[SPARK-9747] [SQL] Avoid starving an unsafe operator in aggregation

This is the sister patch to #8011, but for aggregation.

In a nutshell: create the `TungstenAggregationIterator` before computing the parent partition. Internally this creates a `BytesToBytesMap` which acquires a page in the constructor as of this patch. This ensures that the aggregation operator is not starved since we reserve at least 1 page in advance.

rxin yhuai

Author: Andrew Or <andrew@databricks.com>

Closes #8038 from andrewor14/unsafe-starve-memory-agg.

core/src/main/java/org/apache/spark/unsafe/map/BytesToBytesMap.java

@@ -193,6 +193,11 @@ public BytesToBytesMap(
         TaskMemoryManager.MAXIMUM_PAGE_SIZE_BYTES);
     }
     allocate(initialCapacity);
+
+    // Acquire a new page as soon as we construct the map to ensure that we have at least
+    // one page to work with. Otherwise, other operators in the same task may starve this
+    // map (SPARK-9747).
+    acquireNewPage();
   }
 
   public BytesToBytesMap(
@@ -574,16 +579,9 @@ public boolean putNewKey(
           final long lengthOffsetInPage = currentDataPage.getBaseOffset() + pageCursor;
           Platform.putInt(pageBaseObject, lengthOffsetInPage, END_OF_PAGE_MARKER);
         }
-        final long memoryGranted = shuffleMemoryManager.tryToAcquire(pageSizeBytes);
-        if (memoryGranted != pageSizeBytes) {
-          shuffleMemoryManager.release(memoryGranted);
-          logger.debug("Failed to acquire {} bytes of memory", pageSizeBytes);
+        if (!acquireNewPage()) {
           return false;
         }
-        MemoryBlock newPage = taskMemoryManager.allocatePage(pageSizeBytes);
-        dataPages.add(newPage);
-        pageCursor = 0;
-        currentDataPage = newPage;
         dataPage = currentDataPage;
         dataPageBaseObject = currentDataPage.getBaseObject();
         dataPageInsertOffset = currentDataPage.getBaseOffset();
@@ -642,6 +640,24 @@ public boolean putNewKey(
     }
   }
 
+  /**
+   * Acquire a new page from the {@link ShuffleMemoryManager}.
+   * @return whether there is enough space to allocate the new page.
+   */
+  private boolean acquireNewPage() {
+    final long memoryGranted = shuffleMemoryManager.tryToAcquire(pageSizeBytes);
+    if (memoryGranted != pageSizeBytes) {
+      shuffleMemoryManager.release(memoryGranted);
+      logger.debug("Failed to acquire {} bytes of memory", pageSizeBytes);
+      return false;
+    }
+    MemoryBlock newPage = taskMemoryManager.allocatePage(pageSizeBytes);
+    dataPages.add(newPage);
+    pageCursor = 0;
+    currentDataPage = newPage;
+    return true;
+  }
+
   /**
    * Allocate new data structures for this map. When calling this outside of the constructor,
    * make sure to keep references to the old data structures so that you can free them.
@@ -748,7 +764,7 @@ public long getNumHashCollisions() {
   }
 
   @VisibleForTesting
-  int getNumDataPages() {
+  public int getNumDataPages() {
     return dataPages.size();
   }
 

core/src/main/java/org/apache/spark/util/collection/unsafe/sort/UnsafeExternalSorter.java

@@ -132,16 +132,15 @@ private UnsafeExternalSorter(
 
     if (existingInMemorySorter == null) {
       initializeForWriting();
+      // Acquire a new page as soon as we construct the sorter to ensure that we have at
+      // least one page to work with. Otherwise, other operators in the same task may starve
+      // this sorter (SPARK-9709). We don't need to do this if we already have an existing sorter.
+      acquireNewPage();
     } else {
       this.isInMemSorterExternal = true;
       this.inMemSorter = existingInMemorySorter;
     }
 
-    // Acquire a new page as soon as we construct the sorter to ensure that we have at
-    // least one page to work with. Otherwise, other operators in the same task may starve
-    // this sorter (SPARK-9709).
-    acquireNewPage();
-
     // Register a cleanup task with TaskContext to ensure that memory is guaranteed to be freed at
     // the end of the task. This is necessary to avoid memory leaks in when the downstream operator
     // does not fully consume the sorter's output (e.g. sort followed by limit).

