SPARK-1319: Fix scheduler to account for tasks using > 1 CPUs.

Move CPUS_PER_TASK to TaskSchedulerImpl as the value is a constant and use it in both Mesos and CoarseGrained scheduler backends.

Thanks @kayousterhout for the design discussion

Author: Shivaram Venkataraman <shivaram@eecs.berkeley.edu>

Closes apache#219 from shivaram/multi-cpus and squashes the following commits:

5c7d685 [Shivaram Venkataraman] Don't pass availableCpus to TaskSetManager
260e4d5 [Shivaram Venkataraman] Add a check for non-zero CPUs in TaskSetManager
73fcf6f [Shivaram Venkataraman] Add documentation for spark.task.cpus
647bc45 [Shivaram Venkataraman] Fix scheduler to account for tasks using > 1 CPUs. Move CPUS_PER_TASK to TaskSchedulerImpl as the value is a constant and use it in both Mesos and CoarseGrained scheduler backends.

Author: shivaram

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

@@ -62,6 +62,9 @@ private[spark] class TaskSchedulerImpl(
   // Threshold above which we warn user initial TaskSet may be starved
   val STARVATION_TIMEOUT = conf.getLong("spark.starvation.timeout", 15000)
 
+  // CPUs to request per task
+  val CPUS_PER_TASK = conf.getInt("spark.task.cpus", 1)
+
   // TaskSetManagers are not thread safe, so any access to one should be synchronized
   // on this class.
   val activeTaskSets = new HashMap[String, TaskSetManager]
@@ -228,16 +231,18 @@ private[spark] class TaskSchedulerImpl(
         for (i <- 0 until shuffledOffers.size) {
           val execId = shuffledOffers(i).executorId
           val host = shuffledOffers(i).host
-          for (task <- taskSet.resourceOffer(execId, host, availableCpus(i), maxLocality)) {
-            tasks(i) += task
-            val tid = task.taskId
-            taskIdToTaskSetId(tid) = taskSet.taskSet.id
-            taskIdToExecutorId(tid) = execId
-            activeExecutorIds += execId
-            executorsByHost(host) += execId
-            availableCpus(i) -= taskSet.CPUS_PER_TASK
-            assert (availableCpus(i) >= 0)
-            launchedTask = true
+          if (availableCpus(i) >= CPUS_PER_TASK) {
+            for (task <- taskSet.resourceOffer(execId, host, maxLocality)) {
+              tasks(i) += task
+              val tid = task.taskId
+              taskIdToTaskSetId(tid) = taskSet.taskSet.id
+              taskIdToExecutorId(tid) = execId
+              activeExecutorIds += execId
+              executorsByHost(host) += execId
+              availableCpus(i) -= CPUS_PER_TASK
+              assert (availableCpus(i) >= 0)
+              launchedTask = true
+            }
           }
         }
       } while (launchedTask)

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

@@ -56,9 +56,6 @@ private[spark] class TaskSetManager(
 {
   val conf = sched.sc.conf
 
-  // CPUs to request per task
-  val CPUS_PER_TASK = conf.getInt("spark.task.cpus", 1)
-
   /*
    * Sometimes if an executor is dead or in an otherwise invalid state, the driver
    * does not realize right away leading to repeated task failures. If enabled,
@@ -384,11 +381,10 @@ private[spark] class TaskSetManager(
   def resourceOffer(
       execId: String,
       host: String,
-      availableCpus: Int,
       maxLocality: TaskLocality.TaskLocality)
     : Option[TaskDescription] =
   {
-    if (!isZombie && availableCpus >= CPUS_PER_TASK) {
+    if (!isZombie) {
       val curTime = clock.getTime()
 
       var allowedLocality = getAllowedLocalityLevel(curTime)

core/src/main/scala/org/apache/spark/scheduler/cluster/CoarseGrainedSchedulerBackend.scala

