Invoking SystemML in Hadoop Batch Mode

Overview
Hadoop Batch Mode Invocation Syntax
SystemML with Standalone Hadoop
SystemML with Pseudo-Distributed Hadoop
SystemML with Pseudo-Distributed Hadoop and YARN
SystemML with Distributed Hadoop and YARN
- SystemML with Distributed Hadoop and YARN: Linear Regression Example
- SystemML with Distributed Hadoop and YARN: K-Means Clustering Example
Recommended Hadoop Cluster Configuration Settings

Overview

Given that a primary purpose of SystemML is to perform machine learning on large distributed data sets, two of the most important ways to invoke SystemML are Hadoop Batch and Spark Batch modes. Here, we will look at SystemML’s Hadoop Batch mode in more depth.

We will look at running SystemML with Standalone Hadoop, Pseudo-Distributed Hadoop, and Distributed Hadoop. We will first run SystemML on a single machine with Hadoop running in Standalone mode. Next, we’ll run SystemML on HDFS in Hadoop’s Pseudo-Distributed mode on a single machine, followed by Pseudo-Distributed mode with YARN. After that, we’ll set up a 4-node Hadoop cluster and run SystemML on Distributed Hadoop with YARN.

Note that this tutorial does not address security. For security considerations with regards to Hadoop, please refer to the Hadoop documentation.

Hadoop Batch Mode Invocation Syntax

SystemML can be invoked in Hadoop Batch mode using the following syntax:

hadoop jar SystemML.jar [-? | -help | -f <filename>] (-config <config_filename>) ([-args | -nvargs] <args-list>)

The SystemML.jar file is specified to Hadoop using the jar option. The DML script to invoke is specified after the -f argument. Configuration settings can be passed to SystemML using the optional -config argument. DML scripts can optionally take named arguments (-nvargs) or positional arguments (-args). Named arguments are preferred over positional arguments. Positional arguments are considered to be deprecated. All the primary algorithm scripts included with SystemML use named arguments.

Example #1: DML Invocation with Named Arguments

hadoop jar systemml/SystemML.jar -f systemml/algorithms/Kmeans.dml -nvargs X=X.mtx k=5

Example #2: DML Invocation with Positional Arguments

hadoop jar systemml/SystemML.jar -f example/test/LinearRegression.dml -args "v" "y" 0.00000001 "w"

In a clustered environment, it is highly recommended that SystemML configuration settings are specified in a SystemML-config.xml file. By default, SystemML will look for this file in the current working directory (./SystemML-config.xml). This location can be overridden by the -config argument.

Example #3: DML Invocation with Configuration File Explicitly Specified and Named Arguments

hadoop jar systemml/SystemML.jar -f systemml/algorithms/Kmeans.dml -config /conf/SystemML-config.xml -nvargs X=X.mtx k=5

For recommended SystemML configuration settings in a clustered environment, please see Recommended Hadoop Cluster Configuration Settings.

SystemML with Standalone Hadoop

In Standalone mode, Hadoop runs on a single machine as a single Java process.

To begin, I connected to my Linux server as root and created a hadoop user.

$ ssh root@host1.example.com
[root@host1 ~]# useradd hadoop
[root@host1 ~]# passwd hadoop

Next, I logged on as the hadoop user. I downloaded the version of Hadoop that I wanted to use from an Apache mirror. A list of Hadoop releases can be found at the Apache Hadoop Releases website. After downloading the Hadoop binary release, I unpacked it.

$ ssh hadoop@host1.example.com
[hadoop@host1 ~]$ wget http://mirror.sdunix.com/apache/hadoop/common/hadoop-2.6.2/hadoop-2.6.2.tar.gz
[hadoop@host1 ~]$ tar -xvzf hadoop-2.6.2.tar.gz

My Linux server already had a JDK (Java Development Kit) installed. If you haven’t done so already, you will need Java installed in order to use Hadoop.

I updated my .bash_profile file to export a JAVA_HOME environment variable, which I pointed to my JDK installation directory. I also exported a HADOOP_HOME environment variable, which points to the root directory of the Hadoop release that I unpacked. I updated the PATH variable to include the JAVA_HOME bin directory, the HADOOP_HOME bin directory, and the HADOOP_HOME sbin directory.

[hadoop@host1 ~]# vi .bash_profile

...
export JAVA_HOME=/usr/lib/jvm/java-1.7.0-openjdk.x86_64
export HADOOP_HOME=/home/hadoop/hadoop-2.6.2
PATH=$JAVA_HOME/bin:$PATH:$HOME/bin:$HADOOP_HOME/bin:$HADOOP_HOME/sbin
export PATH
...

[hadoop@host1 ~]$ source ~/.bash_profile

To verify that Java and Hadoop were on the path, I used the java -version and hadoop version commands.

[hadoop@host1 ~]$ java -version
java version "1.7.0_79"
OpenJDK Runtime Environment (rhel-2.5.5.1.el6_6-x86_64 u79-b14)
OpenJDK 64-Bit Server VM (build 24.79-b02, mixed mode)
[hadoop@host1 ~]$ hadoop version
Hadoop 2.6.2
Subversion https://git-wip-us.apache.org/repos/asf/hadoop.git -r 0cfd050febe4a30b1ee1551dcc527589509fb681
Compiled by jenkins on 2015-10-22T00:42Z
Compiled with protoc 2.5.0
From source with checksum f9ebb94bf5bf9bec892825ede28baca
This command was run using /home/hadoop/hadoop-2.6.2/share/hadoop/common/hadoop-common-2.6.2.jar

Next, I downloaded a SystemML release from the downloads page. Following this, I unpacked it.

[hadoop@host1 ~]$ tar -xvzf systemml-1.2.0-SNAPSHOT.tar.gz

Alternatively, we could have built the SystemML distributed release using Apache Maven and unpacked it.

[hadoop@host1 ~]$ git clone https://github.com/apache/systemml.git
[hadoop@host1 ~]$ cd systemml
[hadoop@host1 systemml]$ mvn clean package -P distribution
[hadoop@host1 systemml]$ tar -xvzf target/systemml-1.2.0-SNAPSHOT.tar.gz -C ..
[hadoop@host1 ~]$ cd ..

I downloaded the genLinearRegressionData.dml script that is used in the SystemML README example.

