You are viewing a plain text version of this content. The canonical link for it is here.
Posted to commits@spark.apache.org by pw...@apache.org on 2014/01/04 01:31:31 UTC
[4/6] merge upstream/master
http://git-wip-us.apache.org/repos/asf/incubator-spark/blob/8ddbd531/yarn/alpha/src/main/scala/org/apache/spark/deploy/yarn/YarnAllocationHandler.scala
----------------------------------------------------------------------
diff --cc yarn/alpha/src/main/scala/org/apache/spark/deploy/yarn/YarnAllocationHandler.scala
index 0000000,c8af653..e91257b
mode 000000,100644..100644
--- a/yarn/alpha/src/main/scala/org/apache/spark/deploy/yarn/YarnAllocationHandler.scala
+++ b/yarn/alpha/src/main/scala/org/apache/spark/deploy/yarn/YarnAllocationHandler.scala
@@@ -1,0 -1,680 +1,680 @@@
+ /*
+ * Licensed to the Apache Software Foundation (ASF) under one or more
+ * contributor license agreements. See the NOTICE file distributed with
+ * this work for additional information regarding copyright ownership.
+ * The ASF licenses this file to You under the Apache License, Version 2.0
+ * (the "License"); you may not use this file except in compliance with
+ * the License. You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+ package org.apache.spark.deploy.yarn
+
+ import java.lang.{Boolean => JBoolean}
+ import java.util.{Collections, Set => JSet}
+ import java.util.concurrent.{CopyOnWriteArrayList, ConcurrentHashMap}
+ import java.util.concurrent.atomic.AtomicInteger
+
+ import scala.collection
+ import scala.collection.JavaConversions._
+ import scala.collection.mutable.{ArrayBuffer, HashMap, HashSet}
+
+ import org.apache.spark.{Logging, SparkConf}
+ import org.apache.spark.scheduler.{SplitInfo,TaskSchedulerImpl}
+ import org.apache.spark.scheduler.cluster.CoarseGrainedSchedulerBackend
+ import org.apache.spark.util.Utils
+
+ import org.apache.hadoop.conf.Configuration
+ import org.apache.hadoop.yarn.api.AMRMProtocol
+ import org.apache.hadoop.yarn.api.records.{AMResponse, ApplicationAttemptId}
+ import org.apache.hadoop.yarn.api.records.{Container, ContainerId, ContainerStatus}
+ import org.apache.hadoop.yarn.api.records.{Priority, Resource, ResourceRequest}
+ import org.apache.hadoop.yarn.api.protocolrecords.{AllocateRequest, AllocateResponse}
+ import org.apache.hadoop.yarn.util.{RackResolver, Records}
+
+
+ object AllocationType extends Enumeration {
+ type AllocationType = Value
+ val HOST, RACK, ANY = Value
+ }
+
+ // TODO:
+ // Too many params.
+ // Needs to be mt-safe
+ // Need to refactor this to make it 'cleaner' ... right now, all computation is reactive - should
+ // make it more proactive and decoupled.
+
+ // Note that right now, we assume all node asks as uniform in terms of capabilities and priority
+ // Refer to http://developer.yahoo.com/blogs/hadoop/posts/2011/03/mapreduce-nextgen-scheduler/ for
+ // more info on how we are requesting for containers.
+ private[yarn] class YarnAllocationHandler(
+ val conf: Configuration,
+ val resourceManager: AMRMProtocol,
+ val appAttemptId: ApplicationAttemptId,
+ val maxWorkers: Int,
+ val workerMemory: Int,
+ val workerCores: Int,
+ val preferredHostToCount: Map[String, Int],
+ val preferredRackToCount: Map[String, Int],
+ val sparkConf: SparkConf)
+ extends Logging {
+ // These three are locked on allocatedHostToContainersMap. Complementary data structures
+ // allocatedHostToContainersMap : containers which are running : host, Set<containerid>
+ // allocatedContainerToHostMap: container to host mapping.
+ private val allocatedHostToContainersMap =
+ new HashMap[String, collection.mutable.Set[ContainerId]]()
+
+ private val allocatedContainerToHostMap = new HashMap[ContainerId, String]()
+
+ // allocatedRackCount is populated ONLY if allocation happens (or decremented if this is an
+ // allocated node)
+ // As with the two data structures above, tightly coupled with them, and to be locked on
+ // allocatedHostToContainersMap
+ private val allocatedRackCount = new HashMap[String, Int]()
+
+ // Containers which have been released.
+ private val releasedContainerList = new CopyOnWriteArrayList[ContainerId]()
+ // Containers to be released in next request to RM
+ private val pendingReleaseContainers = new ConcurrentHashMap[ContainerId, Boolean]
+
+ private val numWorkersRunning = new AtomicInteger()
+ // Used to generate a unique id per worker
+ private val workerIdCounter = new AtomicInteger()
+ private val lastResponseId = new AtomicInteger()
+ private val numWorkersFailed = new AtomicInteger()
+
+ def getNumWorkersRunning: Int = numWorkersRunning.intValue
+
+ def getNumWorkersFailed: Int = numWorkersFailed.intValue
+
+ def isResourceConstraintSatisfied(container: Container): Boolean = {
+ container.getResource.getMemory >= (workerMemory + YarnAllocationHandler.MEMORY_OVERHEAD)
+ }
+
+ def allocateContainers(workersToRequest: Int) {
+ // We need to send the request only once from what I understand ... but for now, not modifying
+ // this much.
+
+ // Keep polling the Resource Manager for containers
+ val amResp = allocateWorkerResources(workersToRequest).getAMResponse
+
+ val _allocatedContainers = amResp.getAllocatedContainers()
+
+ if (_allocatedContainers.size > 0) {
+ logDebug("""
+ Allocated containers: %d
+ Current worker count: %d
+ Containers released: %s
+ Containers to be released: %s
+ Cluster resources: %s
+ """.format(
+ _allocatedContainers.size,
+ numWorkersRunning.get(),
+ releasedContainerList,
+ pendingReleaseContainers,
+ amResp.getAvailableResources))
+
+ val hostToContainers = new HashMap[String, ArrayBuffer[Container]]()
+
+ // Ignore if not satisfying constraints {
+ for (container <- _allocatedContainers) {
+ if (isResourceConstraintSatisfied(container)) {
+ // allocatedContainers += container
+
+ val host = container.getNodeId.getHost
+ val containers = hostToContainers.getOrElseUpdate(host, new ArrayBuffer[Container]())
+
+ containers += container
+ }
+ // Add all ignored containers to released list
+ else releasedContainerList.add(container.getId())
+ }
+
+ // Find the appropriate containers to use. Slightly non trivial groupBy ...
+ val dataLocalContainers = new HashMap[String, ArrayBuffer[Container]]()
+ val rackLocalContainers = new HashMap[String, ArrayBuffer[Container]]()
+ val offRackContainers = new HashMap[String, ArrayBuffer[Container]]()
+
+ for (candidateHost <- hostToContainers.keySet)
+ {
+ val maxExpectedHostCount = preferredHostToCount.getOrElse(candidateHost, 0)
+ val requiredHostCount = maxExpectedHostCount - allocatedContainersOnHost(candidateHost)
+
+ var remainingContainers = hostToContainers.get(candidateHost).getOrElse(null)
+ assert(remainingContainers != null)
+
+ if (requiredHostCount >= remainingContainers.size){
+ // Since we got <= required containers, add all to dataLocalContainers
+ dataLocalContainers.put(candidateHost, remainingContainers)
+ // all consumed
+ remainingContainers = null
+ }
+ else if (requiredHostCount > 0) {
+ // Container list has more containers than we need for data locality.
+ // Split into two : data local container count of (remainingContainers.size -
+ // requiredHostCount) and rest as remainingContainer
+ val (dataLocal, remaining) = remainingContainers.splitAt(
+ remainingContainers.size - requiredHostCount)
+ dataLocalContainers.put(candidateHost, dataLocal)
+ // remainingContainers = remaining
+
+ // yarn has nasty habit of allocating a tonne of containers on a host - discourage this :
+ // add remaining to release list. If we have insufficient containers, next allocation
+ // cycle will reallocate (but wont treat it as data local)
+ for (container <- remaining) releasedContainerList.add(container.getId())
+ remainingContainers = null
+ }
+
+ // Now rack local
+ if (remainingContainers != null){
+ val rack = YarnAllocationHandler.lookupRack(conf, candidateHost)
+
+ if (rack != null){
+ val maxExpectedRackCount = preferredRackToCount.getOrElse(rack, 0)
+ val requiredRackCount = maxExpectedRackCount - allocatedContainersOnRack(rack) -
+ rackLocalContainers.get(rack).getOrElse(List()).size
+
+
+ if (requiredRackCount >= remainingContainers.size){
+ // Add all to dataLocalContainers
+ dataLocalContainers.put(rack, remainingContainers)
+ // All consumed
+ remainingContainers = null
+ }
+ else if (requiredRackCount > 0) {
+ // container list has more containers than we need for data locality.
+ // Split into two : data local container count of (remainingContainers.size -
+ // requiredRackCount) and rest as remainingContainer
+ val (rackLocal, remaining) = remainingContainers.splitAt(
+ remainingContainers.size - requiredRackCount)
+ val existingRackLocal = rackLocalContainers.getOrElseUpdate(rack,
+ new ArrayBuffer[Container]())
+
+ existingRackLocal ++= rackLocal
+ remainingContainers = remaining
+ }
+ }
+ }
+
+ // If still not consumed, then it is off rack host - add to that list.
+ if (remainingContainers != null){
+ offRackContainers.put(candidateHost, remainingContainers)
+ }
+ }
+
+ // Now that we have split the containers into various groups, go through them in order :
+ // first host local, then rack local and then off rack (everything else).
+ // Note that the list we create below tries to ensure that not all containers end up within a
+ // host if there are sufficiently large number of hosts/containers.
