bigtop-dist
" + stageIdToStage.size)
-
- sizeBefore = shuffleToMapStage.size
- shuffleToMapStage.clearOldValues(cleanupTime)
- logInfo("shuffleToMapStage " + sizeBefore + " --> " + shuffleToMapStage.size)
-
- sizeBefore = pendingTasks.size
- pendingTasks.clearOldValues(cleanupTime)
- logInfo("pendingTasks " + sizeBefore + " --> " + pendingTasks.size)
-
- sizeBefore = stageToInfos.size
- stageToInfos.clearOldValues(cleanupTime)
- logInfo("stageToInfos " + sizeBefore + " --> " + stageToInfos.size)
+ Map(
+ "stageIdToStage" -> stageIdToStage,
+ "shuffleToMapStage" -> shuffleToMapStage,
+ "pendingTasks" -> pendingTasks,
+ "stageToInfos" -> stageToInfos,
+ "jobIdToStageIds" -> jobIdToStageIds,
+ "stageIdToJobIds" -> stageIdToJobIds).
+ foreach { case(s, t) => {
+ val sizeBefore = t.size
+ t.clearOldValues(cleanupTime)
+ logInfo("%s %d --> %d".format(s, sizeBefore, t.size))
+ }}
}
def removeStages(job: ActiveJob) = {
diff --git a/core/src/main/scala/org/apache/spark/scheduler/DAGSchedulerEvent.scala b/core/src/main/scala/org/apache/spark/scheduler/DAGSchedulerEvent.scala
index 708d221d60caf..9713a889cda45 100644
--- a/core/src/main/scala/org/apache/spark/scheduler/DAGSchedulerEvent.scala
+++ b/core/src/main/scala/org/apache/spark/scheduler/DAGSchedulerEvent.scala
@@ -65,8 +65,7 @@ private[scheduler] case class CompletionEvent(
taskMetrics: TaskMetrics)
extends DAGSchedulerEvent
-private[scheduler]
-case class ExecutorGained(execId: String, host: String) extends DAGSchedulerEvent
+private[scheduler] case class ExecutorGained(execId: String, host: String) extends DAGSchedulerEvent
private[scheduler] case class ExecutorLost(execId: String) extends DAGSchedulerEvent
diff --git a/core/src/main/scala/org/apache/spark/scheduler/SparkListener.scala b/core/src/main/scala/org/apache/spark/scheduler/SparkListener.scala
index a35081f7b10d7..3841b5616dca2 100644
--- a/core/src/main/scala/org/apache/spark/scheduler/SparkListener.scala
+++ b/core/src/main/scala/org/apache/spark/scheduler/SparkListener.scala
@@ -37,7 +37,7 @@ case class SparkListenerTaskGettingResult(
case class SparkListenerTaskEnd(task: Task[_], reason: TaskEndReason, taskInfo: TaskInfo,
taskMetrics: TaskMetrics) extends SparkListenerEvents
-case class SparkListenerJobStart(job: ActiveJob, properties: Properties = null)
+case class SparkListenerJobStart(job: ActiveJob, stageIds: Array[Int], properties: Properties = null)
extends SparkListenerEvents
case class SparkListenerJobEnd(job: ActiveJob, jobResult: JobResult)
diff --git a/core/src/main/scala/org/apache/spark/scheduler/cluster/ClusterScheduler.scala b/core/src/main/scala/org/apache/spark/scheduler/cluster/ClusterScheduler.scala
index 27145ccd5f102..281e65abc9d35 100644
--- a/core/src/main/scala/org/apache/spark/scheduler/cluster/ClusterScheduler.scala
+++ b/core/src/main/scala/org/apache/spark/scheduler/cluster/ClusterScheduler.scala
@@ -180,7 +180,9 @@ private[spark] class ClusterScheduler(val sc: SparkContext)
backend.killTask(tid, execId)
}
}
- tsm.error("Stage %d was cancelled".format(stageId))
+ logInfo("Stage %d was cancelled".format(stageId))
+ tsm.removeAllRunningTasks()
+ taskSetFinished(tsm)
}
}
diff --git a/core/src/main/scala/org/apache/spark/scheduler/cluster/ClusterTaskSetManager.scala b/core/src/main/scala/org/apache/spark/scheduler/cluster/ClusterTaskSetManager.scala
index 4c5eca8537cd6..5afd3f3e0629d 100644
--- a/core/src/main/scala/org/apache/spark/scheduler/cluster/ClusterTaskSetManager.scala
+++ b/core/src/main/scala/org/apache/spark/scheduler/cluster/ClusterTaskSetManager.scala
@@ -573,7 +573,7 @@ private[spark] class ClusterTaskSetManager(
runningTasks = runningTasksSet.size
}
- private def removeAllRunningTasks() {
+ private[cluster] def removeAllRunningTasks() {
val numRunningTasks = runningTasksSet.size
runningTasksSet.clear()
if (parent != null) {
diff --git a/core/src/main/scala/org/apache/spark/scheduler/cluster/CoarseGrainedSchedulerBackend.scala b/core/src/main/scala/org/apache/spark/scheduler/cluster/CoarseGrainedSchedulerBackend.scala
index a45bee536ca9d..d0ba5bf55dcfd 100644
--- a/core/src/main/scala/org/apache/spark/scheduler/cluster/CoarseGrainedSchedulerBackend.scala
+++ b/core/src/main/scala/org/apache/spark/scheduler/cluster/CoarseGrainedSchedulerBackend.scala
@@ -199,6 +199,7 @@ class CoarseGrainedSchedulerBackend(scheduler: ClusterScheduler, actorSystem: Ac
}
override def stop() {
+ stopExecutors()
try {
if (driverActor != null) {
val future = driverActor.ask(StopDriver)(timeout)
diff --git a/core/src/main/scala/org/apache/spark/scheduler/cluster/SimrSchedulerBackend.scala b/core/src/main/scala/org/apache/spark/scheduler/cluster/SimrSchedulerBackend.scala
index 0ea35e2b7a311..e000531a26f7e 100644
--- a/core/src/main/scala/org/apache/spark/scheduler/cluster/SimrSchedulerBackend.scala
+++ b/core/src/main/scala/org/apache/spark/scheduler/cluster/SimrSchedulerBackend.scala
@@ -62,7 +62,6 @@ private[spark] class SimrSchedulerBackend(
val conf = new Configuration()
val fs = FileSystem.get(conf)
fs.delete(new Path(driverFilePath), false)
- super.stopExecutors()
super.stop()
}
}
diff --git a/core/src/main/scala/org/apache/spark/scheduler/local/LocalScheduler.scala b/core/src/main/scala/org/apache/spark/scheduler/local/LocalScheduler.scala
index 2699f0b33e8d3..1c227fefe48d3 100644
--- a/core/src/main/scala/org/apache/spark/scheduler/local/LocalScheduler.scala
+++ b/core/src/main/scala/org/apache/spark/scheduler/local/LocalScheduler.scala
@@ -144,7 +144,8 @@ private[spark] class LocalScheduler(threads: Int, val maxFailures: Int, val sc:
localActor ! KillTask(tid)
}
}
- tsm.error("Stage %d was cancelled".format(stageId))
+ logInfo("Stage %d was cancelled".format(stageId))
+ taskSetFinished(tsm)
}
}
@@ -192,17 +193,19 @@ private[spark] class LocalScheduler(threads: Int, val maxFailures: Int, val sc:
synchronized {
taskIdToTaskSetId.get(taskId) match {
case Some(taskSetId) =>
- val taskSetManager = activeTaskSets(taskSetId)
- taskSetTaskIds(taskSetId) -= taskId
-
- state match {
- case TaskState.FINISHED =>
- taskSetManager.taskEnded(taskId, state, serializedData)
- case TaskState.FAILED =>
- taskSetManager.taskFailed(taskId, state, serializedData)
- case TaskState.KILLED =>
- taskSetManager.error("Task %d was killed".format(taskId))
- case _ => {}
+ val taskSetManager = activeTaskSets.get(taskSetId)
+ taskSetManager.foreach { tsm =>
+ taskSetTaskIds(taskSetId) -= taskId
+
+ state match {
+ case TaskState.FINISHED =>
+ tsm.taskEnded(taskId, state, serializedData)
+ case TaskState.FAILED =>
+ tsm.taskFailed(taskId, state, serializedData)
+ case TaskState.KILLED =>
+ tsm.error("Task %d was killed".format(taskId))
+ case _ => {}
+ }
}
case None =>
logInfo("Ignoring update from TID " + taskId + " because its task set is gone")
diff --git a/core/src/main/scala/org/apache/spark/storage/StoragePerfTester.scala b/core/src/main/scala/org/apache/spark/storage/StoragePerfTester.scala
index 68893a2bf49b1..369a2e536b068 100644
--- a/core/src/main/scala/org/apache/spark/storage/StoragePerfTester.scala
+++ b/core/src/main/scala/org/apache/spark/storage/StoragePerfTester.scala
@@ -1,3 +1,20 @@
+/*
+ * 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.storage
import java.util.concurrent.atomic.AtomicLong
diff --git a/core/src/main/scala/org/apache/spark/ui/jobs/StagePage.scala b/core/src/main/scala/org/apache/spark/ui/jobs/StagePage.scala
index c8625716e9d8e..baccc4281a06c 100644
--- a/core/src/main/scala/org/apache/spark/ui/jobs/StagePage.scala
+++ b/core/src/main/scala/org/apache/spark/ui/jobs/StagePage.scala
@@ -133,7 +133,7 @@ private[spark] class StagePage(parent: JobProgressUI) {
summary ++
Summary Metrics for {numCompleted} Completed Tasks
++
{summaryTable.getOrElse("No tasks have reported metrics yet.")}
++
- Tasks
++ taskTable;
+ Tasks
++ taskTable
headerSparkPage(content, parent.sc, "Details for Stage %d".format(stageId), Stages)
}
diff --git a/core/src/test/scala/org/apache/spark/JobCancellationSuite.scala b/core/src/test/scala/org/apache/spark/JobCancellationSuite.scala
index d8a0e983b228f..1121e06e2e6cc 100644
--- a/core/src/test/scala/org/apache/spark/JobCancellationSuite.scala
+++ b/core/src/test/scala/org/apache/spark/JobCancellationSuite.scala
@@ -114,7 +114,7 @@ class JobCancellationSuite extends FunSuite with ShouldMatchers with BeforeAndAf
// Once A is cancelled, job B should finish fairly quickly.
assert(jobB.get() === 100)
}
-
+/*
test("two jobs sharing the same stage") {
// sem1: make sure cancel is issued after some tasks are launched
// sem2: make sure the first stage is not finished until cancel is issued
@@ -148,7 +148,7 @@ class JobCancellationSuite extends FunSuite with ShouldMatchers with BeforeAndAf
intercept[SparkException] { f1.get() }
intercept[SparkException] { f2.get() }
}
-
+ */
def testCount() {
// Cancel before launching any tasks
{
diff --git a/core/src/test/scala/org/apache/spark/LocalSparkContext.scala b/core/src/test/scala/org/apache/spark/LocalSparkContext.scala
index 6ec124da9c7b1..03f7c0b43a71e 100644
--- a/core/src/test/scala/org/apache/spark/LocalSparkContext.scala
+++ b/core/src/test/scala/org/apache/spark/LocalSparkContext.scala
@@ -30,7 +30,7 @@ trait LocalSparkContext extends BeforeAndAfterEach with BeforeAndAfterAll { self
@transient var sc: SparkContext = _
override def beforeAll() {
- InternalLoggerFactory.setDefaultFactory(new Slf4JLoggerFactory());
+ InternalLoggerFactory.setDefaultFactory(new Slf4JLoggerFactory())
super.beforeAll()
}
diff --git a/core/src/test/scala/org/apache/spark/PartitioningSuite.scala b/core/src/test/scala/org/apache/spark/PartitioningSuite.scala
index 7d938917f2650..1374d01774693 100644
--- a/core/src/test/scala/org/apache/spark/PartitioningSuite.scala
+++ b/core/src/test/scala/org/apache/spark/PartitioningSuite.scala
@@ -142,11 +142,11 @@ class PartitioningSuite extends FunSuite with SharedSparkContext {
.filter(_ >= 0.0)
// Run the partitions, including the consecutive empty ones, through StatCounter
- val stats: StatCounter = rdd.stats();
- assert(abs(6.0 - stats.sum) < 0.01);
- assert(abs(6.0/2 - rdd.mean) < 0.01);
- assert(abs(1.0 - rdd.variance) < 0.01);
- assert(abs(1.0 - rdd.stdev) < 0.01);
+ val stats: StatCounter = rdd.stats()
+ assert(abs(6.0 - stats.sum) < 0.01)
+ assert(abs(6.0/2 - rdd.mean) < 0.01)
+ assert(abs(1.0 - rdd.variance) < 0.01)
+ assert(abs(1.0 - rdd.stdev) < 0.01)
// Add other tests here for classes that should be able to handle empty partitions correctly
}
diff --git a/core/src/test/scala/org/apache/spark/deploy/worker/ExecutorRunnerTest.scala b/core/src/test/scala/org/apache/spark/deploy/worker/ExecutorRunnerTest.scala
index 8f0954122b322..4cb4ddc9cd66a 100644
--- a/core/src/test/scala/org/apache/spark/deploy/worker/ExecutorRunnerTest.scala
+++ b/core/src/test/scala/org/apache/spark/deploy/worker/ExecutorRunnerTest.scala
@@ -1,3 +1,20 @@
+/*
+ * 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.worker
import java.io.File
diff --git a/core/src/test/scala/org/apache/spark/scheduler/DAGSchedulerSuite.scala b/core/src/test/scala/org/apache/spark/scheduler/DAGSchedulerSuite.scala
index 0f9b9992bd4d6..03fc8c020e005 100644
--- a/core/src/test/scala/org/apache/spark/scheduler/DAGSchedulerSuite.scala
+++ b/core/src/test/scala/org/apache/spark/scheduler/DAGSchedulerSuite.scala
@@ -206,6 +206,7 @@ class DAGSchedulerSuite extends FunSuite with BeforeAndAfter with LocalSparkCont
submit(rdd, Array(0))
complete(taskSets(0), List((Success, 42)))
assert(results === Map(0 -> 42))
+ assertDataStructuresEmpty
}
test("local job") {
@@ -219,6 +220,7 @@ class DAGSchedulerSuite extends FunSuite with BeforeAndAfter with LocalSparkCont
val jobId = scheduler.nextJobId.getAndIncrement()
runEvent(JobSubmitted(jobId, rdd, jobComputeFunc, Array(0), true, null, listener))
assert(results === Map(0 -> 42))
+ assertDataStructuresEmpty
}
test("run trivial job w/ dependency") {
@@ -227,6 +229,7 @@ class DAGSchedulerSuite extends FunSuite with BeforeAndAfter with LocalSparkCont
submit(finalRdd, Array(0))
complete(taskSets(0), Seq((Success, 42)))
assert(results === Map(0 -> 42))
+ assertDataStructuresEmpty
}
test("cache location preferences w/ dependency") {
@@ -239,12 +242,14 @@ class DAGSchedulerSuite extends FunSuite with BeforeAndAfter with LocalSparkCont