[hadoop@host1 ~]$ wget https://raw.githubusercontent.com/apache/systemml/master/scripts/datagen/genLinearRegressionData.dml

Next, I invoked the genLinearRegressionData.dml DML script in Hadoop Batch mode. Hadoop was executed with the SystemML.jar file specified by the hadoop jar option. The genLinearRegressionData.dml was specified using the -f option. Named input arguments to the DML script were specified following the -nvargs option.

[hadoop@host1 ~]$ hadoop jar systemml-1.2.0-SNAPSHOT/SystemML.jar -f genLinearRegressionData.dml -nvargs numSamples=1000 numFeatures=50 maxFeatureValue=5 maxWeight=5 addNoise=FALSE b=0 sparsity=0.7 output=linRegData.csv format=csv perc=0.5
15/11/11 15:56:21 INFO api.DMLScript: BEGIN DML run 11/11/2015 15:56:21
15/11/11 15:56:21 INFO api.DMLScript: HADOOP_HOME: /home/hadoop/hadoop-2.6.2
15/11/11 15:56:21 WARN conf.DMLConfig: No default SystemML config file (./SystemML-config.xml) found
15/11/11 15:56:21 WARN conf.DMLConfig: Using default settings in DMLConfig
15/11/11 15:56:22 INFO jvm.JvmMetrics: Initializing JVM Metrics with processName=JobTracker, sessionId=
15/11/11 15:56:22 WARN hops.OptimizerUtils: Auto-disable multi-threaded text read for 'text' and 'csv' due to thread contention on JRE < 1.8 (java.version=1.7.0_79).
15/11/11 15:56:22 INFO api.DMLScript: SystemML Statistics:
Total execution time:		0.288 sec.
Number of executed MR Jobs:	0.

15/11/11 15:56:22 INFO api.DMLScript: END DML run 11/11/2015 15:56:22

In the console output, we see a warning that no default SystemML config file was found in the current working directory. In a distributed environment on a large data set, it is highly advisable to specify configuration settings in a SystemML config file for optimal performance. The location of the SystemML config file can be explicitly specified using the -config argument.

The OptimizerUtils warning occurs because parallel multi-threaded text reads in Java versions less than 1.8 result in thread contention issues, so only a single thread reads matrix data in text formats.

If we examine the contents of the directory, we see that linRegData.csv and perc.csv were written to the file system, along with their corresponding metadata files. The scratch_space directory is used to write temporary matrix files.

[hadoop@host1 ~]$ ls -l
total 197500
-rw-rw-r-- 1 hadoop hadoop      2208 Nov 11 15:45 genLinearRegressionData.dml
drwxr-xr-x 9 hadoop hadoop      4096 Oct 21 17:53 hadoop-2.6.2
-rw-rw-r-- 1 hadoop hadoop 195515434 Oct 30 14:04 hadoop-2.6.2.tar.gz
drwxrwxrwx 2 hadoop hadoop      4096 Nov 11 15:56 linRegData.csv
-rw-r--r-- 1 hadoop hadoop       214 Nov 11 15:56 linRegData.csv.mtd
drwxrwxrwx 2 hadoop hadoop      4096 Nov 11 15:56 perc.csv
-rw-r--r-- 1 hadoop hadoop       206 Nov 11 15:56 perc.csv.mtd
drwxrwxrwx 2 hadoop hadoop      4096 Nov 11 15:56 scratch_space
drwxrwxr-x 4 hadoop hadoop      4096 Nov 11 15:42 systemml-1.2.0-SNAPSHOT
-rw-rw-r-- 1 hadoop hadoop   6683281 Oct 27 21:13 systemml-1.2.0-SNAPSHOT.tar.gz

To clean things up, I’ll delete the files that were generated.

[hadoop@host1 ~]$ rm -r *.csv
[hadoop@host1 ~]$ rm *.csv.mtd
[hadoop@host1 ~]$ rmdir scratch_space/

SystemML with Pseudo-Distributed Hadoop

Next, we’ll look at running SystemML with Hadoop in Pseudo-Distributed mode. In Pseudo-Distributed mode, each Hadoop daemon (such as NameNode and DataNode) runs in a separate Java process on a single machine.

In the previous section about Hadoop Standalone mode, we set up the JAVA_HOME and HADOOP_HOME environment variables and added JAVA_HOME/bin, HADOOP_HOME/bin, and HADOOP_HOME/sbin to the PATH in .bash_profile.

We also need to set the JAVA_HOME value in the hadoop-env.sh file in the Hadoop configuration directory (etc/hadoop).

[hadoop@host1 hadoop]$ pwd
/home/hadoop/hadoop-2.6.2/etc/hadoop
[hadoop@host1 hadoop]$ vi hadoop-env.sh

...
export JAVA_HOME=/usr/lib/jvm/java-1.7.0-openjdk.x86_64
...

We need to be able to passwordlessly ssh to localhost. To do so, I’ll generate a public key/private key pair and add the public key to the hadoop user’s authorized_keys. We can ssh to localhost to verify that we can connect without a password.

[hadoop@host1 ~]$ ssh-keygen -t rsa -b 4096 -C "hadoop example"
Your identification has been saved in /home/hadoop/.ssh/id_rsa.
Your public key has been saved in /home/hadoop/.ssh/id_rsa.pub.
[hadoop@host1 ~]$ cat ~/.ssh/id_rsa.pub >> ~/.ssh/authorized_keys
[hadoop@host1 ~]$ chmod 600 ~/.ssh/authorized_keys
[hadoop@host1 ~]$ ssh localhost
The authenticity of host 'localhost (::1)' can't be established.
RSA key fingerprint is 6b:86:78:86:13:0a:49:d4:c7:a7:15:10:d1:27:88:9e.
Are you sure you want to continue connecting (yes/no)? yes
Warning: Permanently added 'localhost' (RSA) to the list of known hosts.
[hadoop@host1 ~]$ exit
logout
Connection to localhost closed.
[hadoop@host1 ~]$ ls -l .ssh
total 16
-rw------- 1 hadoop hadoop  736 Nov 11 16:44 authorized_keys
-rw------- 1 hadoop hadoop 3243 Nov 11 16:41 id_rsa
-rw-r--r-- 1 hadoop hadoop  736 Nov 11 16:41 id_rsa.pub
-rw-r--r-- 1 hadoop hadoop  391 Nov 11 16:46 known_hosts

In the Hadoop configuration directory (etc/hadoop), in the core-site.xml file, we specify the fs.defaultFS property to be localhost with port 9000.