+
+ val allocatedContainers = new ArrayBuffer[Container](_allocatedContainers.size)
+ allocatedContainers ++= TaskSchedulerImpl.prioritizeContainers(dataLocalContainers)
+ allocatedContainers ++= TaskSchedulerImpl.prioritizeContainers(rackLocalContainers)
+ allocatedContainers ++= TaskSchedulerImpl.prioritizeContainers(offRackContainers)
+
+ // Run each of the allocated containers
+ for (container <- allocatedContainers) {
+ val numWorkersRunningNow = numWorkersRunning.incrementAndGet()
+ val workerHostname = container.getNodeId.getHost
+ val containerId = container.getId
+
+ assert(
+ container.getResource.getMemory >= (workerMemory + YarnAllocationHandler.MEMORY_OVERHEAD))
+
+ if (numWorkersRunningNow > maxWorkers) {
+ logInfo("""Ignoring container %s at host %s, since we already have the required number of
+ containers for it.""".format(containerId, workerHostname))
+ releasedContainerList.add(containerId)
+ // reset counter back to old value.
+ numWorkersRunning.decrementAndGet()
+ }
+ else {
+ // Deallocate + allocate can result in reusing id's wrongly - so use a different counter
+ // (workerIdCounter)
+ val workerId = workerIdCounter.incrementAndGet().toString
+ val driverUrl = "akka.tcp://spark@%s:%s/user/%s".format(
+ sparkConf.get("spark.driver.host"), sparkConf.get("spark.driver.port"),
+ CoarseGrainedSchedulerBackend.ACTOR_NAME)
+
+ logInfo("launching container on " + containerId + " host " + workerHostname)
+ // Just to be safe, simply remove it from pendingReleaseContainers.
+ // Should not be there, but ..
+ pendingReleaseContainers.remove(containerId)
+
+ val rack = YarnAllocationHandler.lookupRack(conf, workerHostname)
+ allocatedHostToContainersMap.synchronized {
+ val containerSet = allocatedHostToContainersMap.getOrElseUpdate(workerHostname,
+ new HashSet[ContainerId]())
+
+ containerSet += containerId
+ allocatedContainerToHostMap.put(containerId, workerHostname)
+ if (rack != null) {
+ allocatedRackCount.put(rack, allocatedRackCount.getOrElse(rack, 0) + 1)
+ }
+ }
+
+ new Thread(
- new WorkerRunnable(container, conf, driverUrl, workerId,
++ new WorkerRunnable(container, conf, sparkConf, driverUrl, workerId,
+ workerHostname, workerMemory, workerCores)
+ ).start()
+ }
+ }
+ logDebug("""
+ Finished processing %d containers.
+ Current number of workers running: %d,
+ releasedContainerList: %s,
+ pendingReleaseContainers: %s
+ """.format(
+ allocatedContainers.size,
+ numWorkersRunning.get(),
+ releasedContainerList,
+ pendingReleaseContainers))
+ }
+
+
+ val completedContainers = amResp.getCompletedContainersStatuses()
+ if (completedContainers.size > 0){
+ logDebug("Completed %d containers, to-be-released: %s".format(
+ completedContainers.size, releasedContainerList))
+ for (completedContainer <- completedContainers){
+ val containerId = completedContainer.getContainerId
+
+ // Was this released by us ? If yes, then simply remove from containerSet and move on.
+ if (pendingReleaseContainers.containsKey(containerId)) {
+ pendingReleaseContainers.remove(containerId)
+ }
+ else {
+ // Simply decrement count - next iteration of ReporterThread will take care of allocating.
+ numWorkersRunning.decrementAndGet()
+ logInfo("Completed container %s (state: %s, exit status: %s)".format(
+ containerId,
+ completedContainer.getState,
+ completedContainer.getExitStatus()))
+ // Hadoop 2.2.X added a ContainerExitStatus we should switch to use
+ // there are some exit status' we shouldn't necessarily count against us, but for
+ // now I think its ok as none of the containers are expected to exit
+ if (completedContainer.getExitStatus() != 0) {
+ logInfo("Container marked as failed: " + containerId)
+ numWorkersFailed.incrementAndGet()
+ }
+ }
+
+ allocatedHostToContainersMap.synchronized {
+ if (allocatedContainerToHostMap.containsKey(containerId)) {
+ val host = allocatedContainerToHostMap.get(containerId).getOrElse(null)
+ assert (host != null)
+
+ val containerSet = allocatedHostToContainersMap.get(host).getOrElse(null)
+ assert (containerSet != null)
+
+ containerSet -= containerId
+ if (containerSet.isEmpty) allocatedHostToContainersMap.remove(host)
+ else allocatedHostToContainersMap.update(host, containerSet)
+
+ allocatedContainerToHostMap -= containerId
+
+ // Doing this within locked context, sigh ... move to outside ?
+ val rack = YarnAllocationHandler.lookupRack(conf, host)
+ if (rack != null) {
+ val rackCount = allocatedRackCount.getOrElse(rack, 0) - 1
+ if (rackCount > 0) allocatedRackCount.put(rack, rackCount)
+ else allocatedRackCount.remove(rack)
+ }
+ }
+ }
+ }
+ logDebug("""
+ Finished processing %d completed containers.
+ Current number of workers running: %d,
+ releasedContainerList: %s,
+ pendingReleaseContainers: %s
+ """.format(
+ completedContainers.size,
+ numWorkersRunning.get(),
+ releasedContainerList,
+ pendingReleaseContainers))
+ }
+ }
+
+ def createRackResourceRequests(hostContainers: List[ResourceRequest]): List[ResourceRequest] = {
+ // First generate modified racks and new set of hosts under it : then issue requests
+ val rackToCounts = new HashMap[String, Int]()
+
+ // Within this lock - used to read/write to the rack related maps too.
+ for (container <- hostContainers) {
+ val candidateHost = container.getHostName
+ val candidateNumContainers = container.getNumContainers
+ assert(YarnAllocationHandler.ANY_HOST != candidateHost)
+
+ val rack = YarnAllocationHandler.lookupRack(conf, candidateHost)
+ if (rack != null) {
+ var count = rackToCounts.getOrElse(rack, 0)
+ count += candidateNumContainers
+ rackToCounts.put(rack, count)
+ }
+ }
+
+ val requestedContainers: ArrayBuffer[ResourceRequest] =
+ new ArrayBuffer[ResourceRequest](rackToCounts.size)
+ for ((rack, count) <- rackToCounts){
+ requestedContainers +=
+ createResourceRequest(AllocationType.RACK, rack, count, YarnAllocationHandler.PRIORITY)
+ }
+
+ requestedContainers.toList
+ }
+
+ def allocatedContainersOnHost(host: String): Int = {
+ var retval = 0
+ allocatedHostToContainersMap.synchronized {
+ retval = allocatedHostToContainersMap.getOrElse(host, Set()).size
+ }
+ retval
+ }
+
+ def allocatedContainersOnRack(rack: String): Int = {
+ var retval = 0
+ allocatedHostToContainersMap.synchronized {
+ retval = allocatedRackCount.getOrElse(rack, 0)
+ }
+ retval
+ }
+
+ private def allocateWorkerResources(numWorkers: Int): AllocateResponse = {
+
+ var resourceRequests: List[ResourceRequest] = null
+
+ // default.
+ if (numWorkers <= 0 || preferredHostToCount.isEmpty) {
+ logDebug("numWorkers: " + numWorkers + ", host preferences: " + preferredHostToCount.isEmpty)
+ resourceRequests = List(
+ createResourceRequest(AllocationType.ANY, null, numWorkers, YarnAllocationHandler.PRIORITY))
+ }
+ else {
+ // request for all hosts in preferred nodes and for numWorkers -
+ // candidates.size, request by default allocation policy.
+ val hostContainerRequests: ArrayBuffer[ResourceRequest] =
+ new ArrayBuffer[ResourceRequest](preferredHostToCount.size)
+ for ((candidateHost, candidateCount) <- preferredHostToCount) {
+ val requiredCount = candidateCount - allocatedContainersOnHost(candidateHost)
+
+ if (requiredCount > 0) {
+ hostContainerRequests += createResourceRequest(
+ AllocationType.HOST,
+ candidateHost,
+ requiredCount,
+ YarnAllocationHandler.PRIORITY)
+ }
+ }
+ val rackContainerRequests: List[ResourceRequest] = createRackResourceRequests(
+ hostContainerRequests.toList)
+
+ val anyContainerRequests: ResourceRequest = createResourceRequest(
+ AllocationType.ANY,
+ resource = null,
+ numWorkers,
+ YarnAllocationHandler.PRIORITY)
+
+ val containerRequests: ArrayBuffer[ResourceRequest] = new ArrayBuffer[ResourceRequest](
+ hostContainerRequests.size + rackContainerRequests.size + 1)
+
+ containerRequests ++= hostContainerRequests
+ containerRequests ++= rackContainerRequests
+ containerRequests += anyContainerRequests
+
+ resourceRequests = containerRequests.toList
+ }
+
+ val req = Records.newRecord(classOf[AllocateRequest])
+ req.setResponseId(lastResponseId.incrementAndGet)
+ req.setApplicationAttemptId(appAttemptId)
+
+ req.addAllAsks(resourceRequests)
+
+ val releasedContainerList = createReleasedContainerList()
+ req.addAllReleases(releasedContainerList)
+
+ if (numWorkers > 0) {
+ logInfo("Allocating %d worker containers with %d of memory each.".format(numWorkers,
+ workerMemory + YarnAllocationHandler.MEMORY_OVERHEAD))
+ }
+ else {
+ logDebug("Empty allocation req .. release : " + releasedContainerList)
+ }
+
+ for (request <- resourceRequests) {
+ logInfo("ResourceRequest (host : %s, num containers: %d, priority = %s , capability : %s)".
+ format(
+ request.getHostName,
+ request.getNumContainers,
+ request.getPriority,
+ request.getCapability))
+ }
+ resourceManager.allocate(req)
+ }
+
+
+ private def createResourceRequest(
+ requestType: AllocationType.AllocationType,
+ resource:String,
+ numWorkers: Int,
+ priority: Int): ResourceRequest = {
+
+ // If hostname specified, we need atleast two requests - node local and rack local.
+ // There must be a third request - which is ANY : that will be specially handled.
