其實 Spark 在某種層面上和 MR 是一樣的,只不過 Spark 更適合於迭代計算的業務。那麼二者的區別有哪些呢?我這裏只列舉幾個大層面的東西,細小的東西我們會在下文繼續分析:
- 當內存足夠的情況下或者數據量相對較小的情況下,Spark RDD之間傳輸的數據不用上傳 HDFS
- Spark 可以將 RDD 存儲在內存,供下一個 task 使用
- Spark 適用於迭代計算,如果不是迭代相關的業務問題,Spark 於 MR 的性能差不多
這裏對 Spark 從集羣啓動到最後的 task 執行完畢並返回之行結果的過程做了一個概述,下面首先從資源調度開始分析。
Spark 資源調度源碼解析
一、Spark 集羣啓動:
首先啓動 Spark 集羣,查看 sbin 下的 start-all.sh 腳本,會先啓動 Master:
# Start Master
"${SPARK_HOME}/sbin"/start-master.sh
# Start Workers
"${SPARK_HOME}/sbin"/start-slaves.sh
查看 sbin/start-master.sh 腳本,發現會去執行 org.apache.spark.deploy.master.Master 類,開始在源碼中跟進Master,從 main 方法開始:
//主方法
def main(argStrings: Array[String]) {
Thread.setDefaultUncaughtExceptionHandler(new SparkUncaughtExceptionHandler(
exitOnUncaughtException = false))
Utils.initDaemon(log)
val conf = new SparkConf
val args = new MasterArguments(argStrings, conf)
/**
* 創建RPC 環境和Endpoint (RPC 遠程過程調用),在Spark中 Driver, Master ,Worker角色都有各自的Endpoint,相當於各自的通信郵箱。
*
*/
val (rpcEnv, _, _) = startRpcEnvAndEndpoint(args.host, args.port, args.webUiPort, conf)
rpcEnv.awaitTermination()
}
在 main 方法中執行了 startRpcEnvAndEndpoint 方法,創建 RPC 環境和 Endpoint(RPC 遠程過程調用),詳細的說 RpcEnv 是用於接收消息和處理消息的遠程通信調用環境,Master 向 RpcEnv 中去註冊,不管是 Master,Driver,Worker,Executor 等都有自己的 Endpoint,相當於是郵箱,其他人想跟我通信先要找到我的郵箱纔可以。Master啓動時會將 Endpoint 註冊在 RpcEnv 裏面,用於接收,處理消息。跟進 startRpcEnvAndEndpoint 方法:
/**
* 創建RPC(Remote Procedure Call )環境 ,Remote Procedure Call
* 這裏只是創建準備好Rpc的環境,後面會向RpcEnv中註冊 角色【Driver,Master,Worker,Executor】
*/
val rpcEnv = RpcEnv.create(SYSTEM_NAME, host, port, conf, securityMgr)
/**
* 向RpcEnv 中 註冊Master
*
* rpcEnv.setupEndpoint(name,new Master)
* 這裏new Master 的Master 是一個伴生類,繼承了 ThreadSafeRpcEndpoint,歸根結底繼承到了 Trait 接口 RpcEndpoint
* 什麼是Endpoint?
* EndPoint中存在
* onstart() :啓動當前Endpoint
* receive() :負責收消息
* receiveAndReply():接受消息並回復
* Endpoint 還有各自的引用,方便其他Endpoint發送消息,直接引用對方的EndpointRef 即可找到對方的Endpoint
* 以下 masterEndpoint 就是Master的Endpoint引用 RpcEndpointRef 。
* RpcEndpointRef中存在:
* send():發送消息
* ask() :請求消息,並等待迴應。
*/
val masterEndpoint: RpcEndpointRef = rpcEnv.setupEndpoint(ENDPOINT_NAME,
new Master(rpcEnv, rpcEnv.address, webUiPort, securityMgr, conf))
val portsResponse = masterEndpoint.askSync[BoundPortsResponse](BoundPortsRequest)
(rpcEnv, portsResponse.webUIPort, portsResponse.restPort)
在該方法有兩個功能:
- 創建 RpcEnv,一些細節已在代碼中說明
- 向RpcEnv 中 註冊Master,一些細節已在代碼中說明
繼續跟進 create 方法:
//創建RPC 環境
create(name, host, host, port, conf, securityManager, 0, clientMode)
繼續跟進 create 方法:
//創建NettyRpc 環境
new NettyRpcEnvFactory().create(config)
在這裏會創建一個 NettyRpcEnvFactory 的對象,並調用 create 方法,看一下在 實例化 NettyRpcEnvFactory 時有哪些操作:
/**
* 創建nettyRPC通信環境。
*/
val nettyEnv =
new NettyRpcEnv(sparkConf, javaSerializerInstance, config.advertiseAddress,
config.securityManager, config.numUsableCores)
該方法的作用是創建nettyRPC通信環境,並且在 new NettyRpcEnv 時會做一些初始化:
- Dispatcher:這個對象中有存放消息的隊列和消息的轉發
- TransportContext:可以創建NettyRpcHandler
private val dispatcher: Dispatcher = new Dispatcher(this, numUsableCores)
private val transportContext = new TransportContext(transportConf,
new NettyRpcHandler(dispatcher, this, streamManager))
Dispatcher 的作用是存放消息的隊列和消息的轉發,首先看 Dispatcher 實例化時執行的邏輯:
private val threadpool: ThreadPoolExecutor = {
val availableCores =
if (numUsableCores > 0) numUsableCores else Runtime.getRuntime.availableProcessors()
val numThreads = nettyEnv.conf.getInt("spark.rpc.netty.dispatcher.numThreads",
math.max(2, availableCores))
val pool = ThreadUtils.newDaemonFixedThreadPool(numThreads, "dispatcher-event-loop")
for (i <- 0 until numThreads) {
pool.execute(new MessageLoop)
}
pool
}
在 Dispatcher 實例化的過程中會創建一個 threadpool,在 threadpool 中會執行 MessageLoop:
private class MessageLoop extends Runnable {
override def run(): Unit = {
try {
while (true) {
try {
//take 出來消息一直處理
val data: EndpointData = receivers.take()
if (data == PoisonPill) {
// Put PoisonPill back so that other MessageLoops can see it.
