Spark Application的调度算法

要想明白spark application调度机制，需要回答一下几个问题：

1.谁来调度？

2.为谁调度？

3.调度什么？

3.何时调度？

4.调度算法

前四个问题可以用如下一句话里来回答:每当集群资源发生变化时，active master 进程 为 所有已注册的并且没有调度完毕的application调度Worker节点上的Executor进程。

"active master" , spark集群可能有多个master，但是只有一个active master 参与调度，standby master不参与调度。

集群资源发生变化是什么意思呢？这里的集群资源指的主要是cores的变化，注册/移除Executor进程使得集群的freeCores变多/变少，添加/移除Worker节点使得集群的freeCores变多/变少... ...，所有导致集群资源发生变化的操作，都会调用schedule()重新为application和driver进行资源调度。

spark提供了两种资源调度算法:spreadOut和非spreadOut。spreadOut算法会尽可能的将一个application 所需要的Executor进程分布在多个worker几点上，从而提高并行度，非spreadOut与之相反，他会把一个worker节点的freeCores都耗尽了才会去下一个worker节点分配。

为了详细说明这两种算法，我们先来以一个具体的例子来介绍，最后再介绍源码。

基本概念

每一个application至少包含以下基本属性:

coresPerExecutor：每一个Executor进程的core个数

memoryPerExecutor：每一个Executor进程的memory大小

maxCores: 这个application最多需要的core个数。

每一个worker至少包含以下基本属性：

freeCores:worker 节点当前可用的core个数

memoryFree:worker节点当前可用的memory大小。

假设一个待注册的application如下：

coresPerExecutor：2

memoryPerExecutor：512M

maxCores: 12

这表示这个application 最多需要12个core，每一个Executor进行都要2个core，512M内存。

假设某一时刻spark集群有如下几个worker节点,他们按照coresFree降序排列:

Worker1:coresFree=10  memoryFree=10G

Worker2:coresFree=7   memoryFree=1G

Worker3:coresFree=3   memoryFree=2G

Worker4:coresFree=2   memoryFree=215M

Worker5:coresFree=1   memoryFree=1G

其中worker5不满足application的要求：worker5.coresFree < application.coresPerExecutor

worker4也不满足application的要求:worker4.memoryFree < application.memoryPerExecutor

因此最终满足调度要求的worker节点只有前三个，我们将这三个节点记作usableWorkers。

spreadOut算法

先介绍spreadOut算法吧。上面已经说了，满足条件的worker只有前三个:

Worker1:coresFree=10  memoryFree=10G

Worker2:coresFree=7   memoryFree=1G

Worker3:coresFree=3   memoryFree=2G

第一次调度之后，worker列表如下:

Worker1:coresFree=8  memoryFree=9.5G  assignedExecutors=1  assignedCores=2

Worker2:coresFree=7   memoryFree=1G    assignedExecutors=0  assignedCores=0

Worker3:coresFree=3   memoryFree=2G    assignedExecutors=0  assignedCores=0

totalExecutors:1,totalCores=2

可以发现，worker1的coresFree和memoryFree都变小了而worker2，worker3并没有发生改变，这是因为我们在worker1上面分配了一个Executor进程(这个Executor进程占用两个2core，512M memory)而没有在workre2和worker3上分配。

接下来，开始去worker2上分配:

Worker1:coresFree=8  memoryFree=9.5G      assignedExecutors=1  assignedCores=2

Worker2:coresFree=5   memoryFree=512M    assignedExecutors=1  assignedCores=2

Worker3:coresFree=3   memoryFree=2G        assignedExecutors=0  assignedCores=0

totalExecutors:2,totalCores=4

此时已经分配了2个Executor进程，4个core。

接下来去worker3上分配:

Worker1:coresFree=8  memoryFree=9.5G      assignedExecutors=1  assignedCores=2

Worker2:coresFree=5   memoryFree=512M    assignedExecutors=1  assignedCores=2

Worker3:coresFree=1   memoryFree=1.5G     assignedExecutors=1  assignedCores=2

totalExecutors:3,totalCores=6

接下来再去worker1分配，然后worker2...  ...以round-robin方式分配，由于worker3.coresFree<application.coresPerExecutor,不会在他上面分配资源了：