@@ -89,7 +89,7 @@ class CoarseGrainedSchedulerBackend(scheduler: TaskSchedulerImpl, actorSystem: A
         scheduler.statusUpdate(taskId, state, data.value)
         if (TaskState.isFinished(state)) {
           if (executorActor.contains(executorId)) {
-            freeCores(executorId) += 1
+            freeCores(executorId) += scheduler.CPUS_PER_TASK
             makeOffers(executorId)
           } else {
             // Ignoring the update since we don't know about the executor.
@@ -140,7 +140,7 @@ class CoarseGrainedSchedulerBackend(scheduler: TaskSchedulerImpl, actorSystem: A
     // Launch tasks returned by a set of resource offers
     def launchTasks(tasks: Seq[Seq[TaskDescription]]) {
       for (task <- tasks.flatten) {
-        freeCores(task.executorId) -= 1
+        freeCores(task.executorId) -= scheduler.CPUS_PER_TASK
         executorActor(task.executorId) ! LaunchTask(task)
       }
     }

core/src/main/scala/org/apache/spark/scheduler/cluster/mesos/MesosSchedulerBackend.scala

@@ -246,7 +246,7 @@ private[spark] class MesosSchedulerBackend(
     val cpuResource = Resource.newBuilder()
       .setName("cpus")
       .setType(Value.Type.SCALAR)
-      .setScalar(Value.Scalar.newBuilder().setValue(1).build())
+      .setScalar(Value.Scalar.newBuilder().setValue(scheduler.CPUS_PER_TASK).build())
       .build()
     MesosTaskInfo.newBuilder()
       .setTaskId(taskId)

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

@@ -80,7 +80,6 @@ class FakeTaskSetManager(
   override def resourceOffer(
       execId: String,
       host: String,
-      availableCpus: Int,
       maxLocality: TaskLocality.TaskLocality)
     : Option[TaskDescription] =
   {
@@ -125,7 +124,7 @@ class TaskSchedulerImplSuite extends FunSuite with LocalSparkContext with Loggin
          manager.parent.name, manager.parent.runningTasks, manager.name, manager.runningTasks))
     }
     for (taskSet <- taskSetQueue) {
-      taskSet.resourceOffer("execId_1", "hostname_1", 1, TaskLocality.ANY) match {
+      taskSet.resourceOffer("execId_1", "hostname_1", TaskLocality.ANY) match {
         case Some(task) =>
           return taskSet.stageId
         case None => {}
@@ -293,4 +292,43 @@ class TaskSchedulerImplSuite extends FunSuite with LocalSparkContext with Loggin
     assert(count > 0)
     assert(count < numTrials)
   }
+
+  test("Scheduler correctly accounts for multiple CPUs per task") {
+    sc = new SparkContext("local", "TaskSchedulerImplSuite")
+    val taskCpus = 2
+
+    sc.conf.set("spark.task.cpus", taskCpus.toString)
+    val taskScheduler = new TaskSchedulerImpl(sc)
+    taskScheduler.initialize(new FakeSchedulerBackend)
+    // Need to initialize a DAGScheduler for the taskScheduler to use for callbacks.
+    val dagScheduler = new DAGScheduler(sc, taskScheduler) {
+      override def taskStarted(task: Task[_], taskInfo: TaskInfo) {}
+      override def executorAdded(execId: String, host: String) {}
+    }
+
+    // Give zero core offers. Should not generate any tasks
+    val zeroCoreWorkerOffers = Seq(new WorkerOffer("executor0", "host0", 0),
+      new WorkerOffer("executor1", "host1", 0))
+    val taskSet = FakeTask.createTaskSet(1)
+    taskScheduler.submitTasks(taskSet)
+    var taskDescriptions = taskScheduler.resourceOffers(zeroCoreWorkerOffers).flatten
+    assert(0 === taskDescriptions.length)
+
+    // No tasks should run as we only have 1 core free.
+    val numFreeCores = 1
+    val singleCoreWorkerOffers = Seq(new WorkerOffer("executor0", "host0", numFreeCores),
+      new WorkerOffer("executor1", "host1", numFreeCores))
+    taskScheduler.submitTasks(taskSet)
+    taskDescriptions = taskScheduler.resourceOffers(singleCoreWorkerOffers).flatten
+    assert(0 === taskDescriptions.length)
+
+    // Now change the offers to have 2 cores in one executor and verify if it
+    // is chosen.
+    val multiCoreWorkerOffers = Seq(new WorkerOffer("executor0", "host0", taskCpus),
+      new WorkerOffer("executor1", "host1", numFreeCores))
+    taskScheduler.submitTasks(taskSet)
+    taskDescriptions = taskScheduler.resourceOffers(multiCoreWorkerOffers).flatten
+    assert(1 === taskDescriptions.length)
+    assert("executor0" === taskDescriptions(0).executorId)
+  }
 }