receivers.offer(PoisonPill)
return
}
//調用process 方法處理消息
data.inbox.process(Dispatcher.this)
} catch {
case NonFatal(e) => logError(e.getMessage, e)
}
}
} catch {
case ie: InterruptedException => // exit
}
}
}
在 MessageLoop 中的 receivers.take() 會一直向 receivers消息隊列中去數據,而 receivers 是在 Dispatcher 初始化時創建的,至此接收消息的程序已經啓動起來:
private val receivers = new LinkedBlockingQueue[EndpointData]
其中會傳入一個 EndpointData 對象,實例化時會實例化一個 Inbox 對象:
private class EndpointData(
val name: String,
val endpoint: RpcEndpoint,
val ref: NettyRpcEndpointRef) {
//將endpoint封裝到Inbox中
val inbox = new Inbox(ref, endpoint)
}
實例化 Inbox 對象,當註冊 endpoint 時都會調用一個異步方法,messages中放一個OnStart樣例類(消息隊列),所以早默認情況下都會調用 OnStart 的匹配方法:
inbox.synchronized {
messages.add(OnStart)
}
在實例化 MessageLoop 時還會調用 process 方法處理消息:
//調用process 方法處理消息
data.inbox.process(Dispatcher.this)
在 process 中就會找到與發送消息所匹配的 case 去執行邏輯,例如:
case OnStart =>
//調用Endpoint 的onStart方法
endpoint.onStart()
if (!endpoint.isInstanceOf[ThreadSafeRpcEndpoint]) {
inbox.synchronized {
if (!stopped) {
enableConcurrent = true
}
}
}
下面來分析 transportContext 對象的作用,在創建完該類的實例之後,會調用 transportContext.createServer 方法去啓動 NettyRpc 的服務,在創建 Rpc 服務的過程中,會創建將處理消息的對象 createChannelHandler:
server = transportContext.createServer(bindAddress, port, bootstraps)
在 createServer 方法中會實例化 TransportServer 的對象,在 try 中會調用 init 方法進行初始化:
try {
//運行初始化init方法
init(hostToBind, portToBind);
}
在 init 方法中,Rpc 的遠程通信對象 bootstrap 會調用 childHandler 方法,會初始化網絡通信管道:
//初始化網絡通信管道
context.initializePipeline(ch, rpcHandler);
在初始化網絡通信管道的過程中,創建處理消息的 channelHandler 對象,該對象的作用是
- 創建並處理客戶端的請求消息和服務消息
//創建處理消息的 channelHandler
TransportChannelHandler channelHandler = createChannelHandler(channel, channelRpcHandler);
private TransportChannelHandler createChannelHandler(Channel channel, RpcHandler rpcHandler) {
TransportResponseHandler responseHandler = new TransportResponseHandler(channel);
TransportClient client = new TransportClient(channel, responseHandler);
TransportRequestHandler requestHandler = new TransportRequestHandler(channel, client,
rpcHandler, conf.maxChunksBeingTransferred());
return new TransportChannelHandler(client, responseHandler, requestHandler,
conf.connectionTimeoutMs(), closeIdleConnections);
}
TransportChannelHandler 由以上 responseHandler,client,requestHandler 三個 handler 構建,並且這個對象中有 channelRead 方法,用於讀取接收到的消息:
@Override
public void channelRead(ChannelHandlerContext ctx, Object request) throws Exception {
//判斷當前消息是請求的消息還是迴應的消息
if (request instanceof RequestMessage) {
requestHandler.handle((RequestMessage) request);
} else if (request instanceof ResponseMessage) {
responseHandler.handle((ResponseMessage) request);
} else {
ctx.fireChannelRead(request);
}
}
以 requestHandler 爲例,調用 headle —> processRpcRequest((RpcRequest) request),會看到 rpcHandler.receive,此時調用的是 NettyRpcHandler 的 receive:
try {
/**
* rpcHandler 是一直傳過來的 NettyRpcHandler
* 這裏的receive 方法 是 NettyRpcHandler 中的方法
*/
rpcHandler.receive(reverseClient, req.body().nioByteBuffer(), new RpcResponseCallback() {
@Override
public void onSuccess(ByteBuffer response) {
respond(new RpcResponse(req.requestId, new NioManagedBuffer(response)));
}
@Override
public void onFailure(Throwable e) {
respond(new RpcFailure(req.requestId, Throwables.getStackTraceAsString(e)));
}
});
}
------------------------------------------------------------------------
override def receive(
client: TransportClient,
message: ByteBuffer,
callback: RpcResponseCallback): Unit = {
val messageToDispatch = internalReceive(client, message)
//dispatcher負責發送遠程的消息,都最終調到postMessage 方法
dispatcher.postRemoteMessage(messageToDispatch, callback)
}
繼續調用 dispatcher.postRemoteMessage 方法:
def postRemoteMessage(message: RequestMessage, callback: RpcResponseCallback): Unit = {
val rpcCallContext =
new RemoteNettyRpcCallContext(nettyEnv, callback, message.senderAddress)
val rpcMessage = RpcMessage(message.senderAddress, message.content, rpcCallContext)
postMessage(message.receiver.name, rpcMessage, (e) => callback.onFailure(e))
}
在 postRemoteMessage 中,無論是請求消息還是迴應消息,都最終會執行到這個 postMessage:
private def postMessage(
endpointName: String,
message: InboxMessage,
callbackIfStopped: (Exception) => Unit): Unit = {
val error = synchronized {
//獲取消息的通信郵箱名稱
val data = endpoints.get(endpointName)
if (stopped) {
Some(new RpcEnvStoppedException())
} else if (data == null) {
Some(new SparkException(s"Could not find $endpointName."))
} else {
//將消息放入通信端的消息隊列中
data.inbox.post(message)
//添加到消息隊列中
receivers.offer(data)
None
}
}
在該方法中會將 message 放入 inbox 中,在處理消息的程序啓動後,處理消息的程序也已經啓動,至此 RpcEnv環境啓動完畢,緊接着 Master 要在RpcEnv 中註冊:
val masterEndpoint: RpcEndpointRef = rpcEnv.setupEndpoint(ENDPOINT_NAME,
new Master(rpcEnv, rpcEnv.address, webUiPort, securityMgr, conf))
Master 註冊了自己 EndPoint,可以接受處理消息,Master啓動成功。
Worker 同理,Spark 集羣啓動成功。
二、Spark 提交任務向 Master 申請啓動 Driver
執行 ./spark-submit…,找到 spark-submit.sh 腳本,找到SparkSubmit 主類:
//提交任務主類運行
override def main(args: Array[String]): Unit = {
// Initialize logging if it hasn't been done yet. Keep track of whether logging needs to
// be reset before the application starts.
