欢迎关注微信公众号“Tim在路上”
Unsafe Shuffle的实现在一定程度上是Tungsten内存管理优化的的主要应用场景。其实现过程实际上和SortShuffleWriter是类似的,但是其中维护和执行的数据结构是不一样的。
UnsafeShuffleWriter 源码解析
@Override
public void write(scala.collection.Iterator<Product2<K, V>> records) throws IOException {
// Keep track of success so we know if we encountered an exception
// We do this rather than a standard try/catch/re-throw to handle
// generic throwables.
// [1] 使用success记录write是否成功,判断是write阶段的异常还是clean阶段
boolean success = false;
try {
// [2] 遍历所有的数据插入ShuffleExternalSorter
while (records.hasNext()) {
insertRecordIntoSorter(records.next());
}
// [3] close排序器使所有数据写出到磁盘,并将多个溢写文件合并到一起
closeAndWriteOutput();
success = true;
} finally {
if (sorter != null) {
try {
// [4] 清除并释放资源
sorter.cleanupResources();
} catch (Exception e) {
// Only throw this error if we won't be masking another
// error.
if (success) {
throw e;
} else {
logger.error("In addition to a failure during writing, we failed during " +
"cleanup.", e);
}
}
}
}
}
从上面的代码可以看出,UnsafeShuffleWriter的write过程如下:
- [1] 使用success记录write是否成功,判断是write阶段的异常还是clean阶段
- [2] 遍历所有的数据插入ShuffleExternalSorter
- [3] close排序器使所有数据写出到磁盘,并将多个溢写文件合并到一起
- [4] 清除并释放资源
// open()方法是在初始化UnsafeShuffleWriter调用的,其中会创建sorter, 并创建一个字节输出流,同时封装序列化流
private void open() throws SparkException {
assert (sorter == null);
sorter = new ShuffleExternalSorter(
memoryManager,
blockManager,
taskContext,
initialSortBufferSize,
partitioner.numPartitions(),
sparkConf,
writeMetrics);
// MyByteArrayOutputStream类是ByteArrayOutputStream的简单封装,只是将内部byte[]数组暴露出来】
//【DEFAULT_INITIAL_SER_BUFFER_SIZE常量值是1024 * 1024,即缓冲区初始1MB大】
serBuffer = new MyByteArrayOutputStream(DEFAULT_INITIAL_SER_BUFFER_SIZE);
serOutputStream = serializer.serializeStream(serBuffer);
}
void insertRecordIntoSorter(Product2<K, V> record) throws IOException {
assert(sorter != null);
// [1] 获取record的key和partitionId
final K key = record._1();
final int partitionId = partitioner.getPartition(key);
// [2] 将record序列化为二进制,并写的字节数组输出流serBuffer中
serBuffer.reset();
serOutputStream.writeKey(key, OBJECT_CLASS_TAG);
serOutputStream.writeValue(record._2(), OBJECT_CLASS_TAG);
serOutputStream.flush();
final int serializedRecordSize = serBuffer.size();
assert (serializedRecordSize > 0);
// [3] 将其插入到ShuffleExternalSorter中
sorter.insertRecord(
serBuffer.getBuf(), Platform.BYTE_ARRAY_OFFSET, serializedRecordSize, partitionId);
}
这一步是将record插入前的准备,现将序列化为二进制存储在内存中。
- [1] 获取record的key和partitionId
- [2] 将record序列化为二进制,并写的字节数组输出流serBuffer中
- [3] 将序列化的二进制数组,分区id, length 作为参数插入到ShuffleExternalSorter中
那么数据在ShuffleExternalSorter中写入过程是怎么样呢?
public void insertRecord(Object recordBase, long recordOffset, int length, int partitionId)
throws IOException {
// [1] 判断inMemSorter中的记录是否到达了溢写阈值(默认是整数最大值),如果满足就先进行spill
// for tests
assert(inMemSorter != null);
if (inMemSorter.numRecords() >= numElementsForSpillThreshold) {
logger.info("Spilling data because number of spilledRecords crossed the threshold " +
numElementsForSpillThreshold);
spill();
}
// [2] 检查inMemSorter是否有额外的空间插入,如果可以获取就扩充空间,否则进行溢写
growPointerArrayIfNecessary();
final int uaoSize = UnsafeAlignedOffset.getUaoSize();
// Need 4 or 8 bytes to store the record length.