+ requestType match {
+ case AllocationType.HOST => {
+ assert(YarnAllocationHandler.ANY_HOST != resource)
+ val hostname = resource
+ val nodeLocal = createResourceRequestImpl(hostname, numWorkers, priority)
+
+ // Add to host->rack mapping
+ YarnAllocationHandler.populateRackInfo(conf, hostname)
+
+ nodeLocal
+ }
+ case AllocationType.RACK => {
+ val rack = resource
+ createResourceRequestImpl(rack, numWorkers, priority)
+ }
+ case AllocationType.ANY => createResourceRequestImpl(
+ YarnAllocationHandler.ANY_HOST, numWorkers, priority)
+ case _ => throw new IllegalArgumentException(
+ "Unexpected/unsupported request type: " + requestType)
+ }
+ }
+
+ private def createResourceRequestImpl(
+ hostname:String,
+ numWorkers: Int,
+ priority: Int): ResourceRequest = {
+
+ val rsrcRequest = Records.newRecord(classOf[ResourceRequest])
+ val memCapability = Records.newRecord(classOf[Resource])
+ // There probably is some overhead here, let's reserve a bit more memory.
+ memCapability.setMemory(workerMemory + YarnAllocationHandler.MEMORY_OVERHEAD)
+ rsrcRequest.setCapability(memCapability)
+
+ val pri = Records.newRecord(classOf[Priority])
+ pri.setPriority(priority)
+ rsrcRequest.setPriority(pri)
+
+ rsrcRequest.setHostName(hostname)
+
+ rsrcRequest.setNumContainers(java.lang.Math.max(numWorkers, 0))
+ rsrcRequest
+ }
+
+ def createReleasedContainerList(): ArrayBuffer[ContainerId] = {
+
+ val retval = new ArrayBuffer[ContainerId](1)
+ // Iterator on COW list ...
+ for (container <- releasedContainerList.iterator()){
+ retval += container
+ }
+ // Remove from the original list.
+ if (! retval.isEmpty) {
+ releasedContainerList.removeAll(retval)
+ for (v <- retval) pendingReleaseContainers.put(v, true)
+ logInfo("Releasing " + retval.size + " containers. pendingReleaseContainers : " +
+ pendingReleaseContainers)
+ }
+
+ retval
+ }
+ }
+
+ object YarnAllocationHandler {
+
+ val ANY_HOST = "*"
+ // All requests are issued with same priority : we do not (yet) have any distinction between
+ // request types (like map/reduce in hadoop for example)
+ val PRIORITY = 1
+
+ // Additional memory overhead - in mb
+ val MEMORY_OVERHEAD = 384
+
+ // Host to rack map - saved from allocation requests
+ // We are expecting this not to change.
+ // Note that it is possible for this to change : and RM will indicate that to us via update
+ // response to allocate. But we are punting on handling that for now.
+ private val hostToRack = new ConcurrentHashMap[String, String]()
+ private val rackToHostSet = new ConcurrentHashMap[String, JSet[String]]()
+
+
+ def newAllocator(
+ conf: Configuration,
+ resourceManager: AMRMProtocol,
+ appAttemptId: ApplicationAttemptId,
+ args: ApplicationMasterArguments,
+ sparkConf: SparkConf): YarnAllocationHandler = {
+
+ new YarnAllocationHandler(
+ conf,
+ resourceManager,
+ appAttemptId,
+ args.numWorkers,
+ args.workerMemory,
+ args.workerCores,
+ Map[String, Int](),
+ Map[String, Int](),
+ sparkConf)
+ }
+
+ def newAllocator(
+ conf: Configuration,
+ resourceManager: AMRMProtocol,
+ appAttemptId: ApplicationAttemptId,
+ args: ApplicationMasterArguments,
+ map: collection.Map[String,
+ collection.Set[SplitInfo]],
+ sparkConf: SparkConf): YarnAllocationHandler = {
+
+ val (hostToCount, rackToCount) = generateNodeToWeight(conf, map)
+ new YarnAllocationHandler(
+ conf,
+ resourceManager,
+ appAttemptId,
+ args.numWorkers,
+ args.workerMemory,
+ args.workerCores,
+ hostToCount,
+ rackToCount,
+ sparkConf)
+ }
+
+ def newAllocator(
+ conf: Configuration,
+ resourceManager: AMRMProtocol,
+ appAttemptId: ApplicationAttemptId,
+ maxWorkers: Int,
+ workerMemory: Int,
+ workerCores: Int,
+ map: collection.Map[String, collection.Set[SplitInfo]],
+ sparkConf: SparkConf): YarnAllocationHandler = {
+
+ val (hostToCount, rackToCount) = generateNodeToWeight(conf, map)
+
+ new YarnAllocationHandler(
+ conf,
+ resourceManager,
+ appAttemptId,
+ maxWorkers,
+ workerMemory,
+ workerCores,
+ hostToCount,
+ rackToCount,
+ sparkConf)
+ }
+
+ // A simple method to copy the split info map.
+ private def generateNodeToWeight(
+ conf: Configuration,
+ input: collection.Map[String, collection.Set[SplitInfo]]) :
+ // host to count, rack to count
+ (Map[String, Int], Map[String, Int]) = {
+
+ if (input == null) return (Map[String, Int](), Map[String, Int]())
+
+ val hostToCount = new HashMap[String, Int]
+ val rackToCount = new HashMap[String, Int]
+
+ for ((host, splits) <- input) {
+ val hostCount = hostToCount.getOrElse(host, 0)
+ hostToCount.put(host, hostCount + splits.size)
+
+ val rack = lookupRack(conf, host)
+ if (rack != null){
+ val rackCount = rackToCount.getOrElse(host, 0)
+ rackToCount.put(host, rackCount + splits.size)
+ }
+ }
+
+ (hostToCount.toMap, rackToCount.toMap)
+ }
+
+ def lookupRack(conf: Configuration, host: String): String = {
+ if (!hostToRack.contains(host)) populateRackInfo(conf, host)
+ hostToRack.get(host)
+ }
+
+ def fetchCachedHostsForRack(rack: String): Option[Set[String]] = {
+ val set = rackToHostSet.get(rack)
+ if (set == null) return None
+
+ // No better way to get a Set[String] from JSet ?
+ val convertedSet: collection.mutable.Set[String] = set
+ Some(convertedSet.toSet)
+ }
+
+ def populateRackInfo(conf: Configuration, hostname: String) {
+ Utils.checkHost(hostname)
+
+ if (!hostToRack.containsKey(hostname)) {
+ // If there are repeated failures to resolve, all to an ignore list ?
+ val rackInfo = RackResolver.resolve(conf, hostname)
+ if (rackInfo != null && rackInfo.getNetworkLocation != null) {
+ val rack = rackInfo.getNetworkLocation
+ hostToRack.put(hostname, rack)
+ if (! rackToHostSet.containsKey(rack)) {
+ rackToHostSet.putIfAbsent(rack,
+ Collections.newSetFromMap(new ConcurrentHashMap[String, JBoolean]()))
+ }
+ rackToHostSet.get(rack).add(hostname)
+
+ // TODO(harvey): Figure out this comment...
+ // Since RackResolver caches, we are disabling this for now ...
+ } /* else {
+ // right ? Else we will keep calling rack resolver in case we cant resolve rack info ...
+ hostToRack.put(hostname, null)
+ } */
+ }
+ }
+ }
http://git-wip-us.apache.org/repos/asf/incubator-spark/blob/8ddbd531/yarn/common/src/main/scala/org/apache/spark/deploy/yarn/ClientArguments.scala
----------------------------------------------------------------------
diff --cc yarn/common/src/main/scala/org/apache/spark/deploy/yarn/ClientArguments.scala
index 0000000,7aac232..1419f21
mode 000000,100644..100644
--- a/yarn/common/src/main/scala/org/apache/spark/deploy/yarn/ClientArguments.scala
+++ b/yarn/common/src/main/scala/org/apache/spark/deploy/yarn/ClientArguments.scala
@@@ -1,0 -1,150 +1,150 @@@
+ /*
+ * Licensed to the Apache Software Foundation (ASF) under one or more
+ * contributor license agreements. See the NOTICE file distributed with
+ * this work for additional information regarding copyright ownership.
+ * The ASF licenses this file to You under the Apache License, Version 2.0
+ * (the "License"); you may not use this file except in compliance with
+ * the License. You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+ package org.apache.spark.deploy.yarn
+
+ import scala.collection.mutable.{ArrayBuffer, HashMap}
+
+ import org.apache.spark.SparkConf
+ import org.apache.spark.scheduler.{InputFormatInfo, SplitInfo}
+ import org.apache.spark.util.IntParam
+ import org.apache.spark.util.MemoryParam
+
+
+ // TODO: Add code and support for ensuring that yarn resource 'tasks' are location aware !
