[GitHub] [spark] otterc commented on a change in pull request #30312: [WIP][SPARK-32917][SHUFFLE][CORE][test-maven][test-hadoop2.7] Adds support for executors to push shuffle blocks after successful map task completion

[GitBox <gi...@apache.org>]

otterc commented on a change in pull request #30312:
URL: https://github.com/apache/spark/pull/30312#discussion_r522265369



##########
File path: core/src/main/scala/org/apache/spark/shuffle/ShuffleWriter.scala
##########
@@ -17,18 +17,466 @@
 
 package org.apache.spark.shuffle
 
-import java.io.IOException
+import java.io.{File, IOException}
+import java.net.ConnectException
+import java.nio.ByteBuffer
+import java.util.concurrent.ExecutorService
 
+import scala.collection.mutable.{ArrayBuffer, HashMap, HashSet, Queue}
+
+import com.google.common.base.Throwables
+
+import org.apache.spark.{ShuffleDependency, SparkConf, SparkEnv}
+import org.apache.spark.annotation.Since
+import org.apache.spark.internal.Logging
+import org.apache.spark.internal.config._
+import org.apache.spark.launcher.SparkLauncher
+import org.apache.spark.network.buffer.{FileSegmentManagedBuffer, ManagedBuffer, NioManagedBuffer}
+import org.apache.spark.network.netty.SparkTransportConf
+import org.apache.spark.network.shuffle.{BlockFetchingListener, BlockStoreClient}
+import org.apache.spark.network.shuffle.ErrorHandler.BlockPushErrorHandler
+import org.apache.spark.network.util.TransportConf
 import org.apache.spark.scheduler.MapStatus
+import org.apache.spark.shuffle.ShuffleWriter._
+import org.apache.spark.storage.{BlockId, BlockManagerId, ShufflePushBlockId}
+import org.apache.spark.util.{ThreadUtils, Utils}
 
 /**
- * Obtained inside a map task to write out records to the shuffle system.
+ * Obtained inside a map task to write out records to the shuffle system, and optionally
+ * initiate the block push process to remote shuffle services if push based shuffle is enabled.
  */
-private[spark] abstract class ShuffleWriter[K, V] {
+private[spark] abstract class ShuffleWriter[K, V] extends Logging {
+  private[this] var maxBytesInFlight = 0L
+  private[this] var maxReqsInFlight = 0
+  private[this] var maxBlocksInFlightPerAddress = 0
+  private[this] var bytesInFlight = 0L
+  private[this] var reqsInFlight = 0
+  private[this] val numBlocksInFlightPerAddress = new HashMap[BlockManagerId, Int]()
+  private[this] val deferredPushRequests = new HashMap[BlockManagerId, Queue[PushRequest]]()
+  private[this] val pushRequests = new Queue[PushRequest]
+  private[this] val errorHandler = createErrorHandler()
+  private[this] val unreachableBlockMgrs = new HashSet[BlockManagerId]()
+
   /** Write a sequence of records to this task's output */
   @throws[IOException]
   def write(records: Iterator[Product2[K, V]]): Unit
 