[hadoop@host1 hadoop]$ vi core-site.xml

...
<configuration>
    <property>
        <name>fs.defaultFS</name>
        <value>hdfs://localhost:9000</value>
    </property>
</configuration>
...

By default, HDFS replicates data on three nodes. Since we’re running on a single machine, we’ll change this to one. We’ll add a dfs.replication property to hdfs-site.xml and set its value to 1.

[hadoop@host1 hadoop]$ vi hdfs-site.xml

...
<configuration>
    <property>
        <name>dfs.replication</name>
        <value>1</value>
    </property>
</configuration>
...

Next, we’ll format HDFS.

[hadoop@host1 ~]$ hdfs namenode -format
15/11/11 17:23:33 INFO namenode.NameNode: STARTUP_MSG:
/************************************************************
STARTUP_MSG: Starting NameNode
STARTUP_MSG:   host = host1.example.com/9.30.252.15
STARTUP_MSG:   args = [-format]
STARTUP_MSG:   version = 2.6.2
...
STARTUP_MSG:   java = 1.7.0_79
************************************************************/
...
15/11/11 17:23:34 INFO common.Storage: Storage directory /tmp/hadoop-hadoop/dfs/name has been successfully formatted.
...
/************************************************************
SHUTDOWN_MSG: Shutting down NameNode at host1.example.com/9.30.252.15
************************************************************/

We’ll start up HDFS using the start-dfs.sh script. This starts the NameNode, DataNode, and SecondaryNameNode daemons on the single machine.

[hadoop@host1 ~]$ start-dfs.sh
Starting namenodes on [localhost]
localhost: starting namenode, logging to /home/hadoop/hadoop-2.6.2/logs/hadoop-hadoop-namenode-host1.out
localhost: starting datanode, logging to /home/hadoop/hadoop-2.6.2/logs/hadoop-hadoop-datanode-host1.out
Starting secondary namenodes [0.0.0.0]
The authenticity of host '0.0.0.0 (0.0.0.0)' can't be established.
RSA key fingerprint is 6b:86:78:86:13:0a:49:d4:c7:a7:15:10:d1:27:88:9e.
Are you sure you want to continue connecting (yes/no)? yes
0.0.0.0: Warning: Permanently added '0.0.0.0' (RSA) to the list of known hosts.
0.0.0.0: starting secondarynamenode, logging to /home/hadoop/hadoop-2.6.2/logs/hadoop-hadoop-secondarynamenode-host1.out

We can see the running Java processes using the jps command.

[hadoop@host1 ~]$ jps
Jps
DataNode
SecondaryNameNode
NameNode

Here, we can see detailed information about the Java processes that were started.

[hadoop@host1 ~]$ ps -C java -f -ww
UID        PID  PPID  C STIME TTY          TIME CMD
hadoop  35722     1  5 17:38 ?        00:00:05 /usr/lib/jvm/java-1.7.0-openjdk.x86_64/bin/java -Dproc_namenode -Xmx1000m -Djava.net.preferIPv4Stack=true -Dhadoop.log.dir=/home/hadoop/hadoop-2.6.2/logs -Dhadoop.log.file=hadoop.log -Dhadoop.home.dir=/home/hadoop/hadoop-2.6.2 -Dhadoop.id.str=hadoop -Dhadoop.root.logger=INFO,console -Djava.library.path=/home/hadoop/hadoop-2.6.2/lib/native -Dhadoop.policy.file=hadoop-policy.xml -Djava.net.preferIPv4Stack=true -Djava.net.preferIPv4Stack=true -Djava.net.preferIPv4Stack=true -Dhadoop.log.dir=/home/hadoop/hadoop-2.6.2/logs -Dhadoop.log.file=hadoop-hadoop-namenode-host1.log -Dhadoop.home.dir=/home/hadoop/hadoop-2.6.2 -Dhadoop.id.str=hadoop -Dhadoop.root.logger=INFO,RFA -Djava.library.path=/home/hadoop/hadoop-2.6.2/lib/native -Dhadoop.policy.file=hadoop-policy.xml -Djava.net.preferIPv4Stack=true -Dhadoop.security.logger=INFO,RFAS -Dhdfs.audit.logger=INFO,NullAppender -Dhadoop.security.logger=INFO,RFAS -Dhdfs.audit.logger=INFO,NullAppender -Dhadoop.security.logger=INFO,RFAS -Dhdfs.audit.logger=INFO,NullAppender -Dhadoop.security.logger=INFO,RFAS org.apache.hadoop.hdfs.server.namenode.NameNode
hadoop  35844     1  4 17:38 ?        00:00:04 /usr/lib/jvm/java-1.7.0-openjdk.x86_64/bin/java -Dproc_datanode -Xmx1000m -Djava.net.preferIPv4Stack=true -Dhadoop.log.dir=/home/hadoop/hadoop-2.6.2/logs -Dhadoop.log.file=hadoop.log -Dhadoop.home.dir=/home/hadoop/hadoop-2.6.2 -Dhadoop.id.str=hadoop -Dhadoop.root.logger=INFO,console -Djava.library.path=/home/hadoop/hadoop-2.6.2/lib/native -Dhadoop.policy.file=hadoop-policy.xml -Djava.net.preferIPv4Stack=true -Djava.net.preferIPv4Stack=true -Djava.net.preferIPv4Stack=true -Dhadoop.log.dir=/home/hadoop/hadoop-2.6.2/logs -Dhadoop.log.file=hadoop-hadoop-datanode-host1.log -Dhadoop.home.dir=/home/hadoop/hadoop-2.6.2 -Dhadoop.id.str=hadoop -Dhadoop.root.logger=INFO,RFA -Djava.library.path=/home/hadoop/hadoop-2.6.2/lib/native -Dhadoop.policy.file=hadoop-policy.xml -Djava.net.preferIPv4Stack=true -server -Dhadoop.security.logger=ERROR,RFAS -Dhadoop.security.logger=ERROR,RFAS -Dhadoop.security.logger=ERROR,RFAS -Dhadoop.security.logger=INFO,RFAS org.apache.hadoop.hdfs.server.datanode.DataNode
hadoop  36007     1  5 17:38 ?        00:00:04 /usr/lib/jvm/java-1.7.0-openjdk.x86_64/bin/java -Dproc_secondarynamenode -Xmx1000m -Djava.net.preferIPv4Stack=true -Dhadoop.log.dir=/home/hadoop/hadoop-2.6.2/logs -Dhadoop.log.file=hadoop.log -Dhadoop.home.dir=/home/hadoop/hadoop-2.6.2 -Dhadoop.id.str=hadoop -Dhadoop.root.logger=INFO,console -Djava.library.path=/home/hadoop/hadoop-2.6.2/lib/native -Dhadoop.policy.file=hadoop-policy.xml -Djava.net.preferIPv4Stack=true -Djava.net.preferIPv4Stack=true -Djava.net.preferIPv4Stack=true -Dhadoop.log.dir=/home/hadoop/hadoop-2.6.2/logs -Dhadoop.log.file=hadoop-hadoop-secondarynamenode-host1.log -Dhadoop.home.dir=/home/hadoop/hadoop-2.6.2 -Dhadoop.id.str=hadoop -Dhadoop.root.logger=INFO,RFA -Djava.library.path=/home/hadoop/hadoop-2.6.2/lib/native -Dhadoop.policy.file=hadoop-policy.xml -Djava.net.preferIPv4Stack=true -Dhadoop.security.logger=INFO,RFAS -Dhdfs.audit.logger=INFO,NullAppender -Dhadoop.security.logger=INFO,RFAS -Dhdfs.audit.logger=INFO,NullAppender -Dhadoop.security.logger=INFO,RFAS -Dhdfs.audit.logger=INFO,NullAppender -Dhadoop.security.logger=INFO,RFAS org.apache.hadoop.hdfs.server.namenode.SecondaryNameNode

