Heuristic to speed up partial aggregates that get larger (NVIDIA#8618)

Signed-off-by: Robert (Bobby) Evans <bobby@apache.org>

build/buildall

@@ -242,7 +242,7 @@ function build_single_shim() {
       -Dskip \
       -Dmaven.scalastyle.skip="$SKIP_CHECKS" \
       -pl aggregator -am > "$LOG_FILE" 2>&1 || {
-        [[ "$LOG_FILE" != "/dev/tty" ]] && tail -20 "$LOG_FILE" || true
+        [[ "$LOG_FILE" != "/dev/tty" ]] && echo "$LOG_FILE:" && tail -20 "$LOG_FILE" || true
         exit 255
       }
 }

docs/additional-functionality/advanced_configs.md

@@ -6,12 +6,12 @@ nav_order: 10
 ---
 <!-- Generated by RapidsConf.help. DO NOT EDIT! -->
 # RAPIDS Accelerator for Apache Spark Advanced Configuration
-Most users will not need to modify the configuration options listed below. 
+Most users will not need to modify the configuration options listed below.
 They are documented here for completeness and advanced usage.
 
 The following configuration options are supported by the RAPIDS Accelerator for Apache Spark.
 
-For commonly used configurations and examples of setting options, please refer to the 
+For commonly used configurations and examples of setting options, please refer to the
 [RAPIDS Accelerator for Configuration](../configs.md) page.
 
 

integration_tests/src/main/python/hash_aggregate_test.py

@@ -1883,3 +1883,15 @@ def test_hash_aggregate_complete_with_grouping_expressions():
         lambda spark : spark.range(10).withColumn("id2", f.col("id")),
         "hash_agg_complete_table",
         "select id, avg(id) from hash_agg_complete_table group by id, id2 + 1")
+
+@ignore_order(local=True)
+@pytest.mark.parametrize('cast_key_to', ["byte", "short", "int",
+    "long", "string", "DECIMAL(38,5)"], ids=idfn)
+def test_hash_agg_force_pre_sort(cast_key_to):
+    def do_it(spark):
+        gen = StructGen([("key", UniqueLongGen()), ("value", long_gen)], nullable=False)
+        df = gen_df(spark, gen)
+        return df.selectExpr("CAST((key div 10) as " + cast_key_to + ") as key", "value").groupBy("key").sum("value")
+    assert_gpu_and_cpu_are_equal_collect(do_it,
+        conf={'spark.rapids.sql.agg.forceSinglePassPartialSort': True,
+            'spark.rapids.sql.agg.singlePassPartialSortEnabled': True})

sql-plugin/src/main/scala/com/nvidia/spark/rapids/GpuBatchUtils.scala

@@ -1,5 +1,5 @@
 /*
- * Copyright (c) 2020-2021, NVIDIA CORPORATION.
+ * Copyright (c) 2020-2023, NVIDIA CORPORATION.
  *
  * Licensed under the Apache License, Version 2.0 (the "License");
  * you may not use this file except in compliance with the License.
@@ -63,6 +63,32 @@ object GpuBatchUtils {
     estimateGpuMemory(field.dataType, field.nullable, rowCount)
   }
 
+  /**
+   * Get the minimum size a column could be that matches these conditions.
+   */
+  def minGpuMemory(dataType:DataType, nullable: Boolean, rowCount: Long): Long = {
+    val validityBufferSize = if (nullable) {
+      calculateValidityBufferSize(rowCount)
+    } else {
+      0
+    }
+
+    val dataSize = dataType match {
+      case DataTypes.BinaryType | DataTypes.StringType | _: MapType | _: ArrayType=>
+        // For nested types (like list or string) the smallest possible size is when
+        // each row is empty (length 0). In that case there is no data, just offsets
+        // and all of the offsets are 0.
+        calculateOffsetBufferSize(rowCount)
+      case dt: StructType =>
+        dt.fields.map { f =>
+          minGpuMemory(f.dataType, f.nullable, rowCount)
+        }.sum
+      case dt =>
+        dt.defaultSize * rowCount
+    }
+    dataSize + validityBufferSize
+  }
+
   def estimateGpuMemory(dataType: DataType, nullable: Boolean, rowCount: Long): Long = {
     val validityBufferSize = if (nullable) {
       calculateValidityBufferSize(rowCount)

sql-plugin/src/main/scala/com/nvidia/spark/rapids/GpuDeviceManager.scala

@@ -61,6 +61,11 @@ object GpuDeviceManager extends Logging {
    */
   def getDeviceId(): Option[Int] = deviceId
 