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

@@ -93,19 +93,16 @@ class TaskSetManagerSuite extends FunSuite with LocalSparkContext with Logging {
     val taskSet = FakeTask.createTaskSet(1)
     val manager = new TaskSetManager(sched, taskSet, MAX_TASK_FAILURES)
 
-    // Offer a host with no CPUs
-    assert(manager.resourceOffer("exec1", "host1", 0, ANY) === None)
-
     // Offer a host with process-local as the constraint; this should work because the TaskSet
     // above won't have any locality preferences
-    val taskOption = manager.resourceOffer("exec1", "host1", 2, TaskLocality.PROCESS_LOCAL)
+    val taskOption = manager.resourceOffer("exec1", "host1", TaskLocality.PROCESS_LOCAL)
     assert(taskOption.isDefined)
     val task = taskOption.get
     assert(task.executorId === "exec1")
     assert(sched.startedTasks.contains(0))
 
     // Re-offer the host -- now we should get no more tasks
-    assert(manager.resourceOffer("exec1", "host1", 2, PROCESS_LOCAL) === None)
+    assert(manager.resourceOffer("exec1", "host1", PROCESS_LOCAL) === None)
 
     // Tell it the task has finished
     manager.handleSuccessfulTask(0, createTaskResult(0))
@@ -121,15 +118,15 @@ class TaskSetManagerSuite extends FunSuite with LocalSparkContext with Logging {
 
     // First three offers should all find tasks
     for (i <- 0 until 3) {
-      val taskOption = manager.resourceOffer("exec1", "host1", 1, PROCESS_LOCAL)
+      val taskOption = manager.resourceOffer("exec1", "host1", PROCESS_LOCAL)
       assert(taskOption.isDefined)
       val task = taskOption.get
       assert(task.executorId === "exec1")
     }
     assert(sched.startedTasks.toSet === Set(0, 1, 2))
 
     // Re-offer the host -- now we should get no more tasks
-    assert(manager.resourceOffer("exec1", "host1", 1, PROCESS_LOCAL) === None)
+    assert(manager.resourceOffer("exec1", "host1", PROCESS_LOCAL) === None)
 
     // Finish the first two tasks
     manager.handleSuccessfulTask(0, createTaskResult(0))
@@ -157,35 +154,35 @@ class TaskSetManagerSuite extends FunSuite with LocalSparkContext with Logging {
     val manager = new TaskSetManager(sched, taskSet, MAX_TASK_FAILURES, clock)
 
     // First offer host1, exec1: first task should be chosen
-    assert(manager.resourceOffer("exec1", "host1", 1, ANY).get.index === 0)
+    assert(manager.resourceOffer("exec1", "host1", ANY).get.index === 0)
 
     // Offer host1, exec1 again: the last task, which has no prefs, should be chosen
-    assert(manager.resourceOffer("exec1", "host1", 1, ANY).get.index === 3)
+    assert(manager.resourceOffer("exec1", "host1", ANY).get.index === 3)
 
     // Offer host1, exec1 again, at PROCESS_LOCAL level: nothing should get chosen
-    assert(manager.resourceOffer("exec1", "host1", 1, PROCESS_LOCAL) === None)
+    assert(manager.resourceOffer("exec1", "host1", PROCESS_LOCAL) === None)
 
     clock.advance(LOCALITY_WAIT)
 