final int required = length + uaoSize;
// [3] 判断当前内存空间currentPage是否有足够的内存,如果不够就申请,申请不下来就需要spill
acquireNewPageIfNecessary(required);
assert(currentPage != null);
// [4] 获取currentPage的base Object和recordAddress
final Object base = currentPage.getBaseObject();
final long recordAddress = taskMemoryManager.encodePageNumberAndOffset(currentPage, pageCursor);
// [5] 根据base, pageCursor, 先向当前内存空间写长度值,并移动指针
UnsafeAlignedOffset.putSize(base, pageCursor, length);
pageCursor += uaoSize;
// [6] 再写序列化之后的数据, 并移动指指
Platform.copyMemory(recordBase, recordOffset, base, pageCursor, length);
pageCursor += length;
// [7] 将recordAddress和partitionId插入inMemSorter进行排序
inMemSorter.insertRecord(recordAddress, partitionId);
}
从上面分析,数据插入ShuffleExternalSorter总共需要7步:
- [1] 判断inMemSorter中的记录是否到达了溢写阈值(默认是整数最大值),如果满足就先进行spill
- [2] 检查inMemSorter是否有额外的空间插入,如果可以获取就扩充空间,否则进行溢写
- [3] 判断当前内存空间currentPage是否有足够的内存,如果不够就申请,申请不下来就需要spill
- [4] 获取currentPage的base Object和recordAddress
- [5] 先向当前内存空间写长度值,并移动指针
- [6] 再写序列化之后的数据, 并移动指指
- [7] 将recordAddress和partitionId插入inMemSorter进行排序
从上面的介绍可以看出在整个插入过程中,主要涉及ShuffleExternalSorter
和 ShuffleInMemorySorter
两个数据结构。我们来简单看了ShuffleExternalSorter
类。
final class ShuffleExternalSorter extends MemoryConsumer implements ShuffleChecksumSupport {
private final int numPartitions;
private final TaskMemoryManager taskMemoryManager;
private final BlockManager blockManager;
private final TaskContext taskContext;
private final ShuffleWriteMetricsReporter writeMetrics;
private final LinkedList<MemoryBlock> allocatedPages = new LinkedList<>();
private final LinkedList<SpillInfo> spills = new LinkedList<>();
/** Peak memory used by this sorter so far, in bytes. **/
private long peakMemoryUsedBytes;
// These variables are reset after spilling:
@Nullable private ShuffleInMemorySorter inMemSorter;
@Nullable private MemoryBlock currentPage = null;
private long pageCursor = -1;
...
}
可见每个ShuffleExternalSorter
中封装着ShuffleInMemorySorter类。同时封装allocatedPages
、spills和currentPage。也就是说ShuffleExternalSorter
使用MemoryBlock存储数据,每条记录包括长度信息和K-V Pair。
另外在 ShuffleInMemorySorter
中,通过LongArray
来存储数据,并实现了SortComparator
排序方法。其中LongArray
存储的record的位置信息,主要有分区id, page id 和offset。
ShuffleExternalSorter | 使用MemoryBlock存储数据,每条记录包括长度信息和K-V Pair |
---|---|
ShuffleInMemorySorter | 使用long数组存储每条记录对应的位置信息(page number + offset),以及其对应的PartitionId,共8 bytes |
从上面的关于ShuffleExternalSorter
和ShuffleInMemorySorter
可以看出,这里其实质上是使用Tungsten实现了类似于BytesToBytesMap的数据结构,不过将其数组部分LongArray用ShuffleInMemorySorter
进行了封装,其余拆分为ShuffleExternalSorter
。
ShuffleExternalSorter
将数据写入了当前的内存空间,将数据的recordAddress和partitionId写入了ShuffleInMemorySorter
,那么其具体是如何实现排序和数据的溢写的?
private void writeSortedFile(boolean isLastFile) {
// [1] 将inMemSorter的数据排序,并返回ShuffleSorterIterator
// This call performs the actual sort.
final ShuffleInMemorySorter.ShuffleSorterIterator sortedRecords =
inMemSorter.getSortedIterator();
// If there are no sorted records, so we don't need to create an empty spill file.
if (!sortedRecords.hasNext()) {
return;
}
final ShuffleWriteMetricsReporter writeMetricsToUse;
...
