Install, Configure, and Run Spark on Top of a Hadoop YARN Cluster

1. Should have a hadoop user that can access all cluster nodes with SSH keys without a password.

2. The path of your Hadoop installation should be in /home/hadoop/hadoop. If it is not, adjust the path in the examples accordingly.

3. Run jps on each of the nodes to confirm that HDFS and YARN are running. If they are not, start the services with:

start-dfs.sh

start-yarn.sh

Note:

This guide is written for a non-root user. Commands that require elevated privileges are prefixed with sudo.

Download and Install Spark Binaries

Spark binaries are available from the Apache Spark download page. Adjust each command below to match the correct version number.

1. Get the download URL from the Spark download page, download it, and uncompress it.

For Spark 2.2.0 with Hadoop 2.7 or later, log on node-master as the hadoop user, and run:

cd /home/hadoop wget https://d3kbcqa49mib13.cloudfront.net/spark-2.2.0-bin-hadoop2.7.tgz tar -xvf spark-2.2.0-bin-hadoop2.7.tgz mv spark-2.2.0-bin-hadoop2.7 spark

2. Add the Spark binaries directory to your PATH. Edit /home/hadoop/.profile and add the following line:

For Debian/Ubuntu systems:

/home/hadoop/.profile PATH=/home/hadoop/spark/bin:$PATH

For RedHat/Fedora/CentOS systems:

/home/hadoop/.profile pathmunge /home/hadoop/spark/bin

Integrate Spark with YARN

To communicate with the YARN Resource Manager, Spark needs to be aware of your Hadoop configuration. This is done via the HADOOP_CONF_DIR environment variable. The SPARK_HOME variable is not mandatory, but is useful when submitting Spark jobs from the command line.

1. Edit the hadoop user profile /home/hadoop/.profile and add the following lines:

/home/hadoop/.profile

export HADOOP_CONF_DIR=/home/hadoop/hadoop/etc/hadoop
export SPARK_HOME=/home/hadoop/spark
export LD_LIBRARY_PATH=/home/hadoop/hadoop/lib/native:$LD_LIBRARY_PATH

2. Restart your session by logging out and logging in again.

3. Rename the spark default template config file:

mv $SPARK_HOME/conf/spark-defaults.conf.template $SPARK_HOME/conf/spark-defaults.conf

4. Edit $SPARK_HOME/conf/spark-defaults.conf and set spark.master to yarn:

$SPARK_HOME/conf/spark-defaults.conf

spark.master    yarn

Spark is now ready to interact with your YARN cluster.

Understand Client and Cluster-Mode

Spark jobs can run on YARN in two modes: cluster mode and client mode. Understanding the difference between the two modes is important for choosing an appropriate memory allocation configuration, and to submit jobs as expected.

A Spark job consists of two parts: Spark Executors that run the actual tasks, and a Spark Driver that schedules the Executors.

• Cluster mode: everything runs inside the cluster. You can start a job from your laptop and the job will continue running even if you close your computer. In this mode, the Spark Driver is encapsulated inside the YARN Application Master.
• Client mode the Spark driver runs on a client, such as your laptop. If the client is shut down, the job fails. Spark Executors still run on the cluster, and to schedule everything, a small YARN Application Master is created.

Client mode is well suited for interactive jobs, but applications will fail if the client stops. For long running jobs, cluster mode is more appropriate.

Configure Memory Allocation

Allocation of Spark containers to run in YARN containers may fail if memory allocation is not configured properly. For nodes with less than 4G RAM, the default configuration is not adequate and may trigger swapping and poor performance, or even the failure of application initialization due to lack of memory.

Be sure to understand how Hadoop YARN manages memory allocation before editing Spark memory settings so that your changes are compatible with your YARN cluster’s limits.

Give Your YARN Containers Maximum Allowed Memory

If the memory requested is above the maximum allowed, YARN will reject the creation of the container, and your Spark application won’t start.

Get the value of yarn.scheduler.maximum-allocation-mb in $HADOOP_CONF_DIR/yarn-site.xml. This is the maximum allowed value, in MB, for a single container.

Make sure that values for Spark memory allocation, configured in the following section, are below the maximum.

Configure the Spark Driver Memory Allocation in Cluster-Mode

In cluster mode, the Spark Driver runs inside YARN Application Master. The amount of memory requested by Spark at initialization is configured either in spark-defaults.conf, or through the command line.