val uninitLog = initializeLogIfNecessary(true, silent = true)
//設置參數
val appArgs = new SparkSubmitArguments(args)
if (appArgs.verbose) {
// scalastyle:off println
printStream.println(appArgs)
// scalastyle:on println
}
appArgs.action match {
//任務提交匹配 submit
case SparkSubmitAction.SUBMIT => submit(appArgs, uninitLog)
case SparkSubmitAction.KILL => kill(appArgs)
case SparkSubmitAction.REQUEST_STATUS => requestStatus(appArgs)
}
}
因爲我們是提交任務,所以會匹配到 SparkSubmitAction.SUBMIT 的類型,繼續執行 submit 方法。
val (childArgs, childClasspath, sparkConf, childMainClass) = prepareSubmitEnvironment(args)
在 submit 中,prepareSubmitEnvironment 會返回一個四元組,重點注意 childMainClass 類,是最後啓動 Driver 的類,這裏以 standalone-cluster 爲例。
//準備提交任務的環境
doPrepareSubmitEnvironment(args, conf)
進入 prepareSubmitEnvironment,找到準備提交任務的環境的方法 doPrepareSubmitEnvironment:
//正常提交方式
// In legacy standalone cluster mode, use Client as a wrapper around the user class
childMainClass = STANDALONE_CLUSTER_SUBMIT_CLASS
因爲這裏是以 standalone-cluster 爲例,所以找到匹配的case,並進入 STANDALONE_CLUSTER_SUBMIT_CLASS 看看代表什麼:
private[deploy] val STANDALONE_CLUSTER_SUBMIT_CLASS = classOf[ClientApp].getName()
這裏的 clientApp 就是我們要啓動的 Driver 部分。
在 submit 方法繼續向下走,因爲方法不會直接執行,所以代碼向下走回執行到 doRunMain():
//運行
runMain(childArgs, childClasspath, sparkConf, childMainClass, args.verbose)
繼續執行 runMain 方法,這裏將 childMainClass 作爲參數傳了進來:
//加載類
mainClass = Utils.classForName(childMainClass)
//------------------------------------------------------------
val app: SparkApplication = if (classOf[SparkApplication].isAssignableFrom(mainClass))
這裏首先將 childMainClass 加載出來,賦給變量 mainClass,之後會將 mainClass 映射成 SparkApplication。
app.start(childArgs.toArray, sparkConf)
接下來調用start方法,這裏調用的是 ClientApp 的 start 方法,因爲 childMainClass 是 clientApp 的類型:
rpcEnv.setupEndpoint("client", new ClientEndpoint(rpcEnv, driverArgs, masterEndpoints, conf))
在 rpc 中設置提交當前任務的 Endpoint,只要設置肯定會運行 new ClientEndpoint 類的 onStart 方法:
sval mainClass = "org.apache.spark.deploy.worker.DriverWrapper"
//將DriverWrapper 這個類封裝到Command中
val command = new Command(mainClass,
Seq("{{WORKER_URL}}", "{{USER_JAR}}", driverArgs.mainClass) ++ driverArgs.driverOptions,
sys.env, classPathEntries, libraryPathEntries, javaOpts)
val driverDescription = new DriverDescription(
driverArgs.jarUrl,
driverArgs.memory,
driverArgs.cores,
driverArgs.supervise,
command)
//向Master申請啓動Driver,Master中的 receiveAndReply 方法會接收此請求消息
asyncSendToMasterAndForwardReply[SubmitDriverResponse](
RequestSubmitDriver(driverDescription))
這裏將 org.apache.spark.deploy.worker.DriverWrapper 封裝成 command,並且將 command 封裝到 driverDescription 中,然後向 Master 申請啓動 Driver,Master 中的 receiveAndReply 方法會接收此請求消息:
case RequestSubmitDriver(description) =>
//判斷Master狀態
if (state != RecoveryState.ALIVE) {
val msg = s"${Utils.BACKUP_STANDALONE_MASTER_PREFIX}: $state. " +
"Can only accept driver submissions in ALIVE state."
context.reply(SubmitDriverResponse(self, false, None, msg))
} else {
logInfo("Driver submitted " + description.command.mainClass)
val driver = createDriver(description)
persistenceEngine.addDriver(driver)
waitingDrivers += driver
drivers.add(driver)
schedule()
// TODO: It might be good to instead have the submission client poll the master to determine
// the current status of the driver. For now it's simply "fire and forget".
context.reply(SubmitDriverResponse(self, true, Some(driver.id),
s"Driver successfully submitted as ${driver.id}"))
}
這裏首先會判斷 Master 的狀態,如果符合要求的話,會使用之前封裝的 description 對象創建 driver,driver 其實是一個 DriverInfo 的類型,裏面封裝了一些 Driver 的信息。之後會在 waitingDrivers (private val waitingDrivers = new ArrayBuffer[DriverInfo]) 中添加剛纔創建完的 DriverInfo對象,進入 schedule() 調度方法:
/**
* schedule() 方法是通用的方法
* 這個方法中當申請啓動Driver的時候也會執行,但是最後一行的startExecutorsOnWorkers 方法中 waitingApp是空的,只是啓動Driver。
* 在提交application時也會執行到這個scheduler方法,這個時候就是要啓動的Driver是空的,但是會直接運行startExecutorsOnWorkers 方法給當前的application分配資源
*
*/
private def schedule(): Unit = {
//判斷Master狀態
if (state != RecoveryState.ALIVE) {
return
}
// Drivers take strict precedence over executors 這裏是打散worker
val shuffledAliveWorkers = Random.shuffle(workers.toSeq.filter(_.state == WorkerState.ALIVE))
//可用的worker數量
val numWorkersAlive = shuffledAliveWorkers.size
var curPos = 0
for (driver <- waitingDrivers.toList) { // iterate over a copy of waitingDrivers
// We assign workers to each waiting driver in a round-robin fashion. For each driver, we
// start from the last worker that was assigned a driver, and continue onwards until we have
// explored all alive workers.