core/src/test/java/org/apache/spark/unsafe/map/AbstractBytesToBytesMapSuite.java

@@ -543,7 +543,7 @@ public void testPeakMemoryUsed() {
           Platform.LONG_ARRAY_OFFSET,
           8);
         newPeakMemory = map.getPeakMemoryUsedBytes();
-        if (i % numRecordsPerPage == 0) {
+        if (i % numRecordsPerPage == 0 && i > 0) {
           // We allocated a new page for this record, so peak memory should change
           assertEquals(previousPeakMemory + pageSizeBytes, newPeakMemory);
         } else {
@@ -561,4 +561,13 @@ public void testPeakMemoryUsed() {
       map.free();
     }
   }
+
+  @Test
+  public void testAcquirePageInConstructor() {
+    final BytesToBytesMap map = new BytesToBytesMap(
+      taskMemoryManager, shuffleMemoryManager, 1, PAGE_SIZE_BYTES);
+    assertEquals(1, map.getNumDataPages());
+    map.free();
+  }
+
 }

sql/core/src/main/java/org/apache/spark/sql/execution/UnsafeFixedWidthAggregationMap.java

@@ -19,6 +19,8 @@
 
 import java.io.IOException;
 
+import com.google.common.annotations.VisibleForTesting;
+
 import org.apache.spark.SparkEnv;
 import org.apache.spark.shuffle.ShuffleMemoryManager;
 import org.apache.spark.sql.catalyst.InternalRow;
@@ -220,6 +222,11 @@ public long getPeakMemoryUsedBytes() {
     return map.getPeakMemoryUsedBytes();
   }
 
+  @VisibleForTesting
+  public int getNumDataPages() {
+    return map.getNumDataPages();
+  }
+
   /**
    * Free the memory associated with this map. This is idempotent and can be called multiple times.
    */

sql/core/src/main/scala/org/apache/spark/sql/execution/aggregate/TungstenAggregate.scala

@@ -17,12 +17,13 @@
 
 package org.apache.spark.sql.execution.aggregate
 
-import org.apache.spark.rdd.RDD
+import org.apache.spark.TaskContext
+import org.apache.spark.rdd.{MapPartitionsWithPreparationRDD, RDD}
 import org.apache.spark.sql.catalyst.InternalRow
 import org.apache.spark.sql.catalyst.errors._
 import org.apache.spark.sql.catalyst.expressions.aggregate.AggregateExpression2
 import org.apache.spark.sql.catalyst.expressions._
-import org.apache.spark.sql.catalyst.plans.physical.{UnspecifiedDistribution, ClusteredDistribution, AllTuples, Distribution}
+import org.apache.spark.sql.catalyst.plans.physical._
 import org.apache.spark.sql.execution.{UnaryNode, SparkPlan}
 import org.apache.spark.sql.execution.metric.SQLMetrics
 
@@ -68,35 +69,56 @@ case class TungstenAggregate(
   protected override def doExecute(): RDD[InternalRow] = attachTree(this, "execute") {
     val numInputRows = longMetric("numInputRows")
     val numOutputRows = longMetric("numOutputRows")
-    child.execute().mapPartitions { iter =>
-      val hasInput = iter.hasNext
-      if (!hasInput && groupingExpressions.nonEmpty) {
-        // This is a grouped aggregate and the input iterator is empty,
-        // so return an empty iterator.
-        Iterator.empty.asInstanceOf[Iterator[UnsafeRow]]
-      } else {
-        val aggregationIterator =
-          new TungstenAggregationIterator(
-            groupingExpressions,
-            nonCompleteAggregateExpressions,
-            completeAggregateExpressions,
-            initialInputBufferOffset,
-            resultExpressions,
-            newMutableProjection,
-            child.output,
-            iter,
-            testFallbackStartsAt,
-            numInputRows,
-            numOutputRows)
-
-        if (!hasInput && groupingExpressions.isEmpty) {
+
+    /**
+     * Set up the underlying unsafe data structures used before computing the parent partition.
+     * This makes sure our iterator is not starved by other operators in the same task.
+     */
+    def preparePartition(): TungstenAggregationIterator = {
+      new TungstenAggregationIterator(
+        groupingExpressions,
+        nonCompleteAggregateExpressions,
+        completeAggregateExpressions,
+        initialInputBufferOffset,
+        resultExpressions,
+        newMutableProjection,
+        child.output,
+        testFallbackStartsAt,
+        numInputRows,
+        numOutputRows)
+    }
+
+    /** Compute a partition using the iterator already set up previously. */
+    def executePartition(
+        context: TaskContext,
+        partitionIndex: Int,
+        aggregationIterator: TungstenAggregationIterator,
+        parentIterator: Iterator[InternalRow]): Iterator[UnsafeRow] = {
+      val hasInput = parentIterator.hasNext
+      if (!hasInput) {
+        // We're not using the underlying map, so we just can free it here
+        aggregationIterator.free()
+        if (groupingExpressions.isEmpty) {
           numOutputRows += 1
           Iterator.single[UnsafeRow](aggregationIterator.outputForEmptyGroupingKeyWithoutInput())
         } else {
-          aggregationIterator
+          // This is a grouped aggregate and the input iterator is empty,
+          // so return an empty iterator.
+          Iterator[UnsafeRow]()
         }
+      } else {
+        aggregationIterator.start(parentIterator)
+        aggregationIterator
       }
     }
+
+    // Note: we need to set up the iterator in each partition before computing the
+    // parent partition, so we cannot simply use `mapPartitions` here (SPARK-9747).
+    val resultRdd = {
+      new MapPartitionsWithPreparationRDD[UnsafeRow, InternalRow, TungstenAggregationIterator](
+        child.execute(), preparePartition, executePartition, preservesPartitioning = true)
+    }
+    resultRdd.asInstanceOf[RDD[InternalRow]]
   }
 
