[SPARK-13704][CORE][YARN] Reduce rack resolution time

core/src/main/scala/org/apache/spark/scheduler/TaskSchedulerImpl.scala

@@ -156,6 +156,12 @@ private[spark] class TaskSchedulerImpl(
 
   private[scheduler] var barrierCoordinator: RpcEndpoint = null
 
+  /**
+   * It can be override in different TaskScheduler, like Yarn.
+   * None by default. This should be initialized before any invocation.
+   */
+  protected val defaultRackValue: Option[String] = None
+
   private def maybeInitBarrierCoordinator(): Unit = {
     if (barrierCoordinator == null) {
       barrierCoordinator = new BarrierCoordinator(barrierSyncTimeout, sc.listenerBus,
@@ -375,9 +381,10 @@ private[spark] class TaskSchedulerImpl(
         executorIdToRunningTaskIds(o.executorId) = HashSet[Long]()
         newExecAvail = true
       }
-      for (rack <- getRackForHost(o.host)) {
-        hostsByRack.getOrElseUpdate(rack, new HashSet[String]()) += o.host
-      }
+    }
+    val hosts = offers.map(_.host).toSet.toSeq
+    for ((host, Some(rack)) <- hosts.zip(getRacksForHosts(hosts))) {
+      hostsByRack.getOrElseUpdate(rack, new HashSet[String]()) += host
     }
 
     // Before making any offers, remove any nodes from the blacklist whose blacklist has expired. Do
@@ -812,7 +819,20 @@ private[spark] class TaskSchedulerImpl(
   }
 
   // By default, rack is unknown
-  def getRackForHost(value: String): Option[String] = None
+  def getRackForHost(hosts: String): Option[String] = {
+    getRacksForHosts(Seq(hosts)).head
+  }
+
+  /**
+   * Get racks info for hosts.
+   *
+   * The returned Sequence must have to be the same length as the hosts
+   * passed in.
+   * Return [[defaultRackValue]] sequence by default.
+   */
+  def getRacksForHosts(hosts: Seq[String]): Seq[Option[String]] = {
+    hosts.map(_ => defaultRackValue)
+  }
 
   private def waitBackendReady(): Unit = {
     if (backend.isReady) {

core/src/main/scala/org/apache/spark/scheduler/TaskSetManager.scala

@@ -186,8 +186,23 @@ private[spark] class TaskSetManager(
 
   // Add all our tasks to the pending lists. We do this in reverse order
   // of task index so that tasks with low indices get launched first.
-  for (i <- (0 until numTasks).reverse) {
-    addPendingTask(i)
+  addPendingTasks()
+
+  private def addPendingTasks(): Unit = {
+    val (_, duration) = Utils.timeTakenMs {
+      for (i <- (0 until numTasks).reverse) {
+        addPendingTask(i, resolveRacks = false)
+      }
+      // Resolve the rack for each host. This can be slow, so de-dupe the list of hosts,
+      // and assign the rack to all relevant task indices.
+      val racks = sched.getRacksForHosts(pendingTasksForHost.keySet.toSeq)
+      racks.zip(pendingTasksForHost.values).foreach {
+        case (Some(rack), indices) =>
+          pendingTasksForRack.getOrElseUpdate(rack, new ArrayBuffer) ++= indices
+        case (None, _) => // no rack, nothing to do
+      }
+    }
+    logDebug(s"Adding pending tasks took $duration ms")
   }
 
   /**
@@ -214,7 +229,9 @@ private[spark] class TaskSetManager(
   private[scheduler] var emittedTaskSizeWarning = false
 
   /** Add a task to all the pending-task lists that it should be on. */
-  private[spark] def addPendingTask(index: Int) {
+  private[spark] def addPendingTask(
+      index: Int,
+      resolveRacks: Boolean = true): Unit = {
     for (loc <- tasks(index).preferredLocations) {
       loc match {
         case e: ExecutorCacheTaskLocation =>
@@ -234,8 +251,11 @@ private[spark] class TaskSetManager(
         case _ =>
       }
       pendingTasksForHost.getOrElseUpdate(loc.host, new ArrayBuffer) += index
-      for (rack <- sched.getRackForHost(loc.host)) {
-        pendingTasksForRack.getOrElseUpdate(rack, new ArrayBuffer) += index
+
+      if (resolveRacks) {
+        sched.getRackForHost(loc.host).foreach { rack =>
+          pendingTasksForRack.getOrElseUpdate(rack, new ArrayBuffer) += index
+        }
       }
     }
 