-class ClientArguments(val args: Array[String]) {
++class ClientArguments(val args: Array[String], val sparkConf: SparkConf) {
+ var addJars: String = null
+ var files: String = null
+ var archives: String = null
+ var userJar: String = null
+ var userClass: String = null
+ var userArgs: Seq[String] = Seq[String]()
+ var workerMemory = 1024 // MB
+ var workerCores = 1
+ var numWorkers = 2
- var amQueue = new SparkConf().get("QUEUE", "default")
++ var amQueue = sparkConf.get("QUEUE", "default")
+ var amMemory: Int = 512 // MB
+ var amClass: String = "org.apache.spark.deploy.yarn.ApplicationMaster"
+ var appName: String = "Spark"
+ // TODO
+ var inputFormatInfo: List[InputFormatInfo] = null
+ // TODO(harvey)
+ var priority = 0
+
+ parseArgs(args.toList)
+
+ private def parseArgs(inputArgs: List[String]): Unit = {
+ val userArgsBuffer: ArrayBuffer[String] = new ArrayBuffer[String]()
+ val inputFormatMap: HashMap[String, InputFormatInfo] = new HashMap[String, InputFormatInfo]()
+
+ var args = inputArgs
+
+ while (!args.isEmpty) {
+ args match {
+ case ("--jar") :: value :: tail =>
+ userJar = value
+ args = tail
+
+ case ("--class") :: value :: tail =>
+ userClass = value
+ args = tail
+
+ case ("--args") :: value :: tail =>
+ userArgsBuffer += value
+ args = tail
+
+ case ("--master-class") :: value :: tail =>
+ amClass = value
+ args = tail
+
+ case ("--master-memory") :: MemoryParam(value) :: tail =>
+ amMemory = value
+ args = tail
+
+ case ("--num-workers") :: IntParam(value) :: tail =>
+ numWorkers = value
+ args = tail
+
+ case ("--worker-memory") :: MemoryParam(value) :: tail =>
+ workerMemory = value
+ args = tail
+
+ case ("--worker-cores") :: IntParam(value) :: tail =>
+ workerCores = value
+ args = tail
+
+ case ("--queue") :: value :: tail =>
+ amQueue = value
+ args = tail
+
+ case ("--name") :: value :: tail =>
+ appName = value
+ args = tail
+
+ case ("--addJars") :: value :: tail =>
+ addJars = value
+ args = tail
+
+ case ("--files") :: value :: tail =>
+ files = value
+ args = tail
+
+ case ("--archives") :: value :: tail =>
+ archives = value
+ args = tail
+
+ case Nil =>
+ if (userJar == null || userClass == null) {
+ printUsageAndExit(1)
+ }
+
+ case _ =>
+ printUsageAndExit(1, args)
+ }
+ }
+
+ userArgs = userArgsBuffer.readOnly
+ inputFormatInfo = inputFormatMap.values.toList
+ }
+
+
+ def printUsageAndExit(exitCode: Int, unknownParam: Any = null) {
+ if (unknownParam != null) {
+ System.err.println("Unknown/unsupported param " + unknownParam)
+ }
+ System.err.println(
+ "Usage: org.apache.spark.deploy.yarn.Client [options] \n" +
+ "Options:\n" +
+ " --jar JAR_PATH Path to your application's JAR file (required)\n" +
+ " --class CLASS_NAME Name of your application's main class (required)\n" +
+ " --args ARGS Arguments to be passed to your application's main class.\n" +
+ " Mutliple invocations are possible, each will be passed in order.\n" +
+ " --num-workers NUM Number of workers to start (Default: 2)\n" +
+ " --worker-cores NUM Number of cores for the workers (Default: 1). This is unsused right now.\n" +
+ " --master-class CLASS_NAME Class Name for Master (Default: spark.deploy.yarn.ApplicationMaster)\n" +
+ " --master-memory MEM Memory for Master (e.g. 1000M, 2G) (Default: 512 Mb)\n" +
+ " --worker-memory MEM Memory per Worker (e.g. 1000M, 2G) (Default: 1G)\n" +
+ " --name NAME The name of your application (Default: Spark)\n" +
+ " --queue QUEUE The hadoop queue to use for allocation requests (Default: 'default')\n" +
+ " --addJars jars Comma separated list of local jars that want SparkContext.addJar to work with.\n" +
+ " --files files Comma separated list of files to be distributed with the job.\n" +
+ " --archives archives Comma separated list of archives to be distributed with the job."
+ )
+ System.exit(exitCode)
+ }
+
+ }
http://git-wip-us.apache.org/repos/asf/incubator-spark/blob/8ddbd531/yarn/common/src/main/scala/org/apache/spark/scheduler/cluster/YarnClientSchedulerBackend.scala
----------------------------------------------------------------------
diff --cc yarn/common/src/main/scala/org/apache/spark/scheduler/cluster/YarnClientSchedulerBackend.scala
index 0000000,4b69f50..324ef46
mode 000000,100644..100644
--- a/yarn/common/src/main/scala/org/apache/spark/scheduler/cluster/YarnClientSchedulerBackend.scala
+++ b/yarn/common/src/main/scala/org/apache/spark/scheduler/cluster/YarnClientSchedulerBackend.scala
@@@ -1,0 -1,110 +1,110 @@@
+ /*
+ * Licensed to the Apache Software Foundation (ASF) under one or more
+ * contributor license agreements. See the NOTICE file distributed with
+ * this work for additional information regarding copyright ownership.
+ * The ASF licenses this file to You under the Apache License, Version 2.0
+ * (the "License"); you may not use this file except in compliance with
+ * the License. You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+ package org.apache.spark.scheduler.cluster
+
+ import org.apache.hadoop.yarn.api.records.{ApplicationId, YarnApplicationState}
+ import org.apache.spark.{SparkException, Logging, SparkContext}
+ import org.apache.spark.deploy.yarn.{Client, ClientArguments}
+ import org.apache.spark.scheduler.TaskSchedulerImpl
+
+ private[spark] class YarnClientSchedulerBackend(
+ scheduler: TaskSchedulerImpl,
+ sc: SparkContext)
+ extends CoarseGrainedSchedulerBackend(scheduler, sc.env.actorSystem)
+ with Logging {
+
+ var client: Client = null
+ var appId: ApplicationId = null
+
+ override def start() {
+ super.start()
+
+ val defalutWorkerCores = "2"
+ val defalutWorkerMemory = "512m"
+ val defaultWorkerNumber = "1"
+
+ val userJar = System.getenv("SPARK_YARN_APP_JAR")
+ var workerCores = System.getenv("SPARK_WORKER_CORES")
+ var workerMemory = System.getenv("SPARK_WORKER_MEMORY")
+ var workerNumber = System.getenv("SPARK_WORKER_INSTANCES")
+
+ if (userJar == null)
+ throw new SparkException("env SPARK_YARN_APP_JAR is not set")
+
+ if (workerCores == null)
+ workerCores = defalutWorkerCores
+ if (workerMemory == null)
+ workerMemory = defalutWorkerMemory
+ if (workerNumber == null)
+ workerNumber = defaultWorkerNumber
+
+ val driverHost = conf.get("spark.driver.host")
+ val driverPort = conf.get("spark.driver.port")
+ val hostport = driverHost + ":" + driverPort
+
+ val argsArray = Array[String](
+ "--class", "notused",
+ "--jar", userJar,
+ "--args", hostport,
+ "--worker-memory", workerMemory,
+ "--worker-cores", workerCores,
+ "--num-workers", workerNumber,
+ "--master-class", "org.apache.spark.deploy.yarn.WorkerLauncher"
+ )
+
- val args = new ClientArguments(argsArray)
- client = new Client(args)
++ val args = new ClientArguments(argsArray, conf)
++ client = new Client(args, conf)
+ appId = client.runApp()
+ waitForApp()
+ }
+
+ def waitForApp() {
+
+ // TODO : need a better way to find out whether the workers are ready or not
+ // maybe by resource usage report?
+ while(true) {
+ val report = client.getApplicationReport(appId)
+
+ logInfo("Application report from ASM: \n" +
+ "\t appMasterRpcPort: " + report.getRpcPort() + "\n" +
+ "\t appStartTime: " + report.getStartTime() + "\n" +
+ "\t yarnAppState: " + report.getYarnApplicationState() + "\n"
+ )
+
+ // Ready to go, or already gone.
+ val state = report.getYarnApplicationState()
+ if (state == YarnApplicationState.RUNNING) {
+ return
+ } else if (state == YarnApplicationState.FINISHED ||
+ state == YarnApplicationState.FAILED ||
+ state == YarnApplicationState.KILLED) {
+ throw new SparkException("Yarn application already ended," +
+ "might be killed or not able to launch application master.")
+ }
+
+ Thread.sleep(1000)
+ }
+ }
+
+ override def stop() {
+ super.stop()
+ client.stop()
+ logInfo("Stoped")
+ }
+
+ }
http://git-wip-us.apache.org/repos/asf/incubator-spark/blob/8ddbd531/yarn/stable/src/main/scala/org/apache/spark/deploy/yarn/ApplicationMaster.scala
----------------------------------------------------------------------
diff --cc yarn/stable/src/main/scala/org/apache/spark/deploy/yarn/ApplicationMaster.scala
index 0000000,7c32e0a..69ae14c
mode 000000,100644..100644
--- a/yarn/stable/src/main/scala/org/apache/spark/deploy/yarn/ApplicationMaster.scala
+++ b/yarn/stable/src/main/scala/org/apache/spark/deploy/yarn/ApplicationMaster.scala
@@@ -1,0 -1,428 +1,432 @@@
+ /*
+ * Licensed to the Apache Software Foundation (ASF) under one or more
+ * contributor license agreements. See the NOTICE file distributed with
+ * this work for additional information regarding copyright ownership.
+ * The ASF licenses this file to You under the Apache License, Version 2.0
+ * (the "License"); you may not use this file except in compliance with
+ * the License. You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+ package org.apache.spark.deploy.yarn
+
+ import java.io.IOException
+ import java.net.Socket
+ import java.util.concurrent.CopyOnWriteArrayList
+ import java.util.concurrent.atomic.{AtomicInteger, AtomicReference}
+
+ import scala.collection.JavaConversions._
+
+ import org.apache.hadoop.conf.Configuration
+ import org.apache.hadoop.fs.{FileSystem, Path}
+ import org.apache.hadoop.net.NetUtils
+ import org.apache.hadoop.security.UserGroupInformation
+ import org.apache.hadoop.util.ShutdownHookManager
+ import org.apache.hadoop.yarn.api._
+ import org.apache.hadoop.yarn.api.protocolrecords._
+ import org.apache.hadoop.yarn.api.records._
+ import org.apache.hadoop.yarn.client.api.AMRMClient
+ import org.apache.hadoop.yarn.client.api.AMRMClient.ContainerRequest
+ import org.apache.hadoop.yarn.conf.YarnConfiguration
+ import org.apache.hadoop.yarn.ipc.YarnRPC
+ import org.apache.hadoop.yarn.util.{ConverterUtils, Records}
+
+ import org.apache.spark.{SparkConf, SparkContext, Logging}
+ import org.apache.spark.util.Utils
+
+
-class ApplicationMaster(args: ApplicationMasterArguments, conf: Configuration) extends Logging {
++class ApplicationMaster(args: ApplicationMasterArguments, conf: Configuration,
++ sparkConf: SparkConf) extends Logging {
+
- def this(args: ApplicationMasterArguments) = this(args, new Configuration())
++ def this(args: ApplicationMasterArguments, sparkConf: SparkConf) =
++ this(args, new Configuration(), sparkConf)
++
++ def this(args: ApplicationMasterArguments) = this(args, new SparkConf())
+
- private var rpc: YarnRPC = YarnRPC.create(conf)
+ private val yarnConf: YarnConfiguration = new YarnConfiguration(conf)
+ private var appAttemptId: ApplicationAttemptId = _
+ private var userThread: Thread = _
+ private val fs = FileSystem.get(yarnConf)
+
+ private var yarnAllocator: YarnAllocationHandler = _
+ private var isFinished: Boolean = false
+ private var uiAddress: String = _
+ private val maxAppAttempts: Int = conf.getInt(
+ YarnConfiguration.RM_AM_MAX_ATTEMPTS, YarnConfiguration.DEFAULT_RM_AM_MAX_ATTEMPTS)
+ private var isLastAMRetry: Boolean = true
+ private var amClient: AMRMClient[ContainerRequest] = _
+
- private val sparkConf = new SparkConf()
+ // Default to numWorkers * 2, with minimum of 3
+ private val maxNumWorkerFailures = sparkConf.getInt("spark.yarn.max.worker.failures",
+ math.max(args.numWorkers * 2, 3))
+
+ def run() {
+ // Setup the directories so things go to YARN approved directories rather
+ // than user specified and /tmp.