   override def simpleString: String = {

sql/core/src/main/scala/org/apache/spark/sql/execution/aggregate/TungstenAggregationIterator.scala

@@ -72,8 +72,6 @@ import org.apache.spark.sql.types.StructType
  *   the function used to create mutable projections.
  * @param originalInputAttributes
  *   attributes of representing input rows from `inputIter`.
- * @param inputIter
- *   the iterator containing input [[UnsafeRow]]s.
  */
 class TungstenAggregationIterator(
     groupingExpressions: Seq[NamedExpression],
@@ -83,12 +81,14 @@ class TungstenAggregationIterator(
     resultExpressions: Seq[NamedExpression],
     newMutableProjection: (Seq[Expression], Seq[Attribute]) => (() => MutableProjection),
     originalInputAttributes: Seq[Attribute],
-    inputIter: Iterator[InternalRow],
     testFallbackStartsAt: Option[Int],
     numInputRows: LongSQLMetric,
     numOutputRows: LongSQLMetric)
   extends Iterator[UnsafeRow] with Logging {
 
+  // The parent partition iterator, to be initialized later in `start`
+  private[this] var inputIter: Iterator[InternalRow] = null
+
   ///////////////////////////////////////////////////////////////////////////
   // Part 1: Initializing aggregate functions.
   ///////////////////////////////////////////////////////////////////////////
@@ -348,11 +348,15 @@ class TungstenAggregationIterator(
     false // disable tracking of performance metrics
   )
 
+  // Exposed for testing
+  private[aggregate] def getHashMap: UnsafeFixedWidthAggregationMap = hashMap
+
   // The function used to read and process input rows. When processing input rows,
   // it first uses hash-based aggregation by putting groups and their buffers in
   // hashMap. If we could not allocate more memory for the map, we switch to
   // sort-based aggregation (by calling switchToSortBasedAggregation).
   private def processInputs(): Unit = {
+    assert(inputIter != null, "attempted to process input when iterator was null")
     while (!sortBased && inputIter.hasNext) {
       val newInput = inputIter.next()
       numInputRows += 1
@@ -372,6 +376,7 @@ class TungstenAggregationIterator(
   // that it switch to sort-based aggregation after `fallbackStartsAt` input rows have
   // been processed.
   private def processInputsWithControlledFallback(fallbackStartsAt: Int): Unit = {
+    assert(inputIter != null, "attempted to process input when iterator was null")
     var i = 0
     while (!sortBased && inputIter.hasNext) {
       val newInput = inputIter.next()
@@ -412,6 +417,7 @@ class TungstenAggregationIterator(
    * Switch to sort-based aggregation when the hash-based approach is unable to acquire memory.
    */
   private def switchToSortBasedAggregation(firstKey: UnsafeRow, firstInput: InternalRow): Unit = {
+    assert(inputIter != null, "attempted to process input when iterator was null")
     logInfo("falling back to sort based aggregation.")
     // Step 1: Get the ExternalSorter containing sorted entries of the map.
     externalSorter = hashMap.destructAndCreateExternalSorter()
@@ -431,6 +437,11 @@ class TungstenAggregationIterator(
       case _ => false
     }
 
