[INTERNAL] Tungsten Cache code review and refactoring #5
There are no files selected for viewing
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
| Original file line number | Diff line number | Diff line change |
|---|---|---|
| @@ -0,0 +1,195 @@ | ||
| /* | ||
| * 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.sql.execution | ||
|
|
||
| import java.io.{ByteArrayInputStream, ByteArrayOutputStream} | ||
|
|
||
| import com.google.common.io.ByteStreams | ||
|
|
||
| import org.apache.spark.storage.StorageLevel | ||
| import org.apache.spark.unsafe.Platform | ||
| import org.apache.spark.unsafe.array.ByteArrayMethods | ||
| import org.apache.spark.unsafe.memory.{TaskMemoryManager, MemoryBlock} | ||
| import org.apache.spark.{TaskContext, SparkEnv} | ||
| import org.apache.spark.annotation.Experimental | ||
| import org.apache.spark.io.CompressionCodec | ||
| import org.apache.spark.rdd.RDD | ||
| import org.apache.spark.sql.catalyst.InternalRow | ||
| import org.apache.spark.sql.catalyst.expressions.UnsafeRow | ||
|
|
||
| /** | ||
| * ::Experimental:: | ||
| * | ||
| * @since 1.6.0 | ||
| */ | ||
| @Experimental | ||
| object TungstenCache { | ||
| def cache( | ||
| child: SparkPlan, | ||
| compressionType: Option[String] = None, | ||
| blockSize: Int = 4 * 1000 * 1000): RDD[InternalRow] = { | ||
| val END_OF_BLOCK: Int = -1 | ||
| val numFields = child.schema.length | ||
| val childRDD: RDD[UnsafeRow] = { | ||
| if (child.outputsUnsafeRows) { | ||
| child.execute().asInstanceOf[RDD[UnsafeRow]] | ||
| } else { | ||
| ConvertToUnsafe(child).execute().asInstanceOf[RDD[UnsafeRow]] | ||
| } | ||
| } | ||
| val cachedRDD: RDD[MemoryBlock] = childRDD.mapPartitions { rowIterator => | ||
| // Note: this buffering is used to let us hold onto rows when rolling across block boundaries. | ||
| val bufferedRowIterator = rowIterator.buffered | ||
| val taskMemoryManager = TaskContext.get().taskMemoryManager() | ||
| val compressionCodec: Option[CompressionCodec] = | ||
| compressionType.map { t => CompressionCodec.createCodec(SparkEnv.get.conf, t) } | ||
| new Iterator[MemoryBlock] { | ||
| // NOTE: This assumes that size of every row < blockSize - 4 | ||
| // TODO(josh): we'll have to figure out how to support large rows. | ||
| def next(): MemoryBlock = { | ||
| // Packs rows into a `blockSize` bytes contiguous block of memory, starting a new block | ||
| // whenever the current fills up. | ||
| // Each row is laid out in memory as [rowSize (int)|rowData (rowSize bytes)]. | ||
| // The end of the block is marked by a special rowSize, END_OF_BLOCK (-1). | ||
| val block = taskMemoryManager.allocateUnchecked(blockSize) | ||
|
|
||
| var currOffset = 0 | ||
| while (bufferedRowIterator.hasNext && | ||
| currOffset + 4 + bufferedRowIterator.head.getSizeInBytes < blockSize - 4) { | ||
| val currRow = bufferedRowIterator.head | ||
| Platform.putInt( | ||
| block.getBaseObject, block.getBaseOffset + currOffset, currRow.getSizeInBytes) | ||
| currRow.writeToMemory(block.getBaseObject, block.getBaseOffset + currOffset + 4) | ||
| bufferedRowIterator.next() | ||
| currOffset += 4 + currRow.getSizeInBytes | ||
| } | ||
| Platform.putInt(block.getBaseObject, block.getBaseOffset + currOffset, END_OF_BLOCK) | ||
|
|
||
| compressionCodec match { | ||
| case Some(codec) => compressBlock(block, codec, TaskContext.get().taskMemoryManager()) | ||
| case None => block | ||
| } | ||
| } | ||
|
|
||
| def hasNext: Boolean = bufferedRowIterator.hasNext | ||
| } | ||
| } | ||
| .setName(compressionType.getOrElse("") + "_" + child.nodeName) | ||
| .persist(StorageLevel.MEMORY_ONLY) | ||
| // TODO(josh): is this the right name? | ||
|
|
||
| cachedRDD.mapPartitions { blockIterator => | ||
| val compressionCodec: Option[CompressionCodec] = | ||
| compressionType.map { t => CompressionCodec.createCodec(SparkEnv.get.conf, t) } | ||
| blockIterator.flatMap { rawBlock => | ||
| val block: MemoryBlock = compressionCodec match { | ||
| case Some(codec) => decompressBlock(rawBlock, blockSize, codec) | ||
| case None => rawBlock | ||
| } | ||
|
|
||
| new Iterator[UnsafeRow] { | ||
| private[this] val unsafeRow = new UnsafeRow() | ||
| private[this] var currOffset: Long = 0 | ||
| private[this] var _nextRowSize = getNextRowSize() | ||
| private[this] def getNextRowSize(): Int = { | ||
| if (currOffset >= block.size()) { | ||
| -1 | ||
| } else { | ||
| Platform.getInt(block.getBaseObject, block.getBaseOffset + currOffset) | ||
| } | ||
