Spark Shuffle(三)Executor是如何fetch shuffle的数据文件(转载)

1. 前言

前面的博客中讨论了Executor, Driver之间如何汇报Executor生成的Shuffle的数据文件,以及Executor获取到Shuffle的数据文件的分布,那么Executor是如何获取到Shuffle的数据文件进行Action的算子的计算呢?

在ResultTask中,Executor通过MapOutPutTracker向Driver获取了ShuffID的Shuffle数据块的结构,整理成以BlockManangerId为Key的结构,这样可以更容易区分究竟是本地的Shuffle还是远端executor的Shuffle

2. Fetch数据

在MapOutputTracker中获取到的BlockID的地址,是以BlockManagerId的seq数组
Seq[(BlockManagerId, Seq[(BlockId, Long)])]   

BlockManagerId结构

class BlockManagerId private (  
    private var executorId_ : String,  
    private var host_ : String,  
    private var port_ : Int,  
    private var topologyInfo_ : Option[String])  
  extends Externalizable   
是以ExecutorId,Executor Host IP, Executor Port 标示从哪个Executor获取Shuffle的数据文件,通过Seq[BlockManagerId, Seq(BlockID,Long)]的结构,当前executor很容易区分究竟哪些是本地的数据文件,哪些是远端的数据,本地的数据可以直接本地读取,而需要不通过网络来获取。

2.1 读取本Executor文件

如何认为是本地数据?
Spark认为区分是通过相同的ExecutorId来区别的,如果ExecutorId和自己的ExecutorId相同,认为是本地Local,可以直接读取文件。
for ((address, blockInfos) <- blocksByAddress) {  
  totalBlocks += blockInfos.size  
  if (address.executorId == blockManager.blockManagerId.executorId) {  
    // Filter out zero-sized blocks  
    localBlocks ++= blockInfos.filter(_._2 != 0).map(_._1)  
    numBlocksToFetch += localBlocks.size  
  }   
}  
这里有两种情况:
  1. 同一个Executor会生成多个Task,单个Executor里的Task运行可以直接获取本地文件,不需要通过网络
  2. 同一台机器多个Executor,在这种情况下,不同的Executor获取相同机器下的其他的Executor的文件,需要通过网络

2.2 读取非本Executor文件

2.2.1 构造FetchRequest请求

获取非本Executor的文件,在Spark里会生成一个FetchRequest,为了避免单个Executor的MapId过多发送多个FetchRequest请求,会合并同一个Executor的多个请求,合并的规则由最大的请求参数控制
spark.reducer.maxSizeInFlight   
val targetRequestSize = math.max(maxBytesInFlight / 5, 1L)  

对同一个Executor,如果请求多个Block请求的数据大小未超过targetRequestSize,将会被分配到同一个FetchRequest中,以避免多次FetchRequest的请求

val iterator = blockInfos.iterator  
        var curRequestSize = 0L  
        var curBlocks = new ArrayBuffer[(BlockId, Long)]  
        while (iterator.hasNext) {  
          val (blockId, size) = iterator.next()  
          // Skip empty blocks  
          if (size > 0) {  
            curBlocks += ((blockId, size))  
            remoteBlocks += blockId  
            numBlocksToFetch += 1  
            curRequestSize += size  
          } else if (size < 0) {  
            throw new BlockException(blockId, "Negative block size " + size)  
          }  
          if (curRequestSize >= targetRequestSize) {  
            // Add this FetchRequest  
            remoteRequests += new FetchRequest(address, curBlocks)  
            curBlocks = new ArrayBuffer[(BlockId, Long)]  
            logDebug(s"Creating fetch request of $curRequestSize at $address")  
            curRequestSize = 0  
          }  
        }  
        // Add in the final request  
        if (curBlocks.nonEmpty) {  
          remoteRequests += new FetchRequest(address, curBlocks)  
        }  
多个FetchRequest会被随机化后放入队列Queue中,每个Executor从Driver端获取的ShuffID对应的BlockManagerID所管理的BlockID的状态是相同的顺序,如果不对FetchRequest进行随机化,那么非常有可能存在多个Executor同时向同一个Executor获取发送FetchRequest的情况,从而导致Executor的负载增高,为了均衡每个Executor的数据获取,随机化FetchRequest是非常有必要的。

2.2.1 发送FetchRequest

FetchRequest并不是并行提交的,对同一个Task来说,在Executor的做combine的时候是一个一个的BlockID块合并的,而Task本身就是一个线程运行的,所以不需要设计FetchRequest成并行提交,当一个BlockID完成计算后,才需要判断是否需要进行下一个FetchRequest请求,因为FetchRequest是多个Block提交的,为了控制Executor获取多个BlockID的shuffle数据的带宽,在提交FetchRequest的时候控制了请求的频率