     // Offer host1, exec1 again, at PROCESS_LOCAL level: nothing should get chosen
-    assert(manager.resourceOffer("exec1", "host1", 1, PROCESS_LOCAL) === None)
+    assert(manager.resourceOffer("exec1", "host1", PROCESS_LOCAL) === None)
 
     // Offer host1, exec1 again, at NODE_LOCAL level: we should choose task 2
-    assert(manager.resourceOffer("exec1", "host1", 1, NODE_LOCAL).get.index == 2)
+    assert(manager.resourceOffer("exec1", "host1", NODE_LOCAL).get.index == 2)
 
     // Offer host1, exec1 again, at NODE_LOCAL level: nothing should get chosen
-    assert(manager.resourceOffer("exec1", "host1", 1, NODE_LOCAL) === None)
+    assert(manager.resourceOffer("exec1", "host1", NODE_LOCAL) === None)
 
     // Offer host1, exec1 again, at ANY level: nothing should get chosen
-    assert(manager.resourceOffer("exec1", "host1", 1, ANY) === None)
+    assert(manager.resourceOffer("exec1", "host1", ANY) === None)
 
     clock.advance(LOCALITY_WAIT)
 
     // Offer host1, exec1 again, at ANY level: task 1 should get chosen
-    assert(manager.resourceOffer("exec1", "host1", 1, ANY).get.index === 1)
+    assert(manager.resourceOffer("exec1", "host1", ANY).get.index === 1)
 
     // Offer host1, exec1 again, at ANY level: nothing should be chosen as we've launched all tasks
-    assert(manager.resourceOffer("exec1", "host1", 1, ANY) === None)
+    assert(manager.resourceOffer("exec1", "host1", ANY) === None)
   }
 
   test("delay scheduling with fallback") {
@@ -203,29 +200,29 @@ class TaskSetManagerSuite extends FunSuite with LocalSparkContext with Logging {
     val manager = new TaskSetManager(sched, taskSet, MAX_TASK_FAILURES, clock)
 
     // First offer host1: first task should be chosen
-    assert(manager.resourceOffer("exec1", "host1", 1, ANY).get.index === 0)
+    assert(manager.resourceOffer("exec1", "host1", ANY).get.index === 0)
 
     // Offer host1 again: nothing should get chosen
-    assert(manager.resourceOffer("exec1", "host1", 1, ANY) === None)
+    assert(manager.resourceOffer("exec1", "host1", ANY) === None)
 
     clock.advance(LOCALITY_WAIT)
 
     // Offer host1 again: second task (on host2) should get chosen
-    assert(manager.resourceOffer("exec1", "host1", 1, ANY).get.index === 1)
+    assert(manager.resourceOffer("exec1", "host1", ANY).get.index === 1)
 
     // Offer host1 again: third task (on host2) should get chosen
-    assert(manager.resourceOffer("exec1", "host1", 1, ANY).get.index === 2)
+    assert(manager.resourceOffer("exec1", "host1", ANY).get.index === 2)
 
     // Offer host2: fifth task (also on host2) should get chosen
-    assert(manager.resourceOffer("exec2", "host2", 1, ANY).get.index === 4)
+    assert(manager.resourceOffer("exec2", "host2", ANY).get.index === 4)
 
     // Now that we've launched a local task, we should no longer launch the task for host3
-    assert(manager.resourceOffer("exec2", "host2", 1, ANY) === None)
+    assert(manager.resourceOffer("exec2", "host2", ANY) === None)
 
     clock.advance(LOCALITY_WAIT)
 
     // After another delay, we can go ahead and launch that task non-locally
-    assert(manager.resourceOffer("exec2", "host2", 1, ANY).get.index === 3)
+    assert(manager.resourceOffer("exec2", "host2", ANY).get.index === 3)
   }
 
   test("delay scheduling with failed hosts") {
@@ -240,24 +237,24 @@ class TaskSetManagerSuite extends FunSuite with LocalSparkContext with Logging {
     val manager = new TaskSetManager(sched, taskSet, MAX_TASK_FAILURES, clock)
 
