A股上市公司传智教育(股票代码 003032)旗下技术交流社区北京昌平校区

 找回密码
 加入黑马

QQ登录

只需一步,快速开始

spark 各个版本的application 调度算法还是有这明显的不同之处的。从spark1.3.0 到 spark 1.6.1、spark2.0 到 现在最新的spark 3.0 ,调度算法有了一定的修改。下面大家一起学习一下,最新的spark 版本spark-3.0的Application 调度机制。

private def startExecutorsOnWorkers(): Unit = {
  // Right now this is a very simple FIFO scheduler. We keep trying to fit in the first app
  // in the queue, then the second app, etc.
  for (app <- waitingApps) {
    //如果在 spark-submmit 脚本中,指定了每个executor 多少个 CPU core,
    // 则每个Executor 分配该个数的 core,
    // 否则 默认每个executor 只分配 1 个 CPU core
    val coresPerExecutor = app.desc.coresPerExecutor.getOrElse(1)
    // If the cores left is less than the coresPerExecutor,the cores left will not be allocated
    //  当前 APP 还需要分配的  core  数 不能  小于 单个 executor 启动 的 CPU core 数
    if (app.coresLeft >= coresPerExecutor) {
      // Filter out workers that don't have enough resources to launch an executo/*ku*/r
      // 过滤出 状态 为 ALIVE,并且还能 发布 Executor 的 worker
      // 按照剩余的 CPU core 数  倒序
      val usableWorkers = workers.toArray.filter(_.state == WorkerState.ALIVE)
        .filter(canLaunchExecutor(_, app.desc))
        .sortBy(_.coresFree).reverse
      if (waitingApps.length == 1 && usableWorkers.isEmpty) {
        logWarning(s"App ${app.id} requires more resource than any of Workers could have.")
      }
    // TODO:  默认采用 spreadOutApps  调度算法, 将 application需要的 executor资源 分派到  多个 worker 上去
      val assignedCores = scheduleExecutorsOnWorkers(app, usableWorkers, spreadOutApps)

      // Now that we've decided how many cores to allocate on each worker, let's allocate them
      for (pos <- 0 until usableWorkers.length if assignedCores(pos) > 0) {
        allocateWorkerResourceToExecutors(
          app, assignedCores(pos), app.desc.coresPerExecutor, usableWorkers(pos))
      }
    }
  }
}
判断一个 worker 是否可以发布 executor
private def canLaunchExecutor(worker: WorkerInfo, desc: ApplicationDescription): Boolean = {
  canLaunch(
    worker,
    desc.memoryPerExecutorMB,
    desc.coresPerExecutor.getOrElse(1),
    desc.resourceReqsPerExecutor)
}
让我们看一看里面的 canlaunch 方法
private def canLaunch(
    worker: WorkerInfo,
    memoryReq: Int,
    coresReq: Int,
    resourceRequirements: Seq[ResourceRequirement])
  : Boolean = {
  // worker 上 空闲的 内存值  要 大于等于  请求的 内存值
  val enoughMem = worker.memoryFree >= memoryReq
  // worker 上 空闲的 core 数  要 大于等于  请求的 core数
  val enoughCores = worker.coresFree >= coresReq
  //  worker 是否满足 executor 请求的资源   
  val enoughResources = ResourceUtils.resourcesMeetRequirements(
    worker.resourcesAmountFree, resourceRequirements)
  enoughMem && enoughCores && enoughResources
}

回到上面的 scheduleExecutorsOnWorkers
private def scheduleExecutorsOnWorkers(
    app: ApplicationInfo,
    usableWorkers: Array[WorkerInfo],
    spreadOutApps: Boolean): Array[Int] = {
  val coresPerExecutor = app.desc.coresPerExecutor
  val minCoresPerExecutor = coresPerExecutor.getOrElse(1)
  // 默认情况下 是 开启  oneExecutorPerWorker 机制的,也就是默认是在 一个 worker 上  只启动 一个 executor的
  //  如果在spark -submit 脚本中设置了coresPerExecutor , 在worker资源充足的时候,则 会在每个worker 上,启动多个executor
  val oneExecutorPerWorker = coresPerExecutor.isEmpty
  val memoryPerExecutor = app.desc.memoryPerExecutorMB
  val resourceReqsPerExecutor = app.desc.resourceReqsPerExecutor
  val numUsable = usableWorkers.length
  val assignedCores = new Array[Int](numUsable) // Number of cores to give to each worker
  val assignedExecutors = new Array[Int](numUsable) // Number of new executors on each worker
  var coresToAssign = math.min(app.coresLeft, usableWorkers.map(_.coresFree).sum)

// 判断  Worker节点是否能够启动Executor
  def canLaunchExecutorForApp(pos: Int): Boolean = {

    val keepScheduling = coresToAssign >= minCoresPerExecutor
    val enoughCores = usableWorkers(pos).coresFree - assignedCores(pos) >= minCoresPerExecutor
    val assignedExecutorNum = assignedExecutors(pos)

    // If we allow multiple executors per worker, then we can always launch new executors.
    // Otherwise, if there is already an executor on this worker, just give it more cores.