var launched = false
var numWorkersVisited = 0
while (numWorkersVisited < numWorkersAlive && !launched) {
//拿到curPos位置的worker
val worker = shuffledAliveWorkers(curPos)
numWorkersVisited += 1
if (worker.memoryFree >= driver.desc.mem && worker.coresFree >= driver.desc.cores) {
//這裏是啓動Driver,啓動Driver之後會爲當前的application 申請資源
launchDriver(worker, driver)
waitingDrivers -= driver
launched = true
}
//curPos 就是一直加一的往後取 Worker ,一直找到滿足資源的worker
curPos = (curPos + 1) % numWorkersAlive
}
}
startExecutorsOnWorkers()
}
這裏會找到滿足條件的 Worker 節點去啓動 Driver,調用 launchDriver(worker, driver) 方法:
private def launchDriver(worker: WorkerInfo, driver: DriverInfo) {
logInfo("Launching driver " + driver.id + " on worker " + worker.id)
worker.addDriver(driver)
driver.worker = Some(worker)
//給Worker發送消息啓動Driver,這裏在Worker中會有receive方法一直匹配LaunchDriver
worker.endpoint.send(LaunchDriver(driver.id, driver.desc))
driver.state = DriverState.RUNNING
}
這裏會向 Worker 發送消息啓動 Driver,這裏在 Worker 中會有 receive 方法一直匹配 LaunchDriver:
case LaunchDriver(driverId, driverDesc) =>
logInfo(s"Asked to launch driver $driverId")
val driver = new DriverRunner(
conf,
driverId,
workDir,
sparkHome,
driverDesc.copy(command = Worker.maybeUpdateSSLSettings(driverDesc.command, conf)),
self,
workerUri,
securityMgr)
drivers(driverId) = driver
//啓動Driver,會初始化 org.apache.spark.deploy.worker.DriverWrapper ,運行main方法
driver.start()
這裏說的啓動的 Driver 就是剛纔說的 val mainClass = "org.apache.spark.deploy.worker.DriverWrapper",Driver 啓動的就是 DriverWrapper 類的啓動,DriverWrapper 的啓動就是在 Worker 中創建一個 Driver 進程。啓動Driver,會初始化 org.apache.spark.deploy.worker.DriverWrapper,運行 main 方法:
//下面的mainClass就是我們真正提交的application
case workerUrl :: userJar :: mainClass :: extraArgs =>
在 DriverWrapper 的 main 方法中的 mainClass,就是我們真正提交的 Application
// Delegate to supplied main class
val clazz = Utils.classForName(mainClass)
//得到提交application的主方法
val mainMethod = clazz.getMethod("main", classOf[Array[String]])
/**
* 啓動提交的application 中的main 方法。
* 這裏啓動application,會先創建SparkConf和SparkContext
* SparkContext中 362行try塊中會創建TaskScheduler(492)
*/
mainMethod.invoke(null, extraArgs.toArray[String])
三、Spark Driver 啓動完畢,並且向 Master 註冊 Applciation
當在 Worker 啓動完 Driver 之後,程序走到在 Driver 端我們手寫的代碼上,首先就是啓動 SparkContext,在與源碼中找到 SparkContext 類,創建 Scala 對象時只有方法不執行,剩下的都執行,找到如下代碼:
/**
* 啓動調度程序,這裏(sched,ts) 分別對應 StandaloneSchedulerBackend 和 TaskSchedulerImpl 兩個對象
* master 是提交任務寫的 spark://node1:7077
*/
val (sched, ts) = SparkContext.createTaskScheduler(this, master, deployMode)
_schedulerBackend = sched
_taskScheduler = ts
_dagScheduler = new DAGScheduler(this)
_heartbeatReceiver.ask[Boolean](TaskSchedulerIsSet)
// start TaskScheduler after taskScheduler sets DAGScheduler reference in DAGScheduler's
// constructor
//TaskSchedulerImpl 對象的start方法
_taskScheduler.start()
我們能看到 createTaskScheduler 方法返回一個元組,且在 SparkContext 中已經創建了 DAGScheduler對象,進入 createTaskScheduler 方法:
//standalone 提交任務都是以 “spark://”這種方式提交
case SPARK_REGEX(sparkUrl) =>
//scheduler 創建TaskSchedulerImpl 對象
val scheduler = new TaskSchedulerImpl(sc)
val masterUrls = sparkUrl.split(",").map("spark://" + _)
/**
* 這裏的 backend 是StandaloneSchedulerBackend 這個類型
*/
val backend = new StandaloneSchedulerBackend(scheduler, sc, masterUrls)
//這裏會調用 TaskSchedulerImpl 對象中的 initialize 方法將 backend 初始化,一會要用到
scheduler.initialize(backend)
//返回了 StandaloneSchedulerBackend 和 TaskSchedulerImpl 兩個對象
(backend, scheduler)
因爲是以Standalone的方式提交任務,所以找到匹配的 case。在這裏創建了 TaskSchedulerImpl 的對象scheduler,並將其傳入了 StandaloneSchedulerBackend(scheduler, sc, masterUrls)中,得知,返回的 backend 是一個 Standalone 大環境的任務調度器,scheduler 則是 TaskScheduler 的調度器。
/*
* 在TaskScheduler中設置進來backend ,
* 這裏的backend 是 StandaloneSchedulerBackend
* StandaloneSchedulerBackend 繼承了CoarseGrainedSchedulerBackend 類
*/
def initialize(backend: SchedulerBackend) {
this.backend = backend
schedulableBuilder = {
schedulingMode match {
case SchedulingMode.FIFO =>
new FIFOSchedulableBuilder(rootPool)
case SchedulingMode.FAIR =>
new FairSchedulableBuilder(rootPool, conf)
case _ =>
throw new IllegalArgumentException(s"Unsupported $SCHEDULER_MODE_PROPERTY: " +
s"$schedulingMode")
}
}
schedulableBuilder.buildPools()
}
在 scheduler.initialize(backend) 中的作用就是將 backend 設置爲 scheduler 的屬性,且StandaloneSchedulerBackend 繼承了CoarseGrainedSchedulerBackend 類。
//TaskSchedulerImpl 對象的start方法
_taskScheduler.start()
ts 調用的是 TaskSchedulerImpl 中的 start() 方法:
/**
* TaskScheduler 啓動
*/
override def start() {
//StandaloneSchedulerBackend 啓動
backend.start()
if (!isLocal && conf.getBoolean("spark.speculation", false)) {
logInfo("Starting speculative execution thread")
speculationScheduler.scheduleWithFixedDelay(new Runnable {
override def run(): Unit = Utils.tryOrStopSparkContext(sc) {
checkSpeculatableTasks()
}
}, SPECULATION_INTERVAL_MS, SPECULATION_INTERVAL_MS, TimeUnit.MILLISECONDS)
}
}
進入 backend.start() 方法,因爲 backend 是 StandaloneSchedulerBackend 類型,所有調用的是 StandaloneSchedulerBackend 中的 start() 方法:
override def start() {
/**
* super.start()中有創建Driver的通信郵箱也就是Driver的引用
* 未來Executor就是向 StandaloneSchedulerBackend中父類 CoarseGrainedSchedulerBackend 中反向註冊信息的.
*/
super.start()
進入 super.start(),在方法中向 RpcEnv 註冊了 Driver 端的 Endpoint:
override def start() {
val properties = new ArrayBuffer[(String, String)]
for ((key, value) <- scheduler.sc.conf.getAll) {
if (key.startsWith("spark.")) {
properties += ((key, value))
}
}
// TODO (prashant) send conf instead of properties
/**
* 創建Driver的Endpoint ,就是創建Driver的通信郵箱,向Rpc中註冊當前DriverEndpoint
* 未來Executor就是向DriverEndpoint中反向註冊信息,這裏Driver中會有receiveAndReply方法一直監聽匹配發過來的信息
*/
driverEndpoint = createDriverEndpointRef(properties)
}
注意在 backend.start() 方法執行創建完 DriverEndpoint 之後,還執行了如下代碼,功能是向 Driver 註冊application 的信息 :
val command = Command("org.apache.spark.executor.CoarseGrainedExecutorBackend", args, sc.executorEnvs, classPathEntries ++ testingClassPath, libraryPathEntries, javaOpts)
......
val appDesc: ApplicationDescription = ApplicationDescription(sc.appName, maxCores, sc.executorMemory, command, webUrl, sc.eventLogDir, sc.eventLogCodec, coresPerExecutor, initialExecutorLimit)
//提交應用程序的描述信息
//封裝 appDesc,這裏已經傳入了StandaloneAppClient 中
client = new StandaloneAppClient(sc.env.rpcEnv, masters, appDesc, this, conf)
//啓動StandaloneAppClient,之後會向Driver註冊application的信息
client.start()
進入 client.start() 之後就是給 rpcEnv 上註冊信息
def start() {
// Just launch an rpcEndpoint; it will call back into the listener.