+ System.setProperty("spark.local.dir", getLocalDirs())
+
+ // set the web ui port to be ephemeral for yarn so we don't conflict with
+ // other spark processes running on the same box
+ System.setProperty("spark.ui.port", "0")
+
+ // Use priority 30 as it's higher then HDFS. It's same priority as MapReduce is using.
+ ShutdownHookManager.get().addShutdownHook(new AppMasterShutdownHook(this), 30)
+
+ appAttemptId = getApplicationAttemptId()
+ isLastAMRetry = appAttemptId.getAttemptId() >= maxAppAttempts
+ amClient = AMRMClient.createAMRMClient()
+ amClient.init(yarnConf)
+ amClient.start()
+
+ // Workaround until hadoop moves to something which has
+ // https://issues.apache.org/jira/browse/HADOOP-8406 - fixed in (2.0.2-alpha but no 0.23 line)
+ // org.apache.hadoop.io.compress.CompressionCodecFactory.getCodecClasses(conf)
+
+ ApplicationMaster.register(this)
+
+ // Start the user's JAR
+ userThread = startUserClass()
+
+ // This a bit hacky, but we need to wait until the spark.driver.port property has
+ // been set by the Thread executing the user class.
+ waitForSparkContextInitialized()
+
+ // Do this after Spark master is up and SparkContext is created so that we can register UI Url.
+ val appMasterResponse: RegisterApplicationMasterResponse = registerApplicationMaster()
+
+ // Allocate all containers
+ allocateWorkers()
+
+ // Wait for the user class to Finish
+ userThread.join()
+
+ System.exit(0)
+ }
+
+ /** Get the Yarn approved local directories. */
+ private def getLocalDirs(): String = {
+ // Hadoop 0.23 and 2.x have different Environment variable names for the
+ // local dirs, so lets check both. We assume one of the 2 is set.
+ // LOCAL_DIRS => 2.X, YARN_LOCAL_DIRS => 0.23.X
+ val localDirs = Option(System.getenv("YARN_LOCAL_DIRS"))
+ .getOrElse(Option(System.getenv("LOCAL_DIRS"))
- .getOrElse(""))
++ .getOrElse(""))
+
+ if (localDirs.isEmpty()) {
+ throw new Exception("Yarn Local dirs can't be empty")
+ }
+ localDirs
+ }
+
+ private def getApplicationAttemptId(): ApplicationAttemptId = {
+ val envs = System.getenv()
+ val containerIdString = envs.get(ApplicationConstants.Environment.CONTAINER_ID.name())
+ val containerId = ConverterUtils.toContainerId(containerIdString)
+ val appAttemptId = containerId.getApplicationAttemptId()
+ logInfo("ApplicationAttemptId: " + appAttemptId)
+ appAttemptId
+ }
+
+ private def registerApplicationMaster(): RegisterApplicationMasterResponse = {
+ logInfo("Registering the ApplicationMaster")
+ amClient.registerApplicationMaster(Utils.localHostName(), 0, uiAddress)
+ }
+
- private def startUserClass(): Thread = {
++ private def startUserClass(): Thread = {
+ logInfo("Starting the user JAR in a separate Thread")
+ val mainMethod = Class.forName(
+ args.userClass,
- false /* initialize */,
++ false /* initialize */ ,
+ Thread.currentThread.getContextClassLoader).getMethod("main", classOf[Array[String]])
+ val t = new Thread {
+ override def run() {
+ var successed = false
+ try {
+ // Copy
+ var mainArgs: Array[String] = new Array[String](args.userArgs.size)
+ args.userArgs.copyToArray(mainArgs, 0, args.userArgs.size)
+ mainMethod.invoke(null, mainArgs)
+ // some job script has "System.exit(0)" at the end, for example SparkPi, SparkLR
+ // userThread will stop here unless it has uncaught exception thrown out
+ // It need shutdown hook to set SUCCEEDED
+ successed = true
+ } finally {
+ logDebug("finishing main")
+ isLastAMRetry = true
+ if (successed) {
+ ApplicationMaster.this.finishApplicationMaster(FinalApplicationStatus.SUCCEEDED)
+ } else {
+ ApplicationMaster.this.finishApplicationMaster(FinalApplicationStatus.FAILED)
+ }
+ }
+ }
+ }
+ t.start()
+ t
+ }
+
+ // This need to happen before allocateWorkers()
+ private def waitForSparkContextInitialized() {
+ logInfo("Waiting for Spark context initialization")
+ try {
+ var sparkContext: SparkContext = null
+ ApplicationMaster.sparkContextRef.synchronized {
+ var numTries = 0
+ val waitTime = 10000L
+ val maxNumTries = sparkConf.getInt("spark.yarn.applicationMaster.waitTries", 10)
+ while (ApplicationMaster.sparkContextRef.get() == null && numTries < maxNumTries) {
+ logInfo("Waiting for Spark context initialization ... " + numTries)
+ numTries = numTries + 1
+ ApplicationMaster.sparkContextRef.wait(waitTime)
+ }
+ sparkContext = ApplicationMaster.sparkContextRef.get()
+ assert(sparkContext != null || numTries >= maxNumTries)
+
+ if (sparkContext != null) {
+ uiAddress = sparkContext.ui.appUIAddress
+ this.yarnAllocator = YarnAllocationHandler.newAllocator(
+ yarnConf,
+ amClient,
+ appAttemptId,
+ args,
+ sparkContext.preferredNodeLocationData,
+ sparkContext.getConf)
+ } else {
+ logWarning("Unable to retrieve SparkContext inspite of waiting for %d, maxNumTries = %d".
+ format(numTries * waitTime, maxNumTries))
+ this.yarnAllocator = YarnAllocationHandler.newAllocator(
+ yarnConf,
+ amClient,
+ appAttemptId,
+ args,
+ sparkContext.getConf)
+ }
+ }
+ } finally {
+ // In case of exceptions, etc - ensure that count is at least ALLOCATOR_LOOP_WAIT_COUNT :
+ // so that the loop (in ApplicationMaster.sparkContextInitialized) breaks.
+ ApplicationMaster.incrementAllocatorLoop(ApplicationMaster.ALLOCATOR_LOOP_WAIT_COUNT)
+ }
+ }
+
+ private def allocateWorkers() {
+ try {
+ logInfo("Allocating " + args.numWorkers + " workers.")
+ // Wait until all containers have finished
+ // TODO: This is a bit ugly. Can we make it nicer?
+ // TODO: Handle container failure
+ yarnAllocator.addResourceRequests(args.numWorkers)
+ // Exits the loop if the user thread exits.
+ while (yarnAllocator.getNumWorkersRunning < args.numWorkers && userThread.isAlive) {
+ if (yarnAllocator.getNumWorkersFailed >= maxNumWorkerFailures) {
+ finishApplicationMaster(FinalApplicationStatus.FAILED,
+ "max number of worker failures reached")
+ }
+ yarnAllocator.allocateResources()
+ ApplicationMaster.incrementAllocatorLoop(1)
+ Thread.sleep(100)
+ }
+ } finally {
+ // In case of exceptions, etc - ensure that count is at least ALLOCATOR_LOOP_WAIT_COUNT,
+ // so that the loop in ApplicationMaster#sparkContextInitialized() breaks.
+ ApplicationMaster.incrementAllocatorLoop(ApplicationMaster.ALLOCATOR_LOOP_WAIT_COUNT)
+ }
+ logInfo("All workers have launched.")
+
+ // Launch a progress reporter thread, else the app will get killed after expiration
+ // (def: 10mins) timeout.
+ if (userThread.isAlive) {
+ // Ensure that progress is sent before YarnConfiguration.RM_AM_EXPIRY_INTERVAL_MS elapses.
+ val timeoutInterval = yarnConf.getInt(YarnConfiguration.RM_AM_EXPIRY_INTERVAL_MS, 120000)
+
+ // we want to be reasonably responsive without causing too many requests to RM.
+ val schedulerInterval =
+ sparkConf.getLong("spark.yarn.scheduler.heartbeat.interval-ms", 5000)
+
+
+ // must be <= timeoutInterval / 2.
+ val interval = math.min(timeoutInterval / 2, schedulerInterval)
+
+ launchReporterThread(interval)
+ }
+ }
+
+ private def launchReporterThread(_sleepTime: Long): Thread = {
- val sleepTime = if (_sleepTime <= 0 ) 0 else _sleepTime
++ val sleepTime = if (_sleepTime <= 0) 0 else _sleepTime
+
+ val t = new Thread {
+ override def run() {
+ while (userThread.isAlive) {
+ if (yarnAllocator.getNumWorkersFailed >= maxNumWorkerFailures) {
+ finishApplicationMaster(FinalApplicationStatus.FAILED,
+ "max number of worker failures reached")
+ }
+ val missingWorkerCount = args.numWorkers - yarnAllocator.getNumWorkersRunning -
+ yarnAllocator.getNumPendingAllocate
+ if (missingWorkerCount > 0) {
+ logInfo("Allocating %d containers to make up for (potentially) lost containers".
+ format(missingWorkerCount))
+ yarnAllocator.addResourceRequests(missingWorkerCount)
+ }
+ sendProgress()
+ Thread.sleep(sleepTime)
+ }
+ }
+ }
+ // Setting to daemon status, though this is usually not a good idea.