   /** Close this writer, passing along whether the map completed */
   def stop(success: Boolean): Option[MapStatus]
+
+  def getPartitionLengths(): Array[Long]
+
+  // VisibleForTesting
+  private[shuffle] def createErrorHandler(): BlockPushErrorHandler = {
+    new BlockPushErrorHandler() {
+      // For a connection exception against a particular host, we will stop pushing any
+      // blocks to just that host and continue push blocks to other hosts. So, here push of
+      // all blocks will only stop when it is "Too Late". Also see updateStateAndCheckIfPushMore.
+      override def shouldRetryError(t: Throwable): Boolean = {
+        // If the block is too late, there is no need to retry it
+        !Throwables.getStackTraceAsString(t).contains(BlockPushErrorHandler.TOO_LATE_MESSAGE_SUFFIX)
+      }
+    }
+  }
+
+  /**
+   * Initiate the block push process. This will be invoked after the shuffle writer
+   * finishes writing the shuffle file if push based shuffle is enabled.
+   *
+   * @param resolver         block resolver used to locate mapper generated shuffle file
+   * @param partitionLengths array of shuffle block size so we can tell shuffle block
+   *                         boundaries within the shuffle file
+   * @param dep              shuffle dependency to get shuffle ID and the location of remote shuffle
+   *                         services to push local shuffle blocks
+   * @param partitionId      map index of the shuffle map task
+   * @param mapId            mapId of the shuffle map task
+   * @param conf             spark configuration
+   */
+  def initiateBlockPush(
+      resolver: IndexShuffleBlockResolver,
+      partitionLengths: Array[Long],
+      dep: ShuffleDependency[_, _, _],
+      partitionId: Int,
+      mapId: Long,
+      conf: SparkConf): Unit = {
+    val numPartitions = dep.partitioner.numPartitions
+    val dataFile = resolver.getDataFile(dep.shuffleId, mapId)
+    val transportConf = SparkTransportConf.fromSparkConf(conf, "shuffle")
+
+    val maxBlockSizeToPush = conf.get(PUSH_SHUFFLE_MAX_BLOCK_SIZE_TO_PUSH) * 1024
+    val maxBlockBatchSize = conf.get(PUSH_SHUFFLE_MAX_BLOCK_BATCH_SIZE) * 1024 * 1024
+    val mergerLocs = dep.getMergerLocs.map(loc =>
+      BlockManagerId("", loc.host, loc.port))
+
+    maxBytesInFlight = conf.getSizeAsMb("spark.reducer.maxSizeInFlight", "48m") * 1024 * 1024
+    maxReqsInFlight = conf.getInt("spark.reducer.maxReqsInFlight", Int.MaxValue)
+    maxBlocksInFlightPerAddress = conf.get(REDUCER_MAX_BLOCKS_IN_FLIGHT_PER_ADDRESS)
+
+    val requests = prepareBlockPushRequests(numPartitions, partitionId, dep.shuffleId, dataFile,
+      partitionLengths, mergerLocs, transportConf, maxBlockSizeToPush, maxBlockBatchSize)
+    // Randomize the orders of the PushRequest, so different mappers pushing blocks at the same
+    // time won't be pushing the same ranges of shuffle partitions.
+    pushRequests ++= Utils.randomize(requests)
+
+    val shuffleClient = SparkEnv.get.blockManager.blockStoreClient
+    submitTask(() => {
+      pushUpToMax(shuffleClient)
+    })
+  }
+
+  /**
+   * Triggers the push. It's a separate method for testing.
+   * VisibleForTesting
+   */
+  protected def submitTask(task: Runnable): Unit = {
+    if (BLOCK_PUSHER_POOL != null) {
+      BLOCK_PUSHER_POOL.execute(task)
+    }
+  }
+
+  /**
+   * Since multiple netty client threads could potentially be calling pushUpToMax for the same
+   * mapper, we synchronize access to this method so that only one thread can push blocks for
+   * a given mapper. This helps to simplify access to the shared states. The down side of this
+   * is that we could unnecessarily block other mappers' block pushes if all netty client threads
+   * are occupied by block pushes from the same mapper.
+   *
+   * This code is similar to ShuffleBlockFetcherIterator#fetchUpToMaxBytes in how it throttles
+   * the data transfer between shuffle client/server.
+   */
+  private def pushUpToMax(shuffleClient: BlockStoreClient): Unit = synchronized {
+    // Process any outstanding deferred push requests if possible.
+    if (deferredPushRequests.nonEmpty) {
+      for ((remoteAddress, defReqQueue) <- deferredPushRequests) {
+        while (isRemoteBlockPushable(defReqQueue) &&
+          !isRemoteAddressMaxedOut(remoteAddress, defReqQueue.front)) {
+          val request = defReqQueue.dequeue()
