赞
踩
原文地址:https://blog.csdn.net/zhanglh046/article/details/78485783
Master是通过schedule方法进行资源调度,告知worker启动executor等。
- private def schedule(): Unit = {
- // 只有alive状态的master才可以进行资源调度,standby是不能够调度的
- if (state != RecoveryState.ALIVE) {
- return
- }
- // 将可用的worker节点打乱,这样有利于driver的均衡
- val shuffledAliveWorkers = Random.shuffle(workers.toSeq.filter(_.state == WorkerState.ALIVE))
- val numWorkersAlive = shuffledAliveWorkers.size
- var curPos = 0
- // 进行driver资源调度,遍历处于等待状态的driver队列
- for (driver <- waitingDrivers.toList) {
- var launched = false
- var numWorkersVisited= 0
- while (numWorkersVisited < numWorkersAlive && !launched) {
- // 获取worker
- val worker = shuffledAliveWorkers(curPos)
- // 记录worker访问数递增
- numWorkersVisited+= 1
- // 判断worker的可使用内存是否大于driver所需要的内存以及worker可使用cpu核数是否大于driver所需要的cpu核数
- if (worker.memoryFree >= driver.desc.mem && worker.coresFree >= driver.desc.cores) {
- // 满足条件发起driver
- launchDriver(worker, driver)
- // 将当前driver从等待队列中移除
- waitingDrivers-= driver
- // 标记该driver发起状态为true
- launched = true
- }
- // 将指针指向下一个worker,当然如果driver已经发起了,则为下一个准备发起下一个处于等待的driver
- curPos = (curPos + 1) % numWorkersAlive
- }
- }
- // 在worker上开启executor进程
- startExecutorsOnWorkers()
- }
每一个executor所需要的核数是可以配置的,一般来讲如果worker有足够的内存和CPU核数,同一个应用程序就可以在该worker启动多个executors;否则就不能再启动新的executor了,则需要到其他worker上去分配executor了
- private def startExecutorsOnWorkers(): Unit = {
- // 遍历处于等待状态的application,且处于等待的状态的application的所需要的cpu核数大于0
- // coresLeft=app请求的核数-已经分配给executor的核数的和
- for (app <- waitingApps if app.coresLeft > 0) {
- // 每一个executor所需要的核数
- val coresPerExecutor: Option[Int] = app.desc.coresPerExecutor
- // 过滤出有效的可用worker
- // 再从worker中过滤出worker剩余内存和CPU核数不小于app对应executor所需要的内存和CPU核数
- // 按照剩余的CPU核数反向排序woker
- val usableWorkers = workers.toArray.filter(_.state == WorkerState.ALIVE)
- .filter(worker => worker.memoryFree >= app.desc.memoryPerExecutorMB &&
- worker.coresFree >= coresPerExecutor.getOrElse(1))
- .sortBy(_.coresFree).reverse
- // 在可用的worker上调度executor,启动executor有两种算法模式:
- // 一:将应用程序尽可能多的分配到不同的worker上
- // 二:和第一种相反,分配到尽可能少的worker上,通常用于计算密集型;
- // 每一个executor所需要的核数是可以配置的,一般来讲如果worker有足够的内存和CPU核数,同一个应用程序就可以
- // 在该worker启动多个executors;否则就不能再启动新的executor了,则需要到其他worker上去分配executor了
- val assignedCores = scheduleExecutorsOnWorkers(app, usableWorkers, spreadOutApps)
-
- // 在可用的worker上分配资源给executor
- for (pos <- 0 until usableWorkers.length if assignedCores(pos) > 0) {
- allocateWorkerResourceToExecutors(
- app, assignedCores(pos), coresPerExecutor, usableWorkers(pos))
- }
- }
- }
判断该worker能不能分配一个或者多个executor,能则分配相对应的executor所需要的CPU核数
- private def scheduleExecutorsOnWorkers(app: ApplicationInfo,
- usableWorkers: Array[WorkerInfo], spreadOutApps: Boolean): Array[Int] = {
- // 如果我们指定executor需要分配的核数,coresPerExecutor表示executor所需要的cpu核数
- val coresPerExecutor = app.desc.coresPerExecutor
- // app中每个executor所需要的最小cpu核数,如果没有默认最小核数为1
- val minCoresPerExecutor = coresPerExecutor.getOrElse(1)
- // 如果我们没有指定executor需要分配的核数,则一个worker上只能启动一个executor
- val oneExecutorPerWorker = coresPerExecutor.isEmpty
- // 每一个executor所需要的内存
- val memoryPerExecutor = app.desc.memoryPerExecutorMB
- // 获取可用worker数量
- val numUsable = usableWorkers.length
- // 构建一个可用worker长度的数组,用于存放每个worker节点分配到的cpu核数(16,16,16,16)
- val assignedCores = new Array[Int](numUsable)
- // 构建一个可用worker长度的数组,用于存放每一个worker上新分配的executor数量(1,2,1,0)
- val assignedExecutors = new Array[Int](numUsable)
- // 针对当前应用程序,还需要分配的cpu核数,它应该是application还需要的cpu核数和worker总共剩余核数之和中最小的
- // 防止超过当前可用的cpu核数