Worker1:coresFree=6  memoryFree=9.0G      assignedExecutors=2  assignedCores=4

Worker2:coresFree=5   memoryFree=512M    assignedExecutors=1  assignedCores=2

Worker3:coresFree=1   memoryFree=1.5G     assignedExecutors=1  assignedCores=2

totalExecutors:4,totalCores=8

Worker1:coresFree=6  memoryFree=9.0G      assignedExecutors=2  assignedCores=4

Worker2:coresFree=3   memoryFree=0M       assignedExecutors=2  assignedCores=4

Worker3:coresFree=1   memoryFree=1.5G     assignedExecutors=1  assignedCores=2

totalExecutors:5,totalCores=10

此时worker2也不满足要求了：worker2.memoryFree<application.memoryPerExecutor

因此，下一次分配就去worker1上了：

Worker1:coresFree=4  memoryFree=8.5G      assignedExecutors=3  assignedCores=6

Worker2:coresFree=3   memoryFree=0M        assignedExecutors=2  assignedCores=4

Worker3:coresFree=1   memoryFree=1.5G     assignedExecutors=1  assignedCores=2

totalExecutors:6,totalCores=12

ok，由于已经分配了12个core，达到了application的要求，所以不在为这个application调度了。

非spreadOUt算法

那么非spraadOut算法呢？他是逮到一个worker如果不把他的资源耗尽了是不会放手的：

Worker1:coresFree=8  memoryFree=9.5G  assignedExecutors=1  assignedCores=2

Worker2:coresFree=7   memoryFree=1G    assignedExecutors=0  assignedCores=0

Worker3:coresFree=3   memoryFree=2G    assignedExecutors=0  assignedCores=0

totalExecutors:1,totalCores=2

Worker1:coresFree=6  memoryFree=9.0G  assignedExecutors=2  assignedCores=4

Worker2:coresFree=7   memoryFree=1G    assignedExecutors=0  assignedCores=0

Worker3:coresFree=3   memoryFree=2G    assignedExecutors=0  assignedCores=0

totalExecutors:2,totalCores=4

Worker1:coresFree=4  memoryFree=8.5    assignedExecutors=3  assignedCores=6

Worker2:coresFree=7   memoryFree=1G    assignedExecutors=0  assignedCores=0

Worker3:coresFree=3   memoryFree=2G    assignedExecutors=0  assignedCores=0

totalExecutors:3,totalCores=6

Worker1:coresFree=2  memoryFree=8.0G  assignedExecutors=4  assignedCores=8

Worker2:coresFree=7   memoryFree=1G    assignedExecutors=0  assignedCores=0

Worker3:coresFree=3   memoryFree=2G    assignedExecutors=0  assignedCores=0

totalExecutors:4,totalCores=8

Worker1:coresFree=0  memoryFree=7.5G  assignedExecutors=5  assignedCores=10

Worker2:coresFree=7   memoryFree=1G    assignedExecutors=0  assignedCores=0

Worker3:coresFree=3   memoryFree=2G    assignedExecutors=0  assignedCores=0

totalExecutors:5,totalCores=10

可以发现，worker1的coresfree已经耗尽了，好可怜。由于application需要12个core，而这里才分配了10个，所以还要继续往下分配：

Worker1:coresFree=0  memoryFree=7.5G      assignedExecutors=5  assignedCores=10

Worker2:coresFree=5   memoryFree=512G    assignedExecutors=1  assignedCores=2

Worker3:coresFree=3   memoryFree=2G        assignedExecutors=0  assignedCores=0

totalExecutors:6,totalCores=12

ok，最终分配来12个core，满足了application的要求。

对比：

spreadOut算法中，是以round-robin方式，轮询的在worker节点上分配Executor进程，即以如下序列分配:worker1,worker2... ... worker n,worker1... ....