@@ -331,7 +351,7 @@ private[spark] class TaskSetManager(
         val executors = prefs.flatMap(_ match {
           case e: ExecutorCacheTaskLocation => Some(e.executorId)
           case _ => None
-        });
+        })
         if (executors.contains(execId)) {
           speculatableTasks -= index
           return Some((index, TaskLocality.PROCESS_LOCAL))

core/src/test/scala/org/apache/spark/scheduler/TaskSetManagerSuite.scala

@@ -22,8 +22,8 @@ import java.util.{Properties, Random}
 import scala.collection.mutable
 import scala.collection.mutable.ArrayBuffer
 
-import org.mockito.ArgumentMatchers.{any, anyInt, anyString}
-import org.mockito.Mockito.{mock, never, spy, times, verify, when}
+import org.mockito.ArgumentMatchers.{any, anyBoolean, anyInt, anyString}
+import org.mockito.Mockito._
 import org.mockito.invocation.InvocationOnMock
 import org.mockito.stubbing.Answer
 
@@ -69,17 +69,23 @@ class FakeDAGScheduler(sc: SparkContext, taskScheduler: FakeTaskScheduler)
 // Get the rack for a given host
 object FakeRackUtil {
   private val hostToRack = new mutable.HashMap[String, String]()
+  var numBatchInvocation = 0
 
   def cleanUp() {
     hostToRack.clear()
+    numBatchInvocation = 0
   }
 
   def assignHostToRack(host: String, rack: String) {
     hostToRack(host) = rack
   }
 
-  def getRackForHost(host: String): Option[String] = {
-    hostToRack.get(host)
+  def getRacksForHosts(hosts: Seq[String]): Seq[Option[String]] = {
+    assert(hosts.toSet.size == hosts.size) // no dups in hosts
+    if (hosts.nonEmpty && hosts.length != 1) {
+      numBatchInvocation += 1
+    }
+    hosts.map(hostToRack.get(_))
   }
 }
 
@@ -99,6 +105,9 @@ class FakeTaskScheduler(sc: SparkContext, liveExecutors: (String, String)* /* ex
   val speculativeTasks = new ArrayBuffer[Int]
 
   val executors = new mutable.HashMap[String, String]
+
+  // this must be initialized before addExecutor
+  override val defaultRackValue: Option[String] = Some("default")
   for ((execId, host) <- liveExecutors) {
     addExecutor(execId, host)
   }
@@ -144,8 +153,9 @@ class FakeTaskScheduler(sc: SparkContext, liveExecutors: (String, String)* /* ex
     }
   }
 
-
-  override def getRackForHost(value: String): Option[String] = FakeRackUtil.getRackForHost(value)
+  override def getRacksForHosts(hosts: Seq[String]): Seq[Option[String]] = {
+    FakeRackUtil.getRacksForHosts(hosts)
+  }
 }
 
 /**
@@ -1316,7 +1326,7 @@ class TaskSetManagerSuite extends SparkFunSuite with LocalSparkContext with Logg
     val taskDesc = taskSetManagerSpy.resourceOffer(exec, host, TaskLocality.ANY)
 
     // Assert the task has been black listed on the executor it was last executed on.
-    when(taskSetManagerSpy.addPendingTask(anyInt())).thenAnswer(
+    when(taskSetManagerSpy.addPendingTask(anyInt(), anyBoolean())).thenAnswer(
       new Answer[Unit] {
         override def answer(invocationOnMock: InvocationOnMock): Unit = {
           val task: Int = invocationOnMock.getArgument(0)
@@ -1330,7 +1340,7 @@ class TaskSetManagerSuite extends SparkFunSuite with LocalSparkContext with Logg
     val e = new ExceptionFailure("a", "b", Array(), "c", None)
     taskSetManagerSpy.handleFailedTask(taskDesc.get.taskId, TaskState.FAILED, e)
 