assertLocations(taskSet, Seq(Seq("hostA", "hostB")))
complete(taskSet, Seq((Success, 42)))
assert(results === Map(0 -> 42))
+ assertDataStructuresEmpty
}
test("trivial job failure") {
submit(makeRdd(1, Nil), Array(0))
failed(taskSets(0), "some failure")
assert(failure.getMessage === "Job aborted: some failure")
+ assertDataStructuresEmpty
}
test("run trivial shuffle") {
@@ -260,6 +265,7 @@ class DAGSchedulerSuite extends FunSuite with BeforeAndAfter with LocalSparkCont
Array(makeBlockManagerId("hostA"), makeBlockManagerId("hostB")))
complete(taskSets(1), Seq((Success, 42)))
assert(results === Map(0 -> 42))
+ assertDataStructuresEmpty
}
test("run trivial shuffle with fetch failure") {
@@ -285,6 +291,7 @@ class DAGSchedulerSuite extends FunSuite with BeforeAndAfter with LocalSparkCont
assert(mapOutputTracker.getServerStatuses(shuffleId, 0).map(_._1.host) === Array("hostA", "hostB"))
complete(taskSets(3), Seq((Success, 43)))
assert(results === Map(0 -> 42, 1 -> 43))
+ assertDataStructuresEmpty
}
test("ignore late map task completions") {
@@ -313,6 +320,7 @@ class DAGSchedulerSuite extends FunSuite with BeforeAndAfter with LocalSparkCont
Array(makeBlockManagerId("hostB"), makeBlockManagerId("hostA")))
complete(taskSets(1), Seq((Success, 42), (Success, 43)))
assert(results === Map(0 -> 42, 1 -> 43))
+ assertDataStructuresEmpty
}
test("run trivial shuffle with out-of-band failure and retry") {
@@ -329,15 +337,16 @@ class DAGSchedulerSuite extends FunSuite with BeforeAndAfter with LocalSparkCont
complete(taskSets(0), Seq(
(Success, makeMapStatus("hostA", 1)),
(Success, makeMapStatus("hostB", 1))))
- // have hostC complete the resubmitted task
- complete(taskSets(1), Seq((Success, makeMapStatus("hostC", 1))))
- assert(mapOutputTracker.getServerStatuses(shuffleId, 0).map(_._1) ===
- Array(makeBlockManagerId("hostC"), makeBlockManagerId("hostB")))
- complete(taskSets(2), Seq((Success, 42)))
- assert(results === Map(0 -> 42))
- }
-
- test("recursive shuffle failures") {
+ // have hostC complete the resubmitted task
+ complete(taskSets(1), Seq((Success, makeMapStatus("hostC", 1))))
+ assert(mapOutputTracker.getServerStatuses(shuffleId, 0).map(_._1) ===
+ Array(makeBlockManagerId("hostC"), makeBlockManagerId("hostB")))
+ complete(taskSets(2), Seq((Success, 42)))
+ assert(results === Map(0 -> 42))
+ assertDataStructuresEmpty
+ }
+
+ test("recursive shuffle failures") {
val shuffleOneRdd = makeRdd(2, Nil)
val shuffleDepOne = new ShuffleDependency(shuffleOneRdd, null)
val shuffleTwoRdd = makeRdd(2, List(shuffleDepOne))
@@ -363,6 +372,7 @@ class DAGSchedulerSuite extends FunSuite with BeforeAndAfter with LocalSparkCont
complete(taskSets(4), Seq((Success, makeMapStatus("hostA", 1))))
complete(taskSets(5), Seq((Success, 42)))
assert(results === Map(0 -> 42))
+ assertDataStructuresEmpty
}
test("cached post-shuffle") {
@@ -394,6 +404,7 @@ class DAGSchedulerSuite extends FunSuite with BeforeAndAfter with LocalSparkCont
complete(taskSets(3), Seq((Success, makeMapStatus("hostD", 1))))
complete(taskSets(4), Seq((Success, 42)))
assert(results === Map(0 -> 42))
+ assertDataStructuresEmpty
}
// test("oneGoodJob") {
@@ -447,16 +458,19 @@ class DAGSchedulerSuite extends FunSuite with BeforeAndAfter with LocalSparkCont
private def makeBlockManagerId(host: String): BlockManagerId =
BlockManagerId("exec-" + host, host, 12345, 0)
- private def assertDagSchedulerEmpty(dagScheduler: DAGScheduler) = {
- assert(dagScheduler.pendingTasks.isEmpty)
- assert(dagScheduler.activeJobs.isEmpty)
- assert(dagScheduler.failed.isEmpty)
- assert(dagScheduler.jobIdToStageIds.isEmpty)
- assert(dagScheduler.stageIdToStage.isEmpty)
- assert(dagScheduler.resultStageToJob.isEmpty)
- assert(dagScheduler.running.isEmpty)
- assert(dagScheduler.shuffleToMapStage.isEmpty)
- assert(dagScheduler.waiting.isEmpty)
+ private def assertDataStructuresEmpty = {
+ assert(scheduler.pendingTasks.isEmpty)
+ assert(scheduler.activeJobs.isEmpty)
+ assert(scheduler.failed.isEmpty)
+ assert(scheduler.idToActiveJob.isEmpty)
+ assert(scheduler.jobIdToStageIds.isEmpty)
+ assert(scheduler.stageIdToJobIds.isEmpty)
+ assert(scheduler.stageIdToStage.isEmpty)
+ assert(scheduler.stageToInfos.isEmpty)
+ assert(scheduler.resultStageToJob.isEmpty)
+ assert(scheduler.running.isEmpty)
+ assert(scheduler.shuffleToMapStage.isEmpty)
+ assert(scheduler.waiting.isEmpty)
}
}
diff --git a/core/src/test/scala/org/apache/spark/scheduler/JobLoggerSuite.scala b/core/src/test/scala/org/apache/spark/scheduler/JobLoggerSuite.scala
index 984881861c9a9..22babe8d2ef78 100644
--- a/core/src/test/scala/org/apache/spark/scheduler/JobLoggerSuite.scala
+++ b/core/src/test/scala/org/apache/spark/scheduler/JobLoggerSuite.scala
@@ -31,6 +31,8 @@ import org.apache.spark.rdd.RDD
class JobLoggerSuite extends FunSuite with LocalSparkContext with ShouldMatchers {
+ /** Length of time to wait while draining listener events. */
+ val WAIT_TIMEOUT_MILLIS = 10000
test("inner method") {
sc = new SparkContext("local", "joblogger")
@@ -91,6 +93,7 @@ class JobLoggerSuite extends FunSuite with LocalSparkContext with ShouldMatchers
sc.addSparkListener(joblogger)
val rdd = sc.parallelize(1 to 1e2.toInt, 4).map{ i => (i % 12, 2 * i) }
rdd.reduceByKey(_+_).collect()
+ assert(sc.dagScheduler.listenerBus.waitUntilEmpty(WAIT_TIMEOUT_MILLIS))
val user = System.getProperty("user.name", SparkContext.SPARK_UNKNOWN_USER)
@@ -119,8 +122,9 @@ class JobLoggerSuite extends FunSuite with LocalSparkContext with ShouldMatchers
}
sc.addSparkListener(joblogger)
val rdd = sc.parallelize(1 to 1e2.toInt, 4).map{ i => (i % 12, 2 * i) }
- rdd.reduceByKey(_+_).collect()
-
+ rdd.reduceByKey(_+_).collect()
+ assert(sc.dagScheduler.listenerBus.waitUntilEmpty(WAIT_TIMEOUT_MILLIS))
+
joblogger.onJobStartCount should be (1)
joblogger.onJobEndCount should be (1)
joblogger.onTaskEndCount should be (8)
diff --git a/core/src/test/scala/org/apache/spark/storage/DiskBlockManagerSuite.scala b/core/src/test/scala/org/apache/spark/storage/DiskBlockManagerSuite.scala
index ef4c4c0f143f9..070982e798f34 100644
--- a/core/src/test/scala/org/apache/spark/storage/DiskBlockManagerSuite.scala
+++ b/core/src/test/scala/org/apache/spark/storage/DiskBlockManagerSuite.scala
@@ -1,3 +1,20 @@
+/*
+ * 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.storage
import java.io.{FileWriter, File}
diff --git a/core/src/test/scala/org/apache/spark/util/collection/OpenHashMapSuite.scala b/core/src/test/scala/org/apache/spark/util/collection/OpenHashMapSuite.scala
index ca3f684668d60..80ee59f26f49e 100644
--- a/core/src/test/scala/org/apache/spark/util/collection/OpenHashMapSuite.scala
+++ b/core/src/test/scala/org/apache/spark/util/collection/OpenHashMapSuite.scala
@@ -1,3 +1,20 @@
+/*
+ * 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.util.collection
import scala.collection.mutable.HashSet
diff --git a/core/src/test/scala/org/apache/spark/util/collection/OpenHashSetSuite.scala b/core/src/test/scala/org/apache/spark/util/collection/OpenHashSetSuite.scala
index 4e11e8a628b44..5ad7ea74f186f 100644
--- a/core/src/test/scala/org/apache/spark/util/collection/OpenHashSetSuite.scala
+++ b/core/src/test/scala/org/apache/spark/util/collection/OpenHashSetSuite.scala
@@ -1,3 +1,20 @@
+/*
+ * 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.util.collection
import org.scalatest.FunSuite
diff --git a/core/src/test/scala/org/apache/spark/util/collection/PrimitiveKeyOpenHashSetSuite.scala b/core/src/test/scala/org/apache/spark/util/collection/PrimitiveKeyOpenHashSetSuite.scala
index dfd6aed2c4bcc..0342195161dc0 100644
--- a/core/src/test/scala/org/apache/spark/util/collection/PrimitiveKeyOpenHashSetSuite.scala
+++ b/core/src/test/scala/org/apache/spark/util/collection/PrimitiveKeyOpenHashSetSuite.scala
@@ -1,3 +1,20 @@
+/*
+ * 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.util.collection
import scala.collection.mutable.HashSet
diff --git a/docs/_layouts/global.html b/docs/_layouts/global.html
index 0c1d657cde470..ad7969d012283 100755
--- a/docs/_layouts/global.html
+++ b/docs/_layouts/global.html
@@ -74,12 +74,12 @@
API Docs
diff --git a/docs/bagel-programming-guide.md b/docs/bagel-programming-guide.md
index 140190a38cfca..de001e6c52f85 100644
--- a/docs/bagel-programming-guide.md
+++ b/docs/bagel-programming-guide.md
@@ -106,7 +106,7 @@ _Example_
## Operations
-Here are the actions and types in the Bagel API. See [Bagel.scala](https://github.com/apache/incubator-spark/blob/master/bagel/src/main/scala/spark/bagel/Bagel.scala) for details.
+Here are the actions and types in the Bagel API. See [Bagel.scala](https://github.com/apache/incubator-spark/blob/master/bagel/src/main/scala/org/apache/spark/bagel/Bagel.scala) for details.
### Actions
diff --git a/docs/building-with-maven.md b/docs/building-with-maven.md
index 19c01e179f35f..b2744b002cca3 100644
--- a/docs/building-with-maven.md
+++ b/docs/building-with-maven.md
@@ -45,6 +45,12 @@ For Apache Hadoop 2.x, 0.23.x, Cloudera CDH MRv2, and other Hadoop versions with
# Cloudera CDH 4.2.0 with MapReduce v2
$ mvn -Phadoop2-yarn -Dhadoop.version=2.0.0-cdh4.2.0 -Dyarn.version=2.0.0-chd4.2.0 -DskipTests clean package
+Hadoop versions 2.2.x and newer can be built by setting the ```new-yarn``` and the ```yarn.version``` as follows:
+
+ # Apache Hadoop 2.2.X and newer
+ $ mvn -Pnew-yarn -Dhadoop.version=2.2.0 -Dyarn.version=2.2.0 -DskipTests clean package
+
+The build process handles Hadoop 2.2.x as a special case that uses the directory ```new-yarn```, which supports the new YARN API. Furthermore, for this version, the build depends on artifacts published by the spark-project to enable Akka 2.0.5 to work with protobuf 2.5.
## Spark Tests in Maven ##
diff --git a/docs/cluster-overview.md b/docs/cluster-overview.md
index 5927f736f3579..e16703292cc22 100644
--- a/docs/cluster-overview.md
+++ b/docs/cluster-overview.md
@@ -45,7 +45,7 @@ The system currently supports three cluster managers:
easy to set up a cluster.
* [Apache Mesos](running-on-mesos.html) -- a general cluster manager that can also run Hadoop MapReduce
and service applications.
-* [Hadoop YARN](running-on-yarn.html) -- the resource manager in Hadoop 2.0.
+* [Hadoop YARN](running-on-yarn.html) -- the resource manager in Hadoop 2.
In addition, Spark's [EC2 launch scripts](ec2-scripts.html) make it easy to launch a standalone
cluster on Amazon EC2.
diff --git a/docs/configuration.md b/docs/configuration.md
index 22abe1c5a9943..2accfe84c606e 100644
--- a/docs/configuration.md
+++ b/docs/configuration.md
@@ -331,10 +331,9 @@ Apart from these, the following properties are also available, and may be useful
spark.shuffle.consolidateFiles |
false |
- If set to "true", consolidates intermediate files created during a shuffle. Creating fewer files can improve filesystem performance if you run shuffles with large numbers of reduce tasks.
+ If set to "true", consolidates intermediate files created during a shuffle. Creating fewer files can improve filesystem performance for shuffles with large numbers of reduce tasks. It is recomended to set this to "true" when using ext4 or xfs filesystems. On ext3, this option might degrade performance on machines with many (>8) cores due to filesystem limitations.
|
-
| spark.speculation |
false |
diff --git a/docs/hadoop-third-party-distributions.md b/docs/hadoop-third-party-distributions.md
index b33af2cf24d77..de6a2b0a43bd5 100644
--- a/docs/hadoop-third-party-distributions.md
+++ b/docs/hadoop-third-party-distributions.md
@@ -10,7 +10,7 @@ with these distributions:
# Compile-time Hadoop Version
When compiling Spark, you'll need to
-[set the SPARK_HADOOP_VERSION flag](http://localhost:4000/index.html#a-note-about-hadoop-versions):
+[set the SPARK_HADOOP_VERSION flag](index.html#a-note-about-hadoop-versions):
SPARK_HADOOP_VERSION=1.0.4 sbt/sbt assembly
@@ -40,6 +40,7 @@ the _exact_ Hadoop version you are running to avoid any compatibility errors.
| HDP 1.2 | 1.1.2 |
| HDP 1.1 | 1.0.3 |
| HDP 1.0 | 1.0.3 |
+ | HDP 2.0 | 2.2.0 |
diff --git a/docs/index.md b/docs/index.md
index bd386a8a8fdb6..d3ac696d1e818 100644
--- a/docs/index.md
+++ b/docs/index.md
@@ -56,14 +56,16 @@ Hadoop, you must build Spark against the same version that your cluster uses.