// [2] 创建缓存数据writeBuffer数组,为了避免DiskBlockObjectWriter的低效的写
// Small writes to DiskBlockObjectWriter will be fairly inefficient. Since there doesn't seem to
// be an API to directly transfer bytes from managed memory to the disk writer, we buffer
// data through a byte array. This array does not need to be large enough to hold a single
// record;
final byte[] writeBuffer = new byte[diskWriteBufferSize];
// Because this output will be read during shuffle, its compression codec must be controlled by
// spark.shuffle.compress instead of spark.shuffle.spill.compress, so we need to use
// createTempShuffleBlock here; see SPARK-3426 for more details.
final Tuple2<TempShuffleBlockId, File> spilledFileInfo =
blockManager.diskBlockManager().createTempShuffleBlock();
final File file = spilledFileInfo._2();
final TempShuffleBlockId blockId = spilledFileInfo._1();
final SpillInfo spillInfo = new SpillInfo(numPartitions, file, blockId);
// Unfortunately, we need a serializer instance in order to construct a DiskBlockObjectWriter.
// Our write path doesn't actually use this serializer (since we end up calling the `write()`
// OutputStream methods), but DiskBlockObjectWriter still calls some methods on it. To work
// around this, we pass a dummy no-op serializer.
final SerializerInstance ser = DummySerializerInstance.INSTANCE;
int currentPartition = -1;
final FileSegment committedSegment;
try (DiskBlockObjectWriter writer =
blockManager.getDiskWriter(blockId, file, ser, fileBufferSizeBytes, writeMetricsToUse)) {
final int uaoSize = UnsafeAlignedOffset.getUaoSize();
// [3] 按分区遍历已经排好序的指针数据, 并未每个分区提交一个FileSegment,并记录分区的大小
while (sortedRecords.hasNext()) {
sortedRecords.loadNext();
final int partition = sortedRecords.packedRecordPointer.getPartitionId();
assert (partition >= currentPartition);
if (partition != currentPartition) {
// Switch to the new partition
if (currentPartition != -1) {
final FileSegment fileSegment = writer.commitAndGet();
spillInfo.partitionLengths[currentPartition] = fileSegment.length();
}
currentPartition = partition;
if (partitionChecksums.length > 0) {
writer.setChecksum(partitionChecksums[currentPartition]);
}
}
// [4] 取得数据的指针,再通过指针取得页号与偏移量
final long recordPointer = sortedRecords.packedRecordPointer.getRecordPointer();
final Object recordPage = taskMemoryManager.getPage(recordPointer);
final long recordOffsetInPage = taskMemoryManager.getOffsetInPage(recordPointer);
// [5] 取得数据前面存储的长度,然后让指针跳过它
int dataRemaining = UnsafeAlignedOffset.getSize(recordPage, recordOffsetInPage);
long recordReadPosition = recordOffsetInPage + uaoSize; // skip over record length
// [6] 数据拷贝到上面创建的缓存中,通过缓存转到DiskBlockObjectWriter, 并写入数据,移动指针
while (dataRemaining > 0) {
final int toTransfer = Math.min(diskWriteBufferSize, dataRemaining);
Platform.copyMemory(
recordPage, recordReadPosition, writeBuffer, Platform.BYTE_ARRAY_OFFSET, toTransfer);
writer.write(writeBuffer, 0, toTransfer);
recordReadPosition += toTransfer;
dataRemaining -= toTransfer;
}
writer.recordWritten();
}
committedSegment = writer.commitAndGet();
}
// If `writeSortedFile()` was called from `closeAndGetSpills()` and no records were inserted,
// then the file might be empty. Note that it might be better to avoid calling
// writeSortedFile() in that case.