/**
* 這裏就是給空的 endpoint[AtomicReference] 設置下 信息,
* 主要是rpcEnv.setupEndpoint 中創建了 ClientEndpoint 只要設置Endpoint 肯定會調用 ClientEndpoint的onStart方法
*/
endpoint.set(rpcEnv.setupEndpoint("AppClient", new ClientEndpoint(rpcEnv)))
}
然後讀取 ClientEndpoint 的 onStart() 方法
//onStart 方法
override def onStart(): Unit = {
try {
//向Master 註冊當前application的信息
registerWithMaster(1)
} catch {
case e: Exception =>
logWarning("Failed to connect to master", e)
markDisconnected()
stop()
}
}
向 Master 註冊當前application的信息的 registerWithMaster() 方法,因爲我們的 Master 會有高可用,所以要給所有的 Master 註冊,進入後看到 tryRegisterAllMasters():
private def registerWithMaster(nthRetry: Int) {
//tryRegisterAllMasters 嘗試向所有的Master 註冊application信息
registerMasterFutures.set(tryRegisterAllMasters())
registrationRetryTimer.set(registrationRetryThread.schedule(new Runnable {
override def run(): Unit = {
if (registered.get) {
registerMasterFutures.get.foreach(_.cancel(true))
registerMasterThreadPool.shutdownNow()
} else if (nthRetry >= REGISTRATION_RETRIES) {
markDead("All masters are unresponsive! Giving up.")
} else {
registerMasterFutures.get.foreach(_.cancel(true))
registerWithMaster(nthRetry + 1)
}
}
}, REGISTRATION_TIMEOUT_SECONDS, TimeUnit.SECONDS))
}
進入 tryRegisterAllMasters() 之後,會獲取 Master 的 Endpoint,向 Master 註冊 application,Master 類中receive 方法中會匹配接收 RegisterApplication 類型,隨機在 Master 中匹配 RegisterApplication 的 case:
//向所有的Master去註冊Application的信息
private def tryRegisterAllMasters(): Array[JFuture[_]] = {
//遍歷所有的Master地址
for (masterAddress <- masterRpcAddresses) yield {
registerMasterThreadPool.submit(new Runnable {
override def run(): Unit = try {
if (registered.get) {
return
}
logInfo("Connecting to master " + masterAddress.toSparkURL + "...")
//獲取Master的通信郵箱
val masterRef = rpcEnv.setupEndpointRef(masterAddress, Master.ENDPOINT_NAME)
//向Master註冊application,Master類中receive方法中會匹配接收 RegisterApplication類型
masterRef.send(RegisterApplication(appDescription, self))
} catch {
case ie: InterruptedException => // Cancelled
case NonFatal(e) => logWarning(s"Failed to connect to master $masterAddress", e)
}
})
}
}
Master 匹配 RegisterApplication 類型消息的處理流程:
//Driver 端提交過來的要註冊Application
case RegisterApplication(description, driver) =>
// TODO Prevent repeated registrations from some driver
//如果Master狀態是standby 忽略不提交任務
if (state == RecoveryState.STANDBY) {
// ignore, don't send response
} else {
logInfo("Registering app " + description.name)
//這裏封裝application信息,注意,在這裏可以跟進去看到默認一個application使用的core的個數就是 Int.MaxValue
val app = createApplication(description, driver)
//註冊app ,這裏面有向 waitingApps中加入當前application
registerApplication(app)
logInfo("Registered app " + description.name + " with ID " + app.id)
persistenceEngine.addApplication(app)
driver.send(RegisteredApplication(app.id, self))
//最終又會執行通用方法schedule()
schedule()
}
至此 Driver 向 Master 註冊 Application 流程結束。
四、Master 發送消息啓動 Executor
在執行 schedule() 之前會向 Driver(StandaloneAppClient) 端發送 接收到Application已經被註冊 的消息,最終又會執行通用方法 schedule() 方法:
/**
* schedule() 方法是通用的方法
* 這個方法中當申請啓動Driver的時候也會執行,但是最後一行的startExecutorsOnWorkers 方法中 waitingApp是空的,只是啓動Driver。
* 在提交application時也會執行到這個scheduler方法,這個時候就是要啓動的Driver是空的,但是會直接運行startExecutorsOnWorkers 方法給當前的application分配資源
*
*/
private def schedule(): Unit = {
//判斷Master狀態
if (state != RecoveryState.ALIVE) {
return
}
// Drivers take strict precedence over executors 這裏是打散worker
val shuffledAliveWorkers = Random.shuffle(workers.toSeq.filter(_.state == WorkerState.ALIVE))
//可用的worker數量
val numWorkersAlive = shuffledAliveWorkers.size
var curPos = 0
for (driver <- waitingDrivers.toList) { // iterate over a copy of waitingDrivers
// We assign workers to each waiting driver in a round-robin fashion. For each driver, we
// start from the last worker that was assigned a driver, and continue onwards until we have
// explored all alive workers.
var launched = false
var numWorkersVisited = 0
while (numWorkersVisited < numWorkersAlive && !launched) {
//拿到curPos位置的worker
val worker = shuffledAliveWorkers(curPos)
numWorkersVisited += 1
if (worker.memoryFree >= driver.desc.mem && worker.coresFree >= driver.desc.cores) {
//這裏是啓動Driver,啓動Driver之後會爲當前的application 申請資源
launchDriver(worker, driver)
waitingDrivers -= driver
launched = true
}
//curPos 就是一直加一的往後取 Worker ,一直找到滿足資源的worker
curPos = (curPos + 1) % numWorkersAlive
}
}
startExecutorsOnWorkers()
}
此時我們又回到了 schedule() 方法,之前是從 launchDriver(worker, driver) 進去的,現在又出來繼續調用 startExecutorsOnWorkers() 方法:
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.