+    // Note: Since we spill the sorter's contents immediately after creating it, we must insert
+    // something into the sorter here to ensure that we acquire at least a page of memory.
+    // This is done through `externalSorter.insertKV`, which will trigger the page allocation.
+    // Otherwise, children operators may steal the window of opportunity and starve our sorter.
+
     if (needsProcess) {
       // First, we create a buffer.
       val buffer = createNewAggregationBuffer()
@@ -588,27 +599,33 @@ class TungstenAggregationIterator(
   //         have not switched to sort-based aggregation.
   ///////////////////////////////////////////////////////////////////////////
 
-  // Starts to process input rows.
-  testFallbackStartsAt match {
-    case None =>
-      processInputs()
-    case Some(fallbackStartsAt) =>
-      // This is the testing path. processInputsWithControlledFallback is same as processInputs
-      // except that it switches to sort-based aggregation after `fallbackStartsAt` input rows
-      // have been processed.
-      processInputsWithControlledFallback(fallbackStartsAt)
-  }
+  /**
+   * Start processing input rows.
+   * Only after this method is called will this iterator be non-empty.
+   */
+  def start(parentIter: Iterator[InternalRow]): Unit = {
+    inputIter = parentIter
+    testFallbackStartsAt match {
+      case None =>
+        processInputs()
+      case Some(fallbackStartsAt) =>
+        // This is the testing path. processInputsWithControlledFallback is same as processInputs
+        // except that it switches to sort-based aggregation after `fallbackStartsAt` input rows
+        // have been processed.
+        processInputsWithControlledFallback(fallbackStartsAt)
+    }
 
-  // If we did not switch to sort-based aggregation in processInputs,
-  // we pre-load the first key-value pair from the map (to make hasNext idempotent).
-  if (!sortBased) {
-    // First, set aggregationBufferMapIterator.
-    aggregationBufferMapIterator = hashMap.iterator()
-    // Pre-load the first key-value pair from the aggregationBufferMapIterator.
-    mapIteratorHasNext = aggregationBufferMapIterator.next()
-    // If the map is empty, we just free it.
-    if (!mapIteratorHasNext) {
-      hashMap.free()
+    // If we did not switch to sort-based aggregation in processInputs,
+    // we pre-load the first key-value pair from the map (to make hasNext idempotent).
+    if (!sortBased) {
+      // First, set aggregationBufferMapIterator.
+      aggregationBufferMapIterator = hashMap.iterator()
+      // Pre-load the first key-value pair from the aggregationBufferMapIterator.
+      mapIteratorHasNext = aggregationBufferMapIterator.next()
+      // If the map is empty, we just free it.
+      if (!mapIteratorHasNext) {
+        hashMap.free()
+      }
     }
   }
 
@@ -673,21 +690,20 @@ class TungstenAggregationIterator(
   }
 
   ///////////////////////////////////////////////////////////////////////////
-  // Part 8: A utility function used to generate a output row when there is no
-  // input and there is no grouping expression.
+  // Part 8: Utility functions
   ///////////////////////////////////////////////////////////////////////////
 
+  /**
+   * Generate a output row when there is no input and there is no grouping expression.
+   */
   def outputForEmptyGroupingKeyWithoutInput(): UnsafeRow = {
-    if (groupingExpressions.isEmpty) {
-      sortBasedAggregationBuffer.copyFrom(initialAggregationBuffer)
-      // We create a output row and copy it. So, we can free the map.
-      val resultCopy =
-        generateOutput(UnsafeRow.createFromByteArray(0, 0), sortBasedAggregationBuffer).copy()
-      hashMap.free()
-      resultCopy
-    } else {
-      throw new IllegalStateException(
-        "This method should not be called when groupingExpressions is not empty.")
-    }
+    assert(groupingExpressions.isEmpty)
+    assert(inputIter == null)
+    generateOutput(UnsafeRow.createFromByteArray(0, 0), initialAggregationBuffer)
+  }
+
+  /** Free memory used in the underlying map. */
+  def free(): Unit = {
+    hashMap.free()
   }
 }