非spreadOut算法中，逮者一个worker就不放手，直到满足一下条件之一：

worker.freeCores<application.coresPerExecutor 或者  worker.memoryFree<application.memoryPerExecutor 。

在上面两个例子中，虽然最终都分配了6个Executor进程和12个core，但是spreadOut方式下，6个Executor进程分散在不同的worker节点上，充分利用了spark集群的worker节点，而非spreadOut方式下，只在worker1和worker2上分配了Executor进程，并没有充分利用spark worker节点。

小插曲，spreadOut + oneExecutorPerWorker 算法

spark还有一个叫做”oneExecutorPerWorker“机制，即一个worker上启动一个Executor进程，下面只是简单的说一下得了：

Worker1:coresFree=8  memoryFree=9.5G  assignedExecutors=1  assignedCores=2

Worker2:coresFree=7   memoryFree=1G    assignedExecutors=0  assignedCores=0

Worker3:coresFree=3   memoryFree=2G    assignedExecutors=0  assignedCores=0

totalExecutors:1,totalCores=2

Worker1:coresFree=8  memoryFree=9.5G      assignedExecutors=1  assignedCores=2

Worker2:coresFree=5   memoryFree=512M    assignedExecutors=1  assignedCores=2

Worker3:coresFree=3   memoryFree=2G        assignedExecutors=0  assignedCores=0

totalExecutors:2,totalCores=4

Worker1:coresFree=8  memoryFree=9.5G      assignedExecutors=1  assignedCores=2

Worker2:coresFree=5   memoryFree=512M    assignedExecutors=1  assignedCores=2

Worker3:coresFree=1   memoryFree=1.5G     assignedExecutors=1  assignedCores=2

totalExecutors:3,totalCores=6

Worker1:coresFree=6  memoryFree=9.0G      assignedExecutors=1  assignedCores=4

Worker2:coresFree=3   memoryFree=512M     assignedExecutors=1  assignedCores=2

Worker3:coresFree=1   memoryFree=1.5G     assignedExecutors=1  assignedCores=2

totalExecutors:3,totalCores=8

Worker1:coresFree=6  memoryFree=9.0G      assignedExecutors=1  assignedCores=4

Worker2:coresFree=2   memoryFree=0   M     assignedExecutors=1  assignedCores=4

Worker3:coresFree=1   memoryFree=1.5G     assignedExecutors=1  assignedCores=2

totalExecutors:3,totalCores=10

Worker1:coresFree=4  memoryFree=9.5G      assignedExecutors=1  assignedCores=6

Worker2:coresFree=2   memoryFree=0   M     assignedExecutors=1  assignedCores=4

Worker3:coresFree=1   memoryFree=1.5G     assignedExecutors=1  assignedCores=2

totalExecutors:3,totalCores=12

和spreadOut+非oneExecutorPerWorker对比发现，唯一的不同就是Executor进程的数量，一个是6，一个是3。

(

这里在额外扩展一下，假设application的maxCores=14，而不是12，那么接着上面那个worker列表来：

Worker1:coresFree=4  memoryFree=9.5G      assignedExecutors=1  assignedCores=6

Worker2:coresFree=0   memoryFree=0   M     assignedExecutors=1  assignedCores=6

Worker3:coresFree=1   memoryFree=1.5G     assignedExecutors=1  assignedCores=2

totalExecutors:3,totalCores=12

虽然worker2.memoryFree=0，但是仍然可以继续在他上面分配core，因为onExecutorPerWorker机制不检查内存的限制。

)