+ t.setDaemon(true)
+ t.start()
+ logInfo("Started progress reporter thread - sleep time : " + sleepTime)
+ t
+ }
+
+ private def sendProgress() {
+ logDebug("Sending progress")
+ // Simulated with an allocate request with no nodes requested.
+ yarnAllocator.allocateResources()
+ }
+
+ /*
+ def printContainers(containers: List[Container]) = {
+ for (container <- containers) {
+ logInfo("Launching shell command on a new container."
+ + ", containerId=" + container.getId()
+ + ", containerNode=" + container.getNodeId().getHost()
+ + ":" + container.getNodeId().getPort()
+ + ", containerNodeURI=" + container.getNodeHttpAddress()
+ + ", containerState" + container.getState()
+ + ", containerResourceMemory"
+ + container.getResource().getMemory())
+ }
+ }
+ */
+
+ def finishApplicationMaster(status: FinalApplicationStatus, diagnostics: String = "") {
+ synchronized {
+ if (isFinished) {
+ return
+ }
+ isFinished = true
+ }
+
+ logInfo("finishApplicationMaster with " + status)
+ // Set tracking URL to empty since we don't have a history server.
- amClient.unregisterApplicationMaster(status, "" /* appMessage */, "" /* appTrackingUrl */)
++ amClient.unregisterApplicationMaster(status, "" /* appMessage */ , "" /* appTrackingUrl */)
+ }
+
+ /**
+ * Clean up the staging directory.
+ */
+ private def cleanupStagingDir() {
+ var stagingDirPath: Path = null
+ try {
+ val preserveFiles = sparkConf.get("spark.yarn.preserve.staging.files", "false").toBoolean
+ if (!preserveFiles) {
+ stagingDirPath = new Path(System.getenv("SPARK_YARN_STAGING_DIR"))
+ if (stagingDirPath == null) {
+ logError("Staging directory is null")
+ return
+ }
+ logInfo("Deleting staging directory " + stagingDirPath)
+ fs.delete(stagingDirPath, true)
+ }
+ } catch {
+ case ioe: IOException =>
+ logError("Failed to cleanup staging dir " + stagingDirPath, ioe)
+ }
+ }
+
+ // The shutdown hook that runs when a signal is received AND during normal close of the JVM.
+ class AppMasterShutdownHook(appMaster: ApplicationMaster) extends Runnable {
+
+ def run() {
+ logInfo("AppMaster received a signal.")
+ // we need to clean up staging dir before HDFS is shut down
+ // make sure we don't delete it until this is the last AM
+ if (appMaster.isLastAMRetry) appMaster.cleanupStagingDir()
+ }
+ }
++
+ }
+
+ object ApplicationMaster {
+ // Number of times to wait for the allocator loop to complete.
+ // Each loop iteration waits for 100ms, so maximum of 3 seconds.
+ // This is to ensure that we have reasonable number of containers before we start
+ // TODO: Currently, task to container is computed once (TaskSetManager) - which need not be
+ // optimal as more containers are available. Might need to handle this better.
+ private val ALLOCATOR_LOOP_WAIT_COUNT = 30
+
+ private val applicationMasters = new CopyOnWriteArrayList[ApplicationMaster]()
+
+ val sparkContextRef: AtomicReference[SparkContext] =
+ new AtomicReference[SparkContext](null /* initialValue */)
+
+ val yarnAllocatorLoop: AtomicInteger = new AtomicInteger(0)
+
+ def incrementAllocatorLoop(by: Int) {
+ val count = yarnAllocatorLoop.getAndAdd(by)
+ if (count >= ALLOCATOR_LOOP_WAIT_COUNT) {
+ yarnAllocatorLoop.synchronized {
+ // to wake threads off wait ...
+ yarnAllocatorLoop.notifyAll()
+ }
+ }
+ }
+
+ def register(master: ApplicationMaster) {
+ applicationMasters.add(master)
+ }
+
+ // TODO(harvey): See whether this should be discarded - it isn't used anywhere atm...
+ def sparkContextInitialized(sc: SparkContext): Boolean = {
+ var modified = false
+ sparkContextRef.synchronized {
+ modified = sparkContextRef.compareAndSet(null, sc)
+ sparkContextRef.notifyAll()
+ }
+
+ // Add a shutdown hook - as a best case effort in case users do not call sc.stop or do
+ // System.exit.
+ // Should not really have to do this, but it helps YARN to evict resources earlier.
+ // Not to mention, prevent the Client from declaring failure even though we exited properly.
+ // Note that this will unfortunately not properly clean up the staging files because it gets
+ // called too late, after the filesystem is already shutdown.
+ if (modified) {
+ Runtime.getRuntime().addShutdownHook(new Thread with Logging {
+ // This is not only logs, but also ensures that log system is initialized for this instance
+ // when we are actually 'run'-ing.
+ logInfo("Adding shutdown hook for context " + sc)
++
+ override def run() {
+ logInfo("Invoking sc stop from shutdown hook")
+ sc.stop()
+ // Best case ...
+ for (master <- applicationMasters) {
+ master.finishApplicationMaster(FinalApplicationStatus.SUCCEEDED)
+ }
+ }
- } )
++ })
+ }
+
+ // Wait for initialization to complete and atleast 'some' nodes can get allocated.
+ yarnAllocatorLoop.synchronized {
+ while (yarnAllocatorLoop.get() <= ALLOCATOR_LOOP_WAIT_COUNT) {
+ yarnAllocatorLoop.wait(1000L)
+ }
+ }
+ modified
+ }
+
+ def main(argStrings: Array[String]) {
+ val args = new ApplicationMasterArguments(argStrings)
+ new ApplicationMaster(args).run()
+ }
+ }
http://git-wip-us.apache.org/repos/asf/incubator-spark/blob/8ddbd531/yarn/stable/src/main/scala/org/apache/spark/deploy/yarn/Client.scala
----------------------------------------------------------------------
diff --cc yarn/stable/src/main/scala/org/apache/spark/deploy/yarn/Client.scala
index 0000000,a750668..440ad5c
mode 000000,100644..100644
--- a/yarn/stable/src/main/scala/org/apache/spark/deploy/yarn/Client.scala
+++ b/yarn/stable/src/main/scala/org/apache/spark/deploy/yarn/Client.scala
@@@ -1,0 -1,523 +1,525 @@@
+ /*
+ * Licensed to the Apache Software Foundation (ASF) under one or more
+ * contributor license agreements. See the NOTICE file distributed with
+ * this work for additional information regarding copyright ownership.
+ * The ASF licenses this file to You under the Apache License, Version 2.0
+ * (the "License"); you may not use this file except in compliance with
+ * the License. You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+ package org.apache.spark.deploy.yarn
+
+ import java.net.{InetAddress, UnknownHostException, URI}
+ import java.nio.ByteBuffer
+
+ import scala.collection.JavaConversions._
+ import scala.collection.mutable.HashMap
+ import scala.collection.mutable.Map
+
+ import org.apache.hadoop.conf.Configuration
+ import org.apache.hadoop.fs.{FileContext, FileStatus, FileSystem, Path, FileUtil}
+ import org.apache.hadoop.fs.permission.FsPermission;
+ import org.apache.hadoop.io.DataOutputBuffer
+ import org.apache.hadoop.mapred.Master
+ import org.apache.hadoop.net.NetUtils
+ import org.apache.hadoop.security.UserGroupInformation
+ import org.apache.hadoop.yarn.api._
+ import org.apache.hadoop.yarn.api.ApplicationConstants.Environment
+ import org.apache.hadoop.yarn.api.protocolrecords._
+ import org.apache.hadoop.yarn.api.records._
+ import org.apache.hadoop.yarn.client.api.impl.YarnClientImpl
+ import org.apache.hadoop.yarn.conf.YarnConfiguration
+ import org.apache.hadoop.yarn.ipc.YarnRPC
+ import org.apache.hadoop.yarn.util.{Apps, Records}
+
+ import org.apache.spark.{Logging, SparkConf}
+ import org.apache.spark.util.Utils
+ import org.apache.spark.deploy.SparkHadoopUtil
+
+
+ /**
+ * The entry point (starting in Client#main() and Client#run()) for launching Spark on YARN. The
+ * Client submits an application to the global ResourceManager to launch Spark's ApplicationMaster,
+ * which will launch a Spark master process and negotiate resources throughout its duration.
+ */
-class Client(conf: Configuration, args: ClientArguments) extends YarnClientImpl with Logging {
++class Client(args: ClientArguments, conf: Configuration, sparkConf: SparkConf)
++ extends YarnClientImpl with Logging {
++
++ def this(args: ClientArguments, sparkConf: SparkConf) =
++ this(args, new Configuration(), sparkConf)
++
++ def this(args: ClientArguments) = this(args, new SparkConf())
+
+ var rpc: YarnRPC = YarnRPC.create(conf)
+ val yarnConf: YarnConfiguration = new YarnConfiguration(conf)
+ val credentials = UserGroupInformation.getCurrentUser().getCredentials()
+ private val SPARK_STAGING: String = ".sparkStaging"
+ private val distCacheMgr = new ClientDistributedCacheManager()
- private val sparkConf = new SparkConf
-
+
+ // Staging directory is private! -> rwx--------
+ val STAGING_DIR_PERMISSION: FsPermission = FsPermission.createImmutable(0700: Short)
+ // App files are world-wide readable and owner writable -> rw-r--r--
+ val APP_FILE_PERMISSION: FsPermission = FsPermission.createImmutable(0644: Short)
+
- def this(args: ClientArguments) = this(new Configuration(), args)
-
+ def runApp(): ApplicationId = {
+ validateArgs()
+ // Initialize and start the client service.
+ init(yarnConf)
+ start()
+
+ // Log details about this YARN cluster (e.g, the number of slave machines/NodeManagers).
+ logClusterResourceDetails()
+
+ // Prepare to submit a request to the ResourcManager (specifically its ApplicationsManager (ASM)
+ // interface).
+
+ // Get a new client application.
+ val newApp = super.createApplication()
+ val newAppResponse = newApp.getNewApplicationResponse()
+ val appId = newAppResponse.getApplicationId()
+
+ verifyClusterResources(newAppResponse)
+
+ // Set up resource and environment variables.