Useful log information is created by default in the hadoop logs directory.

If everything worked correctly, we can hit port 50070 in a browser (http://host1.example.com:50070) to see Hadoop information.

If we look at our HDFS file system, we see that it currently doesn’t contain any files.

[hadoop@host1 ~]$ hdfs dfs -ls
ls: `.': No such file or directory

Let’s go ahead and execute the genLinearRegressionData.dml script in Hadoop Pseudo-Distributed mode.

[hadoop@host1 ~]$ hadoop jar systemml-1.2.0-SNAPSHOT/SystemML.jar -f genLinearRegressionData.dml -nvargs numSamples=1000 numFeatures=50 maxFeatureValue=5 maxWeight=5 addNoise=FALSE b=0 sparsity=0.7 output=linRegData.csv format=csv perc=0.5
15/11/11 18:16:33 INFO api.DMLScript: BEGIN DML run 11/11/2015 18:16:33
15/11/11 18:16:33 INFO api.DMLScript: HADOOP_HOME: /home/hadoop/hadoop-2.6.2
15/11/11 18:16:33 WARN conf.DMLConfig: No default SystemML config file (./SystemML-config.xml) found
15/11/11 18:16:33 WARN conf.DMLConfig: Using default settings in DMLConfig
15/11/11 18:16:33 INFO jvm.JvmMetrics: Initializing JVM Metrics with processName=JobTracker, sessionId=
15/11/11 18:16:33 WARN hops.OptimizerUtils: Auto-disable multi-threaded text read for 'text' and 'csv' due to thread contention on JRE < 1.8 (java.version=1.7.0_79).
15/11/11 18:16:35 INFO api.DMLScript: SystemML Statistics:
Total execution time:		1.484 sec.
Number of executed MR Jobs:	0.

15/11/11 18:16:35 INFO api.DMLScript: END DML run 11/11/2015 18:16:35

If we list the contents of the current directory in our regular file system, we see that no files have been written to the regular file system.

[hadoop@host1 ~]$ ls
genLinearRegressionData.dml  hadoop-2.6.2  hadoop-2.6.2.tar.gz  systemml-1.2.0-SNAPSHOT  systemml-1.2.0-SNAPSHOT.tar.gz

If we list the contents of the HDFS file system, we see that HDFS contains our data files and the corresponding metadata files.

[hadoop@host1 ~]$ hdfs dfs -ls
Found 5 items
drwxr-xr-x   - hadoop supergroup          0 2015-11-11 18:16 linRegData.csv
-rw-r--r--   1 hadoop supergroup        214 2015-11-11 18:16 linRegData.csv.mtd
drwxr-xr-x   - hadoop supergroup          0 2015-11-11 18:16 perc.csv
-rw-r--r--   1 hadoop supergroup        206 2015-11-11 18:16 perc.csv.mtd
drwxr-xr-x   - hadoop supergroup          0 2015-11-11 18:16 scratch_space

If we examine the Hadoop web interface mentioned previously, we see that the files, directories, and blocks in HDFS have increased in number.

Now that we’re done with this example, I’ll clean things up and delete the generated files from HDFS.

[hadoop@host1 hadoop]$ hdfs dfs -rm -r *.csv
[hadoop@host1 hadoop]$ hdfs dfs -rm *.mtd
[hadoop@host1 hadoop]$ hdfs dfs -rmdir scratch_space

I’ll stop HDFS using the stop-dfs.sh script and then verify that the Java processes have stopped.

[hadoop@host1 ~]$ stop-dfs.sh
Stopping namenodes on [localhost]
localhost: stopping namenode
localhost: stopping datanode
Stopping secondary namenodes [0.0.0.0]
0.0.0.0: stopping secondarynamenode

[hadoop@host1 ~]$ jps
37337 Jps

SystemML with Pseudo-Distributed Hadoop and YARN

To add YARN to Pseudo-Distributed Hadoop on the single machine, we need to take our setup from the previous example and update two configuration files and start the ResourceManager and NodeManager daemons.

In the mapred-site.xml configuration file, we specify the mapreduce.framework.name property as yarn.

[hadoop@host1 hadoop]$ pwd
/home/hadoop/hadoop-2.6.2/etc/hadoop
[hadoop@host1 hadoop]$ cp mapred-site.xml.template mapred-site.xml
[hadoop@host1 hadoop]$ vi mapred-site.xml

...
<configuration>
    <property>
        <name>mapreduce.framework.name</name>
        <value>yarn</value>
    </property>
</configuration>
...

In the yarn-site.xml configuration file, we specify the yarn.nodemanager.aux-services property to be mapreduce_shuffle.