-    verify(taskSetManagerSpy, times(1)).addPendingTask(anyInt())
+    verify(taskSetManagerSpy, times(1)).addPendingTask(anyInt(), anyBoolean())
   }
 
   test("SPARK-21563 context's added jars shouldn't change mid-TaskSet") {
@@ -1602,4 +1612,40 @@ class TaskSetManagerSuite extends SparkFunSuite with LocalSparkContext with Logg
     verify(sched.dagScheduler).taskEnded(manager.tasks(3), Success, result.value(),
       result.accumUpdates, info3)
   }
+
+  test("SPARK-13704 Rack Resolution is done with a batch of de-duped hosts") {
+    val conf = new SparkConf()
+      .set(config.LOCALITY_WAIT.key, "0")
+      .set(config.LOCALITY_WAIT_RACK.key, "1s")
+    sc = new SparkContext("local", "test", conf)
+    // Create a cluster with 20 racks, with hosts spread out among them
+    val execAndHost = (0 to 199).map { i =>
+      FakeRackUtil.assignHostToRack("host" + i, "rack" + (i % 20))
+      ("exec" + i, "host" + i)
+    }
+    sched = new FakeTaskScheduler(sc, execAndHost: _*)
+    // make a taskset with preferred locations on the first 100 hosts in our cluster
+    val locations = new ArrayBuffer[Seq[TaskLocation]]()
+    for (i <- 0 to 99) {
+      locations += Seq(TaskLocation("host" + i))
+    }
+    val taskSet = FakeTask.createTaskSet(100, locations: _*)
+    val clock = new ManualClock
+    val manager = new TaskSetManager(sched, taskSet, MAX_TASK_FAILURES, clock = clock)
+    // with rack locality, reject an offer on a host with an unknown rack
+    assert(manager.resourceOffer("otherExec", "otherHost", TaskLocality.RACK_LOCAL).isEmpty)
+    (0 until 20).foreach { rackIdx =>
+      (0 until 5).foreach { offerIdx =>
+        // if we offer hosts which are not in preferred locations,
+        // we'll reject them at NODE_LOCAL level,
+        // but accept them at RACK_LOCAL level if they're on OK racks
+        val hostIdx = 100 + rackIdx
+        assert(manager.resourceOffer("exec" + hostIdx, "host" + hostIdx, TaskLocality.NODE_LOCAL)
+          .isEmpty)
+        assert(manager.resourceOffer("exec" + hostIdx, "host" + hostIdx, TaskLocality.RACK_LOCAL)
+          .isDefined)
+      }
+    }
+    assert(FakeRackUtil.numBatchInvocation === 1)
+  }
 }

resource-managers/yarn/src/main/scala/org/apache/spark/deploy/yarn/LocalityPreferredContainerPlacementStrategy.scala

@@ -138,9 +138,8 @@ private[yarn] class LocalityPreferredContainerPlacementStrategy(
         // Only filter out the ratio which is larger than 0, which means the current host can
         // still be allocated with new container request.
         val hosts = preferredLocalityRatio.filter(_._2 > 0).keys.toArray
-        val racks = hosts.map { h =>
-          resolver.resolve(yarnConf, h)
-        }.toSet
+        val racks = resolver.resolve(hosts).map(_.getNetworkLocation)
+          .filter(_ != null).toSet
         containerLocalityPreferences += ContainerLocalityPreferences(hosts, racks.toArray)
 
         // Minus 1 each time when the host is used. When the current ratio is 0,

resource-managers/yarn/src/main/scala/org/apache/spark/deploy/yarn/SparkRackResolver.scala

@@ -17,24 +17,105 @@
 
 package org.apache.spark.deploy.yarn
 
+import scala.collection.JavaConverters._
+import scala.collection.mutable.ArrayBuffer
+
+import com.google.common.base.Strings
 import org.apache.hadoop.conf.Configuration
+import org.apache.hadoop.fs.CommonConfigurationKeysPublic
+import org.apache.hadoop.net._
+import org.apache.hadoop.util.ReflectionUtils
 import org.apache.hadoop.yarn.util.RackResolver
 import org.apache.log4j.{Level, Logger}
 
+import org.apache.spark.internal.Logging
+
 /**
- * Wrapper around YARN's [[RackResolver]]. This allows Spark tests to easily override the
+ * Re-implement YARN's [[RackResolver]]. This allows Spark tests to easily override the
  * default behavior, since YARN's class self-initializes the first time it's called, and
  * future calls all use the initial configuration.
  */
-private[yarn] class SparkRackResolver {
+private[spark] class SparkRackResolver(conf: Configuration) extends Logging {
 
   // RackResolver logs an INFO message whenever it resolves a rack, which is way too often.
   if (Logger.getLogger(classOf[RackResolver]).getLevel == null) {
     Logger.getLogger(classOf[RackResolver]).setLevel(Level.WARN)
   }
 