By default, Spark links to Hadoop 1.0.4. You can change this by setting the
`SPARK_HADOOP_VERSION` variable when compiling:
- SPARK_HADOOP_VERSION=1.2.1 sbt/sbt assembly
+ SPARK_HADOOP_VERSION=2.2.0 sbt/sbt assembly
-In addition, if you wish to run Spark on [YARN](running-on-yarn.md), set
+In addition, if you wish to run Spark on [YARN](running-on-yarn.html), set
`SPARK_YARN` to `true`:
SPARK_HADOOP_VERSION=2.0.5-alpha SPARK_YARN=true sbt/sbt assembly
-(Note that on Windows, you need to set the environment variables on separate lines, e.g., `set SPARK_HADOOP_VERSION=1.2.1`.)
+Note that on Windows, you need to set the environment variables on separate lines, e.g., `set SPARK_HADOOP_VERSION=1.2.1`.
+
+For this version of Spark (0.8.1) Hadoop 2.2.x (or newer) users will have to build Spark and publish it locally. See [Launching Spark on YARN](running-on-yarn.html). This is needed because Hadoop 2.2 has non backwards compatible API changes.
# Where to Go from Here
diff --git a/docs/job-scheduling.md b/docs/job-scheduling.md
index d304c5497bdb3..dbcb9ae343357 100644
--- a/docs/job-scheduling.md
+++ b/docs/job-scheduling.md
@@ -91,7 +91,7 @@ The fair scheduler also supports grouping jobs into _pools_, and setting differe
(e.g. weight) for each pool. This can be useful to create a "high-priority" pool for more important jobs,
for example, or to group the jobs of each user together and give _users_ equal shares regardless of how
many concurrent jobs they have instead of giving _jobs_ equal shares. This approach is modeled after the
-[Hadoop Fair Scheduler](http://hadoop.apache.org/docs/stable/fair_scheduler.html).
+[Hadoop Fair Scheduler](http://hadoop.apache.org/docs/current/hadoop-yarn/hadoop-yarn-site/FairScheduler.html).
Without any intervention, newly submitted jobs go into a _default pool_, but jobs' pools can be set by
adding the `spark.scheduler.pool` "local property" to the SparkContext in the thread that's submitting them.
diff --git a/docs/running-on-yarn.md b/docs/running-on-yarn.md
index 6fd1d0d150306..9e4c4e123f0af 100644
--- a/docs/running-on-yarn.md
+++ b/docs/running-on-yarn.md
@@ -17,6 +17,7 @@ This can be built by setting the Hadoop version and `SPARK_YARN` environment var
The assembled JAR will be something like this:
`./assembly/target/scala-{{site.SCALA_VERSION}}/spark-assembly_{{site.SPARK_VERSION}}-hadoop2.0.5.jar`.
+The build process now also supports new YARN versions (2.2.x). See below.
# Preparations
@@ -37,12 +38,18 @@ System Properties:
* 'spark.yarn.applicationMaster.waitTries', property to set the number of times the ApplicationMaster waits for the the spark master and then also the number of tries it waits for the Spark Context to be intialized. Default is 10.
* 'spark.yarn.submit.file.replication', the HDFS replication level for the files uploaded into HDFS for the application. These include things like the spark jar, the app jar, and any distributed cache files/archives.
* 'spark.yarn.preserve.staging.files', set to true to preserve the staged files(spark jar, app jar, distributed cache files) at the end of the job rather then delete them.
+* 'spark.yarn.scheduler.heartbeat.interval-ms', the interval in ms in which the Spark application master heartbeats into the YARN ResourceManager. Default is 5 seconds.
+* 'spark.yarn.max.worker.failures', the maximum number of worker failures before failing the application. Default is the number of workers requested times 2 with minimum of 3.
# Launching Spark on YARN
Ensure that HADOOP_CONF_DIR or YARN_CONF_DIR points to the directory which contains the (client side) configuration files for the hadoop cluster.
This would be used to connect to the cluster, write to the dfs and submit jobs to the resource manager.
+There are two scheduler mode that can be used to launch spark application on YARN.
+
+## Launch spark application by YARN Client with yarn-standalone mode.
+
The command to launch the YARN Client is as follows:
SPARK_JAR= ./spark-class org.apache.spark.deploy.yarn.Client \
@@ -50,6 +57,7 @@ The command to launch the YARN Client is as follows:
--class \
--args \
--num-workers \
+ --master-class
--master-memory \
--worker-memory \
--worker-cores \
@@ -83,12 +91,37 @@ For example:
$ cat $YARN_APP_LOGS_DIR/$YARN_APP_ID/container*_000001/stdout
Pi is roughly 3.13794
-The above starts a YARN Client programs which periodically polls the Application Master for status updates and displays them in the console. The client will exit once your application has finished running.
+The above starts a YARN Client programs which start the default Application Master. Then SparkPi will be run as a child thread of Application Master, YARN Client will periodically polls the Application Master for status updates and displays them in the console. The client will exit once your application has finished running.
+
+With this mode, your application is actually run on the remote machine where the Application Master is run upon. Thus application that involve local interaction will not work well, e.g. spark-shell.
+
+## Launch spark application with yarn-client mode.
+
+With yarn-client mode, the application will be launched locally. Just like running application or spark-shell on Local / Mesos / Standalone mode. The launch method is also the similar with them, just make sure that when you need to specify a master url, use "yarn-client" instead. And you also need to export the env value for SPARK_JAR and SPARK_YARN_APP_JAR
+
+In order to tune worker core/number/memory etc. You need to export SPARK_WORKER_CORES, SPARK_WORKER_MEMORY, SPARK_WORKER_INSTANCES e.g. by ./conf/spark-env.sh
+
+For example:
+
+ SPARK_JAR=./assembly/target/scala-{{site.SCALA_VERSION}}/spark-assembly-{{site.SPARK_VERSION}}-hadoop2.0.5-alpha.jar \
+ SPARK_YARN_APP_JAR=examples/target/scala-{{site.SCALA_VERSION}}/spark-examples-assembly-{{site.SPARK_VERSION}}.jar \
+ ./run-example org.apache.spark.examples.SparkPi yarn-client
+
+
+ SPARK_JAR=./assembly/target/scala-{{site.SCALA_VERSION}}/spark-assembly-{{site.SPARK_VERSION}}-hadoop2.0.5-alpha.jar \
+ SPARK_YARN_APP_JAR=examples/target/scala-{{site.SCALA_VERSION}}/spark-examples-assembly-{{site.SPARK_VERSION}}.jar \
+ MASTER=yarn-client ./spark-shell
+
+# Building Spark for Hadoop/YARN 2.2.x
+
+Hadoop 2.2.x users must build Spark and publish it locally. The SBT build process handles Hadoop 2.2.x as a special case. This version of Hadoop has new YARN API changes and depends on a Protobuf version (2.5) that is not compatible with the Akka version (2.0.5) that Spark uses. Therefore, if the Hadoop version (e.g. set through ```SPARK_HADOOP_VERSION```) starts with 2.2.0 or higher then the build process will depend on Akka artifacts distributed by the Spark project compatible with Protobuf 2.5. Furthermore, the build process then uses the directory ```new-yarn``` (instead of ```yarn```), which supports the new YARN API. The build process should seamlessly work out of the box.
+
+See [Building Spark with Maven](building-with-maven.html) for instructions on how to build Spark using the Maven process.
# Important Notes
-- When your application instantiates a Spark context it must use a special "yarn-standalone" master url. This starts the scheduler without forcing it to connect to a cluster. A good way to handle this is to pass "yarn-standalone" as an argument to your program, as shown in the example above.
- We do not requesting container resources based on the number of cores. Thus the numbers of cores given via command line arguments cannot be guaranteed.
- The local directories used for spark will be the local directories configured for YARN (Hadoop Yarn config yarn.nodemanager.local-dirs). If the user specifies spark.local.dir, it will be ignored.
- The --files and --archives options support specifying file names with the # similar to Hadoop. For example you can specify: --files localtest.txt#appSees.txt and this will upload the file you have locally named localtest.txt into HDFS but this will be linked to by the name appSees.txt and your application should use the name as appSees.txt to reference it when running on YARN.
- The --addJars option allows the SparkContext.addJar function to work if you are using it with local files. It does not need to be used if you are using it with HDFS, HTTP, HTTPS, or FTP files.
+- YARN 2.2.x users cannot simply depend on the Spark packages without building Spark, as the published Spark artifacts are compiled to work with the pre 2.2 API. Those users must build Spark and publish it locally.
diff --git a/docs/streaming-programming-guide.md b/docs/streaming-programming-guide.md
index 851e30fe761af..82f42e0b8d16a 100644
--- a/docs/streaming-programming-guide.md
+++ b/docs/streaming-programming-guide.md
@@ -214,7 +214,7 @@ ssc.stop()
{% endhighlight %}
# Example
-A simple example to start off is the [NetworkWordCount](https://github.com/apache/incubator-spark/tree/master/examples/src/main/scala/spark/streaming/examples/NetworkWordCount.scala). This example counts the words received from a network server every second. Given below is the relevant sections of the source code. You can find the full source code in `/streaming/src/main/scala/spark/streaming/examples/NetworkWordCount.scala` .
+A simple example to start off is the [NetworkWordCount](https://github.com/apache/incubator-spark/tree/master/examples/src/main/scala/org/apache/spark/streaming/examples/NetworkWordCount.scala). This example counts the words received from a network server every second. Given below is the relevant sections of the source code. You can find the full source code in `/streaming/src/main/scala/org/apache/spark/streaming/examples/NetworkWordCount.scala` .
{% highlight scala %}
import org.apache.spark.streaming.{Seconds, StreamingContext}
@@ -283,7 +283,7 @@ Time: 1357008430000 ms
-You can find more examples in `/streaming/src/main/scala/spark/streaming/examples/`. They can be run in the similar manner using `./run-example org.apache.spark.streaming.examples....` . Executing without any parameter would give the required parameter list. Further explanation to run them can be found in comments in the files.
+You can find more examples in `/streaming/src/main/scala/org/apache/spark/streaming/examples/`. They can be run in the similar manner using `./run-example org.apache.spark.streaming.examples....` . Executing without any parameter would give the required parameter list. Further explanation to run them can be found in comments in the files.
# DStream Persistence
Similar to RDDs, DStreams also allow developers to persist the stream's data in memory. That is, using `persist()` method on a DStream would automatically persist every RDD of that DStream in memory. This is useful if the data in the DStream will be computed multiple times (e.g., multiple operations on the same data). For window-based operations like `reduceByWindow` and `reduceByKeyAndWindow` and state-based operations like `updateStateByKey`, this is implicitly true. Hence, DStreams generated by window-based operations are automatically persisted in memory, without the developer calling `persist()`.
@@ -483,7 +483,7 @@ Similar to [Spark's Java API](java-programming-guide.html), we also provide a Ja
1. Functions for transformations must be implemented as subclasses of [Function](api/core/index.html#org.apache.spark.api.java.function.Function) and [Function2](api/core/index.html#org.apache.spark.api.java.function.Function2)
1. Unlike the Scala API, the Java API handles DStreams for key-value pairs using a separate [JavaPairDStream](api/streaming/index.html#org.apache.spark.streaming.api.java.JavaPairDStream) class(similar to [JavaRDD and JavaPairRDD](java-programming-guide.html#rdd-classes). DStream functions like `map` and `filter` are implemented separately by JavaDStreams and JavaPairDStream to return DStreams of appropriate types.
-Spark's [Java Programming Guide](java-programming-guide.html) gives more ideas about using the Java API. To extends the ideas presented for the RDDs to DStreams, we present parts of the Java version of the same NetworkWordCount example presented above. The full source code is given at `/examples/src/main/java/spark/streaming/examples/JavaNetworkWordCount.java`
+Spark's [Java Programming Guide](java-programming-guide.html) gives more ideas about using the Java API. To extends the ideas presented for the RDDs to DStreams, we present parts of the Java version of the same NetworkWordCount example presented above. The full source code is given at `/examples/src/main/java/org/apache/spark/streaming/examples/JavaNetworkWordCount.java`
The streaming context and the socket stream from input source is started by using a `JavaStreamingContext`, that has the same parameters and provides the same input streams as its Scala counterpart.
@@ -527,5 +527,5 @@ JavaPairDStream wordCounts = words.map(
# Where to Go from Here
* API docs - [Scala](api/streaming/index.html#org.apache.spark.streaming.package) and [Java](api/streaming/index.html#org.apache.spark.streaming.api.java.package)
-* More examples - [Scala](https://github.com/apache/incubator-spark/tree/master/examples/src/main/scala/spark/streaming/examples) and [Java](https://github.com/apache/incubator-spark/tree/master/examples/src/main/java/spark/streaming/examples)
+* More examples - [Scala](https://github.com/apache/incubator-spark/tree/master/examples/src/main/scala/org/apache/spark/streaming/examples) and [Java](https://github.com/apache/incubator-spark/tree/master/examples/src/main/java/org/apache/spark/streaming/examples)
* [Paper describing Spark Streaming](http://www.eecs.berkeley.edu/Pubs/TechRpts/2012/EECS-2012-259.pdf)
diff --git a/examples/pom.xml b/examples/pom.xml
index 7055c72082506..47ef2d3c9cb10 100644
--- a/examples/pom.xml
+++ b/examples/pom.xml
@@ -40,6 +40,23 @@
+
+ ${akka.group}
+ akka-actor
+
+
+ ${akka.group}
+ akka-remote
+
+
+ ${akka.group}
+ akka-slf4j
+
+
+ ${akka.group}
+ akka-zeromq
+
+
org.apache.spark
spark-core_2.9.3
diff --git a/examples/src/main/scala/org/apache/spark/examples/LocalALS.scala b/examples/src/main/scala/org/apache/spark/examples/LocalALS.scala
index 4af45b2b4a067..83db8b9e26411 100644
--- a/examples/src/main/scala/org/apache/spark/examples/LocalALS.scala
+++ b/examples/src/main/scala/org/apache/spark/examples/LocalALS.scala
@@ -120,7 +120,7 @@ object LocalALS {
System.exit(1)
}
}
- printf("Running with M=%d, U=%d, F=%d, iters=%d\n", M, U, F, ITERATIONS);
+ printf("Running with M=%d, U=%d, F=%d, iters=%d\n", M, U, F, ITERATIONS)
val R = generateR()
diff --git a/examples/src/main/scala/org/apache/spark/examples/SparkTC.scala b/examples/src/main/scala/org/apache/spark/examples/SparkTC.scala
index 5a7a9d1bd8f74..8543ce0e3285e 100644
--- a/examples/src/main/scala/org/apache/spark/examples/SparkTC.scala
+++ b/examples/src/main/scala/org/apache/spark/examples/SparkTC.scala
@@ -65,7 +65,7 @@ object SparkTC {
oldCount = nextCount
// Perform the join, obtaining an RDD of (y, (z, x)) pairs,
// then project the result to obtain the new (x, z) paths.
