[SPARK-11078] Ensure spilling tests actually spill

Author: Andrew Or <[email protected]>

Closes apache#9124 from andrewor14/spilling-tests.

Author: Andrew Or

core/src/main/scala/org/apache/spark/TestUtils.scala

@@ -21,12 +21,16 @@ import java.io.{ByteArrayInputStream, File, FileInputStream, FileOutputStream}
 import java.net.{URI, URL}
 import java.util.jar.{JarEntry, JarOutputStream}
 
-import scala.collection.JavaConversions._
+import scala.collection.JavaConverters._
+import scala.collection.mutable
+import scala.collection.mutable.ArrayBuffer
 
 import com.google.common.base.Charsets.UTF_8
 import com.google.common.io.{ByteStreams, Files}
 import javax.tools.{JavaFileObject, SimpleJavaFileObject, ToolProvider}
 
+import org.apache.spark.executor.TaskMetrics
+import org.apache.spark.scheduler._
 import org.apache.spark.util.Utils
 
 /**
@@ -153,4 +157,51 @@ private[spark] object TestUtils {
       "  @Override public String toString() { return \"" + toStringValue + "\"; }}")
     createCompiledClass(className, destDir, sourceFile, classpathUrls)
   }
+
+  /**
+   * Run some code involving jobs submitted to the given context and assert that the jobs spilled.
+   */
+  def assertSpilled[T](sc: SparkContext, identifier: String)(body: => T): Unit = {
+    val spillListener = new SpillListener
+    sc.addSparkListener(spillListener)
+    body
+    assert(spillListener.numSpilledStages > 0, s"expected $identifier to spill, but did not")
+  }
+
+  /**
+   * Run some code involving jobs submitted to the given context and assert that the jobs
+   * did not spill.
+   */
+  def assertNotSpilled[T](sc: SparkContext, identifier: String)(body: => T): Unit = {
+    val spillListener = new SpillListener
+    sc.addSparkListener(spillListener)
+    body
+    assert(spillListener.numSpilledStages == 0, s"expected $identifier to not spill, but did")
+  }
+
+}
+
+
+/**
+ * A [[SparkListener]] that detects whether spills have occurred in Spark jobs.
+ */
+private class SpillListener extends SparkListener {
+  private val stageIdToTaskMetrics = new mutable.HashMap[Int, ArrayBuffer[TaskMetrics]]
+  private val spilledStageIds = new mutable.HashSet[Int]
+
+  def numSpilledStages: Int = spilledStageIds.size
+
+  override def onTaskEnd(taskEnd: SparkListenerTaskEnd): Unit = {
+    stageIdToTaskMetrics.getOrElseUpdate(
+      taskEnd.stageId, new ArrayBuffer[TaskMetrics]) += taskEnd.taskMetrics
+  }
+
+  override def onStageCompleted(stageComplete: SparkListenerStageCompleted): Unit = {
+    val stageId = stageComplete.stageInfo.stageId
+    val metrics = stageIdToTaskMetrics.remove(stageId).toSeq.flatten
+    val spilled = metrics.map(_.memoryBytesSpilled).sum > 0
+    if (spilled) {
+      spilledStageIds += stageId
+    }
+  }
 }

core/src/main/scala/org/apache/spark/shuffle/ShuffleMemoryManager.scala

@@ -139,8 +139,10 @@ class ShuffleMemoryManager protected (
       throw new SparkException(
         s"Internal error: release called on $numBytes bytes but task only has $curMem")
     }
-    taskMemory(taskAttemptId) -= numBytes
-    memoryManager.releaseExecutionMemory(numBytes)
+    if (taskMemory.contains(taskAttemptId)) {
+      taskMemory(taskAttemptId) -= numBytes
+      memoryManager.releaseExecutionMemory(numBytes)
+    }
     memoryManager.notifyAll() // Notify waiters in tryToAcquire that memory has been freed
   }
 

core/src/main/scala/org/apache/spark/util/collection/ExternalAppendOnlyMap.scala

@@ -95,6 +95,12 @@ class ExternalAppendOnlyMap[K, V, C](
   private val keyComparator = new HashComparator[K]
   private val ser = serializer.newInstance()
 
+  /**
+   * Number of files this map has spilled so far.
+   * Exposed for testing.
+   */
+  private[collection] def numSpills: Int = spilledMaps.size
+
   /**
    * Insert the given key and value into the map.
    */

core/src/main/scala/org/apache/spark/util/collection/Spillable.scala

@@ -43,10 +43,15 @@ private[spark] trait Spillable[C] extends Logging {
   private[this] val shuffleMemoryManager = SparkEnv.get.shuffleMemoryManager
 
   // Initial threshold for the size of a collection before we start tracking its memory usage
-  // Exposed for testing
+  // For testing only
   private[this] val initialMemoryThreshold: Long =
     SparkEnv.get.conf.getLong("spark.shuffle.spill.initialMemoryThreshold", 5 * 1024 * 1024)
 