//從waitingApps中獲取提交的app
for (app <- waitingApps) {
//coresPerExecutor 在application中獲取啓動一個Executor使用幾個core 。參數--executor-core可以指定,下面指明不指定就是1
val coresPerExecutor = app.desc.coresPerExecutor.getOrElse(1)
// If the cores left is less than the coresPerExecutor,the cores left will not be allocated
//判斷是否給application分配夠了core,因爲後面每次給application 分配core後 app.coresLeft 都會相應的減去分配的core數
if (app.coresLeft >= coresPerExecutor) {
// Filter out workers that don't have enough resources to launch an executor
//過濾出可用的worker
val usableWorkers = workers.toArray.filter(_.state == WorkerState.ALIVE)
.filter(worker => worker.memoryFree >= app.desc.memoryPerExecutorMB &&
worker.coresFree >= coresPerExecutor)
.sortBy(_.coresFree).reverse
/**
* 下面就是去worker中劃分每個worker提供多少core和啓動多少Executor,注意:spreadOutApps 是true
* 返回的 assignedCores 就是每個worker節點中應該給當前的application分配多少core
*/
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) {
//在worker中給Executor劃分資源
allocateWorkerResourceToExecutors(
app, assignedCores(pos), app.desc.coresPerExecutor, usableWorkers(pos))
}
}
}
}
在此處應該注意的是 scheduleExecutorsOnWorkers(app, usableWorkers, spreadOutApps) 方法,最後返回最後返回每個Worker上分配多少core,其他解釋已加註釋:
private def scheduleExecutorsOnWorkers(
app: ApplicationInfo,
usableWorkers: Array[WorkerInfo],
spreadOutApps: Boolean): Array[Int] = {
//啓動一個Executor使用多少core,這裏如果提交任務沒有指定 --executor-core這個值就是None
val coresPerExecutor : Option[Int]= app.desc.coresPerExecutor
//這裏指定如果提交任務沒有指定啓動一個Executor使用幾個core,默認就是1
val minCoresPerExecutor = coresPerExecutor.getOrElse(1)
//oneExecutorPerWorker 當前爲true
val oneExecutorPerWorker :Boolean= coresPerExecutor.isEmpty
//默認啓動一個Executor使用的內存就是1024M,這個設置在SparkContext中464行
val memoryPerExecutor = app.desc.memoryPerExecutorMB
//可用worker的個數
val numUsable = usableWorkers.length
//創建兩個重要對象
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
/**
* coresToAssign 指的是當前要給Application分配的core是多少? app.coresLeft 與集羣所有worker剩餘的全部core 取個最小值
* 這裏如果提交application時指定了 --total-executor-core 那麼app.coresLeft 就是指定的值
*/
var coresToAssign = math.min(app.coresLeft, usableWorkers.map(_.coresFree).sum)
/** Return whether the specified worker can launch an executor for this app. */
//判斷某臺worker節點是否還可以啓動Executor
def canLaunchExecutor(pos: Int): Boolean = {
//可分配的core是否大於啓動一個Executor使用的1個core
val keepScheduling = coresToAssign >= minCoresPerExecutor
//是否有足夠的core
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.
//assignedExecutors(pos) == 0 爲true,launchingNewExecutor就是爲true
val launchingNewExecutor = !oneExecutorPerWorker || assignedExecutors(pos) == 0
//啓動新的Executor
if (launchingNewExecutor) {
val assignedMemory = assignedExecutors(pos) * memoryPerExecutor
//是否有足夠的內存
val enoughMemory = usableWorkers(pos).memoryFree - assignedMemory >= memoryPerExecutor
//這裏做安全判斷,說的是要分配啓動的Executor和當前application啓動的使用的Executor總數是否在Application總的Executor限制之下
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
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(one=>canLaunchExecutor(one))
var freeWorkers = (0 until numUsable).filter(canLaunchExecutor)
while (freeWorkers.nonEmpty) {
freeWorkers.foreach { pos =>
var keepScheduling = true
while (keepScheduling && canLaunchExecutor(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) {
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.
if (spreadOutApps) {
keepScheduling = false
}
}
}
freeWorkers = freeWorkers.filter(canLaunchExecutor)
}
//最後返回每個Worker上分配多少core
assignedCores
}
在上述代碼中還有幾個注意點:
爲每個worker節點分配完資源後就開始啓動Executor了,接着執行 Master.schedule() 方法下的 allocateWorkerResourceToExecutors 方法,該方法傳入了當前 application,每個 worker 分配的核數,每個核上要啓動的 Executor 數量以及可用的 usableWorkers :
//在worker中給Executor劃分資源
allocateWorkerResourceToExecutors(
app, assignedCores(pos), app.desc.coresPerExecutor, usableWorkers(pos))
點進入之後 Master 會計算每個 Executor 要分配多少個 core,緊接着去 worker 啓動Executor:
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)
//每個Executor要分配多少個core
val coresToAssign = coresPerExecutor.getOrElse(assignedCores)
for (i <- 1 to numExecutors) {
val exec: ExecutorDesc = app.addExecutor(worker, coresToAssign)
//去worker中啓動Executor
launchExecutor(worker, exec)
app.state = ApplicationState.RUNNING
}
}
在 launchExecutor(worker, exec) 方法中,會獲取 Worker 的通信郵箱,給 Worker 發送啓動 Executor 的消息,具體就是啓動Executor需要多少個 core 和 內存,然後在 Worker 中有 receive 方法一直匹配 LaunchExecutor 類型:
//啓動Executor
private def launchExecutor(worker: WorkerInfo, exec: ExecutorDesc): Unit = {
logInfo("Launching executor " + exec.fullId + " on worker " + worker.id)
worker.addExecutor(exec)
/**
* 獲取Worker的通信郵箱,給Worker發送啓動Executor【多少core和多少內存】
* 在Worker中有receive 方法一直匹配 LaunchExecutor 類型
*/
worker.endpoint.send(LaunchExecutor(masterUrl,
exec.application.id, exec.id, exec.application.desc, exec.cores, exec.memory))
exec.application.driver.send(
ExecutorAdded(exec.id, worker.id, worker.hostPort, exec.cores, exec.memory))
}
在 worker 的處理方法中,會創建 ExecutorRunner 類對象,參數 appDesc 中有 Command("org.apache.spark.executor.CoarseGrainedExecutorBackend",…),其中第一個參數就是Executor類
//創建ExecutorRunner
val manager = new ExecutorRunner(
appId,
execId,
/**