接下来看看源码是怎么实现的：

了解了上面写的，在阅读源码就很轻易了，这里简单说一下。

org.apache.spark.deploy.master.Master收到application发送的RegisterApplication(description, driver)消息后，开始执行注册逻辑:

    case RegisterApplication(description, driver) => {
      // TODO Prevent repeated registrations from some driver
      //standby master不调度
      if (state == RecoveryState.STANDBY) {
        // ignore, don't send response
      } else {
        logInfo("Registering app " + description.name)
        val app = createApplication(description, driver)
        //注册app，即将其加入到waitingApps中
        registerApplication(app)
        logInfo("Registered app " + description.name + " with ID " + app.id)
        //将app加入持久化引擎，主要是为了故障恢复
        persistenceEngine.addApplication(app)
        //向driver发送RegisteredApplication消息表明master已经注册了这个app
        driver.send(RegisteredApplication(app.id, self))
        //为waitingApps中的app调度资源
        schedule()
      }
    }

上面的注释已经写的很清楚了... ...

  /**
   * Schedule the currently available resources among waiting apps. This method will be called
   * every time a new app joins or resource availability changes.
   */
  private def schedule(): Unit = {
    if (state != RecoveryState.ALIVE) { return }
    // Drivers take strict precedence over executors
    //为了避免每次schedule，总是在相同的worker上分配资源，所有这里打乱worker顺序。
    val shuffledWorkers = Random.shuffle(workers) // Randomization helps balance drivers
    //下面这个for循环是为driver调度资源，因为这里只将application的调度，所以driver的调度不说了。
    for (worker <- shuffledWorkers if worker.state == WorkerState.ALIVE) {
      for (driver <- waitingDrivers) {
        if (worker.memoryFree >= driver.desc.mem && worker.coresFree >= driver.desc.cores) {
          launchDriver(worker, driver)
          waitingDrivers -= driver
        }
      }
    }

    //为application调度资源
    startExecutorsOnWorkers()
  }

  /**
   * Schedule and launch executors on workers
   */
  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.
    // 为waitingApps中的app调度资源，app.coresLeft是app还有多少core没有分配
    for (app <- waitingApps if app.coresLeft > 0) {
      val coresPerExecutor: Option[Int] = app.desc.coresPerExecutor
      // Filter out workers that don't have enough resources to launch an executor
      // 筛选出状态为ALIVE并且这个worker剩余内存，剩余core都大于等于app的要求，然后按照coresFree降序排列
      val usableWorkers = workers.toArray.filter(_.state == WorkerState.ALIVE)
        .filter(worker => worker.memoryFree >= app.desc.memoryPerExecutorMB &&
          worker.coresFree >= coresPerExecutor.getOrElse(1))
        .sortBy(_.coresFree).reverse
      //在usableWorkers上为app分配Executor
      val assignedCores = scheduleExecutorsOnWorkers(app, usableWorkers, spreadOutApps)

      // Now that we've decided how many cores to allocate on each worker, let's allocate them
      // 在worker上启动Executor进程
      for (pos <- 0 until usableWorkers.length if assignedCores(pos) > 0) {
        allocateWorkerResourceToExecutors(
          app, assignedCores(pos), coresPerExecutor, usableWorkers(pos))
      }
    }
  }

这个方法做了如下事情：

1.筛选出可用的worker，即usableWorkers，如果一个worker满足以下所有条件，那么这个worker就被添加到usableWorkers中：

Alive

worker.memoryFree >= app.desc.memoryPerExecutorMB

worker.coresFree >= coresPerExecutor

2.assignedCores是一个数组，assignedCores[i]里面存储了需要在usableWorkers[i]上分配的core个数，譬如如果assingedCores[1]=2，那么就需要在usableWorkers[1]上分配2个core。

