Spark

Description

spark website  

Apache Spark is a fast and general-purpose cluster computing system. It provides high-level APIs in Java, Scala, Python and R, and an optimized engine that supports general execution graphs. It also supports a rich set of higher-level tools including Spark SQL for SQL and structured data processing, MLlib for machine learning, GraphX for graph processing, and Spark Streaming.

Environment Modules

Run module spider spark to find out what environment modules are available for this application.

System Variables

• HPC_SPARK_DIR - installation directory
• HPC_SPARK_BIN - executable directory
• HPC_SPARK_SLURM - SLURM job script examples
• SPARK_HOME - examples directory

Running Spark on HiperGator

To run your Spark jobs on HiperGator, two separate steps are required:

1. Create a Spark cluster on HiperGator via SLURM. This section "Spark Cluster on HiPerGator" below shows a simple example how to create a Spark cluster on HiperGator.
2. Submit your job to your Spark cluster. You can do this either interactively at the command line ("Spark Interactive Job" section below) or by submitting a a batch job ("Spark Batch Job" section below)

For details about running Spark jobs on HiPerGator, please refer to Spark Workshop. For Spark parameters used in this section, please refer to Spark's homepage.

Spark cluster on HiperGator

Expand this section to view instructions for creating a spark cluster in HiperGator.

It is assumed that spark-local-cluster.sh is the file name of the SLURM job script for one-worker node Spark cluster in this section. Set SLURM parameters for Spark cluster. spark-local-cluster.sh is available on "Spark_Job_Scripts" page below.

 #!/bin/bash
 #filename: spark-local-cluster.sh
 #SBATCH --job-name=spark_cluster
 #SBATCH --nodes=1 # nodes allocated to the job
 #SBATCH --cpus-per-task=16 # the number of CPUs allocated per task
 #SBATCH --exclusive # not sharing of allocated nodes with other running jobs
 #SBATCH --time=03:00:00
 #SBATCH --output=spark_cluster.log
 #SBATCH --error=spark_cluster.err

 module load spark

 ## Set Spark parameters for Spark cluster
 export SPARK_LOCAL_DIRS=$HOME/spark/tmp
 export SPARK_WORKER_DIR=$SPARK_LOCAL_DIRS
 export SPARK_WORKER_CORES=$SLURM_CPUS_PER_TASK
 export SPARK_MASTER_PORT=7077
 export SPARK_MASTER_WEBUI_PORT=8080
 export SPARK_NO_DAEMONIZE=true
 export SPARK_LOG_DIR=$SPARK_LOCAL_DIRS

 mkdir -p $SPARK_LOCAL_DIRS

 ##Set Spark Master and Workers
 MASTER_HOST=$(scontrol show hostname $SLURM_NODELIST | head -n 1)
 export SPARK_MASTER_NODE=$(host $MASTER_HOST | head -1 | cut -d ' ' -f 4)
 export MAX_SLAVES=$(expr $SLURM_JOB_NUM_NODES - 1)

 ## for starting spark master
 $SPARK_HOME/sbin/start-master.sh & 

 ## use spark defaults for worker resources (all mem -1 GB, all cores) since using exclusive
 ## for starting spark worker
 $SPARK_HOME/sbin/start-slave.sh spark://$SPARK_MASTER_NODE:$SPARK_MASTER_PORT

Submit the SLURM job script to HiperGator

sbatch spark-local-cluster.sh

Check the Spark master launched.

grep "Starting Spark master" spark_cluster.err

This grep command above should end up with information like

18/03/13 14:53:23 INFO Master: Starting Spark master at spark://c29a-s42.ufhpc:7077

Check the Spark worker launched.

grep "Starting Spark worker" spark_cluster.err

This grep command above should end up with information like

18/03/13 14:53:24 INFO Worker: Starting Spark worker 172.16.194.59:42418 with 16 cores, 124.3 GB RAM

Spark interactive job

Expand this section to view instructions for starting preset applications without a job script.

Spark supports interactive job submission through the interactive shells.

Spark interactive shell in Scalar (spark-shell)

First, load spark module in the terminal where you want to submit a spark job.

module load spark

Get the location of the Spark master to connect to it through the interactive shell

SPARK_MASTER=$(grep "Starting Spark master" *.err | cut -d " " -f 9)

Connect to the master using the Spark interactive shell in scalar

spark-shell --master $SPARK_MASTER

Spark interactive shell in Python (pyspark)

Load spark module in the terminal where you want to submit a spark job.

module load spark

Get the location of the Spark master to connect to it through the interactive shell

SPARK_MASTER=$(grep "Starting Spark master" *.err | cut -d " " -f 9)

Connect to the master using the Spark interactive shell in scalar

pyspark --master $SPARK_MASTER

Example - PI estimation via pyspark

SPARK_MASTER=$(grep "Starting Spark master" *.err | cut -d " " -f 9)
pyspark --master $SPARK_MASTER

Example - Pi estimation from file with pyspark

As of Spark 2.0., Spark interactive shell in python does not load python files to run python application. Instead, “PYTHONSTARTUP”, a python environmental variable can be used to run python script with pyspark, which executes the python script before an interactive shell starts.