* appDesc 中有 Command("org.apache.spark.executor.CoarseGrainedExecutorBackend",.......) 中
* 第一個參數就是Executor類
*/
appDesc.copy(command = Worker.maybeUpdateSSLSettings(appDesc.command, conf)),
cores_,
memory_,
self,
workerId,
host,
webUi.boundPort,
publicAddress,
sparkHome,
executorDir,
workerUri,
conf,
appLocalDirs, ExecutorState.RUNNING)
緊接着就是啓動通過 manager 啓動Executor,啓動的就是 CoarseGrainedExecutorBackend 類,下面看 CoarseGrainedExecutorBackend 類中的 main 方法有反向註冊給Driver:
//註冊Executor的通信郵箱,會調用CoarseGrainedExecutorBackend的onstart方法
env.rpcEnv.setupEndpoint("Executor", new CoarseGrainedExecutorBackend(
env.rpcEnv, driverUrl, executorId, hostname, cores, userClassPath, env))
會調用 CoarseGrainedExecutorBackend 類的 onStart() 方法:
override def onStart() {
logInfo("Connecting to driver: " + driverUrl)
//從RPC中拿到Driver的引用,給Driver反向註冊Executor
rpcEnv.asyncSetupEndpointRefByURI(driverUrl).flatMap { ref =>
// This is a very fast action so we can use "ThreadUtils.sameThread"
//拿到Driver的引用
driver = Some(ref)
/**
* 給Driver反向註冊Executor信息,這裏就是註冊給之前看到的 CoarseGrainedSchedulerBackend 類中的DriverEndpoint
* DriverEndpoint類中會有receiveAndReply 方法來匹配RegisterExecutor
*/
ref.ask[Boolean](RegisterExecutor(executorId, self, hostname, cores, extractLogUrls))
}(ThreadUtils.sameThread).onComplete {
// This is a very fast action so we can use "ThreadUtils.sameThread"
case Success(msg) =>
// Always receive `true`. Just ignore it
case Failure(e) =>
exitExecutor(1, s"Cannot register with driver: $driverUrl", e, notifyDriver = false)
}(ThreadUtils.sameThread)
}
在該方法中從 RPC 中拿到 Driver 的引用,將 Executor 反向註冊給 Driver,方法中的 ref 就是 CoarseGrainedSchedulerBackend 類的引用,之後在 CoarseGrainedSchedulerBackend 中找到匹配 RegisterExecutor 的 case,用於反向註冊:
/**
* 拿到Execuotr的通信郵箱,發送消息給ExecutorRef 告訴 Executor已經被註冊。
* 在 CoarseGrainedExecutorBackend 類中 receive方法一直監聽有沒有被註冊,匹配上就會啓動Executor
*
*/
executorRef.send(RegisteredExecutor)
在 Driver 端告訴 Execuotr 端已經被註冊,匹配上就會啓動 Executor。去看 Execuotr 端匹配 RegisteredExecutor 的 case,用於啓動 Executor:
//匹配上Driver端發過來的消息,已經接受註冊Executor了,下面要啓動Executor
case RegisteredExecutor =>
logInfo("Successfully registered with driver")
try {
//下面創建Executor,Executor真正的創建Executor,Executor中有線程池用於task運行
executor = new Executor(executorId, hostname, env, userClassPath, isLocal = false)
} catch {
case NonFatal(e) =>
exitExecutor(1, "Unable to create executor due to " + e.getMessage, e)
}
Executor 在被告知反向註冊成功之後,開始真正的創建 Executor,Executor 中有線程池用於task運行:
/**
* Executor 中的線程池
*/
private val threadPool = {
val threadFactory = new ThreadFactoryBuilder()
.setDaemon(true)
.setNameFormat("Executor task launch worker-%d")
.setThreadFactory(new ThreadFactory {
override def newThread(r: Runnable): Thread =
// Use UninterruptibleThread to run tasks so that we can allow running codes without being
// interrupted by `Thread.interrupt()`. Some issues, such as KAFKA-1894, HADOOP-10622,
// will hang forever if some methods are interrupted.
new UninterruptibleThread(r, "unused") // thread name will be set by ThreadFactoryBuilder
})
.build()
Executors.newCachedThreadPool(threadFactory).asInstanceOf[ThreadPoolExecutor]
}
至此,Executor 創建完畢,開始 Spark 的任務調度過程。
五、spark 任務調度源碼解析
之前的資源調度源碼已經跟進到了 Executor 創建完畢,並且在 Executor 中創建了 ThreadPool,此時就是開始了執行調度,我們隨機找一個 Action 算子,count() 來分析源碼:
def count(): Long = sc.runJob(this, Utils.getIteratorSize _).sum
sc 是 SparkContext 的對象,進入 runJob 方法:
def runJob[T, U: ClassTag](rdd: RDD[T], func: Iterator[T] => U): Array[U] = {
runJob(rdd, func, 0 until rdd.partitions.length)
}
再次進入 runJob 方法:
def runJob[T, U: ClassTag](
rdd: RDD[T],
func: Iterator[T] => U,
partitions: Seq[Int]): Array[U] = {
val cleanedFunc = clean(func)
runJob(rdd, (ctx: TaskContext, it: Iterator[T]) => cleanedFunc(it), partitions)
}
再次進入 runJob 方法:
def runJob[T, U: ClassTag](
rdd: RDD[T],
func: (TaskContext, Iterator[T]) => U,
partitions: Seq[Int]): Array[U] = {
val results = new Array[U](partitions.size)
runJob[T, U](rdd, func, partitions, (index, res) => results(index) = res)
results
}
再進此進入 runJob 方法,發現了 rdd.doCheckpoint() 方法,該方法的功能就是對整個 RDD 進行回溯;在 runJob 中傳了一個參數 rdd,跟進該參數,在 dagScheduler.runJob(rdd, cleanedFunc, partitions, callSite, resultHandler, localProperties.get) 中調用:
def runJob[T, U: ClassTag](
rdd: RDD[T],
func: (TaskContext, Iterator[T]) => U,
partitions: Seq[Int],
resultHandler: (Int, U) => Unit): Unit = {
if (stopped.get()) {
throw new IllegalStateException("SparkContext has been shutdown")
}
val callSite = getCallSite
val cleanedFunc = clean(func)
logInfo("Starting job: " + callSite.shortForm)
if (conf.getBoolean("spark.logLineage", false)) {
logInfo("RDD's recursive dependencies:\n" + rdd.toDebugString)
}
dagScheduler.runJob(rdd, cleanedFunc, partitions, callSite, resultHandler, localProperties.get)
progressBar.foreach(_.finishAll())
rdd.doCheckpoint()
}
進入方法後繼續跟進 rdd 這個參數,在 submitJob 中調用,跟進 submitJob 方法中的 rdd 參數在何時使用:
//提交任務。eventProcessLoop 是 DAGSchedulerEventProcessLoop 對象
eventProcessLoop.post(JobSubmitted(
jobId, rdd, func2, partitions.toArray, callSite, waiter,
SerializationUtils.clone(properties)))