if (currentPartition != -1) {
spillInfo.partitionLengths[currentPartition] = committedSegment.length();
spills.add(spillInfo);
}
if (!isLastFile) { // i.e. this is a spill file
writeMetrics.incRecordsWritten(
((ShuffleWriteMetrics)writeMetricsToUse).recordsWritten());
taskContext.taskMetrics().incDiskBytesSpilled(
((ShuffleWriteMetrics)writeMetricsToUse).bytesWritten());
}
}
溢写排序文件总的来说分为两步:
首先是通过ShuffleInMemorySorter排序,获取对应分区的FileSegment和长度。写文件或溢写前根据数据的PartitionId信息,使用TimSort对ShuffleInMemorySorter的long数组排序,排序的结果为,PartitionId相同的聚集在一起,且PartitionId较小的排在前面,然后按分区写出FileSegment, 并记录每个分区的长度。
其次是基于排好序的指针执行数据的溢写操作。依次读取ShuffleInMemorySorter中long数组的元素,再根据page number和offset信息去ShuffleExternalSorter中读取K-V Pair写入文件, 溢写前先写入writeBuffer,然后在写入DiskBlockObjectWriter中。
具体的步骤见下:
- [1] 将inMemSorter的数据排序,并返回ShuffleSorterIterator
- [2] 创建缓存数据writeBuffer数组,为了避免DiskBlockObjectWriter的低效的写
- [3] 按分区遍历已经排好序的指针数据, 并未每个分区提交一个FileSegment,并记录分区的大小
- [4] 取得数据的指针,再通过指针取得页号与偏移量
- [5] 取得数据前面存储的长度,然后让指针跳过它
- [6] 数据拷贝到上面创建的缓存writeBuffer中,通过缓存转到DiskBlockObjectWriter, 并写入数据,移动指针
最后我们看下,UnsafeShuffleWriter是如何将最后溢写的文件进行合并的?
// UnsafeShuffleWriter
void closeAndWriteOutput() throws IOException {
assert(sorter != null);
updatePeakMemoryUsed();
serBuffer = null;
serOutputStream = null;
// [1] 关闭排序器,并将排序器中的数据全部溢写到磁盘,返回SpillInfo数组
final SpillInfo[] spills = sorter.closeAndGetSpills();
try {
// [2] 将多个溢出文件合并在一起,根据溢出次数和 IO 压缩编解码器选择最快的合并策略
partitionLengths = mergeSpills(spills);
} finally {
sorter = null;
for (SpillInfo spill : spills) {
if (spill.file.exists() && !spill.file.delete()) {
logger.error("Error while deleting spill file {}", spill.file.getPath());
}
}
}
mapStatus = MapStatus$.MODULE$.apply(
blockManager.shuffleServerId(), partitionLengths, mapId);
}
private long[] mergeSpills(SpillInfo[] spills) throws IOException {
long[] partitionLengths;
// [1] 如果根本没有溢写文件,写一个空文件
if (spills.length == 0) {
final ShuffleMapOutputWriter mapWriter = shuffleExecutorComponents
.createMapOutputWriter(shuffleId, mapId, partitioner.numPartitions());
return mapWriter.commitAllPartitions(
ShuffleChecksumHelper.EMPTY_CHECKSUM_VALUE).getPartitionLengths();
// [2] 如果只有一个溢写文件,就直接将它写入输出文件中
} else if (spills.length == 1) {
// [2.1] 创建单个file的map output writer
Optional<SingleSpillShuffleMapOutputWriter> maybeSingleFileWriter =
shuffleExecutorComponents.createSingleFileMapOutputWriter(shuffleId, mapId);
if (maybeSingleFileWriter.isPresent()) {
// Here, we don't need to perform any metrics updates because the bytes written to this
// output file would have already been counted as shuffle bytes written.