From spark-defaults.conf

• Set the default amount of memory allocated to Spark Driver in cluster mode via spark.driver.memory (this value defaults to 1G). To set it to 512MB, edit the file:

$SPARK_HOME/conf/spark-defaults.conf

spark.driver.memory    512m

From the Command Line

• Use the --driver-memory parameter to specify the amount of memory requested by spark-submit. See the following section about application submission for examples.

Note:

Values given from the command line will override whatever has been set in spark-defaults.conf.

Configure the Spark Application Master Memory Allocation in Client Mode

In client mode, the Spark driver will not run on the cluster, so the above configuration will have no effect. A YARN Application Master still needs to be created to schedule the Spark executor, and you can set its memory requirements.

Set the amount of memory allocated to Application Master in client mode with spark.yarn.am.memory (default to 512M)

$SPARK_HOME/conf/spark-defaults.conf

spark.yarn.am.memory    512m 

Configure Spark Executors’ Memory Allocation

The Spark Executors’ memory allocation is calculated based on two parameters inside $SPARK_HOME/conf/spark-defaults.conf:

• spark.executor.memory: sets the base memory used in calculation
• spark.yarn.executor.memoryOverhead: is added to the base memory. It defaults to 7% of base memory, with a minimum of 384MB

Note:

Make sure that Executor requested memory, including overhead memory, is below the YARN container maximum size, otherwise, the Spark application won’t initialize.

To set executor memory to 512MB, edit $SPARK_HOME/conf/spark-defaults.conf and add the following line:

$SPARK_HOME/conf/spark-defaults.conf

spark.executor.memory          512m

How to Submit a Spark Application to the YARN Cluster

Applications are submitted with the spark-submit command. The Spark installation package contains sample applications, like the parallel calculation of Pi, that you can run to practice starting Spark jobs.

To run the sample Pi calculation, use the following command:

 spark-submit --deploy-mode client \
 --class org.apache.spark.examples.SparkPi \
$SPARK_HOME/examples/jars/spark-examples_2.11-2.2.0.jar 10

The first parameter, --deploy-mode, specifies which mode to use, client or cluster.

To run the same application in cluster mode, replace --deploy-mode client with --deploy-mode cluster.

Monitor Your Spark Applications

When you submit a job, Spark Driver automatically starts a web UI on port 4040 that displays information about the application. However, when execution is finished, the Web UI is dismissed with the application driver and can no longer be accessed.

Spark provides a History Server that collects application logs from HDFS and displays them in a persistent web UI. The following steps will enable log persistence in HDFS:

1. Edit $SPARK_HOME/conf/spark-defaults.conf and add the following lines to enable Spark jobs to log in HDFS:

$SPARK_HOME/conf/spark-defaults.conf

spark.eventLog.enabled  true
spark.eventLog.dir hdfs://node-master:9000/spark-logs

2. Create the log directory in HDFS:

hdfs dfs -mkdir /spark-logs

3. Configure History Server related properties in $SPARK_HOME/conf/spark-defaults.conf:

$SPARK_HOME/conf/spark-defaults.conf

spark.history.provider            org.apache.spark.deploy.history.FsHistoryProvider
spark.history.fs.logDirectory     hdfs://node-master:9000/spark-logs
spark.history.fs.update.interval  10s
spark.history.ui.port             18080

You may want to use a different update interval than the default 10s. If you specify a bigger interval, you will have some delay between what you see in the History Server and the real time status of your application. If you use a shorter interval, you will increase I/O on the HDFS.

4. Run the History Server: 

$SPARK_HOME/sbin/start-history-server.sh

5. Repeat steps from previous section to start a job with spark-submit that will generate some logs in the HDFS:

6. Access the History Server by navigating to http://node-master:18080 in a web browser:

Run the Spark ShellPermalink

The Spark shell provides an interactive way to examine and work with your data.

1. Put some data into HDFS for analysis. This example uses the text of Alice In Wonderland from the Gutenberg project:

cd /home/hadoop
wget -O alice.txt https://www.gutenberg.org/files/11/11-0.txt
hdfs dfs -mkdir inputs
hdfs dfs -put alice.txt inputs

2. Start the Spark shell:

spark-shell

var input = spark.read.textFile("inputs/alice.txt")
// Count the number of non blank lines
input.filter(line => line.length()>0).count()