     // First offer host1: first task should be chosen
-    assert(manager.resourceOffer("exec1", "host1", 1, ANY).get.index === 0)
+    assert(manager.resourceOffer("exec1", "host1", ANY).get.index === 0)
 
     // Offer host1 again: third task should be chosen immediately because host3 is not up
-    assert(manager.resourceOffer("exec1", "host1", 1, ANY).get.index === 2)
+    assert(manager.resourceOffer("exec1", "host1", ANY).get.index === 2)
 
     // After this, nothing should get chosen
-    assert(manager.resourceOffer("exec1", "host1", 1, ANY) === None)
+    assert(manager.resourceOffer("exec1", "host1", ANY) === None)
 
     // Now mark host2 as dead
     sched.removeExecutor("exec2")
     manager.executorLost("exec2", "host2")
 
     // Task 1 should immediately be launched on host1 because its original host is gone
-    assert(manager.resourceOffer("exec1", "host1", 1, ANY).get.index === 1)
+    assert(manager.resourceOffer("exec1", "host1", ANY).get.index === 1)
 
     // Now that all tasks have launched, nothing new should be launched anywhere else
-    assert(manager.resourceOffer("exec1", "host1", 1, ANY) === None)
-    assert(manager.resourceOffer("exec2", "host2", 1, ANY) === None)
+    assert(manager.resourceOffer("exec1", "host1", ANY) === None)
+    assert(manager.resourceOffer("exec2", "host2", ANY) === None)
   }
 
   test("task result lost") {
@@ -267,14 +264,14 @@ class TaskSetManagerSuite extends FunSuite with LocalSparkContext with Logging {
     val clock = new FakeClock
     val manager = new TaskSetManager(sched, taskSet, MAX_TASK_FAILURES, clock)
 
-    assert(manager.resourceOffer("exec1", "host1", 1, ANY).get.index === 0)
+    assert(manager.resourceOffer("exec1", "host1", ANY).get.index === 0)
 
     // Tell it the task has finished but the result was lost.
     manager.handleFailedTask(0, TaskState.FINISHED, TaskResultLost)
     assert(sched.endedTasks(0) === TaskResultLost)
 
     // Re-offer the host -- now we should get task 0 again.
-    assert(manager.resourceOffer("exec1", "host1", 1, ANY).get.index === 0)
+    assert(manager.resourceOffer("exec1", "host1", ANY).get.index === 0)
   }
 
   test("repeated failures lead to task set abortion") {
@@ -287,7 +284,7 @@ class TaskSetManagerSuite extends FunSuite with LocalSparkContext with Logging {
     // Fail the task MAX_TASK_FAILURES times, and check that the task set is aborted
     // after the last failure.
     (1 to manager.maxTaskFailures).foreach { index =>
-      val offerResult = manager.resourceOffer("exec1", "host1", 1, ANY)
+      val offerResult = manager.resourceOffer("exec1", "host1", ANY)
       assert(offerResult.isDefined,
         "Expect resource offer on iteration %s to return a task".format(index))
       assert(offerResult.get.index === 0)
@@ -317,7 +314,7 @@ class TaskSetManagerSuite extends FunSuite with LocalSparkContext with Logging {
     val manager = new TaskSetManager(sched, taskSet, 4, clock)
 
     {
-      val offerResult = manager.resourceOffer("exec1", "host1", 1, TaskLocality.PROCESS_LOCAL)
+      val offerResult = manager.resourceOffer("exec1", "host1", TaskLocality.PROCESS_LOCAL)
       assert(offerResult.isDefined, "Expect resource offer to return a task")
 
       assert(offerResult.get.index === 0)
@@ -328,15 +325,15 @@ class TaskSetManagerSuite extends FunSuite with LocalSparkContext with Logging {
       assert(!sched.taskSetsFailed.contains(taskSet.id))
 
