Skip to content

Introduction to Celeborn's Java Columnar Shuffle

Overview

Celeborn presents a Java Columnar Shuffle designed to enhance performance and efficiency in SparkSQL and DataFrame operations. This innovative approach leverages a columnar format for shuffle operations, achieving a higher compression rate than traditional Row-based Shuffle methods. This improvement leads to significant savings in disk space usage during shuffle operations.

Benefits

  • High Compression Rate: By organizing data into a columnar format, this feature significantly increases the compression ratio, reducing the disk space required for Shuffle data. However, enabling this optimization incurs overhead for row-to-column and column-to-row transformations. If disk space is a higher priority, it is recommended to enable this feature. If performance is a higher priority, it is advisable to weigh the trade-offs.

Configuration

To leverage Celeborn's Java Columnar Shuffle, you need to apply a patch and configure certain settings in Spark 3.x. Follow the steps below for implementation:

Step 1: Apply this patch to obtain the schema information for shuffle

  1. Obtain the https://github.com/apache/celeborn/tree/main/assets/spark-patch/Celeborn_Columnar_Shuffle_spark3.patch file that contains the modifications needed for enabling Columnar Shuffle in Spark 3.x.
  2. Navigate to your Spark source directory.
  3. Apply the patch.

Step 2: Configure Celeborn Settings

To enable Columnar Shuffle, adjust the following configurations in your Spark application: Open the Spark configuration file or set these parameters in your Spark application. Add the following configuration settings:

spark.celeborn.columnarShuffle.enabled true
spark.celeborn.columnarShuffle.encoding.dictionary.enabled true

If you require further performance optimization, consider enabling code generation with:

spark.celeborn.columnarShuffle.codegen.enabled true

Each partition buffers rows in memory until spark.celeborn.columnarShuffle.batch.size rows are reached. If rows can be wide, e.g. with long strings, also cap each batch by its estimated size in bytes to limit how much data each partition buffers:

spark.celeborn.columnarShuffle.batch.sizeInBytes.enabled true