+          logDebug(s"Processing deferred push request for $remoteAddress with "
+            + s"${request.blocks.length} blocks")
+          sendRequest(request, shuffleClient)
+          if (defReqQueue.isEmpty) {
+            deferredPushRequests -= remoteAddress
+          }
+        }
+      }
+    }
+
+    // Process any regular push requests if possible.
+    while (isRemoteBlockPushable(pushRequests)) {
+      val request = pushRequests.dequeue()
+      val remoteAddress = request.address
+      if (isRemoteAddressMaxedOut(remoteAddress, request)) {
+        logDebug(s"Deferring push request for $remoteAddress with ${request.blocks.size} blocks")
+        val defReqQueue = deferredPushRequests.getOrElse(remoteAddress, new Queue[PushRequest]())
+        defReqQueue.enqueue(request)
+        deferredPushRequests(remoteAddress) = defReqQueue
+      } else {
+        sendRequest(request, shuffleClient)
+      }
+    }
+
+    def isRemoteBlockPushable(pushReqQueue: Queue[PushRequest]): Boolean = {
+      pushReqQueue.nonEmpty &&
+        (bytesInFlight == 0 ||
+          (reqsInFlight + 1 <= maxReqsInFlight &&
+            bytesInFlight + pushReqQueue.front.size <= maxBytesInFlight))
+    }
+
+    // Checks if sending a new push request will exceed the max no. of blocks being pushed to a
+    // given remote address.
+    def isRemoteAddressMaxedOut(remoteAddress: BlockManagerId, request: PushRequest): Boolean = {
+      (numBlocksInFlightPerAddress.getOrElse(remoteAddress, 0)
+        + request.blocks.size) > maxBlocksInFlightPerAddress
+    }
+  }
+
+  /**
+   * Push blocks to remote shuffle server. The callback listener will invoke #pushUpToMax again
+   * to trigger pushing the next batch of blocks once some block transfer is done in the current
+   * batch. This way, we decouple the map task from the block push process, since it is netty
+   * client thread instead of task execution thread which takes care of majority of the block
+   * pushes.
+   */
+  private def sendRequest(request: PushRequest, shuffleClient: BlockStoreClient): Unit = {
+    bytesInFlight = bytesInFlight + request.size
+    reqsInFlight = reqsInFlight + 1
+    numBlocksInFlightPerAddress(request.address) = numBlocksInFlightPerAddress.getOrElseUpdate(
+      request.address, 0) + request.blocks.length
+
+    val sizeMap = request.blocks.map { case (blockId, size) => (blockId.toString, size) }.toMap
+    val address = request.address
+    val blockIds = request.blocks.map(_._1.toString)
+    val remainingBlocks = new HashSet[String]() ++= blockIds
+
+    val blockPushListener = new BlockFetchingListener {
+      // Initiating a connection and pushing blocks to a remote shuffle service is always handled by
+      // the block-push-threads. We should not initiate the connection creation in the
+      // blockPushListener callbacks which are invoked by the netty eventloop because:
+      // 1. TrasportClient.createConnection(...) blocks for connection to be established and it's
+      // recommended to avoid any blocking operations in the eventloop;
+      // 2. The actual connection creation is a task that gets added to the task queue of another
+      // eventloop which could have eventloops eventually blocking each other.
+      // Once the blockPushListener is notified of the block push success or failure, we
+      // just delegate it to block-push-threads.
+      def handleResult(result: PushResult): Unit = {
+        submitTask(() => {
+          if (updateStateAndCheckIfPushMore(
+            sizeMap(result.blockId), address, remainingBlocks, result)) {
+            pushUpToMax(SparkEnv.get.blockManager.blockStoreClient)
+          }
+        })
+      }
+
+      override def onBlockFetchSuccess(blockId: String, data: ManagedBuffer): Unit = {
+        logTrace(s"Push for block $blockId to $address successful.")
+        handleResult(PushResult(blockId, null))
+      }
+
+      override def onBlockFetchFailure(blockId: String, exception: Throwable): Unit = {
+        // check the message or it's cause to see it needs to be logged.
+        if (!errorHandler.shouldLogError(exception)) {
+          logTrace(s"Pushing block $blockId to $address failed.", exception)
+        } else {
+          logWarning(s"Pushing block $blockId to $address failed.", exception)
+        }
+        handleResult(PushResult(blockId, exception))
