Client/Server use wakeup feature and their own thread

src/main/scala/org/apache/spark/shuffle/compat/spark_2_4/UcxShuffleReader.scala

@@ -50,36 +50,8 @@ class UcxShuffleReader[K, C](handle: BaseShuffleHandle[K, _, C],
         SparkEnv.get.conf.get(config.MAX_REMOTE_BLOCK_SIZE_FETCH_TO_MEM),
         SparkEnv.get.conf.getBoolean("spark.shuffle.detectCorrupt", true))
 
-      // Ucx shuffle logic
-      // Java reflection to get access to private results queue
-      val queueField = wrappedStreams.getClass.getDeclaredField(
-        "org$apache$spark$storage$ShuffleBlockFetcherIterator$$results")
-      queueField.setAccessible(true)
-      val resultQueue = queueField.get(wrappedStreams).asInstanceOf[LinkedBlockingQueue[_]]
-
-      // Do progress if queue is empty before calling next on ShuffleIterator
-      val ucxWrappedStream = new Iterator[(BlockId, InputStream)] {
-        override def next(): (BlockId, InputStream) = {
-          val startTime = System.currentTimeMillis()
-          while (resultQueue.isEmpty) {
-            transport.progress()
-          }
-          shuffleMetrics.incFetchWaitTime(System.currentTimeMillis() - startTime)
-          wrappedStreams.next()
-        }
-
-        override def hasNext: Boolean = {
-          val result = wrappedStreams.hasNext
-          if (!result) {
-            shuffleClient.close()
-          }
-          result
-        }
-      }
-      // End of ucx shuffle logic
-
       val serializerInstance = dep.serializer.newInstance()
-      val recordIter = ucxWrappedStream.flatMap { case (blockId, wrappedStream) =>
+      val recordIter = wrappedStreams.flatMap { case (blockId, wrappedStream) =>
         // Note: the asKeyValueIterator below wraps a key/value iterator inside of a
         // NextIterator. The NextIterator makes sure that close() is called on the
         // underlying InputStream when all records have been read.

src/main/scala/org/apache/spark/shuffle/compat/spark_3_0/UcxLocalDiskShuffleExecutorComponents.scala

@@ -23,9 +23,7 @@ class UcxLocalDiskShuffleExecutorComponents(sparkConf: SparkConf)
 
   override def initializeExecutor(appId: String, execId: String, extraConfigs: util.Map[String, String]): Unit = {
     val ucxShuffleManager = SparkEnv.get.shuffleManager.asInstanceOf[UcxShuffleManager]
-    while (ucxShuffleManager.ucxTransport == null) {
-      Thread.sleep(5)
-    }
+    ucxShuffleManager.awaitUcxTransport()
     blockResolver = ucxShuffleManager.shuffleBlockResolver
   }
 

src/main/scala/org/apache/spark/shuffle/compat/spark_3_0/UcxShuffleClient.scala

@@ -41,7 +41,6 @@ class UcxShuffleClient(val transport: UcxShuffleTransport, mapId2PartitionId: Ma
     }
     val resultBufferAllocator = (size: Long) => transport.hostBounceBufferMemoryPool.get(size)
     transport.fetchBlocksByBlockIds(execId.toLong, ucxBlockIds, resultBufferAllocator, callbacks)
-    transport.progress()
   }
 
   override def close(): Unit = {

src/main/scala/org/apache/spark/shuffle/compat/spark_3_0/UcxShuffleReader.scala

@@ -104,40 +104,10 @@ private[spark] class UcxShuffleReader[K, C](handle: BaseShuffleHandle[K, _, C],
 
     val wrappedStreams = shuffleIterator.toCompletionIterator
 
-
-    // Ucx shuffle logic
-    // Java reflection to get access to private results queue
-    val queueField = shuffleIterator.getClass.getDeclaredField(
-      "org$apache$spark$storage$ShuffleBlockFetcherIterator$$results")
-    queueField.setAccessible(true)
-    val resultQueue = queueField.get(shuffleIterator).asInstanceOf[LinkedBlockingQueue[_]]
-
-    // Do progress if queue is empty before calling next on ShuffleIterator
-    val ucxWrappedStream = new Iterator[(BlockId, InputStream)] {
-      override def next(): (BlockId, InputStream) = {
-        val startTime = System.nanoTime()
-        while (resultQueue.isEmpty) {
-          transport.progress()
-        }
-        val fetchWaitTime = System.nanoTime() - startTime
-        readMetrics.incFetchWaitTime(TimeUnit.NANOSECONDS.toMillis(fetchWaitTime))
-        wrappedStreams.next()
-      }
-
-      override def hasNext: Boolean = {
-        val result = wrappedStreams.hasNext
-        if (!result) {
-          shuffleClient.close()
-        }
-        result
-      }
-    }
-    // End of ucx shuffle logic
-
     val serializerInstance = dep.serializer.newInstance()
 
