[SPARK-4548] []SPARK-4517] improve performance of python broadcast

Re-implement the Python broadcast using file:

1) serialize the python object using cPickle, write into disks.
2) Create a wrapper in JVM (for the dumped file), it read data from during serialization
3) Using TorrentBroadcast or HttpBroadcast to transfer the data (compressed) into executors
4) During deserialization, writing the data into disk.
5) Passing the path into Python worker, read data from disk and unpickle it into python object, until the first access.

It fixes the performance regression introduced in #2659, has similar performance as 1.1, but support object larger than 2G, also improve the memory efficiency (only one compressed copy in driver and executor).

Testing with a 500M broadcast and 4 tasks (excluding the benefit from reused worker in 1.2):

         name |   1.1   | 1.2 with this patch |  improvement
---------|--------|---------|--------
      python-broadcast-w-bytes  |	25.20  |	9.33   |	170.13% |
        python-broadcast-w-set	  |     4.13	   |    4.50  |	-8.35%  |

Testing with 100 tasks (16 CPUs):

         name |   1.1   | 1.2 with this patch |  improvement
---------|--------|---------|--------
     python-broadcast-w-bytes	| 38.16	| 8.40	 | 353.98%
        python-broadcast-w-set	| 23.29	| 9.59 |	142.80%

Author: Davies Liu <davies@databricks.com>

Closes #3417 from davies/pybroadcast and squashes the following commits:

50a58e0 [Davies Liu] address comments
b98de1d [Davies Liu] disable gc while unpickle
e5ee6b9 [Davies Liu] support large string
09303b8 [Davies Liu] read all data into memory
dde02dd [Davies Liu] improve performance of python broadcast

1 files changed, 51 insertions, 22 deletions

diff --git a/core/src/main/scala/org/apache/spark/api/python/PythonRDD.scala b/core/src/main/scala/org/apache/spark/api/python/PythonRDD.scala
index b80c771d58..e0bc00e1eb 100644
--- a/core/src/main/scala/org/apache/spark/api/python/PythonRDD.scala
+++ b/core/src/main/scala/org/apache/spark/api/python/PythonRDD.scala
@@ -19,7 +19,7 @@ package org.apache.spark.api.python
 
 import java.io._
 import java.net._
-import java.util.{List => JList, ArrayList => JArrayList, Map => JMap, Collections}
+import java.util.{List => JList, ArrayList => JArrayList, Map => JMap, UUID, Collections}
 
 import org.apache.spark.input.PortableDataStream
 
@@ -47,7 +47,7 @@ private[spark] class PythonRDD(
     pythonIncludes: JList[String],
     preservePartitoning: Boolean,
     pythonExec: String,
-    broadcastVars: JList[Broadcast[Array[Array[Byte]]]],
+    broadcastVars: JList[Broadcast[PythonBroadcast]],
     accumulator: Accumulator[JList[Array[Byte]]])
   extends RDD[Array[Byte]](parent) {
 
@@ -230,8 +230,7 @@ private[spark] class PythonRDD(
           if (!oldBids.contains(broadcast.id)) {
             // send new broadcast
             dataOut.writeLong(broadcast.id)
-            dataOut.writeLong(broadcast.value.map(_.length.toLong).sum)
-            broadcast.value.foreach(dataOut.write)
+            PythonRDD.writeUTF(broadcast.value.path, dataOut)
             oldBids.add(broadcast.id)
           }
         }
@@ -368,24 +367,8 @@ private[spark] object PythonRDD extends Logging {
     }
   }
 
-  def readBroadcastFromFile(
-      sc: JavaSparkContext,
-      filename: String): Broadcast[Array[Array[Byte]]] = {
-    val size = new File(filename).length()
-    val file = new DataInputStream(new FileInputStream(filename))
-    val blockSize = 1 << 20
-    val n = ((size + blockSize - 1) / blockSize).toInt
-    val obj = new Array[Array[Byte]](n)
-    try {
-      for (i <- 0 until n) {
-        val length = if (i < (n - 1)) blockSize else (size % blockSize).toInt
-        obj(i) = new Array[Byte](length)
-        file.readFully(obj(i))
-      }
-    } finally {
-      file.close()
-    }
-    sc.broadcast(obj)
+  def readBroadcastFromFile(sc: JavaSparkContext, path: String): Broadcast[PythonBroadcast] = {
+    sc.broadcast(new PythonBroadcast(path))
   }
 
   def writeIteratorToStream[T](iter: Iterator[T], dataOut: DataOutputStream) {
@@ -824,3 +807,49 @@ private class PythonAccumulatorParam(@transient serverHost: String, serverPort:
     }
   }
 }
+
+/**
+ * An Wrapper for Python Broadcast, which is written into disk by Python. It also will
+ * write the data into disk after deserialization, then Python can read it from disks.
+ */
+private[spark] class PythonBroadcast(@transient var path: String) extends Serializable {
+
+  /**
+   * Read data from disks, then copy it to `out`
+   */
+  private def writeObject(out: ObjectOutputStream): Unit = Utils.tryOrIOException {
+    val in = new FileInputStream(new File(path))
+    try {
+      Utils.copyStream(in, out)
+    } finally {
+      in.close()
+    }
+  }
+
+  /**
+   * Write data into disk, using randomly generated name.
+   */
+  private def readObject(in: ObjectInputStream): Unit = Utils.tryOrIOException {
+    val dir = new File(Utils.getLocalDir(SparkEnv.get.conf))
+    val file = File.createTempFile("broadcast", "", dir)
+    path = file.getAbsolutePath
+    val out = new FileOutputStream(file)
+    try {
+      Utils.copyStream(in, out)
+    } finally {
+      out.close()
+    }
+  }
+
+  /**
+   * Delete the file once the object is GCed.
+   */
+  override def finalize() {
+    if (!path.isEmpty) {
+      val file = new File(path)
+      if (file.exists()) {
+        file.delete()
+      }
+    }
+  }
+}