in ${(System.currentTimeMillis - cnt_start)/1000} seconds")
    //processing 100000 rows in a single thread in 2+ hrs


      implicit val strategy = Strategy.fromCachedDaemonPool("cachedPool")
      fda_runPar(AQMRPTStream.take(100000).toPar(getIdsThenInsertAction))(8)
        .appendTask(runInsertAction)
        .startRun

      //println(s"processing 10000 rows parallelly  in ${(System.currentTimeMillis - cnt_start)/1000} seconds")
      // processing 10000 rows parallelly  in 316 seconds
      //println(s"processing 20000 rows parallelly  in ${(System.currentTimeMillis - cnt_start)/1000} seconds")
      //processing 20000 rows parallelly  in 614 seconds
      println(s"processing 100000 rows parallelly  in ${(System.currentTimeMillis - cnt_start)/1000} seconds")
      //processing 100000 rows parallelly  in 3885 seconds

}

本文参与腾讯云自媒体分享计划,欢迎正在阅读的你也加入,一起分享。

发表于

我来说两句

0 条评论
登录 后参与评论

相关文章

来自专栏hbbliyong

C#基础知识回顾---你不知道的Lazy<T>

    对象的创建方式,始终代表了软件工业的生产力方向,代表了先进软件技术发展的方向,也代表了广大程序开发者的集体智慧。以new的方式创建,通过工厂方法,利用I...

2603
来自专栏小勇DW3

承接上文

前置通知[Before advice]:在连接点前面执行,前置通知不会影响连接点的执行,除非此处抛出异常。 

722
来自专栏程序员宝库

徒手撸框架---实现 Aop

原文:犀利豆的博客(https://www.xilidou.com/2018/01/13/spring-aop/) 上一讲我们讲解了 Spring 的 IoC ...

28312
来自专栏CodingToDie

FastSql ORM 实现

FastSql 中 ORM 的实现 Table of Contents 原理 实现 1. 使用注解 2. 反射工具类 3. 简单的 model 4. 注解解析 ...

4116
来自专栏函数式编程语言及工具

FunDA(13)- 示范:用户自定义操作函数 - user defined tasks

   FunDA是一种函数式的编程工具,它所产生的程序是由许多功能单一的细小函数组合而成,这些函数就是用户自定义操作函数了。我们在前面曾经提过FunDA的运作原...

1738
来自专栏码匠的流水账

聊聊sentinel的SentinelResourceAspect

com/alibaba/csp/sentinel/annotation/aspectj/SentinelResourceAspect.java

502
来自专栏函数式编程语言及工具

Scalaz(50)- scalaz-stream: 安全的无穷运算-running infinite stream freely

scalaz-stream支持无穷数据流(infinite stream),这本身是它强大的功能之一,试想有多少系统需要通过无穷运算才能得以实现。这是因为外界...

1766
来自专栏Java架构

BeanPostProcessor——连接Spring IOC和AOP的桥梁

1465
来自专栏函数式编程语言及工具

泛函编程(27)-泛函编程模式-Monad Transformer

    经过了一段时间的学习,我们了解了一系列泛函数据类型。我们知道,在所有编程语言中,数据类型是支持软件编程的基础。同样,泛函数据类型Foldable,Mon...

1907
来自专栏程序猿DD

Spring框架中的设计模式(四)​

本文是Spring框架中使用的设计模式第四篇。本文将在此呈现出新的3种模式。一开始,我们会讨论2种结构模式:适配器和装饰器。在第三部分和最后一部分,我们将讨论单...

3386

扫码关注云+社区