Github user pwendell commented on a diff in the pull request:
https://github.com/apache/spark/pull/4450#discussion_r29118576
--- Diff:
core/src/main/scala/org/apache/spark/util/collection/PartitionedSerializedPairBuffer.scala
---
@@ -0,0 +1,254 @@
+/*
+ * 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.util.collection
+
+import java.io.InputStream
+import java.nio.IntBuffer
+import java.util.Comparator
+
+import org.apache.spark.SparkEnv
+import org.apache.spark.serializer.{JavaSerializerInstance,
SerializerInstance}
+import org.apache.spark.storage.BlockObjectWriter
+import org.apache.spark.util.collection.PartitionedSerializedPairBuffer._
+
+/**
+ * Append-only buffer of key-value pairs, each with a corresponding
partition ID, that serializes
+ * its records upon insert and stores them as raw bytes.
+ *
+ * We use two data-structures to store the contents. The serialized
records are stored in a
+ * ChainedBuffer that can expand gracefully as records are added. This
buffer is accompanied by a
+ * metadata buffer that stores pointers into the data buffer as well as
the partition ID of each
+ * record. Each entry in the metadata buffer takes up a fixed amount of
space.
+ *
+ * Sorting the collection means swapping entries in the metadata buffer -
the record buffer need not
+ * be modified at all. Storing the partition IDs in the metadata buffer
means that comparisons can
+ * happen without following any pointers, which should minimize cache
misses.
+ *
+ * Currently, only sorting by partition is supported.
+ *
+ * @param metaInitialRecords The initial number of entries in the metadata
buffer.
+ * @param kvBlockSize The size of each byte buffer in the ChainedBuffer
used to store the records.
+ * @param serializerInstance the serializer used for serializing inserted
records.
+ */
+private[spark] class PartitionedSerializedPairBuffer[K, V](
+ metaInitialRecords: Int,
+ kvBlockSize: Int,
+ serializerInstance: SerializerInstance =
SparkEnv.get.serializer.newInstance)
+ extends WritablePartitionedPairCollection[K, V] {
+
+ if (serializerInstance.isInstanceOf[JavaSerializerInstance]) {
+ throw new IllegalArgumentException("PartitionedSerializedPairBuffer
does not support" +
+ " Java-serialized objects.")
+ }
+
+ private var metaBuffer = IntBuffer.allocate(metaInitialRecords * NMETA)
+
+ private val kvBuffer: ChainedBuffer = new ChainedBuffer(kvBlockSize)
+ private val kvOutputStream = new ChainedBufferOutputStream(kvBuffer)
+ private val kvSerializationStream =
serializerInstance.serializeStream(kvOutputStream)
+
+ def insert(partition: Int, key: K, value: V): Unit = {
+ if (metaBuffer.position == metaBuffer.capacity) {
+ growMetaBuffer()
+ }
+
+ val keyStart = kvBuffer.size
+ if (keyStart < 0) {
+ throw new Exception(s"Can't grow buffer beyond ${1 << 31} bytes")
+ }
+ kvSerializationStream.writeObject[Any](key)
+ kvSerializationStream.flush()
+ val valueStart = kvBuffer.size
+ kvSerializationStream.writeObject[Any](value)
+ kvSerializationStream.flush()
+ val valueEnd = kvBuffer.size
+
+ metaBuffer.put(keyStart)
+ metaBuffer.put(valueStart)
+ metaBuffer.put(valueEnd)
+ metaBuffer.put(partition)
+ }
+
+ /** Double the size of the array because we've reached capacity */
+ private def growMetaBuffer(): Unit = {
+ if (metaBuffer.capacity * 4 >= (1 << 30)) {
+ // Doubling the capacity would create an array bigger than
Int.MaxValue, so don't
+ throw new Exception(
+ s"Can't grow buffer beyond ${(1 << 30) / (NMETA * 4)} elements")
+ }
+ val newMetaBuffer = IntBuffer.allocate(metaBuffer.capacity * 2)
+ newMetaBuffer.put(metaBuffer.array)
+ metaBuffer = newMetaBuffer
+ }
+
+ /** Iterate through the data in a given order. For this class this is
not really destructive. */
+ override def partitionedDestructiveSortedIterator(keyComparator:
Comparator[K])
+ : Iterator[((Int, K), V)] = {
+ sort(keyComparator)
+ val is = orderedInputStream
+ val deserStream = serializerInstance.deserializeStream(is)
+ new Iterator[((Int, K), V)] {
+ var metaBufferPos = 0
+ def hasNext: Boolean = metaBufferPos < metaBuffer.position
+ def next(): ((Int, K), V) = {
+ val key = deserStream.readObject[Any]().asInstanceOf[K]
+ val value = deserStream.readObject[Any]().asInstanceOf[V]
+ val partition = metaBuffer.get(metaBufferPos + PARTITION)
+ metaBufferPos += NMETA
+ ((partition, key), value)
+ }
+ }
+ }
+
+ override def estimateSize: Long = metaBuffer.capacity * 4 +
kvBuffer.capacity
+
+ override def destructiveSortedWritablePartitionedIterator(keyComparator:
Comparator[K])
+ : WritablePartitionedIterator = {
+ sort(keyComparator)
+ writablePartitionedIterator
+ }
+
+ override def writablePartitionedIterator(): WritablePartitionedIterator
= {
+ new WritablePartitionedIterator {
+ // current position in the meta buffer in ints
+ var pos = 0
+
+ def writeNext(writer: BlockObjectWriter): Unit = {
+ val keyStart = metaBuffer.get(pos + KEY_START)
+ val valueEnd = metaBuffer.get(pos + VAL_END)
+ pos += NMETA
+ kvBuffer.read(keyStart, writer, valueEnd - keyStart)
+ }
+ def nextPartition(): Int = metaBuffer.get(pos + PARTITION)
+ def hasNext(): Boolean = pos < metaBuffer.position
+ }
+ }
+
+ // Visible for testing
+ def orderedInputStream: OrderedInputStream = {
+ new OrderedInputStream(metaBuffer, kvBuffer)
+ }
+
+ private def sort(keyComparator: Comparator[K]): Unit = {
+ val comparator = if (keyComparator == null) {
+ new Comparator[Int]() {
+ def compare(partition1: Int, partition2: Int): Int = {
+ partition1 - partition2
+ }
+ }
+ } else {
+ throw new UnsupportedOperationException()
+ }
+
+ val sorter = new Sorter(new SerializedSortDataFormat)
+ sorter.sort(metaBuffer, 0, metaBuffer.position / NMETA, comparator)
+ }
+}
+
+private[spark] class OrderedInputStream(metaBuffer: IntBuffer, kvBuffer:
ChainedBuffer)
+ extends InputStream {
+
+ private var metaBufferPos = 0
+ private var kvBufferPos =
+ if (metaBuffer.position > 0) metaBuffer.get(metaBufferPos + KEY_START)
else 0
+
+ override def read(bytes: Array[Byte]): Int = read(bytes, 0, bytes.length)
+
+ override def read(bytes: Array[Byte], offs: Int, len: Int): Int = {
+ if (metaBufferPos >= metaBuffer.position) {
+ return -1
+ }
+ val bytesRemainingInRecord = metaBuffer.get(metaBufferPos + VAL_END) -
kvBufferPos
+ val toRead = math.min(bytesRemainingInRecord, len)
+ kvBuffer.read(kvBufferPos, bytes, offs, toRead)
+ if (toRead == bytesRemainingInRecord) {
+ metaBufferPos += NMETA
+ if (metaBufferPos < metaBuffer.position) {
+ kvBufferPos = metaBuffer.get(metaBufferPos + KEY_START)
+ }
+ } else {
+ kvBufferPos += toRead
+ }
+ toRead
+ }
+
+ override def read(): Int = {
+ throw new UnsupportedOperationException()
+ }
+}
+
+private[spark] class SerializedSortDataFormat extends SortDataFormat[Int,
IntBuffer] {
+
+ private val META_BUFFER_TMP = new Array[Int](NMETA)
+
+ /** Return the sort key for the element at the given index. */
+ override protected def getKey(metaBuffer: IntBuffer, pos: Int): Int = {
+ metaBuffer.get(pos * NMETA + PARTITION)
+ }
+
+ /** Swap two elements. */
+ override def swap(metaBuffer: IntBuffer, pos0: Int, pos1: Int): Unit = {
+ val iOff = pos0 * NMETA
+ val jOff = pos1 * NMETA
+ System.arraycopy(metaBuffer.array, iOff, META_BUFFER_TMP, 0, NMETA)
+ System.arraycopy(metaBuffer.array, jOff, metaBuffer.array, iOff, NMETA)
+ System.arraycopy(META_BUFFER_TMP, 0, metaBuffer.array, jOff, NMETA)
+ }
+
+ /** Copy a single element from src(srcPos) to dst(dstPos). */
+ override def copyElement(
+ src: IntBuffer,
+ srcPos: Int,
+ dst: IntBuffer,
+ dstPos: Int): Unit = {
+ val srcOff = srcPos * NMETA
+ val dstOff = dstPos * NMETA
+ System.arraycopy(src.array, srcOff, dst.array, dstOff, NMETA)
+ }
+
+ /**
+ * Copy a range of elements starting at src(srcPos) to dst, starting at
dstPos.
+ * Overlapping ranges are allowed.
+ */
+ override def copyRange(
+ src: IntBuffer,
+ srcPos: Int,
+ dst: IntBuffer,
+ dstPos: Int,
+ length: Int): Unit = {
+ val srcOff = srcPos * NMETA
+ val dstOff = dstPos * NMETA
+ System.arraycopy(src.array, srcOff, dst.array, dstOff, NMETA * length)
+ }
+
+ /**
+ * Allocates a Buffer that can hold up to 'length' elements.
+ * All elements of the buffer should be considered invalid until data is
explicitly copied in.
+ */
+ override def allocate(length: Int): IntBuffer = {
+ IntBuffer.allocate(length * NMETA)
+ }
+}
+
+private[spark] object PartitionedSerializedPairBuffer {
+ val KEY_START = 0
+ val VAL_START = 1
+ val VAL_END = 2
+ val PARTITION = 3
+ val NMETA = 4 // num meta ints
--- End diff --
Might be nicer to call this `RECORD_SIZE` or something more clear. Also for
clarity maybe you can just do:
```
val RECORD_SIZE = Seq(KEY_START, VAL_START, VAL_END, PARTITION).size
```
---
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 [email protected] or file a JIRA ticket
with INFRA.
---
---------------------------------------------------------------------
To unsubscribe, e-mail: [email protected]
For additional commands, e-mail: [email protected]