partitionLengths = spills[0].partitionLengths;
logger.debug("Merge shuffle spills for mapId {} with length {}", mapId,
partitionLengths.length);
maybeSingleFileWriter.get()
.transferMapSpillFile(spills[0].file, partitionLengths, sorter.getChecksums());
} else {
partitionLengths = mergeSpillsUsingStandardWriter(spills);
}
// [3] 如果有多个溢写文件,如果启用并支持快速合并,并且启用了transferTo机制,还没有加密, 就使用NIO zero-copy来合并到输出文件, 不启用transferTo或不支持快速合并,就使用压缩的BIO FileStream来合并到输出文件
} else {
partitionLengths = mergeSpillsUsingStandardWriter(spills);
}
return partitionLengths;
}
多个spills的合并的具体的实现在mergeSpillsWithFileStream
方法中,为了减少篇幅的冗长这里就不再展开了。
溢写的文件进行合并,有如下几个步骤:
[1] 关闭排序器,并将排序器中的数据全部溢写到磁盘,返回SpillInfo数组
-
[2] 将多个溢出文件合并在一起,根据溢出次数和 IO 压缩编解码器选择最快的合并策略
- [2.1] 如果根本没有溢写文件,写一个空文件 - [2.2] 如果只有一个溢写文件,就直接将它写入输出文件中 - [2.3] 如果有多个溢写文件,如果启用并支持快速合并,并且启用了transferTo机制,还没有加密, 就使用NIO zero-copy来合并到输出文件, 不启用transferTo或不支持快速合并,就使用压缩的BIO FileStream来合并到输出文件
至此,UnsafeShuffleWriter的实现就介绍完了。
下面我们谈下UnsafeShuffleWriter的优势:
- ShuffleExternalSorter使用UnSafe API操作序列化数据,而不是Java对象,减少了内存占用及因此导致的GC耗时,这个优化需要Serializer支持relocation。 ShuffleExternalSorter存原始数据,ShuffleInMemorySorter使用压缩指针存储元数据,每条记录仅占8 bytes,并且排序时不需要处理原始数据,效率高。
- 溢写 & 合并这一步操作的是同一Partition的数据,因为使用UnSafe API直接操作序列化数据,合并时不需要反序列化数据。
- 溢写 & 合并可以使用fastMerge提升效率(调用NIO的transferTo方法),设置spark.shuffle.unsafe.fastMergeEnabled为true,并且如果使用了压缩,需要压缩算法支持SerializedStreams的连接。
- 排序时并非将数据进行排序,而是将数据的地址指针进行排序
总结,UnsafeShuffleWriter是Tungsten最重要的应用,他的实现原理类似于SortShuffleWriter, 但是基于UnSafe API使用了定义的ShuffleExternalSorter和ShuffleInMemorySorter来存储和维护数据。
其整体流程为,所有的数据在插入前都需要序列化为二进制数组,然后再将其插入到数据结构ShuffleExternalSorter中。在ShuffleExternalSorter定义了ShuffleInMemorySorter主要用于存储数据的partitionId和recordAddress, 另外定义了MemoryBlock页空间数组。
在ShuffleExternalSorter的insertRecord时会先,判断ShuffleInMemorySorter和当前内存空间是否足够新数据的插入,不够需要申请,申请失败则需要spill。
插入数据时会先写入占用内存空间的长度,再写入数据值,最后将recordAddress和partitionId插入ShuffleInMemorySorter中。在进行spill时会将ShuffleInMemorySorter中的数据进行排序,并按照分区生成FileSegment并统计分区的大小,然后遍历指针数组根据地址将对应的数据进行写出。在进行合并时可以直接使用UnSafe API直接操作序列化数据,返回汇总的文件。
通过UnsafeShuffleWriter只会产生两个文件,一个分区的数据文件,一个索引文件。整个UnsafeShuffleWriter过程只会产生2 * M 个中间文件。
今天就先到这里,通过上面的介绍,我们也留下些面试题:
- 为什么UnsafeShuffleWriter无法支持无法支持map端的aggregation?
- 为什么UnsafeShuffleWriter分区数的最大值为 (1 << 24) ?
- ShuffleExternalSorter实现是基于JVM的吗?以及其在排序上有什么优化?
欢迎关注微信公众号“Tim在路上”