otterc commented on a change in pull request #30312: URL: https://github.com/apache/spark/pull/30312#discussion_r535833742
########## File path: core/src/main/scala/org/apache/spark/shuffle/ShuffleBlockPusher.scala ########## @@ -0,0 +1,456 @@ +/* + * 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.shuffle + +import java.io.File +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 +import org.apache.spark.network.shuffle.ErrorHandler.BlockPushErrorHandler +import org.apache.spark.network.util.TransportConf +import org.apache.spark.shuffle.ShuffleBlockPusher._ +import org.apache.spark.storage.{BlockId, BlockManagerId, ShufflePushBlockId} +import org.apache.spark.util.{ThreadUtils, Utils} + +/** + * Used for pushing shuffle blocks to remote shuffle services when push shuffle is enabled. + * When push shuffle is enabled, it is created after the shuffle writer finishes writing the shuffle + * file and initiates the block push process. + * + * @param conf spark configuration + */ +@Since("3.1.0") +private[spark] class ShuffleBlockPusher(conf: SparkConf) extends Logging { + private[this] val maxBlockSizeToPush = conf.get(SHUFFLE_MAX_BLOCK_SIZE_TO_PUSH) + private[this] val maxBlockBatchSize = conf.get(SHUFFLE_MAX_BLOCK_BATCH_SIZE_FOR_PUSH) + private[this] val maxBytesInFlight = + conf.getSizeAsMb("spark.reducer.maxSizeInFlight", "48m") * 1024 * 1024 + private[this] val maxReqsInFlight = conf.getInt("spark.reducer.maxReqsInFlight", Int.MaxValue) + private[this] val maxBlocksInFlightPerAddress = conf.get(REDUCER_MAX_BLOCKS_IN_FLIGHT_PER_ADDRESS) + 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() + // VisibleForTesting + private[shuffle] val unreachableBlockMgrs = new HashSet[BlockManagerId]() + private[this] var stopPushing = false + + // 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) + } + } + } + + /** + * Initiates the block push. + * + * @param dataFile mapper generated shuffle data file + * @param partitionLengths array of shuffle block size so we can tell shuffle block + * @param dep shuffle dependency to get shuffle ID and the location of remote shuffle + * services to push local shuffle blocks + * @param mapIndex map index of the shuffle map task + */ + private[shuffle] def initiateBlockPush( + dataFile: File, + partitionLengths: Array[Long], + dep: ShuffleDependency[_, _, _], + mapIndex: Int): Unit = { + val numPartitions = dep.partitioner.numPartitions + val transportConf = SparkTransportConf.fromSparkConf(conf, "shuffle") + val requests = prepareBlockPushRequests(numPartitions, mapIndex, dep.shuffleId, dataFile, + partitionLengths, dep.getMergerLocs, transportConf) + // 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) + + submitTask(() => { + pushUpToMax() + }) + } + + /** + * 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 block push 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 the 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(): Unit = synchronized { Review comment: If `stopPushing` is `false` this method will never be called. The first time this method is run is in `initiateBlockPush` and after that it is always submitted after a response from the server is received (line 204 - 206) ``` if (updateStateAndCheckIfPushMore( sizeMap(result.blockId), address, remainingBlocks, result)) { pushUpToMax() } ``` Here, `updateStateAndCheckIfPushMore` will return false when it has to stop pushing. ---------------------------------------------------------------- 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: [email protected] --------------------------------------------------------------------- To unsubscribe, e-mail: [email protected] For additional commands, e-mail: [email protected]