- var coresToAssign = math.min(app.coresLeft, usableWorkers.map(_.coresFree).sum)
-
- // 判断我们是否可以为这个application在指定的worker上发起一个executor
- def canLaunchExecutor(pos: Int): Boolean = {
- // 判断当前需要分配的cpu核数是否大于或者等于每个executor所需要的cpu核数,比如总共只能分配8核,但是
- // 每个executor所需要的cpu核数是12,那么就不能发起executor了,因为资源不够用
- val keepScheduling = coresToAssign >= minCoresPerExecutor
- // 当前worker剩余的核数 - 应用程序分配到该worker上的核数是否满足发起一个executor,比如现在worker剩余核数16
- // 然后又给application他分配了12核,即还剩4核可用,但是启动一个executor需要12核,那么4 < 12 表示内核不足使用了
- val enoughCores = usableWorkers(pos).coresFree - assignedCores(pos) >= minCoresPerExecutor
-
- // 如果我们允许每一个worker启动多个executor,然后我们可以启动一个新的executor
- // 否则如果worker已经启动一个新executor,只需要将更多的内核分配给该executor即可
- val launchingNewExecutor = !oneExecutorPerWorker || assignedExecutors(pos) == 0
- // 如果需要发起新的executor,既需要判断cpu核数是否足够,还需要判断 executor是否超过限制总数以及否内存是否足够
- if (launchingNewExecutor) {
- 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 {
- // 否则只是对已经存在的executor添加cpu核数,没必要检查内存和executor限制
- keepScheduling && enoughCores
- }
- }
- // 过滤出那些可用的worker节点
- var freeWorkers = (0 until numUsable).filter(canLaunchExecutor)
- while (freeWorkers.nonEmpty) {
- // 遍历每一个空闲的worker
- freeWorkers.foreach { pos =>
- var keepScheduling = true
- // 检测当前worker是否能够发起executor
- while (keepScheduling && canLaunchExecutor(pos)) {
- // 需要分配的核数减去每个executor所需要的最小核数
- coresToAssign -= minCoresPerExecutor
- // 对应的worker节点需要分配的cpu核数加上要启动该executor所需要的最小CPU核数
- assignedCores(pos) += minCoresPerExecutor
- // 如果每一个worker只允许启动一个executor,那么该worker启动的executor数量只能是1,否则应该加一个
- if (oneExecutorPerWorker) {
- assignedExecutors(pos) = 1
- } else {
- assignedExecutors(pos) += 1
- }
-
- // 如果需要将executor分配到更多的worker,那么就不再从当前worker节点继续分配,而是从下一个worker上继续分配
- if (spreadOutApps) {
- keepScheduling = false
- }
- }
- }
- // 因为进行了一次分配,需要再次从可用的worker节点中过滤可用的worker节点
- freeWorkers = freeWorkers.filter(canLaunchExecutor)
- }
- assignedCores
- }
- private def allocateWorkerResourceToExecutors(app: ApplicationInfo, assignedCores: Int,
- coresPerExecutor: Option[Int], worker: WorkerInfo): Unit = {
- // 获取该worker应该有多少个executor
- val numExecutors = coresPerExecutor.map { assignedCores / _ }.getOrElse(1)
- // 获取每一个executor应该分配的核数,如果没有指定则使用计算的应该分配的核数
- val coresToAssign = coresPerExecutor.getOrElse(assignedCores)
- for (i <- 1 to numExecutors) {
- // 向worker上添加executor,创建ExecutorDesc对象,更新application已经分配到的cpu核数
- val exec = app.addExecutor(worker, coresToAssign)
- // 启动executor
- launchExecutor(worker, exec)
- // 更新application的状态
- app.state = ApplicationState.RUNNING
- }
- }
- private def launchDriver(worker: WorkerInfo, driver: DriverInfo) {
- logInfo("Launching driver " + driver.id + " on worker " + worker.id)
- // worker添加driver
- worker.addDriver(driver)
- driver.worker = Some(worker)
- // 向worker发送LaunchDriver消息
- worker.endpoint.send(LaunchDriver(driver.id, driver.desc))
- // 更新driver状态为RUNNING
- driver.state = DriverState.RUNNING
- }
- private def launchExecutor(worker: WorkerInfo, exec: ExecutorDesc): Unit = {
- logInfo("Launching executor " + exec.fullId + " on worker " + worker.id)
- // worker启动executor,并且更新worker的cpu和内存信息
- worker.addExecutor(exec)
- // 向worker发送LaunchExecutor消息
- worker.endpoint.send(LaunchExecutor(masterUrl,
- exec.application.id, exec.id, exec.application.desc, exec.cores, exec.memory))
- // 向application发送ExecutorAdded消息
- exec.application.driver.send(
- ExecutorAdded(exec.id, worker.id, worker.hostPort, exec.cores, exec.memory))
- }
Copyright © 2003-2013 www.wpsshop.cn 版权所有,并保留所有权利。