    // 如果spark -submit 脚本中设置了coresPerExecutor值,
    // 或者当前 这个worker 还没有为这个 application 分配 过  executor ,
    val launchingNewExecutor = !oneExecutorPerWorker || assignedExecutorNum == 0
      // TODO:  可以启动新的 Executor
    if (launchingNewExecutor) {
      val assignedMemory = assignedExecutorNum * memoryPerExecutor
      val enoughMemory = usableWorkers(pos).memoryFree - assignedMemory >= memoryPerExecutor
      val assignedResources = resourceReqsPerExecutor.map {
        req => req.resourceName -> req.amount * assignedExecutorNum
      }.toMap
      val resourcesFree = usableWorkers(pos).resourcesAmountFree.map {
        case (rName, free) => rName -> (free - assignedResources.getOrElse(rName, 0))
      }
      val enoughResources = ResourceUtils.resourcesMeetRequirements(
        resourcesFree, resourceReqsPerExecutor)
      val underLimit = assignedExecutors.sum + app.executors.size < app.executorLimit
      keepScheduling && enoughCores && enoughMemory && enoughResources && underLimit
    } else {
      // We're adding cores to an existing executor, so no need
      // to check memory and executor limits
      // TODO:  不满足启动新的 Executor条件,则 在 老的 Executor 上 追加  core 数
      keepScheduling && enoughCores
    }
  }

  // Keep launching executors until no more workers can accommodate any
  // more executors, or if we have reached this application's limits

  var freeWorkers = (0 until numUsable).filter(canLaunchExecutorForApp)
  while (freeWorkers.nonEmpty) {
    freeWorkers.foreach { pos =>
      var keepScheduling = true
      while (keepScheduling && canLaunchExecutorForApp(pos)) {
        coresToAssign -= minCoresPerExecutor
        assignedCores(pos) += minCoresPerExecutor

        // If we are launching one executor per worker, then every iteration assigns 1 core
        // to the executor. Otherwise, every iteration assigns cores to a new executor.
        if (oneExecutorPerWorker) {
          //TODO: 如果该Worker节点不能启动新的 Executor,则每次在老的executor 上 分配 minCoresPerExecutor 个 CPU core(此时该值默认 为 1 )
          assignedExecutors(pos) = 1
        } else {
          //TODO: 如果该Worker节点可以启动新的 Executor,则每次在新的executor 上 分配 minCoresPerExecutor 个 CPU core(此时该值为 spark-submit脚本配置的 coresPerExecutor 值)
          assignedExecutors(pos) += 1
        }

        // Spreading out an application means spreading out its executors across as
        // many workers as possible. If we are not spreading out, then we should keep
        // scheduling executors on this worker until we use all of its resources.
        // Otherwise, just move on to the next worker.
        if (spreadOutApps) {
          // TODO: 这里传入 keepScheduling = false , 就是每次 worker上只分配 一次 core ,然后 到 下一个 worker 上  再去 分配 core,直到 worker
          // TODO:  完成一次遍历
          keepScheduling = false
        }
      }
    }
    freeWorkers = freeWorkers.filter(canLaunchExecutorForApp)
  }
  // 返回每个Worker节点分配的CPU核数
  assignedCores
}

再来分析 allocateWorkerResourceToExecutors
private def allocateWorkerResourceToExecutors(
    app: ApplicationInfo,
    assignedCores: Int,
    coresPerExecutor: Option[Int],
    worker: WorkerInfo): Unit = {
  // If the number of cores per executor is specified, we divide the cores assigned
  // to this worker evenly among the executors with no remainder.
  // Otherwise, we launch a single executor that grabs all the assignedCores on this worker.
  val numExecutors = coresPerExecutor.map { assignedCores / _ }.getOrElse(1)
  val coresToAssign = coresPerExecutor.getOrElse(assignedCores)
  for (i <- 1 to numExecutors) {
    val allocated = worker.acquireResources(app.desc.resourceReqsPerExecutor)
    // TODO : 当前 这个 application 追加 一次  Executor
    val exec = app.addExecutor(worker, coresToAssign, allocated)
    //TODO: 给worker 线程 发送 launchExecutor 命令
    launchExecutor(worker, exec)
    app.state = ApplicationState.RUNNING
  }

2 个回复

倒序浏览
任何问题欢迎在评论区留言
回复 使用道具 举报
或者添加学姐维系
DKA-2018
回复 使用道具 举报
您需要登录后才可以回帖 登录 | 加入黑马