- tc = tc.union(tc.join(edges).map(x => (x._2._2, x._2._1))).distinct().cache();
+ tc = tc.union(tc.join(edges).map(x => (x._2._2, x._2._1))).distinct().cache()
nextCount = tc.count()
} while (nextCount != oldCount)
diff --git a/examples/src/main/scala/org/apache/spark/streaming/examples/ActorWordCount.scala b/examples/src/main/scala/org/apache/spark/streaming/examples/ActorWordCount.scala
index cd3423a07b6b7..af52b7e9a12f1 100644
--- a/examples/src/main/scala/org/apache/spark/streaming/examples/ActorWordCount.scala
+++ b/examples/src/main/scala/org/apache/spark/streaming/examples/ActorWordCount.scala
@@ -120,7 +120,7 @@ object FeederActor {
println("Feeder started as:" + feeder)
- actorSystem.awaitTermination();
+ actorSystem.awaitTermination()
}
}
diff --git a/mllib/pom.xml b/mllib/pom.xml
index ed8a076a58937..af3f89a98e730 100644
--- a/mllib/pom.xml
+++ b/mllib/pom.xml
@@ -32,6 +32,18 @@
http://spark.incubator.apache.org/
+
+ ${akka.group}
+ akka-actor
+
+
+ ${akka.group}
+ akka-remote
+
+
+ ${akka.group}
+ akka-slf4j
+
org.apache.spark
spark-core_2.9.3
diff --git a/new-yarn/pom.xml b/new-yarn/pom.xml
new file mode 100644
index 0000000000000..d15404cf2e670
--- /dev/null
+++ b/new-yarn/pom.xml
@@ -0,0 +1,177 @@
+
+
+
+ 4.0.0
+
+ org.apache.spark
+ spark-parent
+ 0.8.1-incubating
+ ../pom.xml
+
+
+ org.apache.spark
+ spark-yarn_2.9.3
+ jar
+ Spark Project YARN Support
+ http://spark.incubator.apache.org/
+
+
+
+ ${akka.group}
+ akka-actor
+
+
+ ${akka.group}
+ akka-remote
+
+
+ ${akka.group}
+ akka-slf4j
+
+
+ ${akka.group}
+ akka-zeromq
+
+
+ org.apache.spark
+ spark-core_2.9.3
+ ${project.version}
+
+
+ org.apache.hadoop
+ hadoop-yarn-api
+
+
+ org.apache.hadoop
+ hadoop-yarn-common
+
+
+ org.apache.hadoop
+ hadoop-yarn-client
+
+
+ org.apache.hadoop
+ hadoop-client
+ ${yarn.version}
+
+
+ org.apache.avro
+ avro
+
+
+ org.apache.avro
+ avro-ipc
+
+
+ org.scalatest
+ scalatest_2.9.3
+ test
+
+
+ org.mockito
+ mockito-all
+ test
+
+
+
+
+ target/scala-${scala.version}/classes
+ target/scala-${scala.version}/test-classes
+
+
+ org.apache.maven.plugins
+ maven-shade-plugin
+
+ false
+ ${project.build.directory}/${project.artifactId}-${project.version}-shaded.jar
+
+
+ *:*
+
+
+
+
+ *:*
+
+ META-INF/*.SF
+ META-INF/*.DSA
+ META-INF/*.RSA
+
+
+
+
+
+
+ package
+
+ shade
+
+
+
+
+
+ reference.conf
+
+
+
+
+
+
+
+ org.apache.maven.plugins
+ maven-antrun-plugin
+
+
+ test
+
+ run
+
+
+ true
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+
+ org.scalatest
+ scalatest-maven-plugin
+
+
+ ${basedir}/..
+ 1
+ ${spark.classpath}
+
+
+
+
+
+
diff --git a/new-yarn/src/main/scala/org/apache/spark/deploy/yarn/ApplicationMaster.scala b/new-yarn/src/main/scala/org/apache/spark/deploy/yarn/ApplicationMaster.scala
new file mode 100644
index 0000000000000..eeeca3ea8a33e
--- /dev/null
+++ b/new-yarn/src/main/scala/org/apache/spark/deploy/yarn/ApplicationMaster.scala
@@ -0,0 +1,446 @@
+/*
+ * 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.{SparkContext, Logging}
+import org.apache.spark.util.Utils
+
+
+class ApplicationMaster(args: ApplicationMasterArguments, conf: Configuration) extends Logging {
+
+ def this(args: ApplicationMasterArguments) = this(args, new Configuration())
+
+ 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] = _
+
+ // Default to numWorkers * 2, with minimum of 3
+ private val maxNumWorkerFailures = System.getProperty("spark.yarn.max.worker.failures",
+ math.max(args.numWorkers * 2, 3).toString()).toInt
+
+ def run() {
+ // Setup the directories so things go to YARN approved directories rather
+ // than user specified and /tmp.
+ System.setProperty("spark.local.dir", getLocalDirs())
+
+ // 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.
+ waitForSparkMaster()
+
+ 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(""))
+
+ 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 waitForSparkMaster() {
+ logInfo("Waiting for Spark driver to be reachable.")
+ var driverUp = false
+ var tries = 0
+ val numTries = System.getProperty("spark.yarn.applicationMaster.waitTries", "10").toInt
+ while (!driverUp && tries < numTries) {
+ val driverHost = System.getProperty("spark.driver.host")
+ val driverPort = System.getProperty("spark.driver.port")
+ try {
+ val socket = new Socket(driverHost, driverPort.toInt)
+ socket.close()
+ logInfo("Driver now available: %s:%s".format(driverHost, driverPort))
+ driverUp = true
+ } catch {
+ case e: Exception => {
+ logWarning("Failed to connect to driver at %s:%s, retrying ...".
+ format(driverHost, driverPort))
+ Thread.sleep(100)
+ tries = tries + 1
+ }
+ }
+ }
+ }
+
+ private def startUserClass(): Thread = {
+ logInfo("Starting the user JAR in a separate Thread")
+ val mainMethod = Class.forName(
+ args.userClass,
+ 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 = System.getProperty("spark.yarn.ApplicationMaster.waitTries", "10").toInt
+ 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)
+ } 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)
+ }
+ }
+ } 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 =
+ System.getProperty("spark.yarn.scheduler.heartbeat.interval-ms", "5000").toLong
+
+ // 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 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 */)
+ }
+
+ /**
+ * Clean up the staging directory.
+ */
+ private def cleanupStagingDir() {
+ var stagingDirPath: Path = null
+ try {
+ val preserveFiles = System.getProperty("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()
+ }
+}
diff --git a/new-yarn/src/main/scala/org/apache/spark/deploy/yarn/ApplicationMasterArguments.scala b/new-yarn/src/main/scala/org/apache/spark/deploy/yarn/ApplicationMasterArguments.scala
new file mode 100644
index 0000000000000..f76a5ddd39e90
--- /dev/null
+++ b/new-yarn/src/main/scala/org/apache/spark/deploy/yarn/ApplicationMasterArguments.scala
@@ -0,0 +1,94 @@
+/*
+ * 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 org.apache.spark.util.IntParam
+import collection.mutable.ArrayBuffer
+
+class ApplicationMasterArguments(val args: Array[String]) {
+ var userJar: String = null
+ var userClass: String = null
+ var userArgs: Seq[String] = Seq[String]()
+ var workerMemory = 1024
+ var workerCores = 1
+ var numWorkers = 2
+
+ parseArgs(args.toList)
+
+ private def parseArgs(inputArgs: List[String]): Unit = {
+ val userArgsBuffer = new ArrayBuffer[String]()
+
+ 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 ("--num-workers") :: IntParam(value) :: tail =>
+ numWorkers = value
+ args = tail
+
+ case ("--worker-memory") :: IntParam(value) :: tail =>
+ workerMemory = value
+ args = tail
+
+ case ("--worker-cores") :: IntParam(value) :: tail =>
+ workerCores = value
+ args = tail
+
+ case Nil =>
+ if (userJar == null || userClass == null) {
+ printUsageAndExit(1)
+ }
+
+ case _ =>
+ printUsageAndExit(1, args)
+ }
+ }
+
+ userArgs = userArgsBuffer.readOnly
+ }
+
+ 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.ApplicationMaster [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)\n" +
+ " --worker-memory MEM Memory per Worker (e.g. 1000M, 2G) (Default: 1G)\n")
+ System.exit(exitCode)
+ }
+}
diff --git a/new-yarn/src/main/scala/org/apache/spark/deploy/yarn/Client.scala b/new-yarn/src/main/scala/org/apache/spark/deploy/yarn/Client.scala
new file mode 100644
index 0000000000000..94678815e806a
--- /dev/null
+++ b/new-yarn/src/main/scala/org/apache/spark/deploy/yarn/Client.scala
@@ -0,0 +1,519 @@
+/*
+ * 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
+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 {
+
+ 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()
+
+ // 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,
+ 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 = System.getProperty("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)
+ 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){
+ 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 = {
+ while (true) {
+ Thread.sleep(1000)
+ 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 args = new ClientArguments(argStrings)
+
+ (new Client(args)).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]) {
+ 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.$() +
+ Path.SEPARATOR + LOG4J_PROP)
+ }
+ // Normally the users app.jar is last in case conflicts with spark jars
+ val userClasspathFirst = System.getProperty("spark.yarn.user.classpath.first", "false")
+ .toBoolean
+ if (userClasspathFirst) {
+ Apps.addToEnvironment(env, Environment.CLASSPATH.name, Environment.PWD.$() +
+ 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 + "*")
+ }
+}
diff --git a/new-yarn/src/main/scala/org/apache/spark/deploy/yarn/ClientArguments.scala b/new-yarn/src/main/scala/org/apache/spark/deploy/yarn/ClientArguments.scala
new file mode 100644
index 0000000000000..9efb28a942672
--- /dev/null
+++ b/new-yarn/src/main/scala/org/apache/spark/deploy/yarn/ClientArguments.scala
@@ -0,0 +1,148 @@
+/*
+ * 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.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]) {
+ 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 = System.getProperty("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
+
+ 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)
+ }
+
+}
diff --git a/new-yarn/src/main/scala/org/apache/spark/deploy/yarn/ClientDistributedCacheManager.scala b/new-yarn/src/main/scala/org/apache/spark/deploy/yarn/ClientDistributedCacheManager.scala
new file mode 100644
index 0000000000000..5f159b073f537
--- /dev/null
+++ b/new-yarn/src/main/scala/org/apache/spark/deploy/yarn/ClientDistributedCacheManager.scala
@@ -0,0 +1,228 @@
+/*
+ * 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.URI
+
+import org.apache.hadoop.conf.Configuration
+import org.apache.hadoop.fs.FileStatus
+import org.apache.hadoop.fs.FileSystem
+import org.apache.hadoop.fs.Path
+import org.apache.hadoop.fs.permission.FsAction
+import org.apache.hadoop.yarn.api.records.LocalResource
+import org.apache.hadoop.yarn.api.records.LocalResourceVisibility
+import org.apache.hadoop.yarn.api.records.LocalResourceType
+import org.apache.hadoop.yarn.util.{Records, ConverterUtils}
+
+import org.apache.spark.Logging
+
+import scala.collection.mutable.HashMap
+import scala.collection.mutable.LinkedHashMap
+import scala.collection.mutable.Map
+
+
+/** Client side methods to setup the Hadoop distributed cache */
+class ClientDistributedCacheManager() extends Logging {
+ private val distCacheFiles: Map[String, Tuple3[String, String, String]] =
+ LinkedHashMap[String, Tuple3[String, String, String]]()
+ private val distCacheArchives: Map[String, Tuple3[String, String, String]] =
+ LinkedHashMap[String, Tuple3[String, String, String]]()
+
+
+ /**
+ * Add a resource to the list of distributed cache resources. This list can
+ * be sent to the ApplicationMaster and possibly the workers so that it can
+ * be downloaded into the Hadoop distributed cache for use by this application.
+ * Adds the LocalResource to the localResources HashMap passed in and saves
+ * the stats of the resources to they can be sent to the workers and verified.