       // Ensure scheduling on exec1 fails after failure 1 due to blacklist
-      assert(manager.resourceOffer("exec1", "host1", 1, TaskLocality.PROCESS_LOCAL).isEmpty)
-      assert(manager.resourceOffer("exec1", "host1", 1, TaskLocality.NODE_LOCAL).isEmpty)
-      assert(manager.resourceOffer("exec1", "host1", 1, TaskLocality.RACK_LOCAL).isEmpty)
-      assert(manager.resourceOffer("exec1", "host1", 1, TaskLocality.ANY).isEmpty)
+      assert(manager.resourceOffer("exec1", "host1", TaskLocality.PROCESS_LOCAL).isEmpty)
+      assert(manager.resourceOffer("exec1", "host1", TaskLocality.NODE_LOCAL).isEmpty)
+      assert(manager.resourceOffer("exec1", "host1", TaskLocality.RACK_LOCAL).isEmpty)
+      assert(manager.resourceOffer("exec1", "host1", TaskLocality.ANY).isEmpty)
     }
 
     // Run the task on exec1.1 - should work, and then fail it on exec1.1
     {
-      val offerResult = manager.resourceOffer("exec1.1", "host1", 1, TaskLocality.NODE_LOCAL)
+      val offerResult = manager.resourceOffer("exec1.1", "host1", TaskLocality.NODE_LOCAL)
       assert(offerResult.isDefined,
         "Expect resource offer to return a task for exec1.1, offerResult = " + offerResult)
 
@@ -348,12 +345,12 @@ class TaskSetManagerSuite extends FunSuite with LocalSparkContext with Logging {
       assert(!sched.taskSetsFailed.contains(taskSet.id))
 
       // Ensure scheduling on exec1.1 fails after failure 2 due to blacklist
-      assert(manager.resourceOffer("exec1.1", "host1", 1, TaskLocality.NODE_LOCAL).isEmpty)
+      assert(manager.resourceOffer("exec1.1", "host1", TaskLocality.NODE_LOCAL).isEmpty)
     }
 
     // Run the task on exec2 - should work, and then fail it on exec2
     {
-      val offerResult = manager.resourceOffer("exec2", "host2", 1, TaskLocality.ANY)
+      val offerResult = manager.resourceOffer("exec2", "host2", TaskLocality.ANY)
       assert(offerResult.isDefined, "Expect resource offer to return a task")
 
       assert(offerResult.get.index === 0)
@@ -364,20 +361,20 @@ class TaskSetManagerSuite extends FunSuite with LocalSparkContext with Logging {
       assert(!sched.taskSetsFailed.contains(taskSet.id))
 
       // Ensure scheduling on exec2 fails after failure 3 due to blacklist
-      assert(manager.resourceOffer("exec2", "host2", 1, TaskLocality.ANY).isEmpty)
+      assert(manager.resourceOffer("exec2", "host2", TaskLocality.ANY).isEmpty)
     }
 
     // After reschedule delay, scheduling on exec1 should be possible.
     clock.advance(rescheduleDelay)
 
     {
-      val offerResult = manager.resourceOffer("exec1", "host1", 1, TaskLocality.PROCESS_LOCAL)
+      val offerResult = manager.resourceOffer("exec1", "host1", TaskLocality.PROCESS_LOCAL)
       assert(offerResult.isDefined, "Expect resource offer to return a task")
 
       assert(offerResult.get.index === 0)
       assert(offerResult.get.executorId === "exec1")
 
-      assert(manager.resourceOffer("exec1", "host1", 1, TaskLocality.PROCESS_LOCAL).isEmpty)
+      assert(manager.resourceOffer("exec1", "host1", TaskLocality.PROCESS_LOCAL).isEmpty)
 
       // Cause exec1 to fail : failure 4
       manager.handleFailedTask(offerResult.get.taskId, TaskState.FINISHED, TaskResultLost)