     // Create a key/value iterator for each stream
-    val recordIter = ucxWrappedStream.flatMap { case (blockId, wrappedStream) =>
+    val recordIter = wrappedStreams.flatMap { case (blockId, wrappedStream) =>
       // Note: the asKeyValueIterator below wraps a key/value iterator inside of a
       // NextIterator. The NextIterator makes sure that close() is called on the
       // underlying InputStream when all records have been read.

src/main/scala/org/apache/spark/shuffle/ucx/ShuffleTransport.scala

@@ -150,13 +150,6 @@ trait ShuffleTransport {
    */
   def unregister(blockId: BlockId): Unit
 
-  /**
-   * Batch version of [[ fetchBlocksByBlockIds ]].
-   */
-  def fetchBlocksByBlockIds(executorId: ExecutorId, blockIds: Seq[BlockId],
-                            resultBufferAllocator: BufferAllocator,
-                            callbacks: Seq[OperationCallback]): Seq[Request]
-
   /**
    * Progress outstanding operations. This routine is blocking (though may poll for event).
    * It's required to call this routine within same thread that submitted [[ fetchBlocksByBlockIds ]].
@@ -167,3 +160,7 @@ trait ShuffleTransport {
   def progress(): Unit
 
 }
+
+class UcxFetchState(val callbacks: Seq[OperationCallback],
+                    val request: UcxRequest,
+                    val timestamp: Long) {}

src/main/scala/org/apache/spark/shuffle/ucx/UcxShuffleTransport.scala

@@ -9,13 +9,15 @@ import org.apache.spark.internal.Logging
 import org.apache.spark.shuffle.ucx.memory.UcxHostBounceBuffersPool
 import org.apache.spark.shuffle.ucx.rpc.GlobalWorkerRpcThread
 import org.apache.spark.shuffle.ucx.utils.{SerializableDirectBuffer, SerializationUtils}
+import org.apache.spark.util.ThreadUtils
 import org.apache.spark.shuffle.utils.UnsafeUtils
 import org.openucx.jucx.UcxException
 import org.openucx.jucx.ucp._
 import org.openucx.jucx.ucs.UcsConstants
 
 import java.net.InetSocketAddress
 import java.nio.ByteBuffer
+import java.util.concurrent.atomic.AtomicInteger
 import scala.collection.concurrent.TrieMap
 import scala.collection.mutable
 
@@ -53,7 +55,7 @@ class UcxStats extends OperationStats {
 }
 
 case class UcxShuffleBockId(shuffleId: Int, mapId: Int, reduceId: Int) extends BlockId {
-  override def serializedSize: Int = 12
+  override def serializedSize: Int = UcxShuffleBockId.serializedSize
 
   override def serialize(byteBuffer: ByteBuffer): Unit = {
     byteBuffer.putInt(shuffleId)
@@ -63,6 +65,8 @@ case class UcxShuffleBockId(shuffleId: Int, mapId: Int, reduceId: Int) extends B
 }
 
 object UcxShuffleBockId {
+  val serializedSize: Int = 12
+
   def deserialize(byteBuffer: ByteBuffer): UcxShuffleBockId = {
     val shuffleId = byteBuffer.getInt
     val mapId = byteBuffer.getInt
@@ -88,6 +92,9 @@ class UcxShuffleTransport(var ucxShuffleConf: UcxShuffleConf = null, var executo
   private var progressThread: Thread = _
   var hostBounceBufferMemoryPool: UcxHostBounceBuffersPool = _
 
+  private[spark] lazy val replyThreadPool = ThreadUtils.newForkJoinPool(
+    "UcxListenerThread", ucxShuffleConf.numListenerThreads)
+
   private val errorHandler = new UcpEndpointErrorHandler {
     override def onError(ucpEndpoint: UcpEndpoint, errorCode: Int, errorString: String): Unit = {
       if (errorCode == UcsConstants.STATUS.UCS_ERR_CONNECTION_RESET) {
@@ -126,7 +133,7 @@ class UcxShuffleTransport(var ucxShuffleConf: UcxShuffleConf = null, var executo
     logInfo(s"Allocating ${ucxShuffleConf.numListenerThreads} server workers")
     for (i <- 0 until ucxShuffleConf.numListenerThreads) {
       val worker = ucxContext.newWorker(ucpWorkerParams)
-      allocatedServerWorkers(i) = UcxWorkerWrapper(worker, this, isClientWorker = false)
+      allocatedServerWorkers(i) = UcxWorkerWrapper(worker, this, isClientWorker = false, i.toLong)
     }
 
     val Array(host, port) = ucxShuffleConf.listenerAddress.split(":")
@@ -150,6 +157,8 @@ class UcxShuffleTransport(var ucxShuffleConf: UcxShuffleConf = null, var executo
       allocatedClientWorkers(i) = UcxWorkerWrapper(worker, this, isClientWorker = true, clientId)
     }
 
