Running GATK4 Spark on Quest

The Broad Institute’s Genome Analysis Toolkit (GATK) is a widely used best practices pipeline for whole genome sequencing and variant calling. As of GATK version 4, many GATK tools are also available to run on Apache Spark, a unified analytics engine for large-scale data processing which can significantly speed up computation time. Note that GATK4’s Spark tools are currently in beta. Because GATK4 SPARK runs on multiple nodes, it must be launched with a job submission script and cannot be run on a login node.

GATK4 SPARK tools

To see a list of available GATK4 tools:

GATK Spark tools have the word “Spark” in their name, and can be explicitly listed with grep:

Additional help is available for each tool with:

Converting FASTA Files for Parallel Runs

Standard FASTA files do not allow for parallel operation and must be converted to 2bit files. To convert these files, use the binary in $SPARK_TOOLS which is in the SPARK module path.

This step only needs to be done once.

Example Job Submission

Below is an example of job submission file for Quest using a converted exampleFASTA.2bit file.

GATKSpark_example.sh

#!/bin/bash
#SBATCH -A <allocation>      # Allocation 
#SBATCH -p <partition_name>  # Partition
#SBATCH -t 00:20:00          # Walltime/duration of job
#SBATCH -N 2                 # Number of nodes
#SBATCH --ntasks-per-node=24 # Number of cores (processors)
#SBATCH --mem-per-cpu=5G     # GB needed per-core for a job
#SBATCH -J "SparkTest"       # Name of job

# Load environment
module purge all
module load spark/2.3.0 gatk/4.0.4

cd $SLURM_SUBMIT_DIR

# Initialize spark cluster on hosts allocated to your job
$SPARK_TOOLS/initialize_spark.sh

# Run GATK HaplotypeCaller in Spark
gatk HaplotypeCallerSpark \
--reference $SLURM_SUBMIT_DIR/exampleFASTA.2bit \
--input $SLURM_SUBMIT_DIR/exampleBAM.bam \
--output $SLURM_SUBMIT_DIR/exampleVCF.txt \
--spark-runner SPARK \
--spark-master spark://`hostname -i`:7077 \
-- --driver-cores=2 --driver-memory=6g \
--executor-cores=22 --executor-memory=114GB 2>&1

# Cleanup spark cluster on hosts allocated to your job

$SPARK_TOOLS/cleanup_spark.sh