+ val appStagingDir = getAppStagingDir(appId)
+ val localResources = prepareLocalResources(appStagingDir)
+ val launchEnv = setupLaunchEnv(localResources, appStagingDir)
+ val amContainer = createContainerLaunchContext(newAppResponse, localResources, launchEnv)
+
+ // Set up an application submission context.
+ val appContext = newApp.getApplicationSubmissionContext()
+ appContext.setApplicationName(args.appName)
+ appContext.setQueue(args.amQueue)
+ appContext.setAMContainerSpec(amContainer)
+
+ // Memory for the ApplicationMaster.
+ val memoryResource = Records.newRecord(classOf[Resource]).asInstanceOf[Resource]
+ memoryResource.setMemory(args.amMemory + YarnAllocationHandler.MEMORY_OVERHEAD)
+ appContext.setResource(memoryResource)
+
+ // Finally, submit and monitor the application.
+ submitApp(appContext)
+ appId
+ }
+
+ def run() {
+ val appId = runApp()
+ monitorApplication(appId)
+ System.exit(0)
+ }
+
+ // TODO(harvey): This could just go in ClientArguments.
+ def validateArgs() = {
+ Map(
+ (System.getenv("SPARK_JAR") == null) -> "Error: You must set SPARK_JAR environment variable!",
+ (args.userJar == null) -> "Error: You must specify a user jar!",
+ (args.userClass == null) -> "Error: You must specify a user class!",
+ (args.numWorkers <= 0) -> "Error: You must specify atleast 1 worker!",
+ (args.amMemory <= YarnAllocationHandler.MEMORY_OVERHEAD) -> ("Error: AM memory size must be" +
+ "greater than: " + YarnAllocationHandler.MEMORY_OVERHEAD),
+ (args.workerMemory <= YarnAllocationHandler.MEMORY_OVERHEAD) -> ("Error: Worker memory size" +
+ "must be greater than: " + YarnAllocationHandler.MEMORY_OVERHEAD.toString)
+ ).foreach { case(cond, errStr) =>
+ if (cond) {
+ logError(errStr)
+ args.printUsageAndExit(1)
+ }
+ }
+ }
+
+ def getAppStagingDir(appId: ApplicationId): String = {
+ SPARK_STAGING + Path.SEPARATOR + appId.toString() + Path.SEPARATOR
+ }
+
+ def logClusterResourceDetails() {
+ val clusterMetrics: YarnClusterMetrics = super.getYarnClusterMetrics
+ logInfo("Got Cluster metric info from ApplicationsManager (ASM), number of NodeManagers: " +
+ clusterMetrics.getNumNodeManagers)
+
+ val queueInfo: QueueInfo = super.getQueueInfo(args.amQueue)
- logInfo("""Queue info ... queueName: %s, queueCurrentCapacity: %s, queueMaxCapacity: %s,
++ logInfo( """Queue info ... queueName: %s, queueCurrentCapacity: %s, queueMaxCapacity: %s,
+ queueApplicationCount = %s, queueChildQueueCount = %s""".format(
+ queueInfo.getQueueName,
+ queueInfo.getCurrentCapacity,
+ queueInfo.getMaximumCapacity,
+ queueInfo.getApplications.size,
+ queueInfo.getChildQueues.size))
+ }
+
+ def verifyClusterResources(app: GetNewApplicationResponse) = {
+ val maxMem = app.getMaximumResourceCapability().getMemory()
+ logInfo("Max mem capabililty of a single resource in this cluster " + maxMem)
+
+ // If we have requested more then the clusters max for a single resource then exit.
+ if (args.workerMemory > maxMem) {
+ logError("Required worker memory (%d MB), is above the max threshold (%d MB) of this cluster.".
+ format(args.workerMemory, maxMem))
+ System.exit(1)
+ }
+ val amMem = args.amMemory + YarnAllocationHandler.MEMORY_OVERHEAD
+ if (amMem > maxMem) {
+ logError("Required AM memory (%d) is above the max threshold (%d) of this cluster".
+ format(args.amMemory, maxMem))
+ System.exit(1)
+ }
+
+ // We could add checks to make sure the entire cluster has enough resources but that involves
+ // getting all the node reports and computing ourselves.
+ }
+
+ /** See if two file systems are the same or not. */
+ private def compareFs(srcFs: FileSystem, destFs: FileSystem): Boolean = {
+ val srcUri = srcFs.getUri()
+ val dstUri = destFs.getUri()
+ if (srcUri.getScheme() == null) {
+ return false
+ }
+ if (!srcUri.getScheme().equals(dstUri.getScheme())) {
+ return false
+ }
+ var srcHost = srcUri.getHost()
+ var dstHost = dstUri.getHost()
+ if ((srcHost != null) && (dstHost != null)) {
+ try {
+ srcHost = InetAddress.getByName(srcHost).getCanonicalHostName()
+ dstHost = InetAddress.getByName(dstHost).getCanonicalHostName()
+ } catch {
+ case e: UnknownHostException =>
+ return false
+ }
+ if (!srcHost.equals(dstHost)) {
+ return false
+ }
+ } else if (srcHost == null && dstHost != null) {
+ return false
+ } else if (srcHost != null && dstHost == null) {
+ return false
+ }
+ //check for ports
+ if (srcUri.getPort() != dstUri.getPort()) {
+ return false
+ }
+ return true
+ }
+
+ /** Copy the file into HDFS if needed. */
+ private def copyRemoteFile(
+ dstDir: Path,
+ originalPath: Path,
+ replication: Short,
+ setPerms: Boolean = false): Path = {
+ val fs = FileSystem.get(conf)
+ val remoteFs = originalPath.getFileSystem(conf)
+ var newPath = originalPath
+ if (! compareFs(remoteFs, fs)) {
+ newPath = new Path(dstDir, originalPath.getName())
+ logInfo("Uploading " + originalPath + " to " + newPath)
+ FileUtil.copy(remoteFs, originalPath, fs, newPath, false, conf)
+ fs.setReplication(newPath, replication)
+ if (setPerms) fs.setPermission(newPath, new FsPermission(APP_FILE_PERMISSION))
+ }
+ // Resolve any symlinks in the URI path so using a "current" symlink to point to a specific
+ // version shows the specific version in the distributed cache configuration
+ val qualPath = fs.makeQualified(newPath)
+ val fc = FileContext.getFileContext(qualPath.toUri(), conf)
+ val destPath = fc.resolvePath(qualPath)
+ destPath
+ }
+
+ def prepareLocalResources(appStagingDir: String): HashMap[String, LocalResource] = {
+ logInfo("Preparing Local resources")
+ // Upload Spark and the application JAR to the remote file system if necessary. Add them as
+ // local resources to the application master.
+ val fs = FileSystem.get(conf)
+
+ val delegTokenRenewer = Master.getMasterPrincipal(conf)
+ if (UserGroupInformation.isSecurityEnabled()) {
+ if (delegTokenRenewer == null || delegTokenRenewer.length() == 0) {
+ logError("Can't get Master Kerberos principal for use as renewer")
+ System.exit(1)
+ }
+ }
+ val dst = new Path(fs.getHomeDirectory(), appStagingDir)
+ val replication = sparkConf.getInt("spark.yarn.submit.file.replication", 3).toShort
+
+ if (UserGroupInformation.isSecurityEnabled()) {
+ val dstFs = dst.getFileSystem(conf)
+ dstFs.addDelegationTokens(delegTokenRenewer, credentials)
+ }
+
+ val localResources = HashMap[String, LocalResource]()
+ FileSystem.mkdirs(fs, dst, new FsPermission(STAGING_DIR_PERMISSION))
+
+ val statCache: Map[URI, FileStatus] = HashMap[URI, FileStatus]()
+
+ Map(
+ Client.SPARK_JAR -> System.getenv("SPARK_JAR"), Client.APP_JAR -> args.userJar,
+ Client.LOG4J_PROP -> System.getenv("SPARK_LOG4J_CONF")
+ ).foreach { case(destName, _localPath) =>
+ val localPath: String = if (_localPath != null) _localPath.trim() else ""
+ if (! localPath.isEmpty()) {
+ var localURI = new URI(localPath)
+ // If not specified assume these are in the local filesystem to keep behavior like Hadoop
+ if (localURI.getScheme() == null) {
+ localURI = new URI(FileSystem.getLocal(conf).makeQualified(new Path(localPath)).toString)
+ }
+ val setPermissions = if (destName.equals(Client.APP_JAR)) true else false
+ val destPath = copyRemoteFile(dst, new Path(localURI), replication, setPermissions)
+ distCacheMgr.addResource(fs, conf, destPath, localResources, LocalResourceType.FILE,
+ destName, statCache)
+ }
+ }
+
+ // Handle jars local to the ApplicationMaster.
+ if ((args.addJars != null) && (!args.addJars.isEmpty())){
+ args.addJars.split(',').foreach { case file: String =>
+ val localURI = new URI(file.trim())
+ val localPath = new Path(localURI)
+ val linkname = Option(localURI.getFragment()).getOrElse(localPath.getName())
+ val destPath = copyRemoteFile(dst, localPath, replication)
+ // Only add the resource to the Spark ApplicationMaster.
+ val appMasterOnly = true
+ distCacheMgr.addResource(fs, conf, destPath, localResources, LocalResourceType.FILE,
+ linkname, statCache, appMasterOnly)
+ }
+ }
+
+ // Handle any distributed cache files
+ if ((args.files != null) && (!args.files.isEmpty())){
+ args.files.split(',').foreach { case file: String =>
+ val localURI = new URI(file.trim())
+ val localPath = new Path(localURI)
+ val linkname = Option(localURI.getFragment()).getOrElse(localPath.getName())
+ val destPath = copyRemoteFile(dst, localPath, replication)
+ distCacheMgr.addResource(fs, conf, destPath, localResources, LocalResourceType.FILE,
+ linkname, statCache)
+ }
+ }
+
+ // Handle any distributed cache archives
+ if ((args.archives != null) && (!args.archives.isEmpty())) {
+ args.archives.split(',').foreach { case file:String =>
+ val localURI = new URI(file.trim())
+ val localPath = new Path(localURI)
+ val linkname = Option(localURI.getFragment()).getOrElse(localPath.getName())
+ val destPath = copyRemoteFile(dst, localPath, replication)
+ distCacheMgr.addResource(fs, conf, destPath, localResources, LocalResourceType.ARCHIVE,
+ linkname, statCache)
+ }
+ }
+
+ UserGroupInformation.getCurrentUser().addCredentials(credentials)
+ localResources
+ }
+
+ def setupLaunchEnv(
+ localResources: HashMap[String, LocalResource],
+ stagingDir: String): HashMap[String, String] = {
+ logInfo("Setting up the launch environment")
+ val log4jConfLocalRes = localResources.getOrElse(Client.LOG4J_PROP, null)
+
+ val env = new HashMap[String, String]()
+
- Client.populateClasspath(yarnConf, log4jConfLocalRes != null, env)
++ Client.populateClasspath(yarnConf, sparkConf, log4jConfLocalRes != null, env)
+ env("SPARK_YARN_MODE") = "true"
+ env("SPARK_YARN_STAGING_DIR") = stagingDir
+
+ // Set the environment variables to be passed on to the Workers.