+  // Force this collection to spill when there are this many elements in memory
+  // For testing only
+  private[this] val numElementsForceSpillThreshold: Long =
+    SparkEnv.get.conf.getLong("spark.shuffle.spill.numElementsForceSpillThreshold", Long.MaxValue)
+
   // Threshold for this collection's size in bytes before we start tracking its memory usage
   // To avoid a large number of small spills, initialize this to a value orders of magnitude > 0
   private[this] var myMemoryThreshold = initialMemoryThreshold
@@ -69,27 +74,27 @@ private[spark] trait Spillable[C] extends Logging {
    * @return true if `collection` was spilled to disk; false otherwise
    */
   protected def maybeSpill(collection: C, currentMemory: Long): Boolean = {
+    var shouldSpill = false
     if (elementsRead % 32 == 0 && currentMemory >= myMemoryThreshold) {
       // Claim up to double our current memory from the shuffle memory pool
       val amountToRequest = 2 * currentMemory - myMemoryThreshold
       val granted = shuffleMemoryManager.tryToAcquire(amountToRequest)
       myMemoryThreshold += granted
-      if (myMemoryThreshold <= currentMemory) {
-        // We were granted too little memory to grow further (either tryToAcquire returned 0,
-        // or we already had more memory than myMemoryThreshold); spill the current collection
-        _spillCount += 1
-        logSpillage(currentMemory)
-
-        spill(collection)
-
-        _elementsRead = 0
-        // Keep track of spills, and release memory
-        _memoryBytesSpilled += currentMemory
-        releaseMemoryForThisThread()
-        return true
-      }
+      // If we were granted too little memory to grow further (either tryToAcquire returned 0,
+      // or we already had more memory than myMemoryThreshold), spill the current collection
+      shouldSpill = currentMemory >= myMemoryThreshold
+    }
+    shouldSpill = shouldSpill || _elementsRead > numElementsForceSpillThreshold
+    // Actually spill
+    if (shouldSpill) {
+      _spillCount += 1
+      logSpillage(currentMemory)
+      spill(collection)
+      _elementsRead = 0
+      _memoryBytesSpilled += currentMemory
+      releaseMemoryForThisThread()
     }
-    false
+    shouldSpill
   }
 
   /**

core/src/test/scala/org/apache/spark/DistributedSuite.scala

@@ -203,22 +203,35 @@ class DistributedSuite extends SparkFunSuite with Matchers with LocalSparkContex
   }
 
   test("compute without caching when no partitions fit in memory") {
-    sc = new SparkContext(clusterUrl, "test")
-    // data will be 4 million * 4 bytes = 16 MB in size, but our memoryFraction set the cache
-    // to only 50 KB (0.0001 of 512 MB), so no partitions should fit in memory
-    val data = sc.parallelize(1 to 4000000, 2).persist(StorageLevel.MEMORY_ONLY_SER)
-    assert(data.count() === 4000000)
-    assert(data.count() === 4000000)
-    assert(data.count() === 4000000)
+    val size = 10000
+    val conf = new SparkConf()
+      .set("spark.storage.unrollMemoryThreshold", "1024")
+      .set("spark.testing.memory", (size / 2).toString)
+    sc = new SparkContext(clusterUrl, "test", conf)
+    val data = sc.parallelize(1 to size, 2).persist(StorageLevel.MEMORY_ONLY)
+    assert(data.count() === size)
+    assert(data.count() === size)
+    assert(data.count() === size)
+    // ensure only a subset of partitions were cached
+    val rddBlocks = sc.env.blockManager.master.getMatchingBlockIds(_.isRDD, askSlaves = true)
+    assert(rddBlocks.size === 0, s"expected no RDD blocks, found ${rddBlocks.size}")
   }
 
   test("compute when only some partitions fit in memory") {
-    sc = new SparkContext(clusterUrl, "test", new SparkConf)
-    // TODO: verify that only a subset of partitions fit in memory (SPARK-11078)
-    val data = sc.parallelize(1 to 4000000, 20).persist(StorageLevel.MEMORY_ONLY_SER)
-    assert(data.count() === 4000000)
-    assert(data.count() === 4000000)
-    assert(data.count() === 4000000)
+    val size = 10000
+    val numPartitions = 10
+    val conf = new SparkConf()
+      .set("spark.storage.unrollMemoryThreshold", "1024")
+      .set("spark.testing.memory", (size * numPartitions).toString)
+    sc = new SparkContext(clusterUrl, "test", conf)
+    val data = sc.parallelize(1 to size, numPartitions).persist(StorageLevel.MEMORY_ONLY)
+    assert(data.count() === size)
+    assert(data.count() === size)
+    assert(data.count() === size)
+    // ensure only a subset of partitions were cached
+    val rddBlocks = sc.env.blockManager.master.getMatchingBlockIds(_.isRDD, askSlaves = true)
+    assert(rddBlocks.size > 0, "no RDD blocks found")
+    assert(rddBlocks.size < numPartitions, s"too many RDD blocks found, expected <$numPartitions")
   }
 
   test("passing environment variables to cluster") {