  /**
   * Schedule executors to be launched on the workers.
   * Returns an array containing number of cores assigned to each worker.
   *
   * There are two modes of launching executors. The first attempts to spread out an application's
   * executors on as many workers as possible, while the second does the opposite (i.e. launch them
   * on as few workers as possible). The former is usually better for data locality purposes and is
   * the default.
   *
   * The number of cores assigned to each executor is configurable. When this is explicitly set,
   * multiple executors from the same application may be launched on the same worker if the worker
   * has enough cores and memory. Otherwise, each executor grabs all the cores available on the
   * worker by default, in which case only one executor may be launched on each worker.
   *
   * It is important to allocate coresPerExecutor on each worker at a time (instead of 1 core
   * at a time). Consider the following example: cluster has 4 workers with 16 cores each.
   * User requests 3 executors (spark.cores.max = 48, spark.executor.cores = 16). If 1 core is
   * allocated at a time, 12 cores from each worker would be assigned to each executor.
   * Since 12 < 16, no executors would launch [SPARK-8881].
   */
  private def scheduleExecutorsOnWorkers(
      app: ApplicationInfo,
      usableWorkers: Array[WorkerInfo],
      spreadOutApps: Boolean): Array[Int] = {
    val coresPerExecutor = app.desc.coresPerExecutor
    val minCoresPerExecutor = coresPerExecutor.getOrElse(1)
    val oneExecutorPerWorker = coresPerExecutor.isEmpty
    val memoryPerExecutor = app.desc.memoryPerExecutorMB
    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)

    /** Return whether the specified worker can launch an executor for this app. */
    //是否可以在一个worker上分配Executor
    def canLaunchExecutor(pos: Int): Boolean = {
      val keepScheduling = coresToAssign >= minCoresPerExecutor
      val enoughCores = usableWorkers(pos).coresFree - assignedCores(pos) >= minCoresPerExecutor

      // 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.
      val launchingNewExecutor = !oneExecutorPerWorker || assignedExecutors(pos) == 0
      if (launchingNewExecutor) {
        //在不里，需要检查worker的空闲core和内存是否够用
        val assignedMemory = assignedExecutors(pos) * memoryPerExecutor
        val enoughMemory = usableWorkers(pos).memoryFree - assignedMemory >= memoryPerExecutor
        val underLimit = assignedExecutors.sum + app.executors.size < app.executorLimit
        keepScheduling && enoughCores && enoughMemory && underLimit
      } else {
        // We're adding cores to an existing executor, so no need
        // to check memory and executor limits
        //尤其需要注意的是，oneExecutorPerWorker机制下，不检测内存限制，很重要。
        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(canLaunchExecutor)
    while (freeWorkers.nonEmpty) {
      freeWorkers.foreach { pos =>
        var keepScheduling = true
        while (keepScheduling && canLaunchExecutor(pos)) {
          //要分配的cores
          coresToAssign -= minCoresPerExecutor
          //已分配的cores
          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.
          //一个worker只启动一个Executor
          if (oneExecutorPerWorker) {
            assignedExecutors(pos) = 1
          } else {
            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.
          //如果没有开启spreadOUt算法，就一直在一个worker上分配，直到不能再分配为止。
          if (spreadOutApps) {
            keepScheduling = false
          }
        }
      }
      freeWorkers = freeWorkers.filter(canLaunchExecutor)
    }
    assignedCores
  }

  /**
   * Allocate a worker's resources to one or more executors.
   * @param app the info of the application which the executors belong to
   * @param assignedCores number of cores on this worker for this application
   * @param coresPerExecutor number of cores per executor
   * @param worker the worker info
   */
  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.
    //计算要创建多少个Executor进程，默认值是1.

    val numExecutors = coresPerExecutor.map { assignedCores / _ }.getOrElse(1)

    val coresToAssign = coresPerExecutor.getOrElse(assignedCores)
    for (i <- 1 to numExecutors) {
      val exec = app.addExecutor(worker, coresToAssign)
      //真正的启动Executor进程了。
      launchExecutor(worker, exec)
      app.state = ApplicationState.RUNNING
    }
  }

由于本人接触spark时间不长，如有错误或者任何意见可以在留言或者发送邮件到franciswbs@163.com，让我们一起交流。

作者：FrancisWang

邮箱：franciswbs@163.com
出处：http://www.cnblogs.com/francisYoung/
本文地址：http://www.cnblogs.com/francisYoung/
本文版权归作者和博客园共有，欢迎转载，但未经作者同意必须保留此段声明，且在文章页面明显位置给出原文连接，否则保留追究法律责任的权利。