在满足下面以下条件下,才允许提交下个FetchRequest

  1. 当正在请求的所有BlockId的内容和下一个FetchRequest的请求内容之和小于maxBytesInFlight的时候,才能进行下一个FetchRequest 的请求
  2. 当正在请求的数量小于所设置的最大的允许请求数量的时候,才能进行下一个FetchRequest的请求,控制参数如下:
spark.reducer.maxReqsInFlight  

2.2.2 完整的FetchRequest流程

  1.  Executor A 通过ExternalShuffleClient 进行fetchBlocks的操作,如果配置了io.maxRetries  
  2. 最大重试参数的话,将启动一个能重试RetryingBlockFetcher的获取器
  3. 初始化TransportClient,OneForOneBlockFetcher获取器
  4. 在OneForOneBlockFetcher里首先向另一个Executor B发送了OpenBlocks的询问请求,里面告知ExecutorID, APPID和BlockID的集合
  5. Executor B获取到BlockIDs,后通过BlockManager获取相关的BlockID的文件(通过mapid, reduceid获取相关的索引和数据文件),构建FileSegmentManagedBuffer
  6. 通过StreamManager(OneForOneStreamManager) registerStream 生成streamId,和StreamState(多个ManagedBuffer,AppID)的缓存返回所生成的StreamId
  7. Executor B 返回给 StreamHandle的消息,里面包含了StreamId和Chunk的数量,这里chunk的数量其实就是Block的数量
  8. Executor A 获取到 StreamHandle的消息,一个一个的发送ChunkFetchRequest里面包含了StreamId, Chunk index,去真实的获取Executor B的shuffle数据文件
  9. Executor B 通过传递的ChunkFetchRequest消息获取到StreamId, Chunk index, 通过缓存获取到对应的FileSgementManagedBuffer,返回chunkFetchSuccess消息,里面包含着streamID, 和FileSegmentManagedBuffer
  10. 在步骤3-6步骤里是堵塞在Task线程里,而步骤7一个一个发送ChunkFetchRequest后,并不堵塞等待返回结果,结果是通过回调函数来实现的,在调用前注册了一个回调函数
client.fetchChunk(streamHandle.streamId, i, chunkCallback);  
  private class ChunkCallback implements ChunkReceivedCallback {  
    @Override  
    public void onSuccess(int chunkIndex, ManagedBuffer buffer) {  
      // On receipt of a chunk, pass it upwards as a block.  
      listener.onBlockFetchSuccess(blockIds[chunkIndex], buffer);  
    }  
  
    @Override  
    public void onFailure(int chunkIndex, Throwable e) {  
      // On receipt of a failure, fail every block from chunkIndex onwards.  
      String[] remainingBlockIds = Arrays.copyOfRange(blockIds, chunkIndex, blockIds.length);  
      failRemainingBlocks(remainingBlockIds, e);  
    }  
  }  

在这里的listener就是前面fetchBlocks里注入的BlockFetchingListener

new BlockFetchingListener {  
       override def onBlockFetchSuccess(blockId: String, buf: ManagedBuffer): Unit = {  
         // Only add the buffer to results queue if the iterator is not zombie,  
         // i.e. cleanup() has not been called yet.  
         ShuffleBlockFetcherIterator.this.synchronized {  
           if (!isZombie) {  
             // Increment the ref count because we need to pass this to a different thread.  
             // This needs to be released after use.  
             buf.retain()  
             remainingBlocks -= blockId  
             results.put(new SuccessFetchResult(BlockId(blockId), address, sizeMap(blockId), buf,  
               remainingBlocks.isEmpty))  
             logDebug("remainingBlocks: " + remainingBlocks)  
           }  
         }  
         logTrace("Got remote block " + blockId + " after " + Utils.getUsedTimeMs(startTime))  
       }  
  
       override def onBlockFetchFailure(blockId: String, e: Throwable): Unit = {  
         logError(s"Failed to get block(s) from ${req.address.host}:${req.address.port}", e)  
         results.put(new FailureFetchResult(BlockId(blockId), address, e))  
       }  
     }  
  1. 如果获取成功将封装SuccessFetchResult里面保存着blockId,地址,数据大小,以及ManagedBuffer,并保存到results的queue中

2.2.3 Fetch 迭代获取数据文件

Executor在BlockStoreShuffeReader的read函数中构建ShuffleBlockFetcherIterator,ShuffleBlockFetcherIterator是个InputStream的迭代器,每个BlockID生成一个InputStream,在设计里并没有区分是本地的还是远端的,每一次迭代都是从堵塞的Queue里获取到BlockID的ManagerBuffer,通过调用ManagerBuffer.createInputStream获取每个InputStream,进行读取并且反序列话,进行KV的combine.
 
