Increase maxConcurrencyTask to 10.

Signed-off-by: Pascal Spörri <[email protected]>

README.md

@@ -40,7 +40,7 @@ Changing these values might have an impact on performance.
 
 - `spark.shuffle.s3.bufferSize`: Default buffer size when writing (default: `8388608`)
 - `spark.shuffle.s3.maxBufferSizeTask`: Maximum size of the buffered output streams per task (default: `134217728`)
-- `spark.shuffle.s3.maxConcurrencyTask`: Maximum per task concurrency. Computed by analysing the IO latencies (default: `25`).
+- `spark.shuffle.s3.maxConcurrencyTask`: Maximum per task concurrency. Computed by analysing the IO latencies (default: `10`).
 - `spark.shuffle.s3.cachePartitionLengths`: Cache partition lengths in memory (default: `true`)
 - `spark.shuffle.s3.cacheChecksums`: Cache checksums in memory (default: `true`)
 - `spark.shuffle.s3.cleanup`: Cleanup the shuffle files (default: `true`)

src/main/scala/org/apache/spark/shuffle/helper/S3ShuffleDispatcher.scala

@@ -50,7 +50,7 @@ class S3ShuffleDispatcher extends Logging {
   // Optional
   val bufferSize: Int = conf.getInt("spark.shuffle.s3.bufferSize", defaultValue = 8 * 1024 * 1024)
   val maxBufferSizeTask: Int = conf.getInt("spark.shuffle.s3.maxBufferSizeTask", defaultValue = 128 * 1024 * 1024)
-  val maxConcurrencyTask: Int = conf.getInt("spark.shuffle.s3.maxConcurrencyTask", defaultValue = 5)
+  val maxConcurrencyTask: Int = conf.getInt("spark.shuffle.s3.maxConcurrencyTask", defaultValue = 10)
   val cachePartitionLengths: Boolean = conf.getBoolean("spark.shuffle.s3.cachePartitionLengths", defaultValue = true)
   val cacheChecksums: Boolean = conf.getBoolean("spark.shuffle.s3.cacheChecksums", defaultValue = true)
   val cleanupShuffleFiles: Boolean = conf.getBoolean("spark.shuffle.s3.cleanup", defaultValue = true)