HowTo:spark

Apache Spark is a fast and general engine for big data processing, with built-in modules for streaming, SQL, machine learning and graph processing. This page is a how to guide on using Apache Spark on the SW cluster.

Running a Spark job requires several steps:

• To load the Spark libraries and scripts to your path, run the command: use spark.

• To setup scratch disks appropriately, run source /opt/global/spark/setup.sh

• Run your application in Spark using spark-submit. Make sure to set proper values for executor and driver memory or you will likely experience memory-related errors! --master local[#] sets up and uses a local standalone Spark cluster with # workers for your job (using "$NSLOTS" for "#" automatically sets the number of workers to the number of cores requested by your job).

Note that HDFS is not installed on the SW cluster. To use the cluster filesystem instead, use file:///path/to/file instead of hdfs://host:8020/path/to/file in your scripts.

Template job

The following template script (found at /opt/global/spark/spark-job.sh) will work as a good job template for a typical Spark job. You must edit this script to reflect the proper number of cores and memory you will use, or your job may be scheduled with sub-optimal resources. Note that shm.pe can be substituted for glinux.pe for access to larger machines (in terms of memory and CPUs), but this comes at a cost of not having local scratch disks, which may dramatically slow down performance for jobs requiring lots of I/O (Spark also performs caching on disk).

This particular job will start a Spark application on a node with 55GB of memory, 4 cpus, and local scratch disks for improved I/O performance. STDOUT and STDERR will be written to nameOfJob.oJob# and nameOfJob.eJob#, respectively. The "anaconda2" python distribution will be used for "pyspark".

#!/bin/bash
#$ -S /bin/bash
#$ -o $JOB_NAME.o$JOB_ID
#$ -e $JOB_NAME.e$JOB_ID
#$ -cwd
#$ -V
#$ -q abaqus.q
#$ -l qname=abaqus.q

# This line should be equal to the total executor and driver memory used.
#$ -l mf=55G

# Number of worker cores you want to use
#$ -pe glinux.pe 4

# Setup your environment to use Spark
use spark
source /opt/global/spark/setup.sh

# If using pyspark, you will likely want to use one of the Anaconda python distributions installed:
use anaconda3
use anaconda2

# Edit this line to change how your job is run:
spark-submit --master local[$NSLOTS] --executor-memory 50G --driver-memory 5G example-spark-application.py

Using Python 3 instead of Python 2

You can use Python 3 instead of Python 2 by setting the environment variable PYSPARK_PYTHON.

hpc####@swlogin1$ use spark
hpc####@swlogin1$ use anaconda3
hpc####@swlogin1$ export PYSPARK_PYTHON=python3
hpc####@swlogin1$ pyspark
Python 3.4.5 |Anaconda 2.3.0 (64-bit)| (default, Jul  2 2016, 17:47:47) 
[GCC 4.4.7 20120313 (Red Hat 4.4.7-1)] on linux
Type "help", "copyright", "credits" or "license" for more information.
Using Spark's default log4j profile: org/apache/spark/log4j-defaults.properties
Setting default log level to "WARN".
To adjust logging level use sc.setLogLevel(newLevel).
17/01/13 15:56:48 WARN NativeCodeLoader: Unable to load native-hadoop library for your platform... using builtin-java classes where applicable
Welcome to
      ____              __
     / __/__  ___ _____/ /__
    _\ \/ _ \/ _ `/ __/  '_/
   /__ / .__/\_,_/_/ /_/\_\   version 2.0.1
      /_/

Using Python version 3.4.5 (default, Jul  2 2016 17:47:47)
SparkSession available as 'spark'.
>>> 

Troubleshooting

If you see out-of-memory errors, increase --executor-memory and --driver-memory accordingly.