 
如何判断所有的BlockID已经读取完了?
override def hasNext: Boolean = numBlocksProcessed < numBlocksToFetch  
在hasNext里判断当前的是否已经达到需要读取的block数量了,每一次读取下一个block的时候都会在numBlocksProcessed+1,在读取失败的情况下会直接抛出异常。

3. Fetch 交互协议

在前面的博客里描述了很多交互协议都使用了Java的原生态的反序列化,但在上文描述的Fetch协议中,是Spark单独定义的一套协议标准,自己实现encoder和decoder
ChunkFetchRequest, ChunkFetchSuccess,  RpcRequest, RpcResponse.... 这些都是直接使用Java进行封装,在Network-Commmon的包里,所有的消息最后都实现了基本的接口。

3.1 Message Encoder

public interface Message extends Encodable{}  

而核心的是Encodable,有点类似Java的Serializable接口,需要自己实现Encoder和Decoder的方法

public interface Encodable {  
  /** Number of bytes of the encoded form of this object. */  
  int encodedLength();  
  
  /** 
   * Serializes this object by writing into the given ByteBuf. 
   * This method must write exactly encodedLength() bytes. 
   */  
  void encode(ByteBuf buf);  
}  
核心的序列话的encode的入参数是ByteBuf 很符合Netty里的NIO所暴露出的接口,同时也要注意这是Netty的ByteBuf 和Netty是耦合了
 
如何让Netty调用Encodable encode方法呢?
在Netty里暴露出的类MessageToMessageEncoder,里暴露encode的抽象方法,这是一个可以允许对传递的消息进行一次自定义的编码
MessageToMessageEncoder

protected abstract void encode(ChannelHandlerContext paramChannelHandlerContext, I paramI, List<Object> paramList) /* */ throws Exception;

在Spark里自己实现MessageToMessageEncoder的encoder的方法

public final class MessageEncoder extends MessageToMessageEncoder<Message> {  
  
  private static final Logger logger = LoggerFactory.getLogger(MessageEncoder.class);  
  
  /*** 
   * Encodes a Message by invoking its encode() method. For non-data messages, we will add one 
   * ByteBuf to 'out' containing the total frame length, the message type, and the message itself. 
   * In the case of a ChunkFetchSuccess, we will also add the ManagedBuffer corresponding to the 
   * data to 'out', in order to enable zero-copy transfer. 
   */  
  @Override  
  public void encode(ChannelHandlerContext ctx, Message in, List<Object> out) throws Exception {  
    Object body = null;  
    long bodyLength = 0;  
    boolean isBodyInFrame = false;  
  
    // If the message has a body, take it out to enable zero-copy transfer for the payload.  
    if (in.body() != null) {  
      try {  
        bodyLength = in.body().size();  
        body = in.body().convertToNetty();  
        isBodyInFrame = in.isBodyInFrame();  
      } catch (Exception e) {  
        in.body().release();  
        if (in instanceof AbstractResponseMessage) {  
          AbstractResponseMessage resp = (AbstractResponseMessage) in;  
          // Re-encode this message as a failure response.  
          String error = e.getMessage() != null ? e.getMessage() : "null";  
          logger.error(String.format("Error processing %s for client %s",  
            in, ctx.channel().remoteAddress()), e);  
          encode(ctx, resp.createFailureResponse(error), out);  
        } else {  
          throw e;  
        }  
        return;  
      }  
    }  
  
    Message.Type msgType = in.type();  
    // All messages have the frame length, message type, and message itself. The frame length  
    // may optionally include the length of the body data, depending on what message is being  
    // sent.  
    int headerLength = 8 + msgType.encodedLength() + in.encodedLength();  
    long frameLength = headerLength + (isBodyInFrame ? bodyLength : 0);  
    ByteBuf header = ctx.alloc().heapBuffer(headerLength);  
    header.writeLong(frameLength);  
    msgType.encode(header);  
    in.encode(header);  
    assert header.writableBytes() == 0;  
  
    if (body != null) {  
      // We transfer ownership of the reference on in.body() to MessageWithHeader.  
      // This reference will be freed when MessageWithHeader.deallocate() is called.  
      out.add(new MessageWithHeader(in.body(), header, body, bodyLength));  
    } else {  
      out.add(header);  
    }  
  }  
  
}  
在encoder的方法里去对Message进行了编码

3.2 Message Decoder

和3.1类似,Spark 针对Netty 封装了MessageDecoder 
public final class MessageDecoder extends MessageToMessageDecoder<ByteBuf>   