+ *
+ * @param fs FileSystem
+ * @param conf Configuration
+ * @param destPath path to the resource
+ * @param localResources localResource hashMap to insert the resource into
+ * @param resourceType LocalResourceType
+ * @param link link presented in the distributed cache to the destination
+ * @param statCache cache to store the file/directory stats
+ * @param appMasterOnly Whether to only add the resource to the app master
+ */
+ def addResource(
+ fs: FileSystem,
+ conf: Configuration,
+ destPath: Path,
+ localResources: HashMap[String, LocalResource],
+ resourceType: LocalResourceType,
+ link: String,
+ statCache: Map[URI, FileStatus],
+ appMasterOnly: Boolean = false) = {
+ val destStatus = fs.getFileStatus(destPath)
+ val amJarRsrc = Records.newRecord(classOf[LocalResource]).asInstanceOf[LocalResource]
+ amJarRsrc.setType(resourceType)
+ val visibility = getVisibility(conf, destPath.toUri(), statCache)
+ amJarRsrc.setVisibility(visibility)
+ amJarRsrc.setResource(ConverterUtils.getYarnUrlFromPath(destPath))
+ amJarRsrc.setTimestamp(destStatus.getModificationTime())
+ amJarRsrc.setSize(destStatus.getLen())
+ if (link == null || link.isEmpty()) throw new Exception("You must specify a valid link name")
+ localResources(link) = amJarRsrc
+
+ if (appMasterOnly == false) {
+ val uri = destPath.toUri()
+ val pathURI = new URI(uri.getScheme(), uri.getAuthority(), uri.getPath(), null, link)
+ if (resourceType == LocalResourceType.FILE) {
+ distCacheFiles(pathURI.toString()) = (destStatus.getLen().toString(),
+ destStatus.getModificationTime().toString(), visibility.name())
+ } else {
+ distCacheArchives(pathURI.toString()) = (destStatus.getLen().toString(),
+ destStatus.getModificationTime().toString(), visibility.name())
+ }
+ }
+ }
+
+ /**
+ * Adds the necessary cache file env variables to the env passed in
+ * @param env
+ */
+ def setDistFilesEnv(env: Map[String, String]) = {
+ val (keys, tupleValues) = distCacheFiles.unzip
+ val (sizes, timeStamps, visibilities) = tupleValues.unzip3
+
+ if (keys.size > 0) {
+ env("SPARK_YARN_CACHE_FILES") = keys.reduceLeft[String] { (acc,n) => acc + "," + n }
+ env("SPARK_YARN_CACHE_FILES_TIME_STAMPS") =
+ timeStamps.reduceLeft[String] { (acc,n) => acc + "," + n }
+ env("SPARK_YARN_CACHE_FILES_FILE_SIZES") =
+ sizes.reduceLeft[String] { (acc,n) => acc + "," + n }
+ env("SPARK_YARN_CACHE_FILES_VISIBILITIES") =
+ visibilities.reduceLeft[String] { (acc,n) => acc + "," + n }
+ }
+ }
+
+ /**
+ * Adds the necessary cache archive env variables to the env passed in
+ * @param env
+ */
+ def setDistArchivesEnv(env: Map[String, String]) = {
+ val (keys, tupleValues) = distCacheArchives.unzip
+ val (sizes, timeStamps, visibilities) = tupleValues.unzip3
+
+ if (keys.size > 0) {
+ env("SPARK_YARN_CACHE_ARCHIVES") = keys.reduceLeft[String] { (acc,n) => acc + "," + n }
+ env("SPARK_YARN_CACHE_ARCHIVES_TIME_STAMPS") =
+ timeStamps.reduceLeft[String] { (acc,n) => acc + "," + n }
+ env("SPARK_YARN_CACHE_ARCHIVES_FILE_SIZES") =
+ sizes.reduceLeft[String] { (acc,n) => acc + "," + n }
+ env("SPARK_YARN_CACHE_ARCHIVES_VISIBILITIES") =
+ visibilities.reduceLeft[String] { (acc,n) => acc + "," + n }
+ }
+ }
+
+ /**
+ * Returns the local resource visibility depending on the cache file permissions
+ * @param conf
+ * @param uri
+ * @param statCache
+ * @return LocalResourceVisibility
+ */
+ def getVisibility(conf: Configuration, uri: URI, statCache: Map[URI, FileStatus]):
+ LocalResourceVisibility = {
+ if (isPublic(conf, uri, statCache)) {
+ return LocalResourceVisibility.PUBLIC
+ }
+ return LocalResourceVisibility.PRIVATE
+ }
+
+ /**
+ * Returns a boolean to denote whether a cache file is visible to all(public)
+ * or not
+ * @param conf
+ * @param uri
+ * @param statCache
+ * @return true if the path in the uri is visible to all, false otherwise
+ */
+ def isPublic(conf: Configuration, uri: URI, statCache: Map[URI, FileStatus]): Boolean = {
+ val fs = FileSystem.get(uri, conf)
+ val current = new Path(uri.getPath())
+ //the leaf level file should be readable by others
+ if (!checkPermissionOfOther(fs, current, FsAction.READ, statCache)) {
+ return false
+ }
+ return ancestorsHaveExecutePermissions(fs, current.getParent(), statCache)
+ }
+
+ /**
+ * Returns true if all ancestors of the specified path have the 'execute'
+ * permission set for all users (i.e. that other users can traverse
+ * the directory heirarchy to the given path)
+ * @param fs
+ * @param path
+ * @param statCache
+ * @return true if all ancestors have the 'execute' permission set for all users
+ */
+ def ancestorsHaveExecutePermissions(fs: FileSystem, path: Path,
+ statCache: Map[URI, FileStatus]): Boolean = {
+ var current = path
+ while (current != null) {
+ //the subdirs in the path should have execute permissions for others
+ if (!checkPermissionOfOther(fs, current, FsAction.EXECUTE, statCache)) {
+ return false
+ }
+ current = current.getParent()
+ }
+ return true
+ }
+
+ /**
+ * Checks for a given path whether the Other permissions on it
+ * imply the permission in the passed FsAction
+ * @param fs
+ * @param path
+ * @param action
+ * @param statCache
+ * @return true if the path in the uri is visible to all, false otherwise
+ */
+ def checkPermissionOfOther(fs: FileSystem, path: Path,
+ action: FsAction, statCache: Map[URI, FileStatus]): Boolean = {
+ val status = getFileStatus(fs, path.toUri(), statCache)
+ val perms = status.getPermission()
+ val otherAction = perms.getOtherAction()
+ if (otherAction.implies(action)) {
+ return true
+ }
+ return false
+ }
+
+ /**
+ * Checks to see if the given uri exists in the cache, if it does it
+ * returns the existing FileStatus, otherwise it stats the uri, stores
+ * it in the cache, and returns the FileStatus.
+ * @param fs
+ * @param uri
+ * @param statCache
+ * @return FileStatus
+ */
+ def getFileStatus(fs: FileSystem, uri: URI, statCache: Map[URI, FileStatus]): FileStatus = {
+ val stat = statCache.get(uri) match {
+ case Some(existstat) => existstat
+ case None =>
+ val newStat = fs.getFileStatus(new Path(uri))
+ statCache.put(uri, newStat)
+ newStat
+ }
+ return stat
+ }
+}
diff --git a/new-yarn/src/main/scala/org/apache/spark/deploy/yarn/WorkerLauncher.scala b/new-yarn/src/main/scala/org/apache/spark/deploy/yarn/WorkerLauncher.scala
new file mode 100644
index 0000000000000..c38f33e212fbf
--- /dev/null
+++ b/new-yarn/src/main/scala/org/apache/spark/deploy/yarn/WorkerLauncher.scala
@@ -0,0 +1,223 @@
+/*
+ * 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.Socket
+import org.apache.hadoop.conf.Configuration
+import org.apache.hadoop.net.NetUtils
+import org.apache.hadoop.yarn.api._
+import org.apache.hadoop.yarn.api.records._
+import org.apache.hadoop.yarn.api.protocolrecords._
+import org.apache.hadoop.yarn.conf.YarnConfiguration
+import org.apache.hadoop.yarn.util.{ConverterUtils, Records}
+import akka.actor._
+import akka.remote.{RemoteClientShutdown, RemoteClientDisconnected, RemoteClientLifeCycleEvent}
+import akka.actor.Terminated
+import org.apache.spark.{SparkContext, Logging}
+import org.apache.spark.util.{Utils, AkkaUtils}
+import org.apache.spark.scheduler.cluster.CoarseGrainedSchedulerBackend
+import org.apache.spark.scheduler.SplitInfo
+import org.apache.hadoop.yarn.client.api.AMRMClient
+import org.apache.hadoop.yarn.client.api.AMRMClient.ContainerRequest
+
+class WorkerLauncher(args: ApplicationMasterArguments, conf: Configuration) extends Logging {
+
+ def this(args: ApplicationMasterArguments) = this(args, new Configuration())
+
+ private var appAttemptId: ApplicationAttemptId = _
+ private var reporterThread: Thread = _
+ private val yarnConf: YarnConfiguration = new YarnConfiguration(conf)
+
+ private var yarnAllocator: YarnAllocationHandler = _
+ private var driverClosed:Boolean = false
+
+ private var amClient: AMRMClient[ContainerRequest] = _
+
+ val actorSystem : ActorSystem = AkkaUtils.createActorSystem("sparkYarnAM", Utils.localHostName, 0)._1
+ var actor: ActorRef = _
+
+ // This actor just working as a monitor to watch on Driver Actor.
+ class MonitorActor(driverUrl: String) extends Actor {
+
+ var driver: ActorRef = null
+
+ override def preStart() {
+ logInfo("Listen to driver: " + driverUrl)
+ driver = context.actorFor(driverUrl)
+ context.system.eventStream.subscribe(self, classOf[RemoteClientLifeCycleEvent])
+ context.watch(driver) // Doesn't work with remote actors, but useful for testing
+ }
+
+ override def receive = {
+ case Terminated(_) | RemoteClientDisconnected(_, _) | RemoteClientShutdown(_, _) =>
+ logInfo("Driver terminated or disconnected! Shutting down.")
+ driverClosed = true
+ }
+ }
+
+ def run() {
+
+ amClient = AMRMClient.createAMRMClient()
+ amClient.init(yarnConf)
+ amClient.start()
+
+ appAttemptId = getApplicationAttemptId()
+ registerApplicationMaster()
+
+ waitForSparkMaster()
+
+ // Allocate all containers
+ allocateWorkers()
+
+ // Launch a progress reporter thread, else app will get killed after expiration (def: 10mins) timeout
+ // ensure that progress is sent before YarnConfiguration.RM_AM_EXPIRY_INTERVAL_MS elapse.
+
+ val timeoutInterval = yarnConf.getInt(YarnConfiguration.RM_AM_EXPIRY_INTERVAL_MS, 120000)
+ // must be <= timeoutInterval/ 2.
+ // On other hand, also ensure that we are reasonably responsive without causing too many requests to RM.
+ // so atleast 1 minute or timeoutInterval / 10 - whichever is higher.
+ val interval = math.min(timeoutInterval / 2, math.max(timeoutInterval/ 10, 60000L))
+ reporterThread = launchReporterThread(interval)
+
+ // Wait for the reporter thread to Finish.
+ reporterThread.join()
+
+ finishApplicationMaster(FinalApplicationStatus.SUCCEEDED)
+ actorSystem.shutdown()
+
+ logInfo("Exited")
+ System.exit(0)
+ }
+
+ 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")
+ // TODO:(Raymond) Find out Spark UI address and fill in here?
+ amClient.registerApplicationMaster(Utils.localHostName(), 0, "")
+ }
+
+ private def waitForSparkMaster() {
+ logInfo("Waiting for Spark driver to be reachable.")
+ var driverUp = false
+ val hostport = args.userArgs(0)
+ val (driverHost, driverPort) = Utils.parseHostPort(hostport)
+ while(!driverUp) {
+ try {
+ val socket = new Socket(driverHost, driverPort)
+ socket.close()
+ logInfo("Driver now available: %s:%s".format(driverHost, driverPort))
+ driverUp = true
+ } catch {
+ case e: Exception =>
+ logError("Failed to connect to driver at %s:%s, retrying ...".
+ format(driverHost, driverPort))
+ Thread.sleep(100)
+ }
+ }
+ System.setProperty("spark.driver.host", driverHost)
+ System.setProperty("spark.driver.port", driverPort.toString)
+
+ val driverUrl = "akka://spark@%s:%s/user/%s".format(
+ driverHost, driverPort.toString, CoarseGrainedSchedulerBackend.ACTOR_NAME)
+
+ actor = actorSystem.actorOf(Props(new MonitorActor(driverUrl)), name = "YarnAM")
+ }
+
+
+ private def allocateWorkers() {
+
+ // Fixme: should get preferredNodeLocationData from SparkContext, just fake a empty one for now.
+ val preferredNodeLocationData: scala.collection.Map[String, scala.collection.Set[SplitInfo]] =
+ scala.collection.immutable.Map()
+
+ yarnAllocator = YarnAllocationHandler.newAllocator(
+ yarnConf,
+ amClient,
+ appAttemptId,
+ args,
+ preferredNodeLocationData)
+
+ 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)
+ while(yarnAllocator.getNumWorkersRunning < args.numWorkers) {
+ yarnAllocator.allocateResources()
+ Thread.sleep(100)
+ }
+
+ logInfo("All workers have launched.")
+
+ }
+
+ // TODO: We might want to extend this to allocate more containers in case they die !
+ private def launchReporterThread(_sleepTime: Long): Thread = {
+ val sleepTime = if (_sleepTime <= 0 ) 0 else _sleepTime
+
+ val t = new Thread {
+ override def run() {
+ while (!driverClosed) {
+ 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 finishApplicationMaster(status: FinalApplicationStatus) {
+ logInfo("finish ApplicationMaster with " + status)
+ amClient.unregisterApplicationMaster(status, "" /* appMessage */, "" /* appTrackingUrl */)
+ }
+
+}
+
+
+object WorkerLauncher {
+ def main(argStrings: Array[String]) {
+ val args = new ApplicationMasterArguments(argStrings)
+ new WorkerLauncher(args).run()
+ }
+}
diff --git a/new-yarn/src/main/scala/org/apache/spark/deploy/yarn/WorkerRunnable.scala b/new-yarn/src/main/scala/org/apache/spark/deploy/yarn/WorkerRunnable.scala
new file mode 100644
index 0000000000000..9f5523c4b97a8
--- /dev/null
+++ b/new-yarn/src/main/scala/org/apache/spark/deploy/yarn/WorkerRunnable.scala
@@ -0,0 +1,209 @@
+/*
+ * 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.URI
+import java.nio.ByteBuffer
+import java.security.PrivilegedExceptionAction
+
+import scala.collection.JavaConversions._
+import scala.collection.mutable.HashMap
+
+import org.apache.hadoop.conf.Configuration
+import org.apache.hadoop.fs.Path
+import org.apache.hadoop.io.DataOutputBuffer
+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.records._
+import org.apache.hadoop.yarn.api.records.impl.pb.ProtoUtils
+import org.apache.hadoop.yarn.api.protocolrecords._
+import org.apache.hadoop.yarn.client.api.NMClient
+import org.apache.hadoop.yarn.conf.YarnConfiguration
+import org.apache.hadoop.yarn.ipc.YarnRPC
+import org.apache.hadoop.yarn.util.{Apps, ConverterUtils, Records}
+
+import org.apache.spark.Logging
+
+
+class WorkerRunnable(
+ container: Container,
+ conf: Configuration,
+ masterAddress: String,
+ slaveId: String,
+ hostname: String,
+ workerMemory: Int,
+ workerCores: Int)
+ extends Runnable with Logging {
+
+ var rpc: YarnRPC = YarnRPC.create(conf)
+ var nmClient: NMClient = _
+ val yarnConf: YarnConfiguration = new YarnConfiguration(conf)
+
+ def run = {
+ logInfo("Starting Worker Container")
+ nmClient = NMClient.createNMClient()
+ nmClient.init(yarnConf)
+ nmClient.start()
+ startContainer
+ }
+
+ def startContainer = {
+ logInfo("Setting up ContainerLaunchContext")
+
+ val ctx = Records.newRecord(classOf[ContainerLaunchContext])
+ .asInstanceOf[ContainerLaunchContext]
+
+ val localResources = prepareLocalResources
+ ctx.setLocalResources(localResources)
+
+ val env = prepareEnvironment
+ ctx.setEnvironment(env)
+
+ // Extra options for the JVM
+ var JAVA_OPTS = ""
+ // Set the JVM memory
+ val workerMemoryString = workerMemory + "m"
+ JAVA_OPTS += "-Xms" + workerMemoryString + " -Xmx" + workerMemoryString + " "
+ if (env.isDefinedAt("SPARK_JAVA_OPTS")) {
+ JAVA_OPTS += env("SPARK_JAVA_OPTS") + " "
+ }
+
+ JAVA_OPTS += " -Djava.io.tmpdir=" +
+ new Path(Environment.PWD.$(), YarnConfiguration.DEFAULT_CONTAINER_TEMP_DIR) + " "
+
+ // Commenting it out for now - so that people can refer to the properties if required. Remove
+ // it once cpuset version is pushed out.