+  @volatile private var poolSizeLimit = 0L
+
+  // Never split below 100 MiB (but this is really just for testing)
+  def getSplitUntilSize: Long = Math.max(poolSizeLimit / 8, 100 * 1024 * 1024)
+
   // Attempt to set and acquire the gpu, return true if acquired, false otherwise
   def tryToSetGpuDeviceAndAcquire(addr: Int): Boolean = {
     try {
@@ -149,6 +154,7 @@ object GpuDeviceManager extends Logging {
 
     chunkedPackMemoryResource.foreach(_.close)
     chunkedPackMemoryResource = None
+    poolSizeLimit = 0L
 
     RapidsBufferCatalog.close()
     GpuShuffleEnv.shutdown()
@@ -338,6 +344,7 @@ object GpuDeviceManager extends Logging {
 
       Cuda.setDevice(gpuId)
       try {
+        poolSizeLimit = poolAllocation
         Rmm.initialize(init, logConf, poolAllocation)
       } catch {
         case firstEx: CudfException if ((init & RmmAllocationMode.CUDA_ASYNC) != 0) => {
@@ -351,6 +358,7 @@ object GpuDeviceManager extends Logging {
               logError(
                 "Failed to initialize RMM with either ASYNC or ARENA allocators. Exiting...")
               secondEx.addSuppressed(firstEx)
+              poolSizeLimit = 0L
               throw secondEx
             }
           }

sql-plugin/src/main/scala/com/nvidia/spark/rapids/GpuSortExec.scala

@@ -136,8 +136,7 @@ case class GpuSortExec(
     val singleBatch = sortType == FullSortSingleBatch
     child.executeColumnar().mapPartitions { cbIter =>
       if (outOfCore) {
-        val cpuOrd = new LazilyGeneratedOrdering(sorter.cpuOrdering)
-        val iter = GpuOutOfCoreSortIterator(cbIter, sorter, cpuOrd,
+        val iter = GpuOutOfCoreSortIterator(cbIter, sorter,
           targetSize, opTime, sortTime, outputBatch, outputRows)
         TaskContext.get().addTaskCompletionListener(_ -> iter.close())
         iter
@@ -239,14 +238,14 @@ class Pending(cpuOrd: LazilyGeneratedOrdering) extends AutoCloseable {
 case class GpuOutOfCoreSortIterator(
     iter: Iterator[ColumnarBatch],
     sorter: GpuSorter,
-    cpuOrd: LazilyGeneratedOrdering,
     targetSize: Long,
     opTime: GpuMetric,
     sortTime: GpuMetric,
     outputBatches: GpuMetric,
     outputRows: GpuMetric) extends Iterator[ColumnarBatch]
     with AutoCloseable {
 
+  private val cpuOrd = new LazilyGeneratedOrdering(sorter.cpuOrdering)
   // A priority queue of data that is not merged yet.
   private val pending = new Pending(cpuOrd)
 
@@ -328,16 +327,16 @@ case class GpuOutOfCoreSortIterator(
         targetRowCount until rows by targetRowCount
       }
       // Get back the first row so we can sort the batches
-      val gatherIndexes = if (hasFullySortedData) {
+      val lowerGatherIndexes = if (hasFullySortedData) {
         // The first batch is sorted so don't gather a row for it
         splitIndexes
       } else {
         Seq(0) ++ splitIndexes
       }
 