| } | ||
| override def hasNext: Boolean = _nextRowSize != END_OF_BLOCK | ||
| override def next(): UnsafeRow = { | ||
| assert(_nextRowSize > 0) | ||
| currOffset += 4 | ||
|
There was a problem hiding this comment. Should document this is advancing offset over memory for "_nextRowSize" |
||
| unsafeRow.pointTo( | ||
| block.getBaseObject, block.getBaseOffset + currOffset, numFields, _nextRowSize) | ||
| currOffset += _nextRowSize | ||
| _nextRowSize = getNextRowSize() | ||
| unsafeRow | ||
| } | ||
| } | ||
| } | ||
| } | ||
| } | ||
|
|
||
| private def compressBlock( | ||
| memoryBlock: MemoryBlock, | ||
| compressionCodec: CompressionCodec, | ||
| taskMemoryManager: TaskMemoryManager): MemoryBlock = { | ||
| // Compress the block using an on-heap byte array | ||
| val compressedBlockArray: Array[Byte] = { | ||
|
There was a problem hiding this comment. Can we also reuse this |
||
| val blockArray = new Array[Byte](memoryBlock.size().toInt) | ||
| Platform.copyMemory( | ||
| memoryBlock.getBaseObject, | ||
| memoryBlock.getBaseOffset, | ||
| blockArray, | ||
| Platform.BYTE_ARRAY_OFFSET, | ||
| memoryBlock.size()) | ||
| val baos = new ByteArrayOutputStream(memoryBlock.size().toInt) | ||
| val compressedBaos = compressionCodec.compressedOutputStream(baos) | ||
| compressedBaos.write(blockArray) | ||
| compressedBaos.flush() | ||
| compressedBaos.close() | ||
| baos.toByteArray | ||
| } | ||
|
|
||
| // Allocate a new block with compressed byte array padded to word boundary | ||
| val totalRecordSize = compressedBlockArray.length + 4 // data + int to store size of padding | ||
| val nearestWordBoundary = | ||
| ByteArrayMethods.roundNumberOfBytesToNearestWord(totalRecordSize) | ||
| val padding = nearestWordBoundary - totalRecordSize | ||
| val compressedBlock = taskMemoryManager.allocateUnchecked(totalRecordSize + padding) | ||
| Platform.putInt( | ||
| compressedBlock.getBaseObject, | ||
| compressedBlock.getBaseOffset, | ||
| padding) | ||
| Platform.copyMemory( | ||
| compressedBlockArray, | ||
| Platform.BYTE_ARRAY_OFFSET, | ||
| compressedBlock.getBaseObject, | ||
| compressedBlock.getBaseOffset + 4, | ||
| compressedBlockArray.length) | ||
| taskMemoryManager.freeUnchecked(memoryBlock) | ||
| compressedBlock | ||
| } | ||
|
|
||
| private def decompressBlock( | ||
| compressedMemoryBlock: MemoryBlock, | ||
| decompressedSize: Int, | ||
| compressionCodec: CompressionCodec): MemoryBlock = { | ||
| // Copy compressed block (excluding padding) to on-heap byte array | ||
| val padding = | ||
| Platform.getInt(compressedMemoryBlock.getBaseObject, compressedMemoryBlock.getBaseOffset) | ||
| val compressedBlockArray = new Array[Byte](compressedMemoryBlock.size().toInt - padding) | ||
|
There was a problem hiding this comment. I think @JoshRosen mentioned earlier that we don't need to reinitialize here and can reuse this |
||
| Platform.copyMemory( | ||
| compressedMemoryBlock.getBaseObject, | ||
| compressedMemoryBlock.getBaseOffset + 4, | ||
| compressedBlockArray, | ||
| Platform.BYTE_ARRAY_OFFSET, | ||
| compressedMemoryBlock.size() - padding) | ||
|
|
||
| // Decompress into MemoryBlock backed by on-heap byte array | ||
| val decompressionStream = | ||
| compressionCodec.compressedInputStream(new ByteArrayInputStream(compressedBlockArray)) | ||
| val decompressedBlock = new Array[Byte](decompressedSize) | ||
| ByteStreams.readFully(decompressionStream, decompressedBlock) | ||
| decompressionStream.close() | ||
| MemoryBlock.fromByteArray(decompressedBlock) | ||
| } | ||
| } | ||
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Oops, something went wrong.
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
I'm wondering what we actually gain by using the TaskMemoryManager here. The TMM would allow us to cache off-heap but it looks like we don't / won't currently do that if we use compression. For the initial version of this task, I'd like to focus on making the basic caching work really well and defer the question of how to integrate offheap caching with block manager to a followup PR.
There was a problem hiding this comment.
Choose a reason for hiding this comment
The reason will be displayed to describe this comment to others. Learn more.
Actually, I take that back: I see that we do actually end up allocating off-heap for the compressed data as well. We still have complexities in how to ensure cleanup when things are evicted, but that's out of scope for perf. benchmarking studies for now.