[hadoop@host1 hadoop]$ vi yarn-site.xml

...
<configuration>
    <property>
        <name>yarn.nodemanager.aux-services</name>
        <value>mapreduce_shuffle</value>
    </property>
</configuration>
...

Next, we’ll start HDFS using the start-dfs.sh script.

[hadoop@host1 hadoop]$ start-dfs.sh
Starting namenodes on [localhost]
localhost: starting namenode, logging to /home/hadoop/hadoop-2.6.2/logs/hadoop-hadoop-namenode-host1.out
localhost: starting datanode, logging to /home/hadoop/hadoop-2.6.2/logs/hadoop-hadoop-datanode-host1.out
Starting secondary namenodes [0.0.0.0]
0.0.0.0: starting secondarynamenode, logging to /home/hadoop/hadoop-2.6.2/logs/hadoop-hadoop-secondarynamenode-host1.out

After that, we’ll start YARN using the start-yarn.sh script.

[hadoop@host1 hadoop]$ start-yarn.sh
starting yarn daemons
starting resourcemanager, logging to /home/hadoop/hadoop-2.6.2/logs/yarn-hadoop-resourcemanager-host1.out
localhost: starting nodemanager, logging to /home/hadoop/hadoop-2.6.2/logs/yarn-hadoop-nodemanager-host1.out

We can use the jps command to verify that the HDFS daemons (NameNode, DataNode, and SecondaryNameNode) and YARN daemons (ResourceManager and NodeManager) are running.

[hadoop@host1 hadoop]$ jps
ResourceManager
Jps
NodeManager
NameNode
DataNode
SecondaryNameNode

We can now view YARN information via the web interface on port 8088 (http://host1.example.com:8088).

I’ll execute the genLinearRegressionData.dml example that we’ve previously considered.

[hadoop@host1 hadoop]$ cd ~
[hadoop@host1 ~]$ hadoop jar systemml-1.2.0-SNAPSHOT/SystemML.jar -f genLinearRegressionData.dml -nvargs numSamples=1000 numFeatures=50 maxFeatureValue=5 maxWeight=5 addNoise=FALSE b=0 sparsity=0.7 output=linRegData.csv format=csv perc=0.5
15/11/12 11:57:04 INFO api.DMLScript: BEGIN DML run 11/12/2015 11:57:04
15/11/12 11:57:04 INFO api.DMLScript: HADOOP_HOME: /home/hadoop/hadoop-2.6.2
15/11/12 11:57:04 WARN conf.DMLConfig: No default SystemML config file (./SystemML-config.xml) found
15/11/12 11:57:04 WARN conf.DMLConfig: Using default settings in DMLConfig
15/11/12 11:57:05 INFO client.RMProxy: Connecting to ResourceManager at /0.0.0.0:8032
15/11/12 11:57:06 WARN hops.OptimizerUtils: Auto-disable multi-threaded text read for 'text' and 'csv' due to thread contention on JRE < 1.8 (java.version=1.7.0_79).
15/11/12 11:57:07 INFO api.DMLScript: SystemML Statistics:
Total execution time:		1.265 sec.
Number of executed MR Jobs:	0.

15/11/12 11:57:07 INFO api.DMLScript: END DML run 11/12/2015 11:57:07

If we examine the HDFS file system, we see the files generated by the execution of the DML script by SystemML on Hadoop.

[hadoop@host1 ~]$ hdfs dfs -ls
Found 5 items
drwxr-xr-x   - hadoop supergroup          0 2015-11-12 11:57 linRegData.csv
-rw-r--r--   1 hadoop supergroup        214 2015-11-12 11:57 linRegData.csv.mtd
drwxr-xr-x   - hadoop supergroup          0 2015-11-12 11:57 perc.csv
-rw-r--r--   1 hadoop supergroup        206 2015-11-12 11:57 perc.csv.mtd
drwxr-xr-x   - hadoop supergroup          0 2015-11-12 11:57 scratch_space

I’ll go ahead and delete the generated example files from HDFS.

[hadoop@host1 ~]$ hdfs dfs -rm -r *.csv
[hadoop@host1 ~]$ hdfs dfs -rm *.mtd
[hadoop@host1 ~]$ hdfs dfs -rmdir scratch_space

We’ll stop the YARN daemons using the stop-yarn.sh script.

[hadoop@host1 ~]$ stop-yarn.sh
stopping yarn daemons
stopping resourcemanager
localhost: stopping nodemanager
no proxyserver to stop

We can stop HDFS with the stop-dfs.sh script.

[hadoop@host1 ~]$ stop-dfs.sh
Stopping namenodes on [localhost]
localhost: stopping namenode
localhost: stopping datanode
Stopping secondary namenodes [0.0.0.0]
0.0.0.0: stopping secondarynamenode

If we list the running Java processes, we see all the YARN daemons and HDFS daemons have stopped.

[hadoop@host1 ~]$ jps
53459 Jps

For cleanliness, I’ll also delete the /tmp/hadoop-hadoop files created by Hadoop before proceeding to the next example.

SystemML with Distributed Hadoop and YARN

In our previous example, we ran SystemML on Hadoop in Pseudo-Distributed mode with YARN on a single machine. This example will look at Distributed Hadoop with YARN on a 4-node cluster. Each server is running Red Hat Enterprise Linux Server, release 6.6.

I have 4 nodes: host1, host2, host3, and host4. The host1 node that we previously set up will act as the master for both HDFS and YARN, and host2, bd150, and host4 will be slaves. For more information regarding network configurations, please see the Hadoop documentation.

First, I created a hadoop user on each slave node.

[root@host1 ~]$ ssh root@host2.example.com
[root@host2 ~]# useradd hadoop
[root@host2 ~]# passwd hadoop
[root@host2 ~]# exit

[root@host1 ~]$ ssh root@host3.example.com
[root@host2 ~]# useradd hadoop
[root@host2 ~]# passwd hadoop
[root@host2 ~]# exit

[root@host1 ~]$ ssh root@host4.example.com
[root@host2 ~]# useradd hadoop
[root@host2 ~]# passwd hadoop
[root@host2 ~]# exit

Next, I set up passwordless login from the hadoop user on the master node (host1) to each of the slave nodes. The ssh-copy-id command copied the master node’s hadoop user’s public key value to the ~/.ssh/authorized_keys file of each of the slave nodes. I tested the passwordless login from the master node to each of the slave nodes for the hadoop user.