SPARK_MASTER=$(grep "Starting Spark master" *.err | cut -d " " -f 9)
PYTHONSTARTUP=pi_with_pythonstartup.py pyspark --master $SPARK_MASTER

pi_with_pythonstartup.py script is avaialble on "Spark_Job_Scripts" page below.

Spark batch job

Expand this section to view instructions for starting preset applications without a job script.

Spark supports batch job submission through spark-submit which provides unified interface for Spark jobs

$SPARK_HOME/bin/spark-submit  \
             --class <main-class>  --master <master-url> \
             --deploy-mode <deploy-mode>  --conf <key>=<value>  \
            ... # other options  <application-jar>  [application-arguments]

--class: The entry point for your application (e.g. org.apache.spark.examples.SparkPi)
--master: The master URL for the cluster (e.g. spark://123.45.67.890:7077)
--deploy-mode: Whether to deploy your driver on the worker nodes (cluster) or locally as an external client (client) (default: client)
--conf: Arbitrary Spark configuration property in key=value format. For values that contain spaces wrap “key=value” in quotes (as shown).
<application-jar>: Path to a bundled jar including your application and all dependencies. The URL must be globally visible inside of your cluster, 
  for instance, an hdfs:// path or a file:// path that is present on all nodes.
<application-arguments>: Arguments passed to the main method of your main class, if any

For further details about spark-submit, refer to https://spark.apache.org/docs/2.2.0/submitting-applications.html.

Example - Pi estimation via Spark-submit

SPARK_MASTER=$(grep "Starting Spark master" *.err | cut -d " " -f 9)
spark-submit --master $SPARK_MASTER $SPARK_HOME/examples/src/main/python/pi.py 10

Job Script Examples

Expand this section to view spark-local-cluster.sh

#!/bin/bash
#filename: spark-local-cluster.sh

#SBATCH --job-name=spark_cluster
#SBATCH --nodes=1 # nodes allocated to the job
#SBATCH --cpus-per-task=16 # the number of CPUs allocated per task
#SBATCH --exclusive # not sharing of allocated nodes with other running jobs
#SBATCH --time=03:00:00
#SBATCH --output=spark_cluster.log
#SBATCH --error=spark_cluster.err

###SBATCH --ntasks= # tasks to be created for the job
###SBATCH --ntasks-per-core= # max number of tasks per allocated core
###SBATCH --ntasks-per-node= # max number of tasks per allocated node
###SBATCH --mail-type=END,FAIL
###SBATCH --mail-user=<yourID>@ufl.edu

module load spark
### Set Spark variables
export SPARK_LOCAL_DIRS=$HOME/spark/tmp
export SPARK_WORKER_DIR=$SPARK_LOCAL_DIRS
export SPARK_WORKER_CORES=$SLURM_CPUS_PER_TASK
export SPARK_MASTER_PORT=7077
export SPARK_MASTER_WEBUI_PORT=8080
export SPARK_NO_DAEMONIZE=true
export SPARK_LOG_DIR=$SPARK_LOCAL_DIRS
#export SPARK_CONF_DIR=$SPARK_LOCAL_DIRS
mkdir -p $SPARK_LOCAL_DIRS

MASTER_HOST=$(scontrol show hostname $SLURM_NODELIST | head -n 1)
export SPARK_MASTER_NODE=$(host $MASTER_HOST | head -1 | cut -d ' ' -f 4)
export MAX_SLAVES=$(expr $SLURM_JOB_NUM_NODES - 1)

# start master
$SPARK_HOME/sbin/start-master.sh &

# start workers
# use spark defaults for worker resources (all mem -1 GB, all cores) since using exclusive

$SPARK_HOME/sbin/start-slave.sh spark://$SPARK_MASTER_NODE:$SPARK_MASTER_PORT

Expand this section to view pi_with_pythonstartup.py

from operator import add
from random import random

partitions =10
n = 100000 * partitions

def f(_):
    x = random() * 2 - 1
    y = random() * 2 - 1
    return 1 if x ** 2 + y ** 2 <= 1 else 0

count = sc.parallelize(range(1, n + 1), partitions).map(f).reduce(add)
print("Pi is roughly %f" % (4.0 * count / n))