+ // The context is, default gc for server class machines end up using all cores to do gc - hence
+ // if there are multiple containers in same node, spark gc effects all other containers
+ // performance (which can also be other spark containers)
+ // Instead of using this, rely on cpusets by YARN to enforce spark behaves 'properly' in
+ // multi-tenant environments. Not sure how default java gc behaves if it is limited to subset
+ // of cores on a node.
+/*
+ else {
+ // If no java_opts specified, default to using -XX:+CMSIncrementalMode
+ // It might be possible that other modes/config is being done in SPARK_JAVA_OPTS, so we dont
+ // want to mess with it.
+ // In our expts, using (default) throughput collector has severe perf ramnifications in
+ // multi-tennent machines
+ // The options are based on
+ // http://www.oracle.com/technetwork/java/gc-tuning-5-138395.html#0.0.0.%20When%20to%20Use%20the%20Concurrent%20Low%20Pause%20Collector|outline
+ JAVA_OPTS += " -XX:+UseConcMarkSweepGC "
+ JAVA_OPTS += " -XX:+CMSIncrementalMode "
+ JAVA_OPTS += " -XX:+CMSIncrementalPacing "
+ JAVA_OPTS += " -XX:CMSIncrementalDutyCycleMin=0 "
+ JAVA_OPTS += " -XX:CMSIncrementalDutyCycle=10 "
+ }
+*/
+
+ val credentials = UserGroupInformation.getCurrentUser().getCredentials()
+ val dob = new DataOutputBuffer()
+ credentials.writeTokenStorageToStream(dob)
+ ctx.setTokens(ByteBuffer.wrap(dob.getData()))
+
+ 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 " +
+ // Kill if OOM is raised - leverage yarn's failure handling to cause rescheduling.
+ // Not killing the task leaves various aspects of the worker and (to some extent) the jvm in
+ // an inconsistent state.
+ // TODO: If the OOM is not recoverable by rescheduling it on different node, then do
+ // 'something' to fail job ... akin to blacklisting trackers in mapred ?
+ " -XX:OnOutOfMemoryError='kill %p' " +
+ JAVA_OPTS +
+ " org.apache.spark.executor.CoarseGrainedExecutorBackend " +
+ masterAddress + " " +
+ slaveId + " " +
+ hostname + " " +
+ workerCores +
+ " 1> " + ApplicationConstants.LOG_DIR_EXPANSION_VAR + "/stdout" +
+ " 2> " + ApplicationConstants.LOG_DIR_EXPANSION_VAR + "/stderr")
+ logInfo("Setting up worker with commands: " + commands)
+ ctx.setCommands(commands)
+
+ // Send the start request to the ContainerManager
+ nmClient.startContainer(container, ctx)
+ }
+
+ private def setupDistributedCache(
+ file: String,
+ rtype: LocalResourceType,
+ localResources: HashMap[String, LocalResource],
+ timestamp: String,
+ size: String,
+ vis: String) = {
+ val uri = new URI(file)
+ val amJarRsrc = Records.newRecord(classOf[LocalResource]).asInstanceOf[LocalResource]
+ amJarRsrc.setType(rtype)
+ amJarRsrc.setVisibility(LocalResourceVisibility.valueOf(vis))
+ amJarRsrc.setResource(ConverterUtils.getYarnUrlFromURI(uri))
+ amJarRsrc.setTimestamp(timestamp.toLong)
+ amJarRsrc.setSize(size.toLong)
+ localResources(uri.getFragment()) = amJarRsrc
+ }
+
+ def prepareLocalResources: HashMap[String, LocalResource] = {
+ logInfo("Preparing Local resources")
+ val localResources = HashMap[String, LocalResource]()
+
+ if (System.getenv("SPARK_YARN_CACHE_FILES") != null) {
+ val timeStamps = System.getenv("SPARK_YARN_CACHE_FILES_TIME_STAMPS").split(',')
+ val fileSizes = System.getenv("SPARK_YARN_CACHE_FILES_FILE_SIZES").split(',')
+ val distFiles = System.getenv("SPARK_YARN_CACHE_FILES").split(',')
+ val visibilities = System.getenv("SPARK_YARN_CACHE_FILES_VISIBILITIES").split(',')
+ for( i <- 0 to distFiles.length - 1) {
+ setupDistributedCache(distFiles(i), LocalResourceType.FILE, localResources, timeStamps(i),
+ fileSizes(i), visibilities(i))
+ }
+ }
+
+ if (System.getenv("SPARK_YARN_CACHE_ARCHIVES") != null) {
+ val timeStamps = System.getenv("SPARK_YARN_CACHE_ARCHIVES_TIME_STAMPS").split(',')
+ val fileSizes = System.getenv("SPARK_YARN_CACHE_ARCHIVES_FILE_SIZES").split(',')
+ val distArchives = System.getenv("SPARK_YARN_CACHE_ARCHIVES").split(',')
+ val visibilities = System.getenv("SPARK_YARN_CACHE_ARCHIVES_VISIBILITIES").split(',')
+ for( i <- 0 to distArchives.length - 1) {
+ setupDistributedCache(distArchives(i), LocalResourceType.ARCHIVE, localResources,
+ timeStamps(i), fileSizes(i), visibilities(i))
+ }
+ }
+
+ logInfo("Prepared Local resources " + localResources)
+ localResources
+ }
+
+ def prepareEnvironment: HashMap[String, String] = {
+ val env = new HashMap[String, String]()
+
+ Client.populateClasspath(yarnConf, System.getenv("SPARK_YARN_LOG4J_PATH") != null, env)
+
+ // Allow users to specify some environment variables
+ Apps.setEnvFromInputString(env, System.getenv("SPARK_YARN_USER_ENV"))
+
+ System.getenv().filterKeys(_.startsWith("SPARK")).foreach { case (k,v) => env(k) = v }
+ env
+ }
+
+}
diff --git a/new-yarn/src/main/scala/org/apache/spark/deploy/yarn/YarnAllocationHandler.scala b/new-yarn/src/main/scala/org/apache/spark/deploy/yarn/YarnAllocationHandler.scala
new file mode 100644
index 0000000000000..dba0f7640e67c
--- /dev/null
+++ b/new-yarn/src/main/scala/org/apache/spark/deploy/yarn/YarnAllocationHandler.scala
@@ -0,0 +1,687 @@
+/*
+ * 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
+import org.apache.spark.scheduler.SplitInfo
+import org.apache.spark.scheduler.cluster.{ClusterScheduler, CoarseGrainedSchedulerBackend}
+import org.apache.spark.util.Utils
+
+import org.apache.hadoop.conf.Configuration
+import org.apache.hadoop.yarn.api.ApplicationMasterProtocol
+import org.apache.hadoop.yarn.api.records.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.client.api.AMRMClient
+import org.apache.hadoop.yarn.client.api.AMRMClient.ContainerRequest
+import org.apache.hadoop.yarn.util.{RackResolver, Records}
+
+
+object AllocationType extends Enumeration ("HOST", "RACK", "ANY") {
+ 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 amClient: AMRMClient[ContainerRequest],
+ val appAttemptId: ApplicationAttemptId,
+ val maxWorkers: Int,
+ val workerMemory: Int,
+ val workerCores: Int,
+ val preferredHostToCount: Map[String, Int],
+ val preferredRackToCount: Map[String, Int])
+ extends Logging {
+ // These three are locked on allocatedHostToContainersMap. Complementary data structures
+ // allocatedHostToContainersMap : containers which are running : host, Set
+ // 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]
+
+ // Number of container requests that have been sent to, but not yet allocated by the
+ // ApplicationMaster.
+ private val numPendingAllocate = new AtomicInteger()
+ 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 getNumPendingAllocate: Int = numPendingAllocate.intValue
+
+ def getNumWorkersRunning: Int = numWorkersRunning.intValue
+
+ def getNumWorkersFailed: Int = numWorkersFailed.intValue
+
+ def isResourceConstraintSatisfied(container: Container): Boolean = {
+ container.getResource.getMemory >= (workerMemory + YarnAllocationHandler.MEMORY_OVERHEAD)
+ }
+
+ def releaseContainer(container: Container) {
+ val containerId = container.getId
+ pendingReleaseContainers.put(containerId, true)
+ amClient.releaseAssignedContainer(containerId)
+ }
+
+ def allocateResources() {
+ // We have already set the container request. Poll the ResourceManager for a response.
+ // This doubles as a heartbeat if there are no pending container requests.
+ val progressIndicator = 0.1f
+ val allocateResponse = amClient.allocate(progressIndicator)
+
+ val allocatedContainers = allocateResponse.getAllocatedContainers()
+ if (allocatedContainers.size > 0) {
+ var numPendingAllocateNow = numPendingAllocate.addAndGet(-1 * allocatedContainers.size)
+
+ if (numPendingAllocateNow < 0) {
+ numPendingAllocateNow = numPendingAllocate.addAndGet(-1 * numPendingAllocateNow)
+ }
+
+ 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,
+ allocateResponse.getAvailableResources))
+
+ val hostToContainers = new HashMap[String, ArrayBuffer[Container]]()
+
+ for (container <- allocatedContainers) {
+ if (isResourceConstraintSatisfied(container)) {
+ // Add the accepted `container` to the host's list of already accepted,
+ // allocated containers
+ val host = container.getNodeId.getHost
+ val containersForHost = hostToContainers.getOrElseUpdate(host,
+ new ArrayBuffer[Container]())
+ containersForHost += container
+ } else {
+ // Release container, since it doesn't satisfy resource constraints.
+ releaseContainer(container)
+ }
+ }
+
+ // Find the appropriate containers to use.
+ // TODO: Cleanup this group-by...
+ 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)
+
+ val remainingContainersOpt = hostToContainers.get(candidateHost)
+ assert(remainingContainersOpt.isDefined)
+ var remainingContainers = remainingContainersOpt.get
+
+ if (requiredHostCount >= remainingContainers.size) {
+ // Since we have <= required containers, add all remaining containers to
+ // `dataLocalContainers`.
+ dataLocalContainers.put(candidateHost, remainingContainers)
+ // There are no more free containers remaining.
+ remainingContainers = null
+ } else if (requiredHostCount > 0) {
+ // Container list has more containers than we need for data locality.
+ // Split the list into two: one based on the data local container count,
+ // (`remainingContainers.size` - `requiredHostCount`), and the other to hold remaining
+ // containers.
+ val (dataLocal, remaining) = remainingContainers.splitAt(
+ remainingContainers.size - requiredHostCount)
+ dataLocalContainers.put(candidateHost, dataLocal)
+
+ // Invariant: remainingContainers == remaining
+
+ // YARN has a nasty habit of allocating a ton of containers on a host - discourage this.
+ // Add each container in `remaining` to list of containers to release. If we have an
+ // insufficient number of containers, then the next allocation cycle will reallocate
+ // (but won't treat it as data local).
+ // TODO(harvey): Rephrase this comment some more.
+ for (container <- remaining) releaseContainer(container)
+ remainingContainers = null
+ }
+
+ // For rack local containers
+ if (remainingContainers != null) {
+ val rack = YarnAllocationHandler.lookupRack(conf, candidateHost)
+ if (rack != null) {
+ val maxExpectedRackCount = preferredRackToCount.getOrElse(rack, 0)
+ val requiredRackCount = maxExpectedRackCount - allocatedContainersOnRack(rack) -
+ rackLocalContainers.getOrElse(rack, List()).size
+
+ if (requiredRackCount >= remainingContainers.size) {
+ // Add all remaining containers to to `dataLocalContainers`.
+ dataLocalContainers.put(rack, remainingContainers)
+ remainingContainers = null
+ } else if (requiredRackCount > 0) {
+ // Container list has more containers that we need for data locality.
+ // Split the list into two: one based on the data local container count,
+ // (`remainingContainers.size` - `requiredHostCount`), and the other to hold remaining
+ // containers.
+ val (rackLocal, remaining) = remainingContainers.splitAt(
+ remainingContainers.size - requiredRackCount)
+ val existingRackLocal = rackLocalContainers.getOrElseUpdate(rack,
+ new ArrayBuffer[Container]())
+
+ existingRackLocal ++= rackLocal
+
+ remainingContainers = remaining
+ }
+ }
+ }
+
+ if (remainingContainers != null) {
+ // Not all containers have been consumed - add them to the list of off-rack containers.
+ 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 finally off-rack.
+ // Note that the list we create below tries to ensure that not all containers end up within
+ // a host if there is a sufficiently large number of hosts/containers.
+ val allocatedContainersToProcess = new ArrayBuffer[Container](allocatedContainers.size)
+ allocatedContainersToProcess ++= ClusterScheduler.prioritizeContainers(dataLocalContainers)
+ allocatedContainersToProcess ++= ClusterScheduler.prioritizeContainers(rackLocalContainers)
+ allocatedContainersToProcess ++= ClusterScheduler.prioritizeContainers(offRackContainers)
+
+ // Run each of the allocated containers.
+ for (container <- allocatedContainersToProcess) {
+ val numWorkersRunningNow = numWorkersRunning.incrementAndGet()
+ val workerHostname = container.getNodeId.getHost
+ val containerId = container.getId
+
+ val workerMemoryOverhead = (workerMemory + YarnAllocationHandler.MEMORY_OVERHEAD)
+ assert(container.getResource.getMemory >= workerMemoryOverhead)
+
+ if (numWorkersRunningNow > maxWorkers) {
+ logInfo("""Ignoring container %s at host %s, since we already have the required number of
+ containers for it.""".format(containerId, workerHostname))
+ releaseContainer(container)
+ numWorkersRunning.decrementAndGet()
+ } else {
+ val workerId = workerIdCounter.incrementAndGet().toString
+ val driverUrl = "akka://spark@%s:%s/user/%s".format(
+ System.getProperty("spark.driver.host"),
+ System.getProperty("spark.driver.port"),
+ CoarseGrainedSchedulerBackend.ACTOR_NAME)
+
+ logInfo("Launching container %s for on host %s".format(containerId, workerHostname))
+
+ // To be safe, remove the container from `pendingReleaseContainers`.
+ 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)
+ }
+ }
+ logInfo("Launching WorkerRunnable. driverUrl: %s, workerHostname: %s".format(driverUrl, workerHostname))
+ val workerRunnable = new WorkerRunnable(
+ container,
+ conf,
+ driverUrl,
+ workerId,
+ workerHostname,
+ workerMemory,
+ workerCores)
+ new Thread(workerRunnable).start()
+ }
+ }
+ logDebug("""
+ Finished allocating %s containers (from %s originally).
+ Current number of workers running: %d,
+ releasedContainerList: %s,
+ pendingReleaseContainers: %s
+ """.format(
+ allocatedContainersToProcess,
+ allocatedContainers,
+ numWorkersRunning.get(),
+ releasedContainerList,
+ pendingReleaseContainers))
+ }
+
+ val completedContainers = allocateResponse.getCompletedContainersStatuses()
+ if (completedContainers.size > 0) {
+ logDebug("Completed %d containers".format(completedContainers.size))
+
+ for (completedContainer <- completedContainers) {
+ val containerId = completedContainer.getContainerId
+
+ if (pendingReleaseContainers.containsKey(containerId)) {
+ // YarnAllocationHandler already marked the container for release, so remove it from
+ // `pendingReleaseContainers`.