$ ssh hadoop@host1.example.com

[hadoop@host1 ~]$ ssh-copy-id host2.example.com
[hadoop@host1 ~]$ ssh hadoop@host2.example.com
Last login: Thu Nov 12 14:16:21 2015
[hadoop@host2 ~]$ exit

[hadoop@host1 ~]$ ssh-copy-id host3.example.com
[hadoop@host1 ~]$ ssh hadoop@host3.example.com
Last login: Thu Nov 12 14:16:40 2015
[hadoop@host3 ~]$ exit

[hadoop@host1 ~]$ ssh-copy-id host4.example.com
[hadoop@host1 ~]$ ssh hadoop@host4.example.com
Last login: Thu Nov 12 14:17:10 2015
[hadoop@host4 ~]$ exit

On the master node, I specified the slave nodes in the Hadoop slaves configuration file.

[hadoop@host1 hadoop]$ pwd
/home/hadoop/hadoop-2.6.2/etc/hadoop
[hadoop@host1 hadoop]$ more slaves
host2.example.com
host3.example.com
host4.example.com

In the core-site.xml file, I specified the fs.defaultFS property to reference the master node.

[hadoop@host1 hadoop]$ more core-site.xml

...
<configuration>
    <property>
        <name>fs.defaultFS</name>
        <value>hdfs://host1.example.com:9000</value>
    </property>
</configuration>
...

In the hdfs-site.xml configuration file, I removed the previous dfs.replication property, since we will use the default replication value (of 3).

[hadoop@host1 hadoop]$ more hdfs-site.xml

...
<configuration>
</configuration>
...

We’ll be using YARN, so our mapred-site.xml will have the mapreduce.framework.name property set to yarn, as in the previous example. Additionally, we’ll set the mapreduce.map.java.opts and mapreduce.reduce.java.opts properties to -Xmx2g -Xms2g -Xmn200m. The -Xmn parameter fixes the size of the young generation and typically is set to 10% of the maximum heap, which we have set to 2g. Furthermore, we’ll set mapreduce.map.memory.mb and mapreduce.reduce.memory.mb to 3072. Typically these values are set to at least 1.5 times the value of the maximum heap size.

[hadoop@host1 hadoop]$ more mapred-site.xml

...
<configuration>
    <property>
        <name>mapreduce.framework.name</name>
        <value>yarn</value>
    </property>
    <property>
        <name>mapreduce.map.java.opts</name>
        <value>-Xmx2g -Xms2g -Xmn200m</value>
    </property>
    <property>
        <name>mapreduce.reduce.java.opts</name>
        <value>-Xmx2g -Xms2g -Xmn200m</value>
    </property>
    <property>
        <name>mapreduce.map.memory.mb</name>
        <value>3072</value>
    </property>
    <property>
        <name>mapreduce.reduce.memory.mb</name>
        <value>3072</value>
    </property>
</configuration>
...

In the yarn-site.xml configuration file, I added a yarn.resourcemanager.hostname property and specified the master node as the host.

[hadoop@host1 hadoop]$ more yarn-site.xml

...
<configuration>
    <property>
        <name>yarn.nodemanager.aux-services</name>
        <value>mapreduce_shuffle</value>
    </property>
    <property>
        <name>yarn.resourcemanager.hostname</name>
        <value>host1.example.com</value>
    </property>
</configuration>
...

In the previous example, we specified the JAVA_HOME in the hadoop-env.sh configuration script. We will use that same value.

[hadoop@host1 hadoop]$ more hadoop-env.sh

...
export JAVA_HOME=/usr/lib/jvm/java-1.7.0-openjdk.x86_64
...

Next, I copied my hadoop installation (which includes all of the mentioned configuration settings) to each slave node.

[hadoop@host1 ~]$ pwd
/home/hadoop
[hadoop@host1 ~]$ scp -r hadoop-2.6.2 hadoop@host2.example.com:~/
[hadoop@host1 ~]$ scp -r hadoop-2.6.2 hadoop@host3.example.com:~/
[hadoop@host1 ~]$ scp -r hadoop-2.6.2 hadoop@host4.example.com:~/

My master node .bash_profile contains JAVA_HOME and HADOOP_HOME environment variables and adds JAVA_HOME/bin, HADOOP_HOME/bin and HADOOP_HOME/sbin to the PATH.

...
export JAVA_HOME=/usr/lib/jvm/java-1.7.0-openjdk.x86_64
export HADOOP_HOME=/home/hadoop/hadoop-2.6.2
PATH=$JAVA_HOME/bin:$PATH:$HOME/bin:$HADOOP_HOME/bin:$HADOOP_HOME/sbin
export PATH
...

I copied the .bash_profile file to the slave nodes.

[hadoop@host1 ~]$ pwd
/home/hadoop
[hadoop@host1 ~]$ scp .bash_profile hadoop@host2.example.com:~/.bash_profile
[hadoop@host1 ~]$ scp .bash_profile hadoop@host3.example.com:~/.bash_profile
[hadoop@host1 ~]$ scp .bash_profile hadoop@host4.example.com:~/.bash_profile

On the master, I formatted HDFS.

[hadoop@host1 ~]$ hdfs namenode -format

Next, on the master, I started HDFS using start-dfs.sh. We can see that the master NameNode and the slave DataNodes started up.

[hadoop@host1 ~]$ start-dfs.sh
Starting namenodes on [host1.example.com]
host1.example.com: starting namenode, logging to /home/hadoop/hadoop-2.6.2/logs/hadoop-hadoop-namenode-host1.out
host4.example.com: starting datanode, logging to /home/hadoop/hadoop-2.6.2/logs/hadoop-hadoop-datanode-host4.out
host2.example.com: starting datanode, logging to /home/hadoop/hadoop-2.6.2/logs/hadoop-hadoop-datanode-host2.out
host3.example.com: starting datanode, logging to /home/hadoop/hadoop-2.6.2/logs/hadoop-hadoop-datanode-host3.out
Starting secondary namenodes [0.0.0.0]
0.0.0.0: starting secondarynamenode, logging to /home/hadoop/hadoop-2.6.2/logs/hadoop-hadoop-secondarynamenode-host1.out

Next I started YARN using the start-yarn.sh script. We see the master ResourceManager and the slave NodeManagers started up.

