Merge pull request #47 from Eugene-Mark/pmof-fix

[PMEM-SHUFFLE-46] Fix the bug that off-heap memory is over used in shuffle reduce stage.
oap-project · Aug 20, 2021 · 7bc311c · 7bc311c
2 parents 5e8ab2d + 7f18715
commit 7bc311c
Show file tree

Hide file tree

Showing 6 changed files with 15 additions and 14 deletions.
diff --git a/core/src/main/scala/org/apache/spark/shuffle/pmof/BaseShuffleReader.scala b/core/src/main/scala/org/apache/spark/shuffle/pmof/BaseShuffleReader.scala
@@ -90,7 +90,7 @@ private[spark] class BaseShuffleReader[K, C](handle: BaseShuffleHandle[K, _, C],
     /**
     *   Force iterator to traverse itself and update internal counter
     **/
-    wrappedStreams.size
+    //wrappedStreams.size
 
     val serializerInstance = dep.serializer.newInstance()
 

diff --git a/core/src/main/scala/org/apache/spark/shuffle/pmof/PmofShuffleManager.scala b/core/src/main/scala/org/apache/spark/shuffle/pmof/PmofShuffleManager.scala
@@ -62,7 +62,6 @@ private[spark] class PmofShuffleManager(conf: SparkConf) extends ShuffleManager
   override def getReader[K, C](handle: _root_.org.apache.spark.shuffle.ShuffleHandle, startMapIndex: Int, endMapIndex: Int, startPartition: Int, endPartition: Int, context: _root_.org.apache.spark.TaskContext, readMetrics: ShuffleReadMetricsReporter): _root_.org.apache.spark.shuffle.ShuffleReader[K, C] = {
     val blocksByAddress = SparkEnv.get.mapOutputTracker.getMapSizesByExecutorId(
       handle.shuffleId, startMapIndex, endMapIndex, startPartition, endPartition)
-    val env: SparkEnv = SparkEnv.get 
     if (pmofConf.enableRemotePmem) {
       new RpmpShuffleReader(
         handle.asInstanceOf[BaseShuffleHandle[K, _, C]],
@@ -73,6 +72,7 @@ private[spark] class PmofShuffleManager(conf: SparkConf) extends ShuffleManager
         context,
         pmofConf)
     } else if (pmofConf.enableRdma) {
+      val env: SparkEnv = SparkEnv.get 
       metadataResolver = MetadataResolver.getMetadataResolver(pmofConf)
       PmofTransferService.getTransferServiceInstance(pmofConf, env.blockManager, this)
       new RdmaShuffleReader(

diff --git a/core/src/main/scala/org/apache/spark/storage/pmof/PmemBlockInputStream.scala b/core/src/main/scala/org/apache/spark/storage/pmof/PmemBlockInputStream.scala
@@ -14,10 +14,10 @@ trait PmemBlockInputStream[K, C] {
 }
 
 class LocalPmemBlockInputStream[K, C](
-    blockId: BlockId,
-    total_records: Long,
+    pmemBlockOutputStream: PmemBlockOutputStream,
     serializer: Serializer)
     extends PmemBlockInputStream[K, C] {
+  val blockId: BlockId = pmemBlockOutputStream.getBlockId()
   val serializerManager: SerializerManager = SparkEnv.get.serializerManager
   val serInstance: SerializerInstance = serializer.newInstance()
   val persistentMemoryWriter: PersistentMemoryHandler =
@@ -27,7 +27,9 @@ class LocalPmemBlockInputStream[K, C](
   var inObjStream: DeserializationStream = serInstance.deserializeStream(wrappedStream)
 
   var indexInBatch: Int = 0
+  var total_records: Long = 0
   var closing: Boolean = false
+  total_records = pmemBlockOutputStream.getTotalRecords()
 
   def readNextItem(): (K, C) = {
     if (closing == true) {
@@ -126,4 +128,4 @@ class RemotePmemBlockInputStream[K, C](
     }
   }
 
-}
+}
diff --git a/core/src/main/scala/org/apache/spark/storage/pmof/PmemBlockOutputStream.scala b/core/src/main/scala/org/apache/spark/storage/pmof/PmemBlockOutputStream.scala
@@ -95,13 +95,13 @@ private[spark] class PmemBlockOutputStream(
     if ((pmofConf.spill_throttle != -1 && pmemOutputStream.bufferRemainingSize >= pmofConf.spill_throttle) || force == true) {
       val start = System.nanoTime()
       flush()
-      //pmemOutputStream.doFlush()
+      pmemOutputStream.doFlush()
       val bufSize = pmemOutputStream.flushedSize
       mapStatus += ((pmemOutputStream.flushed_block_id, bufSize, recordsPerBlock))
       if (bufSize > 0) {
         recordsArray += recordsPerBlock
         recordsPerBlock = 0
-        size += bufSize
+        size = bufSize
 
         if (blockId.isShuffle == true) {
           val writeMetrics = taskMetrics.shuffleWriteMetrics

diff --git a/core/src/main/scala/org/apache/spark/storage/pmof/PmemOutputStream.scala b/core/src/main/scala/org/apache/spark/storage/pmof/PmemOutputStream.scala
@@ -19,7 +19,7 @@ class PmemOutputStream(
   var is_closed = false
   var key_id = 0
 
-  val length: Int = 1024 * 1024 * 6
+  val length: Int = bufferSize
   var bufferFlushedSize: Int = 0
   var bufferRemainingSize: Int = 0
   val buf: ByteBuf = NettyByteBufferPool.allocateFlexibleNewBuffer(length)
@@ -37,6 +37,10 @@ class PmemOutputStream(
   }
 
   override def flush(): Unit = {
+
+  }
+
+  def doFlush(): Unit = {
     if (bufferRemainingSize > 0) {
       if (remotePersistentMemoryPool != null) {
         logDebug(s" [PUT Started]${cur_block_id}-${bufferRemainingSize}")
@@ -73,10 +77,6 @@ class PmemOutputStream(
     }
   }
 
-  def doFlush(): Unit = {
-
-  }
-
   def flushedSize(): Int = {
     bufferFlushedSize
   }

diff --git a/core/src/main/scala/org/apache/spark/util/collection/pmof/PmemExternalSorter.scala b/core/src/main/scala/org/apache/spark/util/collection/pmof/PmemExternalSorter.scala
@@ -337,8 +337,7 @@ private[spark] class PmemExternalSorter[K, V, C](
     // which is different from spark original codes (relate to one spill file)
     val pmemBlockInputStream = if (!pmofConf.enableRemotePmem) {
       new LocalPmemBlockInputStream[K, C](
-        pmemBlockOutputStream.getBlockId,
-        pmemBlockOutputStream.getTotalRecords,
+        pmemBlockOutputStream,
         serializer)
     } else {
       new RemotePmemBlockInputStream[K, C](