-      val boundaries =
-        withResource(new NvtxRange("boundaries", NvtxColor.ORANGE)) { _ =>
-          withResource(ColumnVector.fromInts(gatherIndexes: _*)) { gatherMap =>
+      val lowerBoundaries =
+        withResource(new NvtxRange("lower boundaries", NvtxColor.ORANGE)) { _ =>
+          withResource(ColumnVector.fromInts(lowerGatherIndexes: _*)) { gatherMap =>
             withResource(sortedTbl.gather(gatherMap)) { boundariesTab =>
               convertBoundaries(boundariesTab)
             }
@@ -355,9 +354,9 @@ case class GpuOutOfCoreSortIterator(
         }
 
         closeOnExcept(sortedCb) { _ =>
-          assert(boundaries.length == stillPending.length)
+          assert(lowerBoundaries.length == stillPending.length)
           closeOnExcept(pendingObs) { _ =>
-            stillPending.zip(boundaries).foreach {
+            stillPending.zip(lowerBoundaries).foreach {
               case (ct: ContiguousTable, lower: UnsafeRow) =>
                 if (ct.getRowCount > 0) {
                   val sp = SpillableColumnarBatch(ct, sorter.projectedBatchTypes,

sql-plugin/src/main/scala/com/nvidia/spark/rapids/RapidsConf.scala

@@ -1273,6 +1273,23 @@ object RapidsConf {
     .booleanConf
     .createWithDefault(true)
 
+  val ENABLE_SINGLE_PASS_PARTIAL_SORT_AGG: ConfEntryWithDefault[Boolean] =
+    conf("spark.rapids.sql.agg.singlePassPartialSortEnabled")
+    .doc("Enable or disable a single pass partial sort optimization where if a heuristic " +
+        "indicates it would be good we pre-sort the data before a partial agg and then " +
+        "do the agg in a single pass with no merge, so there is no spilling")
+    .internal()
+    .booleanConf
+    .createWithDefault(true)
+
+  val FORCE_SINGLE_PASS_PARTIAL_SORT_AGG: ConfEntryWithDefault[Boolean] =
+    conf("spark.rapids.sql.agg.forceSinglePassPartialSort")
+    .doc("Force a single pass partial sort agg to happen in all cases that it could, " +
+        "no matter what the heuristic says. This is really just for testing.")
+    .internal()
+    .booleanConf
+    .createWithDefault(false)
+
   val ENABLE_REGEXP = conf("spark.rapids.sql.regexp.enabled")
     .doc("Specifies whether supported regular expressions will be evaluated on the GPU. " +
       "Unsupported expressions will fall back to CPU. However, there are some known edge cases " +
@@ -2007,12 +2024,12 @@ object RapidsConf {
       println(s"<!-- Generated by RapidsConf.help. DO NOT EDIT! -->")
       // scalastyle:off line.size.limit
       println("""# RAPIDS Accelerator for Apache Spark Advanced Configuration
-        |Most users will not need to modify the configuration options listed below. 
+        |Most users will not need to modify the configuration options listed below.
         |They are documented here for completeness and advanced usage.
         |
         |The following configuration options are supported by the RAPIDS Accelerator for Apache Spark.
         |
-        |For commonly used configurations and examples of setting options, please refer to the 
+        |For commonly used configurations and examples of setting options, please refer to the
         |[RAPIDS Accelerator for Configuration](../configs.md) page.
         |""".stripMargin)
       // scalastyle:on line.size.limit
@@ -2554,6 +2571,10 @@ class RapidsConf(conf: Map[String, String]) extends Logging {
 
   lazy val isRangeWindowDecimalEnabled: Boolean = get(ENABLE_RANGE_WINDOW_DECIMAL)
 
+  lazy val allowSinglePassPartialSortAgg: Boolean = get(ENABLE_SINGLE_PASS_PARTIAL_SORT_AGG)
+
+  lazy val forceSinglePassPartialSortAgg: Boolean = get(FORCE_SINGLE_PASS_PARTIAL_SORT_AGG)
+
   lazy val isRegExpEnabled: Boolean = get(ENABLE_REGEXP)
 
   lazy val maxRegExpStateMemory: Long =  {

sql-plugin/src/main/scala/com/nvidia/spark/rapids/RapidsMeta.scala

@@ -1418,7 +1418,8 @@ final class RuleNotFoundExprMeta[INPUT <: Expression](
     willNotWorkOnGpu(s"GPU does not currently support the operator ${expr.getClass}")
 
   override def convertToGpu(): GpuExpression =
-    throw new IllegalStateException("Cannot be converted to GPU")
+    throw new IllegalStateException(s"Cannot be converted to GPU ${expr.getClass} " +
+        s"${expr.dataType} $expr")
 }
 
 /** Base class for metadata around `RunnableCommand`. */