[hadoop@host1 ~]$ start-yarn.sh
starting yarn daemons
starting resourcemanager, logging to /home/hadoop/hadoop-2.6.2/logs/yarn-hadoop-resourcemanager-host1.out
host3.example.com: starting nodemanager, logging to /home/hadoop/hadoop-2.6.2/logs/yarn-hadoop-nodemanager-host3.out
host2.example.com: starting nodemanager, logging to /home/hadoop/hadoop-2.6.2/logs/yarn-hadoop-nodemanager-host2.out
host4.example.com: starting nodemanager, logging to /home/hadoop/hadoop-2.6.2/logs/yarn-hadoop-nodemanager-host4.out

On the master, we see that the NameNode, SecondaryNameNode, and ResourceManager daemons are running.

[hadoop@host1 ~]$ jps
NameNode
SecondaryNameNode
Jps
ResourceManager

On the slaves, we see that the DataNode and NodeManager daemons are running.

[hadoop@host2 ~]$ jps
Jps
NodeManager
DataNode

[hadoop@host3 ~]$ jps
Jps
DataNode
NodeManager

[hadoop@host4 ~]$ jps
Jps
DataNode
NodeManager

If we look at the Hadoop (on port 50070) and YARN (on port 8088) web interfaces, we can see information about our running cluster.

SystemML with Distributed Hadoop and YARN: Linear Regression Example

Let’s go ahead and run the SystemML example from the GitHub README.

[hadoop@host1 ~]$ hadoop jar systemml-1.2.0-SNAPSHOT/SystemML.jar -f genLinearRegressionData.dml -nvargs numSamples=1000 numFeatures=50 maxFeatureValue=5 maxWeight=5 addNoise=FALSE b=0 sparsity=0.7 output=linRegData.csv format=csv perc=0.5

[hadoop@host1 ~]$ hadoop jar systemml-1.2.0-SNAPSHOT/SystemML.jar -f systemml-1.2.0-SNAPSHOT/algorithms/utils/sample.dml -nvargs X=linRegData.csv sv=perc.csv O=linRegDataParts ofmt=csv

[hadoop@host1 ~]$ hadoop jar systemml-1.2.0-SNAPSHOT/SystemML.jar -f systemml-1.2.0-SNAPSHOT/algorithms/utils/splitXY.dml -nvargs X=linRegDataParts/1 y=51 OX=linRegData.train.data.csv OY=linRegData.train.labels.csv ofmt=csv

[hadoop@host1 ~]$ hadoop jar systemml-1.2.0-SNAPSHOT/SystemML.jar -f systemml-1.2.0-SNAPSHOT/algorithms/utils/splitXY.dml -nvargs X=linRegDataParts/2 y=51 OX=linRegData.test.data.csv OY=linRegData.test.labels.csv ofmt=csv

[hadoop@host1 ~]$ hadoop jar systemml-1.2.0-SNAPSHOT/SystemML.jar -f systemml-1.2.0-SNAPSHOT/algorithms/LinearRegDS.dml -nvargs X=linRegData.train.data.csv Y=linRegData.train.labels.csv B=betas.csv fmt=csv
...
BEGIN LINEAR REGRESSION SCRIPT
Reading X and Y...
Calling the Direct Solver...
Computing the statistics...
AVG_TOT_Y,-0.051722694902638956
STDEV_TOT_Y,54.132787822718356
AVG_RES_Y,1.5905895170230406E-10
STDEV_RES_Y,2.0668015575844624E-8
DISPERSION,4.262683023432828E-16
R2,1.0
ADJUSTED_R2,1.0
R2_NOBIAS,1.0
ADJUSTED_R2_NOBIAS,1.0
R2_VS_0,1.0
ADJUSTED_R2_VS_0,1.0
Writing the output matrix...
END LINEAR REGRESSION SCRIPT
15/11/17 15:50:34 INFO api.DMLScript: SystemML Statistics:
Total execution time:		0.480 sec.
...

[hadoop@host1 ~]$ hadoop jar systemml-1.2.0-SNAPSHOT/SystemML.jar -f systemml-1.2.0-SNAPSHOT/algorithms/GLM-predict.dml -nvargs X=linRegData.test.data.csv Y=linRegData.test.labels.csv B=betas.csv fmt=csv
...
LOGLHOOD_Z,,FALSE,NaN
LOGLHOOD_Z_PVAL,,FALSE,NaN
PEARSON_X2,,FALSE,2.5039962709907123E-13
PEARSON_X2_BY_DF,,FALSE,5.703863943031236E-16
PEARSON_X2_PVAL,,FALSE,1.0
DEVIANCE_G2,,FALSE,0.0
DEVIANCE_G2_BY_DF,,FALSE,0.0
DEVIANCE_G2_PVAL,,FALSE,1.0
LOGLHOOD_Z,,TRUE,NaN
LOGLHOOD_Z_PVAL,,TRUE,NaN
PEARSON_X2,,TRUE,2.5039962709907123E-13
PEARSON_X2_BY_DF,,TRUE,5.703863943031236E-16
PEARSON_X2_PVAL,,TRUE,1.0
DEVIANCE_G2,,TRUE,0.0
DEVIANCE_G2_BY_DF,,TRUE,0.0
DEVIANCE_G2_PVAL,,TRUE,1.0
AVG_TOT_Y,1,,0.9381218622147646
STDEV_TOT_Y,1,,55.6116696631821
AVG_RES_Y,1,,2.5577864570734575E-10
STDEV_RES_Y,1,,2.390848397359923E-8
PRED_STDEV_RES,1,TRUE,1.0
R2,1,,1.0
ADJUSTED_R2,1,,1.0
R2_NOBIAS,1,,1.0
ADJUSTED_R2_NOBIAS,1,,1.0
15/11/17 15:51:17 INFO api.DMLScript: SystemML Statistics:
Total execution time:		0.269 sec.
...

If we look at HDFS, we can see the files that were generated by the SystemML DML script executions.