-  def resolve(conf: Configuration, hostName: String): String = {
-    RackResolver.resolve(conf, hostName).getNetworkLocation()
+  private val dnsToSwitchMapping: DNSToSwitchMapping = {
+    val dnsToSwitchMappingClass =
+      conf.getClass(CommonConfigurationKeysPublic.NET_TOPOLOGY_NODE_SWITCH_MAPPING_IMPL_KEY,
+        classOf[ScriptBasedMapping], classOf[DNSToSwitchMapping])
+    ReflectionUtils.newInstance(dnsToSwitchMappingClass, conf)
+        .asInstanceOf[DNSToSwitchMapping] match {
+      case c: CachedDNSToSwitchMapping => c
+      case o => new CachedDNSToSwitchMapping(o)
+    }
+  }
+
+  def resolve(hostName: String): String = {
+    coreResolve(Seq(hostName)).head.getNetworkLocation
+  }
+
+  /**
+   * Added in SPARK-13704.
+   * This should be changed to `RackResolver.resolve(conf, hostNames)`
+   * in hadoop releases with YARN-9332.
+   */
+  def resolve(hostNames: Seq[String]): Seq[Node] = {
+    coreResolve(hostNames)
+  }
+
+  private def coreResolve(hostNames: Seq[String]): Seq[Node] = {
+    val nodes = new ArrayBuffer[Node]
+    // dnsToSwitchMapping is thread-safe
+    val rNameList = dnsToSwitchMapping.resolve(hostNames.toList.asJava).asScala
+    if (rNameList == null || rNameList.isEmpty) {
+      hostNames.foreach(nodes += new NodeBase(_, NetworkTopology.DEFAULT_RACK))
+      logInfo(s"Got an error when resolving hostNames. " +
+        s"Falling back to ${NetworkTopology.DEFAULT_RACK} for all")
+    } else {
+      for ((hostName, rName) <- hostNames.zip(rNameList)) {
+        if (Strings.isNullOrEmpty(rName)) {
+          nodes += new NodeBase(hostName, NetworkTopology.DEFAULT_RACK)
+          logDebug(s"Could not resolve $hostName. " +
+            s"Falling back to ${NetworkTopology.DEFAULT_RACK}")
+        } else {
+          nodes += new NodeBase(hostName, rName)
+        }
+      }
+    }
+    nodes.toList
+  }
+}
+
+/**
+ * Utility to resolve the rack for hosts in an efficient manner.
+ * It will cache the rack for individual hosts to avoid
+ * repeatedly performing the same expensive lookup.
+ *
+ * Its logic refers [[org.apache.hadoop.yarn.util.RackResolver]] and enhanced.
+ * This will be unnecessary in hadoop releases with YARN-9332.
+ * With that, we could just directly use [[org.apache.hadoop.yarn.util.RackResolver]].
+ * In the meantime, this is a re-implementation for spark's use.
+ */
+object SparkRackResolver extends Logging {
+  @volatile private var instance: SparkRackResolver = _
+
+  /**
+   * It will return the static resolver instance.  If there is already an instance, the passed
+   * conf is entirely ignored.  If there is not a shared instance, it will create one with the
+   * given conf.
+   */
+  def get(conf: Configuration): SparkRackResolver = {
+    if (instance == null) {
+      synchronized {
+        if (instance == null) {
+          instance = new SparkRackResolver(conf)
+        }
+      }
+    }
+    instance
   }
 
 }

resource-managers/yarn/src/main/scala/org/apache/spark/deploy/yarn/YarnAllocator.scala

@@ -420,7 +420,7 @@ private[yarn] class YarnAllocator(
     // Match remaining by rack
     val remainingAfterRackMatches = new ArrayBuffer[Container]
     for (allocatedContainer <- remainingAfterHostMatches) {
-      val rack = resolver.resolve(conf, allocatedContainer.getNodeId.getHost)
+      val rack = resolver.resolve(allocatedContainer.getNodeId.getHost)
       matchContainerToRequest(allocatedContainer, rack, containersToUse,
         remainingAfterRackMatches)
     }

resource-managers/yarn/src/main/scala/org/apache/spark/deploy/yarn/YarnRMClient.scala

@@ -83,7 +83,7 @@ private[spark] class YarnRMClient extends Logging {
       localResources: Map[String, LocalResource]): YarnAllocator = {
     require(registered, "Must register AM before creating allocator.")
     new YarnAllocator(driverUrl, driverRef, conf, sparkConf, amClient, appAttemptId, securityMgr,
-      localResources, new SparkRackResolver())
+      localResources, new SparkRackResolver(conf))
   }
 
   /**