Github user andrewor14 commented on a diff in the pull request:
https://github.com/apache/spark/pull/1090#discussion_r13820032
--- Diff: core/src/main/scala/org/apache/spark/rdd/SortedParitionsRDD.scala
---
@@ -0,0 +1,337 @@
+/*
+ * 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.rdd
+
+import scala.reflect.ClassTag
+import scala.collection.mutable.ArrayBuffer
+import java.io.{InputStream, BufferedInputStream, FileInputStream, File,
Serializable, EOFException}
+import org.apache.spark.{Partition, TaskContext}
+import org.apache.spark.{Logging, SparkEnv}
+import org.apache.spark.serializer.Serializer
+import org.apache.spark.storage.{BlockId, BlockManager}
+import org.apache.spark.util.SizeEstimator
+
+/**
+ * An RDD that sorts each of it's partitions independently.
+ *
+ * If partitions are too large to fit in memory, they are externally
sorted.
+ *
+ * Two parameters control the memory threshold for external sort:
+ *
+ * `spark.shuffle.memoryFraction` specifies the collective amount of
memory used for storing
+ * sub lists as a fraction of the executor's total memory. Since each
concurrently running
+ * task maintains one map, the actual threshold for each map is this
quantity divided by the
+ * number of running tasks.
+ *
+ * `spark.shuffle.safetyFraction` specifies an additional margin of
safety as a fraction of
+ * this threshold, in case sub list size estimation is not sufficiently
accurate.
+ */
+private[spark] class SortedPartitionsRDD[T: ClassTag](
+ prev: RDD[T],
+ lt: (T, T) => Boolean)
+ extends RDD[T](prev) {
+
+ override def getPartitions: Array[Partition] = firstParent[T].partitions
+
+ // Since sorting partitions cannot change a partition's keys
+ override val partitioner = prev.partitioner
+
+ override def compute(split: Partition, context: TaskContext) = {
+ new SortedIterator(firstParent[T].iterator(split, context), lt)
+ }
+}
+
+/**
+ * An iterator that sorts a supplied iterator, either in-memory or
externally.
+ */
+private[spark] class SortedIterator[T](iter: Iterator[T], lt: (T, T) =>
Boolean)
+ extends Iterator[T] with Logging {
+ private val sparkConf = SparkEnv.get.conf
+ // Collective memory threshold shared across all running tasks
+ private val maxMemoryThreshold = {
+ val memoryFraction =
sparkConf.getDouble("spark.shuffle.memoryFraction", 0.3)
+ val safetyFraction =
sparkConf.getDouble("spark.shuffle.safetyFraction", 0.8)
+ (Runtime.getRuntime.maxMemory * memoryFraction * safetyFraction).toLong
+ }
+
+ // Number of list elements before tracking memory usage
+ private val trackMemoryThreshold = 1000
+
+ private val sorted = doSort()
+
+ def hasNext : Boolean = {
+ sorted.hasNext
+ }
+
+ def next : T = {
+ sorted.next
+ }
+
+ /**
+ * Sort the incoming iterator.
+ * Any input that cannot fit in memory is split into sorted sub-lists
and spilled to disk.
+ * Any spilled sub-lists are merge sorted and written back to disk.
+ */
+ private def doSort() : Iterator[T] = {
+ val subLists = new ArrayBuffer[Iterator[T]]()
+
+ // keep the first sub-list in memory
+ subLists += nextSubList
+
+ while (iter.hasNext) {
+ // spill remaining sub-lists to disk
+ var diskBuffer = new DiskBuffer[T]()
+ diskBuffer ++= nextSubList
+ subLists += diskBuffer.iterator
+ }
+ logInfo("Merge sorting one in-memory list with %d external
list(s)".format(subLists.size - 1))
+
+ merge(subLists)
+ }
+
+ /**
+ * Gets a sorted sub-list that can fit in memory.
+ */
+ private def nextSubList() : Iterator[T] = {
+ var subList = new SizeTrackingArrayBuffer[T](1000)
+ while (fitsInMemory(subList) && iter.hasNext) {
+ subList += iter.next
+ }
+ return subList.sortWith(lt).iterator
+ }
+
+ /**
+ * Determines if a given list can fit in memory.
+ * This algorithm is similar to that found in ExternalAppendOnlyMap.
+ */
+ private def fitsInMemory(list : SizeTrackingArrayBuffer[T]) : Boolean = {
+ if (list.size > trackMemoryThreshold && list.atNextSampleSize) {
+ val listSize = list.estimateSize()
+ val shuffleMemoryMap = SparkEnv.get.shuffleMemoryMap
--- End diff --
Same here. `ExternalAppendOnlyMap` is supposedly the only user of this map.
We should probably rename this if we want to use it for sorting too.
---
If your project is set up for it, you can reply to this email and have your
reply appear on GitHub as well. If your project does not have this feature
enabled and wishes so, or if the feature is enabled but not working, please
contact infrastructure at infrastruct...@apache.org or file a JIRA ticket
with INFRA.
---