+      }
+    }
+    shuffleClient.pushBlocks(address.host, address.port, blockIds.toArray,
+      sliceReqBufferIntoBlockBuffers(request.reqBuffer, request.blocks.map(_._2)),
+      blockPushListener)
+  }
+
+  /**
+   * Given the ManagedBuffer representing all the continuous blocks inside the shuffle data file
+   * for a PushRequest and an array of individual block sizes, load the buffer from disk into
+   * memory and slice it into multiple smaller buffers representing each block.
+   *
+   * With nio ByteBuffer, the individual block buffers share data with the initial in memory
+   * buffer loaded from disk. Thus only one copy of the block data is kept in memory.
+   * @param reqBuffer A {{FileSegmentManagedBuffer}} representing all the continuous blocks in
+   *                  the shuffle data file for a PushRequest
+   * @param blockSizes Array of block sizes
+   * @return Array of in memory buffer for each individual block
+   */
+  private def sliceReqBufferIntoBlockBuffers(
+      reqBuffer: ManagedBuffer,
+      blockSizes: Seq[Long]): Array[ManagedBuffer] = {
+    if (blockSizes.size == 1) {
+      Seq(reqBuffer).toArray
+    } else {
+      val inMemoryBuffer = reqBuffer.nioByteBuffer()
+      val blockOffsets = new Array[Long](blockSizes.size)
+      var offset = 0L
+      for (index <- blockSizes.indices) {
+        blockOffsets(index) = offset
+        offset += blockSizes(index)
+      }
+      blockOffsets.zip(blockSizes).map {
+        case (offset, size) =>
+          new NioManagedBuffer(inMemoryBuffer.duplicate()
+            .position(offset.toInt)
+            .limit((offset + size).toInt).asInstanceOf[ByteBuffer].slice())
+      }.toArray
+    }
+  }
+
+  /**
+   * Updates the stats and based on the previous push result decides whether to push more blocks
+   * or stop.
+   *
+   * @param bytesPushed     number of bytes pushed.
+   * @param address         address of the remote service
+   * @param remainingBlocks remaining blocks
+   * @param pushResult      result of the last push
+   * @return true if more blocks should be pushed; false otherwise.
+   */
+  private def updateStateAndCheckIfPushMore(
+      bytesPushed: Long,
+      address: BlockManagerId,
+      remainingBlocks: HashSet[String],
+      pushResult: PushResult): Boolean = synchronized {
+    remainingBlocks -= pushResult.blockId
+    bytesInFlight = bytesInFlight - bytesPushed
+    numBlocksInFlightPerAddress(address) = numBlocksInFlightPerAddress(address) - 1
+    if (remainingBlocks.isEmpty) {
+      reqsInFlight = reqsInFlight - 1
+    }
+    if (pushResult.failure != null && pushResult.failure.getCause != null &&
+      pushResult.failure.getCause.isInstanceOf[ConnectException]) {
+      // Remove all the blocks for this address just once because removing from pushRequests
+      // is expensive. If there is a ConnectException for the first block, all the subsequent
+      // blocks to that address will fail, so should avoid removing multiple times.
+      if (!unreachableBlockMgrs.contains(address)) {
+        var removed = 0
+        unreachableBlockMgrs.add(address)
+        removed += pushRequests.dequeueAll(req => req.address == address).length
+        val droppedReq = deferredPushRequests.remove(address)
+        if (droppedReq.isDefined) {
+          removed += droppedReq.get.length
+        }
+        logWarning(s"Received a ConnectException from $address. " +
+          s"Dropping push of $removed blocks and " +
+          s"not pushing any more blocks to this address.")
+      }
+    }
+    if (pushResult.failure != null && !errorHandler.shouldRetryError(pushResult.failure)) {
+      logDebug(s"Received after merge is finalized from $address. Not pushing any more blocks.")
+      return false
+    } else {
+      remainingBlocks.isEmpty && (pushRequests.nonEmpty || deferredPushRequests.nonEmpty)
+    }
+  }
+
+  /**
+   * Convert the shuffle data file of the current mapper into a list of PushRequest. Basically,
+   * continuous blocks in the shuffle file are grouped into a single request to allow more
+   * efficient read of the block data. Each mapper for a given shuffle will receive the same
+   * list of BlockManagerIds as the target location to push the blocks to. All mappers in the
+   * same shuffle will map shuffle partition ranges to individual target locations in a consistent