在decode方法里,直接对ByteBuf进行decode会Message

public void decode(ChannelHandlerContext ctx, ByteBuf in, List<Object> out) {  
  Message.Type msgType = Message.Type.decode(in);  
  Message decoded = decode(msgType, in);  
  assert decoded.type() == msgType;  
  logger.trace("Received message {}: {}", msgType, decoded);  
  out.add(decoded);  
}  
对每个不同的Message 分别调用了各自的decode的方法。

3.3 传递文件

3.3.1 发送文件

还记的前面fetch文件的返回结果么?
respond(new ChunkFetchSuccess(req.streamChunkId, buf));  

在buf里的ManagedBuffer是FileSegmentManagedBuffer,而在刚才的encode函数里

body = in.body().convertToNetty();  

对ChunkFetchSuccess来说in.body是FileSegmentManagedBuffer,而它封装的方法里

public Object convertToNetty() throws IOException {  
  if (conf.lazyFileDescriptor()) {  
    return new DefaultFileRegion(file, offset, length);  
  } else {  
    FileChannel fileChannel = new FileInputStream(file).getChannel();  
    return new DefaultFileRegion(fileChannel, offset, length);  
  }  
}  

使用了DefaultFileRegion,这是一个Netty里传递文件使用零拷贝的方式,在FileRegion里是调用TransferTo进行零拷贝复制文件,关于零拷贝在这里不介绍了

public abstract long transferTo(WritableByteChannel paramWritableByteChannel, long paramLong)  
  throws IOException;  

但是问题是encode的方法里返回的MessageWithHeader对象,并不是DefaultFileRegion

if (body != null) {  
      // We transfer ownership of the reference on in.body() to MessageWithHeader.  
      // This reference will be freed when MessageWithHeader.deallocate() is called.  
      out.add(new MessageWithHeader(in.body(), header, body, bodyLength));  
    }  

我们来看看什么是MessageWithHeader

class MessageWithHeader extends AbstractReferenceCounted implements FileRegion   

原来是FileRegion,对Netty来说FileRegion最后调用的TransferTo进行传递

public long transferTo(final WritableByteChannel target, final long position) throws IOException {  
    Preconditions.checkArgument(position == totalBytesTransferred, "Invalid position.");  
    // Bytes written for header in this call.  
    long writtenHeader = 0;  
    if (header.readableBytes() > 0) {  
      writtenHeader = copyByteBuf(header, target);  
      totalBytesTransferred += writtenHeader;  
      if (header.readableBytes() > 0) {  
        return writtenHeader;  
      }  
    }  
  
    // Bytes written for body in this call.  
    long writtenBody = 0;  
    if (body instanceof FileRegion) {  
      writtenBody = ((FileRegion) body).transferTo(target, totalBytesTransferred - headerLength);  
    } else if (body instanceof ByteBuf) {  
      writtenBody = copyByteBuf((ByteBuf) body, target);  
    }  
    totalBytesTransferred += writtenBody;  
  
    return writtenHeader + writtenBody;  
  }  

在这里巧妙的将Header和文件封装成了一个文件的region,在TransferTo的函数里先传递头,然后在调用

writtenBody = ((FileRegion) body).transferTo(target, totalBytesTransferred - headerLength);  
来传递文件,而其他的ByteBuf 直接写到Write的channel 里。

3.3.2 接收文件

在3.2章节里介绍了如何decode message的方法,对消息ChunkFetchSuccess进行decode生成ChunkFetchSuccess对象
public static ChunkFetchSuccess decode(ByteBuf buf) {  
  StreamChunkId streamChunkId = StreamChunkId.decode(buf);  
  buf.retain();  
  NettyManagedBuffer managedBuf = new NettyManagedBuffer(buf.duplicate());  
  return new ChunkFetchSuccess(streamChunkId, managedBuf);  
}  
注意:这里的ManagedBuffer不在是FileSegmentManagedBuffer,而是NettyManagedBuffer,里面的ByteBuf才是文件的内容

4. 总结

  1. Fetch Shuffle data 数据区分本地数据,远端数据,本地数据和远端数据的区分依据是ExecutorID
  2. 单个Task线程Fetch Shuffle Data数据是以Block为最小单位,串行获取并进行运算
  3. 远端Fetch的多个Block 数据,是异步发送请求,通过回调函数来异步获取返回结果提交到堵塞的队列中,让Task线程获取、读取,运算
  4. Fetch的交互协议,并没有使用Java的默认反序列的协议,而是自己独立封装Encode、Decode,进行编码和解码
posted @ 2018-06-21 12:41  大葱拌豆腐  阅读(661)  评论(0编辑  收藏  举报