rdd 傳進了 eventProcessLoop.post 方法,進入 post 方法,該方法內部調用了 eventQueue.put(event),功能是將
提交的任務放入隊列:
//將提交的任務放入隊列
def post(event: E): Unit = {
eventQueue.put(event)
}
觀察 eventQueue 類型:
private val eventQueue: BlockingQueue[E] = new LinkedBlockingDeque[E]()
eventQueue 存在於 EventLoop 類中,該類有一個 run() 方法,內部邏輯是處理剛剛被 eventProcessLoop.post 進去的新的任務,處理邏輯被封裝到 onReceive(event) 方法中,進入該方法無法看到代碼邏輯,此時去找 eventProcessLoop 的父類,跟進代碼後發現 eventProcessLoop 是 DAGSchedulerEventProcessLoop 實例對象:
private[scheduler] val eventProcessLoop = new DAGSchedulerEventProcessLoop(this)
在 DAGSchedulerEventProcessLoop 類中找到 onReceive 方法,觀察邏輯:
override def onReceive(event: DAGSchedulerEvent): Unit = {
val timerContext = timer.time()
try {
doOnReceive(event)
} finally {
timerContext.stop()
}
}
在 doOnReceive(event) 中運行 handleJobSubmitted 方法:
// 運行handleJobSubmitted 方法
dagScheduler.handleJobSubmitted(jobId, rdd, func, partitions, callSite, listener, properties)
在 handleJobSubmitted 方法中通過 submitStage(finalStage) 根據寬窄依賴遞歸尋找 Stage 並提交:
//遞歸尋找stage
private def submitStage(stage: Stage) {
val jobId = activeJobForStage(stage)
if (jobId.isDefined) {
logDebug("submitStage(" + stage + ")")
if (!waitingStages(stage) && !runningStages(stage) && !failedStages(stage)) {
val missing = getMissingParentStages(stage).sortBy(_.id)
logDebug("missing: " + missing)
if (missing.isEmpty) {
logInfo("Submitting " + stage + " (" + stage.rdd + "), which has no missing parents")
//最終會執行 submitMissingTasks 方法
submitMissingTasks(stage, jobId.get)
} else {
for (parent <- missing) {
submitStage(parent)
}
waitingStages += stage
}
}
} else {
abortStage(stage, "No active job for stage " + stage.id, None)
}
}
劃分完 Stage 之後,submitMissingTasks(stage, jobId.get) 方法會將 stage 劃分成 task 發送到 Exeuctor 中執行,並以以 TaskSet 形式提交任務:
//以taskSet形式提交任務
taskScheduler.submitTasks(new TaskSet(
tasks.toArray, stage.id, stage.latestInfo.attemptNumber, jobId, properties))
提交task, 最後會執行 backend.reviveOffers() 調用的是 CoarseGrainedSchedulerBackend 對象中的方法:
backend.reviveOffers()
給 Driver 提交 task,在當前類中的 DriverEndpoint 中有 receive 方法來接收數據:
override def reviveOffers() {
//給Driver 提交task,在當前類中的DriverEndpoint中 有receive方法來接收數據
driverEndpoint.send(ReviveOffers)
}
找到 makeOffers():
case ReviveOffers =>
makeOffers()
在該方法內部會去執行 launchTasks(taskDescs) 就是去 Executor 中啓動Task:
private def makeOffers() {
// Make sure no executor is killed while some task is launching on it
val taskDescs = CoarseGrainedSchedulerBackend.this.synchronized {
// Filter out executors under killing
val activeExecutors = executorDataMap.filterKeys(executorIsAlive)
val workOffers = activeExecutors.map {
case (id, executorData) =>
new WorkerOffer(id, executorData.executorHost, executorData.freeCores)
}.toIndexedSeq
scheduler.resourceOffers(workOffers)
}
if (!taskDescs.isEmpty) {
//去Executor中啓動Task
launchTasks(taskDescs)
}
}
在 launchTasks(taskDescs) 中就是給 Executor 發送 task 執行:
executorData.executorEndpoint.send(LaunchTask(new SerializableBuffer(serializedTask)))
在 CoarseGrainedExecutorBackend 中會有receive方法匹配 LaunchTask:
//啓動Task
case LaunchTask(data) =>
if (executor == null) {
exitExecutor(1, "Received LaunchTask command but executor was null")
} else {
val taskDesc = TaskDescription.decode(data.value)
logInfo("Got assigned task " + taskDesc.taskId)
//Executor 啓動Task
executor.launchTask(this, taskDesc)
}
得到 task,在線程池中之行:
// 得到task 在線程池中執行
def launchTask(context: ExecutorBackend, taskDescription: TaskDescription): Unit = {
val tr = new TaskRunner(context, taskDescription)
runningTasks.put(taskDescription.taskId, tr)
threadPool.execute(tr)
}
至此,任務調度完畢。
六、總結
一、Spark資源調度流程圖:
流程詳解如下:
- 集羣啓動之後,Worker節點會向Master節點彙報資源情況,Master就掌握了集羣資源情況。
- 當Spark提交一個Application後,會根據RDD之間的依賴關係將Application形成一個DAG有向無環圖。任務提交之後,Spark會在Driver端創建兩個對象:DAGScheduler和TaskScheduler
- DAGScheduler是任務調度的高層調度器,是一個對象。DAGScheduler的主要作用就是將DAG根據RDD之間的寬窄依賴關係劃分爲一個個的Stage,然後將這些Stage以TaskSet的形式提交給TaskScheduler(TaskScheduler是任務調度的低層調度器,這裏TaskSet其實就是一個集合,裏面封裝的就是一個個的task任務,也就是stage中的並行度task任務)
- TaskSchedule會遍歷TaskSet集合,拿到每個task後會將task發送到計算節點Executor中去執行(其實就是發送到Executor中的線程池ThreadPool去執行)。
- task在Executor線程池中的運行情況會向TaskScheduler反饋,當task執行失敗時,則由TaskScheduler負責重試,將task重新發送給Executor去執行,默認重試3次。如果重試3次依然失敗,那麼這個task所在的stage就失敗了。stage失敗了則由DAGScheduler來負責重試,重新發送TaskSet到TaskSchdeuler,Stage默認重試4次。如果重試4次以後依然失敗,那麼這個job就失敗了。job失敗了,Application就失敗了。因此一個task默認情況下重試3*4=12次。
- TaskScheduler不僅能重試失敗的task,還會重試straggling(落後,緩慢)task(也就是執行速度比其他task慢太多的task)。如果有運行緩慢的task那麼TaskScheduler會啓動一個新的task來與這個運行緩慢的task執行相同的處理邏輯。兩個task哪個先執行完,就以哪個task的執行結果爲準。這就是Spark的推測執行機制。在Spark中推測執行默認是關閉的。推測執行可以通過spark.speculation屬性來配置。
注意:
1.對於ETL類型要入數據庫的業務要關閉推測執行機制,這樣就不會有重複的數據入庫
2.如果遇到數據傾斜的情況,開啓推測機制則有可能導致一直會有task重新啓動處理相同的邏輯,任務可能一直處於處理不完的狀態
二、粗粒度資源申請和細粒度資源申請
粗粒度資源申請(Spark):
在Applicatioin執行之前,將所有的資源申請完畢,當資源申請成功後,纔會進行任務的調度,當所有的task執行完成後,纔會釋放這部分資源。
優點:在Application執行之前,所有的資源都申請完畢,每一個task直接使用資源就可以了,不需要task在執行前自己去申請資源,task啓動就快了,stage執行就快了,job就快了,application執行就快了
缺點:直到最後一個task執行完成纔會釋放資源,集羣的資源無法充分利用
細粒度資源申請(MapReduce):
Application執行之前不需要先去申請資源,而是直接執行,讓job中的每一個task在執行前自己去申請資源,task執行完成後,就釋放資源
優點:集羣資源可以充分利用
缺點:task自己去申請資源,task啓動變慢,Application的運行就相應的變慢了。