+ pendingReleaseContainers.remove(containerId)
+ } else {
+ // Decrement the number of workers running. The next iteration of the ApplicationMaster's
+ // reporting thread 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 hostOpt = allocatedContainerToHostMap.get(containerId)
+ assert(hostOpt.isDefined)
+ val host = hostOpt.get
+
+ val containerSetOpt = allocatedHostToContainersMap.get(host)
+ assert(containerSetOpt.isDefined)
+ val containerSet = containerSetOpt.get
+
+ containerSet.remove(containerId)
+ if (containerSet.isEmpty) {
+ allocatedHostToContainersMap.remove(host)
+ } else {
+ allocatedHostToContainersMap.update(host, containerSet)
+ }
+
+ allocatedContainerToHostMap.remove(containerId)
+
+ // TODO: Move this part outside the synchronized block?
+ 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: ArrayBuffer[ContainerRequest]
+ ): ArrayBuffer[ContainerRequest] = {
+ // Generate modified racks and new set of hosts under it before issuing requests.
+ val rackToCounts = new HashMap[String, Int]()
+
+ for (container <- hostContainers) {
+ val candidateHost = container.getNodes.last
+ assert(YarnAllocationHandler.ANY_HOST != candidateHost)
+
+ val rack = YarnAllocationHandler.lookupRack(conf, candidateHost)
+ if (rack != null) {
+ var count = rackToCounts.getOrElse(rack, 0)
+ count += 1
+ rackToCounts.put(rack, count)
+ }
+ }
+
+ val requestedContainers = new ArrayBuffer[ContainerRequest](rackToCounts.size)
+ for ((rack, count) <- rackToCounts) {
+ requestedContainers ++= createResourceRequests(
+ AllocationType.RACK,
+ rack,
+ count,
+ YarnAllocationHandler.PRIORITY)
+ }
+
+ requestedContainers
+ }
+
+ 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
+ }
+
+ def addResourceRequests(numWorkers: Int) {
+ val containerRequests: List[ContainerRequest] =
+ if (numWorkers <= 0 || preferredHostToCount.isEmpty) {
+ logDebug("numWorkers: " + numWorkers + ", host preferences: " +
+ preferredHostToCount.isEmpty)
+ createResourceRequests(
+ AllocationType.ANY,
+ resource = null,
+ numWorkers,
+ YarnAllocationHandler.PRIORITY).toList
+ } else {
+ // Request for all hosts in preferred nodes and for numWorkers -
+ // candidates.size, request by default allocation policy.
+ val hostContainerRequests = new ArrayBuffer[ContainerRequest](preferredHostToCount.size)
+ for ((candidateHost, candidateCount) <- preferredHostToCount) {
+ val requiredCount = candidateCount - allocatedContainersOnHost(candidateHost)
+
+ if (requiredCount > 0) {
+ hostContainerRequests ++= createResourceRequests(
+ AllocationType.HOST,
+ candidateHost,
+ requiredCount,
+ YarnAllocationHandler.PRIORITY)
+ }
+ }
+ val rackContainerRequests: List[ContainerRequest] = createRackResourceRequests(
+ hostContainerRequests).toList
+
+ val anyContainerRequests = createResourceRequests(
+ AllocationType.ANY,
+ resource = null,
+ numWorkers,
+ YarnAllocationHandler.PRIORITY)
+
+ val containerRequestBuffer = new ArrayBuffer[ContainerRequest](
+ hostContainerRequests.size + rackContainerRequests.size() + anyContainerRequests.size)
+
+ containerRequestBuffer ++= hostContainerRequests
+ containerRequestBuffer ++= rackContainerRequests
+ containerRequestBuffer ++= anyContainerRequests
+ containerRequestBuffer.toList
+ }
+
+ for (request <- containerRequests) {
+ amClient.addContainerRequest(request)
+ }
+
+ if (numWorkers > 0) {
+ numPendingAllocate.addAndGet(numWorkers)
+ logInfo("Will Allocate %d worker containers, each with %d memory".format(
+ numWorkers,
+ (workerMemory + YarnAllocationHandler.MEMORY_OVERHEAD)))
+ } else {
+ logDebug("Empty allocation request ...")
+ }
+
+ for (request <- containerRequests) {
+ val nodes = request.getNodes
+ var hostStr = if (nodes == null || nodes.isEmpty) {
+ "Any"
+ } else {
+ nodes.last
+ }
+ logInfo("Container request (host: %s, priority: %s, capability: %s".format(
+ hostStr,
+ request.getPriority().getPriority,
+ request.getCapability))
+ }
+ }
+
+ private def createResourceRequests(
+ requestType: AllocationType.AllocationType,
+ resource: String,
+ numWorkers: Int,
+ priority: Int
+ ): ArrayBuffer[ContainerRequest] = {
+
+ // If hostname is specified, then we need at least 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 = constructContainerRequests(
+ Array(hostname),
+ racks = null,
+ numWorkers,
+ priority)
+
+ // Add `hostname` to the global (singleton) host->rack mapping in YarnAllocationHandler.
+ YarnAllocationHandler.populateRackInfo(conf, hostname)
+ nodeLocal
+ }
+ case AllocationType.RACK => {
+ val rack = resource
+ constructContainerRequests(hosts = null, Array(rack), numWorkers, priority)
+ }
+ case AllocationType.ANY => constructContainerRequests(
+ hosts = null, racks = null, numWorkers, priority)
+ case _ => throw new IllegalArgumentException(
+ "Unexpected/unsupported request type: " + requestType)
+ }
+ }
+
+ private def constructContainerRequests(
+ hosts: Array[String],
+ racks: Array[String],
+ numWorkers: Int,
+ priority: Int
+ ): ArrayBuffer[ContainerRequest] = {
+
+ val memoryResource = Records.newRecord(classOf[Resource])
+ memoryResource.setMemory(workerMemory + YarnAllocationHandler.MEMORY_OVERHEAD)
+
+ val prioritySetting = Records.newRecord(classOf[Priority])
+ prioritySetting.setPriority(priority)
+
+ val requests = new ArrayBuffer[ContainerRequest]()
+ for (i <- 0 until numWorkers) {
+ requests += new ContainerRequest(memoryResource, hosts, racks, prioritySetting)
+ }
+ requests
+ }
+}
+
+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 ResurceManager 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,
+ amClient: AMRMClient[ContainerRequest],
+ appAttemptId: ApplicationAttemptId,
+ args: ApplicationMasterArguments
+ ): YarnAllocationHandler = {
+ new YarnAllocationHandler(
+ conf,
+ amClient,
+ appAttemptId,
+ args.numWorkers,
+ args.workerMemory,
+ args.workerCores,
+ Map[String, Int](),
+ Map[String, Int]())
+ }
+
+ def newAllocator(
+ conf: Configuration,
+ amClient: AMRMClient[ContainerRequest],
+ appAttemptId: ApplicationAttemptId,
+ args: ApplicationMasterArguments,
+ map: collection.Map[String,
+ collection.Set[SplitInfo]]
+ ): YarnAllocationHandler = {
+ val (hostToSplitCount, rackToSplitCount) = generateNodeToWeight(conf, map)
+ new YarnAllocationHandler(
+ conf,
+ amClient,
+ appAttemptId,
+ args.numWorkers,
+ args.workerMemory,
+ args.workerCores,
+ hostToSplitCount,
+ rackToSplitCount)
+ }
+
+ def newAllocator(
+ conf: Configuration,
+ amClient: AMRMClient[ContainerRequest],
+ appAttemptId: ApplicationAttemptId,
+ maxWorkers: Int,
+ workerMemory: Int,
+ workerCores: Int,
+ map: collection.Map[String, collection.Set[SplitInfo]]
+ ): YarnAllocationHandler = {
+ val (hostToCount, rackToCount) = generateNodeToWeight(conf, map)
+ new YarnAllocationHandler(
+ conf,
+ amClient,
+ appAttemptId,
+ maxWorkers,
+ workerMemory,
+ workerCores,
+ hostToCount,
+ rackToCount)
+ }
+
+ // A simple method to copy the split info map.
+ private def generateNodeToWeight(
+ conf: Configuration,
+ input: collection.Map[String, collection.Set[SplitInfo]]
+ ): (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]] = {
+ Option(rackToHostSet.get(rack)).map { set =>
+ val convertedSet: collection.mutable.Set[String] = set
+ // TODO: Better way to get a Set[String] from JSet.
+ 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 what this comment means...
+ // 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)
+ } */
+ }
+ }
+}
diff --git a/new-yarn/src/main/scala/org/apache/spark/deploy/yarn/YarnSparkHadoopUtil.scala b/new-yarn/src/main/scala/org/apache/spark/deploy/yarn/YarnSparkHadoopUtil.scala
new file mode 100644
index 0000000000000..2ba2366ead171
--- /dev/null
+++ b/new-yarn/src/main/scala/org/apache/spark/deploy/yarn/YarnSparkHadoopUtil.scala
@@ -0,0 +1,43 @@
+/*
+ * 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 org.apache.spark.deploy.SparkHadoopUtil
+import org.apache.hadoop.mapred.JobConf
+import org.apache.hadoop.security.UserGroupInformation
+import org.apache.hadoop.yarn.conf.YarnConfiguration
+import org.apache.hadoop.conf.Configuration
+
+/**
+ * Contains util methods to interact with Hadoop from spark.
+ */
+class YarnSparkHadoopUtil extends SparkHadoopUtil {
+
+ // Note that all params which start with SPARK are propagated all the way through, so if in yarn mode, this MUST be set to true.
+ override def isYarnMode(): Boolean = { true }
+
+ // Return an appropriate (subclass) of Configuration. Creating config can initializes some hadoop subsystems
+ // Always create a new config, dont reuse yarnConf.
+ override def newConfiguration(): Configuration = new YarnConfiguration(new Configuration())
+
+ // add any user credentials to the job conf which are necessary for running on a secure Hadoop cluster
+ override def addCredentials(conf: JobConf) {
+ val jobCreds = conf.getCredentials()
+ jobCreds.mergeAll(UserGroupInformation.getCurrentUser().getCredentials())
+ }
+}
diff --git a/new-yarn/src/main/scala/org/apache/spark/scheduler/cluster/YarnClientClusterScheduler.scala b/new-yarn/src/main/scala/org/apache/spark/scheduler/cluster/YarnClientClusterScheduler.scala
new file mode 100644
index 0000000000000..63a0449e5a073
--- /dev/null
+++ b/new-yarn/src/main/scala/org/apache/spark/scheduler/cluster/YarnClientClusterScheduler.scala
@@ -0,0 +1,47 @@
+/*
+ * 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.spark._
+import org.apache.hadoop.conf.Configuration
+import org.apache.spark.deploy.yarn.YarnAllocationHandler
+import org.apache.spark.util.Utils
+
+/**
+ *
+ * This scheduler launch worker through Yarn - by call into Client to launch WorkerLauncher as AM.
+ */
+private[spark] class YarnClientClusterScheduler(sc: SparkContext, conf: Configuration) extends ClusterScheduler(sc) {
+
+ def this(sc: SparkContext) = this(sc, new Configuration())
+
+ // By default, rack is unknown
+ override def getRackForHost(hostPort: String): Option[String] = {
+ val host = Utils.parseHostPort(hostPort)._1
+ val retval = YarnAllocationHandler.lookupRack(conf, host)
+ if (retval != null) Some(retval) else None
+ }
+
+ override def postStartHook() {
+
+ // The yarn application is running, but the worker might not yet ready
+ // Wait for a few seconds for the slaves to bootstrap and register with master - best case attempt
+ Thread.sleep(2000L)
+ logInfo("YarnClientClusterScheduler.postStartHook done")
+ }
+}
diff --git a/new-yarn/src/main/scala/org/apache/spark/scheduler/cluster/YarnClientSchedulerBackend.scala b/new-yarn/src/main/scala/org/apache/spark/scheduler/cluster/YarnClientSchedulerBackend.scala
new file mode 100644
index 0000000000000..b206780c7806e
--- /dev/null
+++ b/new-yarn/src/main/scala/org/apache/spark/scheduler/cluster/YarnClientSchedulerBackend.scala
@@ -0,0 +1,109 @@
+/*
+ * 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}
+
+private[spark] class YarnClientSchedulerBackend(
+ scheduler: ClusterScheduler,
+ 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 = System.getProperty("spark.driver.host")
+ val driverPort = System.getProperty("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)
+ 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")
+ }
+
+}
diff --git a/new-yarn/src/main/scala/org/apache/spark/scheduler/cluster/YarnClusterScheduler.scala b/new-yarn/src/main/scala/org/apache/spark/scheduler/cluster/YarnClusterScheduler.scala
new file mode 100644
index 0000000000000..29b3f22e13697
--- /dev/null
+++ b/new-yarn/src/main/scala/org/apache/spark/scheduler/cluster/YarnClusterScheduler.scala
@@ -0,0 +1,55 @@
+/*
+ * 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.spark._
+import org.apache.spark.deploy.yarn.{ApplicationMaster, YarnAllocationHandler}
+import org.apache.spark.util.Utils
+import org.apache.hadoop.conf.Configuration
+
+/**
+ *
+ * This is a simple extension to ClusterScheduler - to ensure that appropriate initialization of ApplicationMaster, etc is done
+ */
+private[spark] class YarnClusterScheduler(sc: SparkContext, conf: Configuration) extends ClusterScheduler(sc) {
+
+ logInfo("Created YarnClusterScheduler")
+
+ def this(sc: SparkContext) = this(sc, new Configuration())
+
+ // Nothing else for now ... initialize application master : which needs sparkContext to determine how to allocate
+ // Note that only the first creation of SparkContext influences (and ideally, there must be only one SparkContext, right ?)
+ // Subsequent creations are ignored - since nodes are already allocated by then.
