[SPARK-20014] Optimize mergeSpillsWithFileStream method

## What changes were proposed in this pull request?

When the individual partition size in a spill is small, mergeSpillsWithTransferTo method does many small disk ios which is really inefficient. One way to improve the performance will be to use mergeSpillsWithFileStream method by turning off transfer to and using buffered file read/write to improve the io throughput.
However, the current implementation of mergeSpillsWithFileStream does not do a buffer read/write of the files and in addition to that it unnecessarily flushes the output files for each partitions.

## How was this patch tested?

Tested this change by running a job on the cluster and the map stage run time was reduced by around 20%.

Author: Sital Kedia <skedia@fb.com>

Closes #17343 from sitalkedia/upstream_mergeSpillsWithFileStream.

core/src/main/java/org/apache/spark/shuffle/sort/UnsafeShuffleWriter.java

@@ -40,6 +40,7 @@ import org.apache.spark.annotation.Private;
 import org.apache.spark.executor.ShuffleWriteMetrics;
 import org.apache.spark.io.CompressionCodec;
 import org.apache.spark.io.CompressionCodec$;
+import org.apache.spark.io.NioBufferedFileInputStream;
 import org.apache.commons.io.output.CloseShieldOutputStream;
 import org.apache.commons.io.output.CountingOutputStream;
 import org.apache.spark.memory.TaskMemoryManager;
@@ -98,6 +99,18 @@ public class UnsafeShuffleWriter<K, V> extends ShuffleWriter<K, V> {
    */
   private boolean stopping = false;
 
+  private class CloseAndFlushShieldOutputStream extends CloseShieldOutputStream {
+
+    CloseAndFlushShieldOutputStream(OutputStream outputStream) {
+      super(outputStream);
+    }
+
+    @Override
+    public void flush() {
+      // do nothing
+    }
+  }
+
   public UnsafeShuffleWriter(
       BlockManager blockManager,
       IndexShuffleBlockResolver shuffleBlockResolver,
@@ -321,11 +334,15 @@ public class UnsafeShuffleWriter<K, V> extends ShuffleWriter<K, V> {
   }
 
   /**
-   * Merges spill files using Java FileStreams. This code path is slower than the NIO-based merge,
-   * {@link UnsafeShuffleWriter#mergeSpillsWithTransferTo(SpillInfo[], File)}, so it's only used in
-   * cases where the IO compression codec does not support concatenation of compressed data, when
-   * encryption is enabled, or when users have explicitly disabled use of {@code transferTo} in
-   * order to work around kernel bugs.
+   * Merges spill files using Java FileStreams. This code path is typically slower than
+   * the NIO-based merge, {@link UnsafeShuffleWriter#mergeSpillsWithTransferTo(SpillInfo[],
+   * File)}, and it's mostly used in cases where the IO compression codec does not support
+   * concatenation of compressed data, when encryption is enabled, or when users have
+   * explicitly disabled use of {@code transferTo} in order to work around kernel bugs.
+   * This code path might also be faster in cases where individual partition size in a spill
+   * is small and UnsafeShuffleWriter#mergeSpillsWithTransferTo method performs many small
+   * disk ios which is inefficient. In those case, Using large buffers for input and output
+   * files helps reducing the number of disk ios, making the file merging faster.
    *
    * @param spills the spills to merge.
    * @param outputFile the file to write the merged data to.
@@ -339,23 +356,28 @@ public class UnsafeShuffleWriter<K, V> extends ShuffleWriter<K, V> {
     assert (spills.length >= 2);
     final int numPartitions = partitioner.numPartitions();
     final long[] partitionLengths = new long[numPartitions];
-    final InputStream[] spillInputStreams = new FileInputStream[spills.length];
+    final InputStream[] spillInputStreams = new InputStream[spills.length];
 
+    final OutputStream bos = new BufferedOutputStream(
+            new FileOutputStream(outputFile),
+            (int) sparkConf.getSizeAsKb("spark.shuffle.unsafe.file.output.buffer", "32k") * 1024);
     // Use a counting output stream to avoid having to close the underlying file and ask
     // the file system for its size after each partition is written.
-    final CountingOutputStream mergedFileOutputStream = new CountingOutputStream(
-      new FileOutputStream(outputFile));
+    final CountingOutputStream mergedFileOutputStream = new CountingOutputStream(bos);
+    final int inputBufferSizeInBytes = (int) sparkConf.getSizeAsKb("spark.shuffle.file.buffer", "32k") * 1024;
 
     boolean threwException = true;
     try {
       for (int i = 0; i < spills.length; i++) {
-        spillInputStreams[i] = new FileInputStream(spills[i].file);
+        spillInputStreams[i] = new NioBufferedFileInputStream(
+            spills[i].file,
+            inputBufferSizeInBytes);
       }
       for (int partition = 0; partition < numPartitions; partition++) {
         final long initialFileLength = mergedFileOutputStream.getByteCount();
-        // Shield the underlying output stream from close() calls, so that we can close the higher
+        // Shield the underlying output stream from close() and flush() calls, so that we can close the higher
         // level streams to make sure all data is really flushed and internal state is cleaned.
-        OutputStream partitionOutput = new CloseShieldOutputStream(
+        OutputStream partitionOutput = new CloseAndFlushShieldOutputStream(
           new TimeTrackingOutputStream(writeMetrics, mergedFileOutputStream));
         partitionOutput = blockManager.serializerManager().wrapForEncryption(partitionOutput);
         if (compressionCodec != null) {