+    allocatedServerWorkers.foreach(_.progressStart())
+    allocatedClientWorkers.foreach(_.progressStart())
     initialized = true
     logInfo(s"Started listener on ${listener.getAddress}")
     SerializationUtils.serializeInetAddress(listener.getAddress)
@@ -166,7 +175,6 @@ class UcxShuffleTransport(var ucxShuffleConf: UcxShuffleConf = null, var executo
       hostBounceBufferMemoryPool.close()
 
       allocatedClientWorkers.foreach(_.close())
-      allocatedServerWorkers.foreach(_.close())
 
       if (listener != null) {
         listener.close()
@@ -183,6 +191,8 @@ class UcxShuffleTransport(var ucxShuffleConf: UcxShuffleConf = null, var executo
         globalWorker = null
       }
 
+      allocatedServerWorkers.foreach(_.close())
+
       if (ucxContext != null) {
         ucxContext.close()
         ucxContext = null
@@ -196,10 +206,7 @@ class UcxShuffleTransport(var ucxShuffleConf: UcxShuffleConf = null, var executo
    */
   override def addExecutor(executorId: ExecutorId, workerAddress: ByteBuffer): Unit = {
     executorAddresses.put(executorId, workerAddress)
-    allocatedClientWorkers.foreach(w => {
-      w.getConnection(executorId)
-      w.progressConnect()
-    })
+    allocatedClientWorkers.foreach(_.getConnection(executorId))
   }
 
   def addExecutors(executorIdsToAddress: Map[ExecutorId, SerializableDirectBuffer]): Unit = {
@@ -265,36 +272,35 @@ class UcxShuffleTransport(var ucxShuffleConf: UcxShuffleConf = null, var executo
   /**
    * Batch version of [[ fetchBlocksByBlockIds ]].
    */
-  override def fetchBlocksByBlockIds(executorId: ExecutorId, blockIds: Seq[BlockId],
+  def fetchBlocksByBlockIds(executorId: ExecutorId, blockIds: Seq[BlockId],
                                      resultBufferAllocator: BufferAllocator,
-                                     callbacks: Seq[OperationCallback]): Seq[Request] = {
-    allocatedClientWorkers((Thread.currentThread().getId % allocatedClientWorkers.length).toInt)
-      .fetchBlocksByBlockIds(executorId, blockIds, resultBufferAllocator, callbacks)
+                                     callbacks: Seq[OperationCallback]): Unit = {
+    selectClientWorker.fetchBlocksByBlockIds(executorId, blockIds,
+                                             resultBufferAllocator, callbacks)
   }
 
   def connectServerWorkers(executorId: ExecutorId, workerAddress: ByteBuffer): Unit = {
     allocatedServerWorkers.foreach(w => w.connectByWorkerAddress(executorId, workerAddress))
   }
 
-  def handleFetchBlockRequest(replyTag: Int, amData: UcpAmData, replyExecutor: Long): Unit = {
-    val buffer = UnsafeUtils.getByteBufferView(amData.getDataAddress, amData.getLength.toInt)
-    val blockIds = mutable.ArrayBuffer.empty[BlockId]
-
-    // 1. Deserialize blockIds from header
-    while (buffer.remaining() > 0) {
-      val blockId = UcxShuffleBockId.deserialize(buffer)
-      if (!registeredBlocks.contains(blockId)) {
-        throw new UcxException(s"$blockId is not registered")
+  def handleFetchBlockRequest(replyTag: Int, blockIds: Seq[BlockId],
+                              replyExecutor: Long): Unit = {
+    replyThreadPool.submit(new Runnable {
+      override def run(): Unit = {
+        val blocks = blockIds.map(bid => registeredBlocks(bid))
+        selectServerWorker.handleFetchBlockRequest(blocks, replyTag,
+                                                   replyExecutor)
       }
-      blockIds += blockId
-    }
-
-    val blocks = blockIds.map(bid => registeredBlocks(bid))
-    amData.close()
-    allocatedServerWorkers((Thread.currentThread().getId % allocatedServerWorkers.length).toInt)
-      .handleFetchBlockRequest(blocks, replyTag, replyExecutor)
+    })
   }
 
+  @inline
+  def selectClientWorker(): UcxWorkerWrapper = allocatedClientWorkers(
+    (Thread.currentThread().getId % allocatedClientWorkers.length).toInt)
+
+  @inline
+  def selectServerWorker(): UcxWorkerWrapper = allocatedServerWorkers(
+    (Thread.currentThread().getId % allocatedServerWorkers.length).toInt)
 
   /**
    * Progress outstanding operations. This routine is blocking (though may poll for event).
@@ -304,10 +310,5 @@ class UcxShuffleTransport(var ucxShuffleConf: UcxShuffleConf = null, var executo
    * But not guaranteed that at least one [[ fetchBlocksByBlockIds ]] completed!
    */
   override def progress(): Unit = {
-    allocatedClientWorkers((Thread.currentThread().getId % allocatedClientWorkers.length).toInt).progress()
   }
-
-  def progressConnect(): Unit = {
-    allocatedClientWorkers.par.foreach(_.progressConnect())
-  }
-}
+}