一、引子
在Worker Actor中,每次LaunchExecutor会创建一个CoarseGrainedExecutorBackend进程,Executor和CoarseGrainedExecutorBackend是1对1的关系。也就是说集群里启动多少Executor实例就有多少CoarseGrainedExecutorBackend进程。
那么到底是如何分配Executor的呢?怎么控制调节Executor的个数呢?
二、Driver和Executor资源调度
下面主要介绍一下Spark Executor分配策略:
我们仅看,当Application提交注册到Master后,Master会返回RegisteredApplication,之后便会调用schedule()这个方法,来分配Driver的资源,和启动Executor的资源。
schedule()方法是来调度当前可用资源的调度方法,它管理还在排队等待的Apps资源的分配,这个方法是每次在集群资源发生变动的时候都会调用,根据当前集群最新的资源来进行Apps的资源分配。
Driver资源调度:
- // First schedule drivers, they take strict precedence over applications
- val shuffledWorkers = Random.shuffle(workers) // 把当前workers这个HashSet的顺序随机打乱
- for (worker <- shuffledWorkers if worker.state == WorkerState.ALIVE) { //遍历活着的workers
- for (driver <- waitingDrivers) { //在等待队列中的Driver们会进行资源分配
- if (worker.memoryFree >= driver.desc.mem && worker.coresFree >= driver.desc.cores) { //当前的worker内存和cpu均大于当前driver请求的mem和cpu,则启动
- launchDriver(worker, driver) //启动Driver 内部实现是发送启动Driver命令给指定Worker,Worker来启动Driver。
- waitingDrivers -= driver //把启动过的Driver从队列移除
- }
- }
- }
Executor资源调度:
- val spreadOutApps = conf.getBoolean("spark.deploy.spreadOut", true)
在介绍之前我们先介绍一个概念,
- /**
- * Can an app use the given worker? True if the worker has enough memory and we haven't already
- * launched an executor for the app on it (right now the standalone backend doesn't like having
- * two executors on the same worker).
- */
- def canUse(app: ApplicationInfo, worker: WorkerInfo): Boolean = {
- worker.memoryFree >= app.desc.memoryPerSlave && !worker.hasExecutor(app)
- }
SpreadOut分配策略:
- // 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.
- if (spreadOutApps) {
- // Try to spread out each app among all the nodes, until it has all its cores
- for (app <- waitingApps if app.coresLeft > 0) { //对还未被完全分配资源的apps处理
- val usableWorkers = workers.toArray.filter(_.state == WorkerState.ALIVE)
- .filter(canUse(app, _)).sortBy(_.coresFree).reverse //根据core Free对可用Worker进行降序排序。
- val numUsable = usableWorkers.length //可用worker的个数 eg:可用5个worker
- val assigned = new Array[Int](numUsable) //候选Worker,每个Worker一个下标,是一个数组,初始化默认都是0
- var toAssign = math.min(app.coresLeft, usableWorkers.map(_.coresFree).sum)//还要分配的cores = 集群中可用Worker的可用cores总和(10), 当前未分配core(5)中找最小的
- var pos = 0
- while (toAssign > 0) {
- if (usableWorkers(pos).coresFree - assigned(pos) > 0) { //以round robin方式在所有可用Worker里判断当前worker空闲cpu是否大于当前数组已经分配core值
- toAssign -= 1
- assigned(pos) += 1 //当前下标pos的Worker分配1个core +1
- }
- pos = (pos + 1) % numUsable //round-robin轮询寻找有资源的Worker
- }
- // Now that we've decided how many cores to give on each node, let's actually give them
- for (pos <- 0 until numUsable) {
- if (assigned(pos) > 0) { //如果assigned数组中的值>0,将启动一个executor在,指定下标的机器上。
- val exec = app.addExecutor(usableWorkers(pos), assigned(pos)) //更新app里的Executor信息
- launchExecutor(usableWorkers(pos), exec) //通知可用Worker去启动Executor
- app.state = ApplicationState.RUNNING
- }
- }
- }
- } else {
非SpreadOut分配策略:
- } else {
- // Pack each app into as few nodes as possible until we've assigned all its cores
- for (worker <- workers if worker.coresFree > 0 && worker.state == WorkerState.ALIVE) {
- for (app <- waitingApps if app.coresLeft > 0) {
- if (canUse(app, worker)) { //直接问当前worker是有空闲的core
- val coresToUse = math.min(worker.coresFree, app.coresLeft) //有则取,不管多少
- if (coresToUse > 0) { //有
- val exec = app.addExecutor(worker, coresToUse) //直接启动
- launchExecutor(worker, exec)
- app.state = ApplicationState.RUNNING
- }
- }
- }
- }
- }
- }
三、总结:
2、针对同一个App,每个Worker里只能有一个针对该App的Executor存在,切记。如果想让整个App的Executor变多,设置SPARK_WORKER_INSTANCES,让Worker变多。
3、Executor的资源分配有2种策略:
3.1、SpreadOut :一种以round-robin方式遍历集群所有可用Worker,分配Worker资源,来启动创建Executor的策略,好处是尽可能的将cores分配到各个节点,最大化负载均衡和高并行。
3.2、非SpreadOut:会尽可能的根据每个Worker的剩余资源来启动Executor,这样启动的Executor可能只在集群的一小部分机器的Worker上。这样做对node较少的集群还可以,集群规模大了,Executor的并行度和机器负载均衡就不能够保证了。
行文仓促,如有不正之处,请指出,欢迎讨论 :)
补充:
1、关于: 一个App一个Worker为什么只有允许有针对该App的一个Executor 到底这样设计为何? 的讨论:
连城404:Spark是线程级并行模型,为什么需要一个worker为一个app启动多个executor呢?
朴动_zju:一个worker对应一个executorbackend是从mesos那一套迁移过来的,mesos下也是一个slave一个executorbackend。我理解这里是可以实现起多个,但起多个貌似没什么好处,而且增加了复杂度。
CrazyJvm:@CodingCat 做了一个patch可以启动多个,但是还没有被merge。 从Yarn的角度考虑的话,一个Worker可以对应多个executorbackend,正如一个nodemanager对应多个container。 @OopsOutOfMemory
OopsOutOfMemory:回复@连城404: 如果一个executor太大且装的对象太多,会导致GC很慢,多几个Executor会减少full gc慢的问题。 see this post http://t.cn/RP1bVO4(今天 11:25)
连城404:回复@OopsOutOfMemory:哦,这个考虑是有道理的。一个workaround是单台机器部署多个worker,worker相对来说比较廉价。
JerryLead:回复@OopsOutOfMemory:看来都还在变化当中,standalone 和 YARN 还是有很多不同,我们暂不下结论 (今天 11:35)
JerryLead:问题开始变得复杂了,是提高线程并行度还是提高进程并行度?我想 Spark 还是优先选择前者,这样 task 好管理,而且 broadcast,cache 的效率高些。后者有一些道理,但参数配置会变得更复杂,各有利弊吧 (今天 11:40)
未完待续。。。
传送门:@JerrLead https://github.com/JerryLead/SparkInternals/blob/master/markdown/1-Overview.md
——EOF——
原创文章,转载请注明出自:http://blog.csdn.net/oopsoom/article/details/38763985