+ distCacheMgr.setDistFilesEnv(env)
+ distCacheMgr.setDistArchivesEnv(env)
+
+ // Allow users to specify some environment variables.
+ Apps.setEnvFromInputString(env, System.getenv("SPARK_YARN_USER_ENV"))
+
+ // Add each SPARK_* key to the environment.
+ System.getenv().filterKeys(_.startsWith("SPARK")).foreach { case (k,v) => env(k) = v }
+
+ env
+ }
+
+ def userArgsToString(clientArgs: ClientArguments): String = {
+ val prefix = " --args "
+ val args = clientArgs.userArgs
+ val retval = new StringBuilder()
- for (arg <- args){
++ for (arg <- args) {
+ retval.append(prefix).append(" '").append(arg).append("' ")
+ }
+ retval.toString
+ }
+
+ def createContainerLaunchContext(
+ newApp: GetNewApplicationResponse,
+ localResources: HashMap[String, LocalResource],
+ env: HashMap[String, String]): ContainerLaunchContext = {
+ logInfo("Setting up container launch context")
+ val amContainer = Records.newRecord(classOf[ContainerLaunchContext])
+ amContainer.setLocalResources(localResources)
+ amContainer.setEnvironment(env)
+
+ // TODO: Need a replacement for the following code to fix -Xmx?
+ // val minResMemory: Int = newApp.getMinimumResourceCapability().getMemory()
+ // var amMemory = ((args.amMemory / minResMemory) * minResMemory) +
+ // ((if ((args.amMemory % minResMemory) == 0) 0 else minResMemory) -
+ // YarnAllocationHandler.MEMORY_OVERHEAD)
+
+ // Extra options for the JVM
+ var JAVA_OPTS = ""
+
+ // Add Xmx for AM memory
+ JAVA_OPTS += "-Xmx" + args.amMemory + "m"
+
+ val tmpDir = new Path(Environment.PWD.$(), YarnConfiguration.DEFAULT_CONTAINER_TEMP_DIR)
+ JAVA_OPTS += " -Djava.io.tmpdir=" + tmpDir
+
+ // TODO: Remove once cpuset version is pushed out.
+ // The context is, default gc for server class machines ends up using all cores to do gc -
+ // hence if there are multiple containers in same node, Spark GC affects all other containers'
+ // performance (which can be that of other Spark containers)
+ // Instead of using this, rely on cpusets by YARN to enforce "proper" Spark behavior in
+ // multi-tenant environments. Not sure how default Java GC behaves if it is limited to subset
+ // of cores on a node.
+ val useConcurrentAndIncrementalGC = env.isDefinedAt("SPARK_USE_CONC_INCR_GC") &&
+ java.lang.Boolean.parseBoolean(env("SPARK_USE_CONC_INCR_GC"))
+ if (useConcurrentAndIncrementalGC) {
+ // In our expts, using (default) throughput collector has severe perf ramifications in
+ // multi-tenant machines
+ JAVA_OPTS += " -XX:+UseConcMarkSweepGC "
+ JAVA_OPTS += " -XX:+CMSIncrementalMode "
+ JAVA_OPTS += " -XX:+CMSIncrementalPacing "
+ JAVA_OPTS += " -XX:CMSIncrementalDutyCycleMin=0 "
+ JAVA_OPTS += " -XX:CMSIncrementalDutyCycle=10 "
+ }
+
+ if (env.isDefinedAt("SPARK_JAVA_OPTS")) {
+ JAVA_OPTS += " " + env("SPARK_JAVA_OPTS")
+ }
+
+ // Command for the ApplicationMaster
+ var javaCommand = "java"
+ val javaHome = System.getenv("JAVA_HOME")
+ if ((javaHome != null && !javaHome.isEmpty()) || env.isDefinedAt("JAVA_HOME")) {
+ javaCommand = Environment.JAVA_HOME.$() + "/bin/java"
+ }
+
+ val commands = List[String](
+ javaCommand +
+ " -server " +
+ JAVA_OPTS +
+ " " + args.amClass +
+ " --class " + args.userClass +
+ " --jar " + args.userJar +
+ userArgsToString(args) +
+ " --worker-memory " + args.workerMemory +
+ " --worker-cores " + args.workerCores +
+ " --num-workers " + args.numWorkers +
+ " 1> " + ApplicationConstants.LOG_DIR_EXPANSION_VAR + "/stdout" +
+ " 2> " + ApplicationConstants.LOG_DIR_EXPANSION_VAR + "/stderr")
+
+ logInfo("Command for starting the Spark ApplicationMaster: " + commands(0))
+ amContainer.setCommands(commands)
+
+ // Setup security tokens.
+ val dob = new DataOutputBuffer()
+ credentials.writeTokenStorageToStream(dob)
+ amContainer.setTokens(ByteBuffer.wrap(dob.getData()))
+
+ amContainer
+ }
+
+ def submitApp(appContext: ApplicationSubmissionContext) = {
+ // Submit the application to the applications manager.
+ logInfo("Submitting application to ASM")
+ super.submitApplication(appContext)
+ }
+
+ def monitorApplication(appId: ApplicationId): Boolean = {
+ val interval = sparkConf.getLong("spark.yarn.report.interval", 1000)
+
+ while (true) {
+ Thread.sleep(interval)
+ val report = super.getApplicationReport(appId)
+
+ logInfo("Application report from ASM: \n" +
+ "\t application identifier: " + appId.toString() + "\n" +
+ "\t appId: " + appId.getId() + "\n" +
+ "\t clientToAMToken: " + report.getClientToAMToken() + "\n" +
+ "\t appDiagnostics: " + report.getDiagnostics() + "\n" +
+ "\t appMasterHost: " + report.getHost() + "\n" +
+ "\t appQueue: " + report.getQueue() + "\n" +
+ "\t appMasterRpcPort: " + report.getRpcPort() + "\n" +
+ "\t appStartTime: " + report.getStartTime() + "\n" +
+ "\t yarnAppState: " + report.getYarnApplicationState() + "\n" +
+ "\t distributedFinalState: " + report.getFinalApplicationStatus() + "\n" +
+ "\t appTrackingUrl: " + report.getTrackingUrl() + "\n" +
+ "\t appUser: " + report.getUser()
+ )
+
+ val state = report.getYarnApplicationState()
+ val dsStatus = report.getFinalApplicationStatus()
+ if (state == YarnApplicationState.FINISHED ||
+ state == YarnApplicationState.FAILED ||
+ state == YarnApplicationState.KILLED) {
+ return true
+ }
+ }
+ true
+ }
+ }
+
+ object Client {
+ val SPARK_JAR: String = "spark.jar"
+ val APP_JAR: String = "app.jar"
+ val LOG4J_PROP: String = "log4j.properties"
+
+ def main(argStrings: Array[String]) {
+ // Set an env variable indicating we are running in YARN mode.
+ // Note: anything env variable with SPARK_ prefix gets propagated to all (remote) processes -
+ // see Client#setupLaunchEnv().
+ System.setProperty("SPARK_YARN_MODE", "true")
++ val sparkConf = new SparkConf()
++ val args = new ClientArguments(argStrings, sparkConf)
+
- val args = new ClientArguments(argStrings)
-
- (new Client(args)).run()
++ new Client(args, sparkConf).run()
+ }
+
+ // Based on code from org.apache.hadoop.mapreduce.v2.util.MRApps
+ def populateHadoopClasspath(conf: Configuration, env: HashMap[String, String]) {
+ for (c <- conf.getStrings(YarnConfiguration.YARN_APPLICATION_CLASSPATH)) {
+ Apps.addToEnvironment(env, Environment.CLASSPATH.name, c.trim)
+ }
+ }
+
- def populateClasspath(conf: Configuration, addLog4j: Boolean, env: HashMap[String, String]) {
++ def populateClasspath(conf: Configuration, sparkConf: SparkConf, addLog4j: Boolean, env: HashMap[String, String]) {
+ Apps.addToEnvironment(env, Environment.CLASSPATH.name, Environment.PWD.$())
+ // If log4j present, ensure ours overrides all others
+ if (addLog4j) {
+ Apps.addToEnvironment(env, Environment.CLASSPATH.name, Environment.PWD.$() +
+ Path.SEPARATOR + LOG4J_PROP)
+ }
+ // Normally the users app.jar is last in case conflicts with spark jars
- val userClasspathFirst = new SparkConf().get("spark.yarn.user.classpath.first", "false")
++ val userClasspathFirst = sparkConf.get("spark.yarn.user.classpath.first", "false")
+ .toBoolean
+ if (userClasspathFirst) {
+ Apps.addToEnvironment(env, Environment.CLASSPATH.name, Environment.PWD.$() +
+ Path.SEPARATOR + APP_JAR)
+ }
+ Apps.addToEnvironment(env, Environment.CLASSPATH.name, Environment.PWD.$() +
+ Path.SEPARATOR + SPARK_JAR)
+ Client.populateHadoopClasspath(conf, env)
+
+ if (!userClasspathFirst) {
+ Apps.addToEnvironment(env, Environment.CLASSPATH.name, Environment.PWD.$() +
+ Path.SEPARATOR + APP_JAR)
+ }
+ Apps.addToEnvironment(env, Environment.CLASSPATH.name, Environment.PWD.$() +
+ Path.SEPARATOR + "*")
+ }
+ }