+
+
+ // By default, rack is unknown
+ override def getRackForHost(hostPort: String): Option[String] = {
+ val host = Utils.parseHostPort(hostPort)._1
+ val retval = YarnAllocationHandler.lookupRack(conf, host)
+ if (retval != null) Some(retval) else None
+ }
+
+ override def postStartHook() {
+ val sparkContextInitialized = ApplicationMaster.sparkContextInitialized(sc)
+ if (sparkContextInitialized){
+ // Wait for a few seconds for the slaves to bootstrap and register with master - best case attempt
+ Thread.sleep(3000L)
+ }
+ logInfo("YarnClusterScheduler.postStartHook done")
+ }
+}
diff --git a/new-yarn/src/test/scala/org/apache/spark/deploy/yarn/ClientDistributedCacheManagerSuite.scala b/new-yarn/src/test/scala/org/apache/spark/deploy/yarn/ClientDistributedCacheManagerSuite.scala
new file mode 100644
index 0000000000000..2941356bc55f9
--- /dev/null
+++ b/new-yarn/src/test/scala/org/apache/spark/deploy/yarn/ClientDistributedCacheManagerSuite.scala
@@ -0,0 +1,220 @@
+/*
+ * 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.URI
+
+import org.scalatest.FunSuite
+import org.scalatest.mock.MockitoSugar
+import org.mockito.Mockito.when
+
+import org.apache.hadoop.conf.Configuration
+import org.apache.hadoop.fs.FileStatus
+import org.apache.hadoop.fs.FileSystem
+import org.apache.hadoop.fs.Path
+import org.apache.hadoop.fs.permission.FsAction
+import org.apache.hadoop.yarn.api.records.LocalResource
+import org.apache.hadoop.yarn.api.records.LocalResourceVisibility
+import org.apache.hadoop.yarn.api.records.LocalResourceType
+import org.apache.hadoop.yarn.util.{Records, ConverterUtils}
+
+import scala.collection.mutable.HashMap
+import scala.collection.mutable.Map
+
+
+class ClientDistributedCacheManagerSuite extends FunSuite with MockitoSugar {
+
+ class MockClientDistributedCacheManager extends ClientDistributedCacheManager {
+ override def getVisibility(conf: Configuration, uri: URI, statCache: Map[URI, FileStatus]):
+ LocalResourceVisibility = {
+ return LocalResourceVisibility.PRIVATE
+ }
+ }
+
+ test("test getFileStatus empty") {
+ val distMgr = new ClientDistributedCacheManager()
+ val fs = mock[FileSystem]
+ val uri = new URI("/tmp/testing")
+ when(fs.getFileStatus(new Path(uri))).thenReturn(new FileStatus())
+ val statCache: Map[URI, FileStatus] = HashMap[URI, FileStatus]()
+ val stat = distMgr.getFileStatus(fs, uri, statCache)
+ assert(stat.getPath() === null)
+ }
+
+ test("test getFileStatus cached") {
+ val distMgr = new ClientDistributedCacheManager()
+ val fs = mock[FileSystem]
+ val uri = new URI("/tmp/testing")
+ val realFileStatus = new FileStatus(10, false, 1, 1024, 10, 10, null, "testOwner",
+ null, new Path("/tmp/testing"))
+ when(fs.getFileStatus(new Path(uri))).thenReturn(new FileStatus())
+ val statCache: Map[URI, FileStatus] = HashMap[URI, FileStatus](uri -> realFileStatus)
+ val stat = distMgr.getFileStatus(fs, uri, statCache)
+ assert(stat.getPath().toString() === "/tmp/testing")
+ }
+
+ test("test addResource") {
+ val distMgr = new MockClientDistributedCacheManager()
+ val fs = mock[FileSystem]
+ val conf = new Configuration()
+ val destPath = new Path("file:///foo.invalid.com:8080/tmp/testing")
+ val localResources = HashMap[String, LocalResource]()
+ val statCache: Map[URI, FileStatus] = HashMap[URI, FileStatus]()
+ when(fs.getFileStatus(destPath)).thenReturn(new FileStatus())
+
+ distMgr.addResource(fs, conf, destPath, localResources, LocalResourceType.FILE, "link",
+ statCache, false)
+ val resource = localResources("link")
+ assert(resource.getVisibility() === LocalResourceVisibility.PRIVATE)
+ assert(ConverterUtils.getPathFromYarnURL(resource.getResource()) === destPath)
+ assert(resource.getTimestamp() === 0)
+ assert(resource.getSize() === 0)
+ assert(resource.getType() === LocalResourceType.FILE)
+
+ val env = new HashMap[String, String]()
+ distMgr.setDistFilesEnv(env)
+ assert(env("SPARK_YARN_CACHE_FILES") === "file:/foo.invalid.com:8080/tmp/testing#link")
+ assert(env("SPARK_YARN_CACHE_FILES_TIME_STAMPS") === "0")
+ assert(env("SPARK_YARN_CACHE_FILES_FILE_SIZES") === "0")
+ assert(env("SPARK_YARN_CACHE_FILES_VISIBILITIES") === LocalResourceVisibility.PRIVATE.name())
+
+ distMgr.setDistArchivesEnv(env)
+ assert(env.get("SPARK_YARN_CACHE_ARCHIVES") === None)
+ assert(env.get("SPARK_YARN_CACHE_ARCHIVES_TIME_STAMPS") === None)
+ assert(env.get("SPARK_YARN_CACHE_ARCHIVES_FILE_SIZES") === None)
+ assert(env.get("SPARK_YARN_CACHE_ARCHIVES_VISIBILITIES") === None)
+
+ //add another one and verify both there and order correct
+ val realFileStatus = new FileStatus(20, false, 1, 1024, 10, 30, null, "testOwner",
+ null, new Path("/tmp/testing2"))
+ val destPath2 = new Path("file:///foo.invalid.com:8080/tmp/testing2")
+ when(fs.getFileStatus(destPath2)).thenReturn(realFileStatus)
+ distMgr.addResource(fs, conf, destPath2, localResources, LocalResourceType.FILE, "link2",
+ statCache, false)
+ val resource2 = localResources("link2")
+ assert(resource2.getVisibility() === LocalResourceVisibility.PRIVATE)
+ assert(ConverterUtils.getPathFromYarnURL(resource2.getResource()) === destPath2)
+ assert(resource2.getTimestamp() === 10)
+ assert(resource2.getSize() === 20)
+ assert(resource2.getType() === LocalResourceType.FILE)
+
+ val env2 = new HashMap[String, String]()
+ distMgr.setDistFilesEnv(env2)
+ val timestamps = env2("SPARK_YARN_CACHE_FILES_TIME_STAMPS").split(',')
+ val files = env2("SPARK_YARN_CACHE_FILES").split(',')
+ val sizes = env2("SPARK_YARN_CACHE_FILES_FILE_SIZES").split(',')
+ val visibilities = env2("SPARK_YARN_CACHE_FILES_VISIBILITIES") .split(',')
+ assert(files(0) === "file:/foo.invalid.com:8080/tmp/testing#link")
+ assert(timestamps(0) === "0")
+ assert(sizes(0) === "0")
+ assert(visibilities(0) === LocalResourceVisibility.PRIVATE.name())
+
+ assert(files(1) === "file:/foo.invalid.com:8080/tmp/testing2#link2")
+ assert(timestamps(1) === "10")
+ assert(sizes(1) === "20")
+ assert(visibilities(1) === LocalResourceVisibility.PRIVATE.name())
+ }
+
+ test("test addResource link null") {
+ val distMgr = new MockClientDistributedCacheManager()
+ val fs = mock[FileSystem]
+ val conf = new Configuration()
+ val destPath = new Path("file:///foo.invalid.com:8080/tmp/testing")
+ val localResources = HashMap[String, LocalResource]()
+ val statCache: Map[URI, FileStatus] = HashMap[URI, FileStatus]()
+ when(fs.getFileStatus(destPath)).thenReturn(new FileStatus())
+
+ intercept[Exception] {
+ distMgr.addResource(fs, conf, destPath, localResources, LocalResourceType.FILE, null,
+ statCache, false)
+ }
+ assert(localResources.get("link") === None)
+ assert(localResources.size === 0)
+ }
+
+ test("test addResource appmaster only") {
+ val distMgr = new MockClientDistributedCacheManager()
+ val fs = mock[FileSystem]
+ val conf = new Configuration()
+ val destPath = new Path("file:///foo.invalid.com:8080/tmp/testing")
+ val localResources = HashMap[String, LocalResource]()
+ val statCache: Map[URI, FileStatus] = HashMap[URI, FileStatus]()
+ val realFileStatus = new FileStatus(20, false, 1, 1024, 10, 30, null, "testOwner",
+ null, new Path("/tmp/testing"))
+ when(fs.getFileStatus(destPath)).thenReturn(realFileStatus)
+
+ distMgr.addResource(fs, conf, destPath, localResources, LocalResourceType.ARCHIVE, "link",
+ statCache, true)
+ val resource = localResources("link")
+ assert(resource.getVisibility() === LocalResourceVisibility.PRIVATE)
+ assert(ConverterUtils.getPathFromYarnURL(resource.getResource()) === destPath)
+ assert(resource.getTimestamp() === 10)
+ assert(resource.getSize() === 20)
+ assert(resource.getType() === LocalResourceType.ARCHIVE)
+
+ val env = new HashMap[String, String]()
+ distMgr.setDistFilesEnv(env)
+ assert(env.get("SPARK_YARN_CACHE_FILES") === None)
+ assert(env.get("SPARK_YARN_CACHE_FILES_TIME_STAMPS") === None)
+ assert(env.get("SPARK_YARN_CACHE_FILES_FILE_SIZES") === None)
+ assert(env.get("SPARK_YARN_CACHE_FILES_VISIBILITIES") === None)
+
+ distMgr.setDistArchivesEnv(env)
+ assert(env.get("SPARK_YARN_CACHE_ARCHIVES") === None)
+ assert(env.get("SPARK_YARN_CACHE_ARCHIVES_TIME_STAMPS") === None)
+ assert(env.get("SPARK_YARN_CACHE_ARCHIVES_FILE_SIZES") === None)
+ assert(env.get("SPARK_YARN_CACHE_ARCHIVES_VISIBILITIES") === None)
+ }
+
+ test("test addResource archive") {
+ val distMgr = new MockClientDistributedCacheManager()
+ val fs = mock[FileSystem]
+ val conf = new Configuration()
+ val destPath = new Path("file:///foo.invalid.com:8080/tmp/testing")
+ val localResources = HashMap[String, LocalResource]()
+ val statCache: Map[URI, FileStatus] = HashMap[URI, FileStatus]()
+ val realFileStatus = new FileStatus(20, false, 1, 1024, 10, 30, null, "testOwner",
+ null, new Path("/tmp/testing"))
+ when(fs.getFileStatus(destPath)).thenReturn(realFileStatus)
+
+ distMgr.addResource(fs, conf, destPath, localResources, LocalResourceType.ARCHIVE, "link",
+ statCache, false)
+ val resource = localResources("link")
+ assert(resource.getVisibility() === LocalResourceVisibility.PRIVATE)
+ assert(ConverterUtils.getPathFromYarnURL(resource.getResource()) === destPath)
+ assert(resource.getTimestamp() === 10)
+ assert(resource.getSize() === 20)
+ assert(resource.getType() === LocalResourceType.ARCHIVE)
+
+ val env = new HashMap[String, String]()
+
+ distMgr.setDistArchivesEnv(env)
+ assert(env("SPARK_YARN_CACHE_ARCHIVES") === "file:/foo.invalid.com:8080/tmp/testing#link")
+ assert(env("SPARK_YARN_CACHE_ARCHIVES_TIME_STAMPS") === "10")
+ assert(env("SPARK_YARN_CACHE_ARCHIVES_FILE_SIZES") === "20")
+ assert(env("SPARK_YARN_CACHE_ARCHIVES_VISIBILITIES") === LocalResourceVisibility.PRIVATE.name())
+
+ distMgr.setDistFilesEnv(env)
+ assert(env.get("SPARK_YARN_CACHE_FILES") === None)
+ assert(env.get("SPARK_YARN_CACHE_FILES_TIME_STAMPS") === None)
+ assert(env.get("SPARK_YARN_CACHE_FILES_FILE_SIZES") === None)
+ assert(env.get("SPARK_YARN_CACHE_FILES_VISIBILITIES") === None)
+ }
+
+
+}
diff --git a/pom.xml b/pom.xml
index 0d89a357483c5..f85a3b5fd3e47 100644
--- a/pom.xml
+++ b/pom.xml
@@ -102,8 +102,10 @@
1.5
2.9.3
0.13.0
+ com.typesafe.akka
2.0.5
1.7.5
+ 2.4.1
1.2.17
4.1.2
cdh${cdh.version}
@@ -204,6 +206,145 @@
commons-io
2.4
+
+ org.apache.spark
+ spark-core_2.9.3
+ ${project.version}
+
+
+
+ com.typesafe.akka
+ akka-actor
+
+
+ com.typesafe.akka
+ akka-remote
+
+
+ com.typesafe.akka
+ akka-slf4j
+
+
+ com.typesafe.akka
+ akka-zeromq
+
+
+
+
+ org.apache.spark
+ spark-bagel_2.9.3
+ ${project.version}
+
+
+ com.typesafe.akka
+ akka-actor
+
+
+ com.typesafe.akka
+ akka-remote
+
+
+ com.typesafe.akka
+ akka-slf4j
+
+
+ com.typesafe.akka
+ akka-zeromq
+
+
+
+
+ org.apache.spark
+ spark-mllib_2.9.3
+ ${project.version}
+
+
+ com.typesafe.akka
+ akka-actor
+
+
+ com.typesafe.akka
+ akka-remote
+
+
+ com.typesafe.akka
+ akka-slf4j
+
+
+ com.typesafe.akka
+ akka-zeromq
+
+
+
+
+ org.apache.spark
+ spark-repl_2.9.3
+ ${project.version}
+
+
+ com.typesafe.akka
+ akka-actor
+
+
+ com.typesafe.akka
+ akka-remote
+
+
+ com.typesafe.akka
+ akka-slf4j
+
+
+ com.typesafe.akka
+ akka-zeromq
+
+
+
+
+ org.apache.spark
+ spark-streaming_2.9.3
+ ${project.version}
+
+
+ com.typesafe.akka
+ akka-actor
+
+
+ com.typesafe.akka
+ akka-remote
+
+
+ com.typesafe.akka
+ akka-slf4j
+
+
+ com.typesafe.akka
+ akka-zeromq
+
+
+
+
+ org.apache.spark
+ spark-yarn_2.9.3
+ ${project.version}
+
+
+ com.typesafe.akka
+ akka-actor
+
+
+ com.typesafe.akka
+ akka-remote
+
+
+ com.typesafe.akka
+ akka-slf4j
+
+
+ com.typesafe.akka
+ akka-zeromq
+
+
+
org.eclipse.jetty
jetty-server
@@ -257,7 +398,7 @@
com.google.protobuf
protobuf-java
- 2.4.1
+ ${protobuf.version}
com.twitter
@@ -270,20 +411,31 @@
0.3.1
- com.typesafe.akka
+ ${akka.group}
akka-actor
${akka.version}
- com.typesafe.akka
+ ${akka.group}
akka-remote
${akka.version}
- com.typesafe.akka
+ ${akka.group}
akka-slf4j
${akka.version}
+
+ ${akka.group}
+ akka-zeromq
+ ${akka.version}
+
+
+ org.jboss.netty
+ netty
+
+
+
it.unimi.dsi
fastutil
@@ -294,11 +446,6 @@
colt
1.2.0
-
- com.github.scala-incubator.io
- scala-io-file_2.9.2
- 0.4.1
-
org.apache.mesos
mesos
@@ -791,6 +938,41 @@
+
+