+   * manner to make sure each target location receives shuffle blocks belonging to the same set
+   * of partition ranges. 0-length blocks and blocks that are large enough will be skipped.
+   *
+   * @param numPartitions Number of shuffle partitions in the shuffle file
+   * @param partitionId map index of the current mapper
+   * @param shuffleId ShuffleId of current shuffle
+   * @param dataFile Shuffle data file
+   * @param partitionLengths array of sizes of blocks in the shuffle data file
+   * @param mergerLocs Target locations to push blocks to
+   * @param transportConf TransportConf used to create FileSegmentManagedBuffer
+   * @param maxBlockSizeToPush Max size of individual blocks that will be pushed. Blocks larger
+   *                           than this threshold will be skipped.
+   * @param maxBlockBatchSize Max size of a batch of shuffle blocks to be grouped into a single
+   *                          request
+   * @return List of the PushRequest, randomly shuffled.
+   */
+  private def prepareBlockPushRequests(
+      numPartitions: Int,
+      partitionId: Int,
+      shuffleId: Int,
+      dataFile: File,
+      partitionLengths: Array[Long],
+      mergerLocs: Seq[BlockManagerId],
+      transportConf: TransportConf,
+      maxBlockSizeToPush: Long,
+      maxBlockBatchSize: Long): Seq[PushRequest] = {
+    var offset = 0L
+    var currentReqSize = 0L
+    var currentReqOffset = 0L
+    var currentMergerId = 0
+    val numMergers = mergerLocs.length
+    val requests = new ArrayBuffer[PushRequest]
+    var blocks = new ArrayBuffer[(BlockId, Long)]
+    for (reduceId <- 0 until numPartitions) {
+      val blockSize = partitionLengths(reduceId)
+      logDebug(
+        s"Block ${ShufflePushBlockId(shuffleId, partitionId, reduceId)} is of size $blockSize")
+      // Skip 0-length blocks and blocks that are large enough
+      if (blockSize > 0) {
+        val mergerId = math.min(math.floor(reduceId * 1.0 / numPartitions * numMergers),
+          numMergers - 1).asInstanceOf[Int]
+        // Start a new PushRequest if the current request goes beyond the max batch size,
+        // or the number of blocks in the current request goes beyond the limit per destination,
+        // or the next block push location is for a different shuffle service, or the next block
+        // exceeds the max block size to push limit. This guarantees that each PushReqeust
+        // represents continuous blocks in the shuffle file to be pushed to the same shuffle
+        // service, and does not go beyond existing limitations.
+        if (currentReqSize + blockSize <= maxBlockBatchSize
+          && blocks.size < maxBlocksInFlightPerAddress
+          && mergerId == currentMergerId && blockSize <= maxBlockSizeToPush) {
+          // Add current block to current batch
+          currentReqSize += blockSize
+        } else {
+          if (blocks.nonEmpty) {
+            // Convert the previous batch into a PushRequest
+            requests += PushRequest(mergerLocs(currentMergerId), blocks.toSeq,
+              createRequestBuffer(transportConf, dataFile, currentReqOffset, currentReqSize))
+          }
+          // Start a new batch
+          currentReqSize = 0
+          // Set currentReqffset to -1 so we are able to distinguish between the initial value
+          // of currentReqOffset and when we are about to start a new batch
+          currentReqOffset = -1
+          currentMergerId = mergerId
+          blocks = new ArrayBuffer[(BlockId, Long)]

Review comment:
       `blocks` are added to the `PushRequest`  in line 392 and 419. 
   `PushRequest(mergerLocs(currentMergerId), blocks.toSeq,
           createRequestBuffer(transportConf, dataFile, currentReqOffset, currentReqSize))`
   Clearing the `blocks` changes the PushRequest. I verified this by running one of the UTs.




----------------------------------------------------------------
This is an automated message from the Apache Git Service.
To respond to the message, please log on to GitHub and use the
URL above to go to the specific comment.

For queries about this service, please contact Infrastructure at:
users@infra.apache.org



---------------------------------------------------------------------
To unsubscribe, e-mail: reviews-unsubscribe@spark.apache.org
For additional commands, e-mail: reviews-help@spark.apache.org