[hadoop@host1 ~]$ hdfs dfs -ls
Found 16 items
drwxr-xr-x   - hadoop supergroup          0 2015-11-17 15:50 betas.csv
-rw-r--r--   3 hadoop supergroup        208 2015-11-17 15:50 betas.csv.mtd
drwxr-xr-x   - hadoop supergroup          0 2015-11-17 15:21 linRegData.csv
-rw-r--r--   3 hadoop supergroup        214 2015-11-17 15:21 linRegData.csv.mtd
drwxr-xr-x   - hadoop supergroup          0 2015-11-17 15:50 linRegData.test.data.csv
-rw-r--r--   3 hadoop supergroup        213 2015-11-17 15:50 linRegData.test.data.csv.mtd
drwxr-xr-x   - hadoop supergroup          0 2015-11-17 15:50 linRegData.test.labels.csv
-rw-r--r--   3 hadoop supergroup        210 2015-11-17 15:50 linRegData.test.labels.csv.mtd
drwxr-xr-x   - hadoop supergroup          0 2015-11-17 15:49 linRegData.train.data.csv
-rw-r--r--   3 hadoop supergroup        213 2015-11-17 15:49 linRegData.train.data.csv.mtd
drwxr-xr-x   - hadoop supergroup          0 2015-11-17 15:49 linRegData.train.labels.csv
-rw-r--r--   3 hadoop supergroup        210 2015-11-17 15:49 linRegData.train.labels.csv.mtd
drwxr-xr-x   - hadoop supergroup          0 2015-11-17 15:49 linRegDataParts
drwxr-xr-x   - hadoop supergroup          0 2015-11-17 15:21 perc.csv
-rw-r--r--   3 hadoop supergroup        206 2015-11-17 15:21 perc.csv.mtd
drwxr-xr-x   - hadoop supergroup          0 2015-11-17 15:21 scratch_space

Before the next example, I’ll delete the files created in HDFS by this example.

[hadoop@host1 ~]$ hdfs dfs -rm -r linRegData*
[hadoop@host1 ~]$ hdfs dfs -rm -r *.csv
[hadoop@host1 ~]$ hdfs dfs -rm -r *.mtd

SystemML with Distributed Hadoop and YARN: K-Means Clustering Example

Our previous example showed SystemML running in Hadoop Batch mode on a 4-node cluster with YARN. However, the size of the data used was trivial. In this example, we’ll generate a slightly larger set of data and then analyze that data with the Kmeans.dml and Kmeans-predict.dml scripts. Information about the SystemML K-means clustering algorithm can be found in the K-Means Clustering section of the SystemML Algorithms Reference.

I’m going to modify my SystemML-config.xml file. I updated the numreducers property to be 6, which is twice my number of data nodes. The numreducers property specifies the number of reduce tasks per MR job.

<numreducers>6</numreducers>

To begin, I’ll download the genRandData4Kmeans.dml script that I’ll use to generate a set of data.

[hadoop@host1 ~]$ wget https://raw.githubusercontent.com/apache/systemml/master/scripts/datagen/genRandData4Kmeans.dml

A description of the named arguments that can be passed in to this script can be found in the comment section at the top of the genRandData4Kmeans.dml file. For data, I’ll generate a matrix X.mtx consisting of 1 million rows and 100 features. I’ll explicitly reference my SystemML-config.xml file, since I’m executing SystemML in Hadoop from my home directory rather than from the SystemML project root directory.

[hadoop@host1 ~]$ hadoop jar systemml-1.2.0-SNAPSHOT/SystemML.jar -f genRandData4Kmeans.dml -config systemml-1.2.0-SNAPSHOT/SystemML-config.xml -nvargs nr=1000000 nf=100 nc=10 dc=10.0 dr=1.0 fbf=100.0 cbf=100.0 X=X.mtx C=C.mtx Y=Y.mtx YbyC=YbyC.mtx

After the data generation has finished, I’ll check HDFS for the amount of space used. The 1M-row matrix X.mtx requires about 2.8GB of space.

[hadoop@host1 ~]$ hdfs dfs -df -h
Filesystem                           Size   Used  Available  Use%
hdfs://host1.example.com:9000  400.7 G  2.8 G    318.7 G    1%

Here we can see the data files that were generated.

[hadoop@host1 ~]$ hdfs dfs -ls
Found 9 items
drwxr-xr-x   - hadoop supergroup          0 2015-11-19 11:53 C.mtx
-rw-r--r--   3 hadoop supergroup        176 2015-11-19 11:53 C.mtx.mtd
drwxr-xr-x   - hadoop supergroup          0 2015-11-19 11:56 X.mtx
-rw-r--r--   3 hadoop supergroup        186 2015-11-19 11:56 X.mtx.mtd
drwxr-xr-x   - hadoop supergroup          0 2015-11-19 11:53 Y.mtx
-rw-r--r--   3 hadoop supergroup        182 2015-11-19 11:53 Y.mtx.mtd
drwxr-xr-x   - hadoop supergroup          0 2015-11-19 11:56 YbyC.mtx
-rw-r--r--   3 hadoop supergroup        182 2015-11-19 11:56 YbyC.mtx.mtd
drwxr-xr-x   - hadoop supergroup          0 2015-11-19 11:56 scratch_space

Here we can see the X.mtx data files.

[hadoop@host1 ~]$ hdfs dfs -ls X.mtx
Found 6 items
-rw-r--r--   1 hadoop supergroup  484418384 2015-11-19 11:56 X.mtx/2-r-00000
-rw-r--r--   1 hadoop supergroup  481626112 2015-11-19 11:56 X.mtx/2-r-00001
-rw-r--r--   1 hadoop supergroup  475834931 2015-11-19 11:56 X.mtx/2-r-00002
-rw-r--r--   1 hadoop supergroup  478519922 2015-11-19 11:56 X.mtx/2-r-00003
-rw-r--r--   1 hadoop supergroup  481624723 2015-11-19 11:56 X.mtx/2-r-00004
-rw-r--r--   1 hadoop supergroup  481624048 2015-11-19 11:56 X.mtx/2-r-00005

Next, I’ll run the Kmeans.dml algorithm on the 1M-row matrix X.mtx.

[hadoop@host1 ~]$ hadoop jar systemml-1.2.0-SNAPSHOT/SystemML.jar -f systemml-1.2.0-SNAPSHOT/algorithms/Kmeans.dml -config /systemml-1.2.0-SNAPSHOT/SystemML-config.xml -nvargs X=X.mtx k=5 C=Centroids.mtx

We can see the Centroids.mtx data file has been written to HDFS.

[hadoop@host1 ~]$ hdfs dfs -ls
Found 11 items
drwxr-xr-x   - hadoop supergroup          0 2015-11-19 11:53 C.mtx
-rw-r--r--   3 hadoop supergroup        176 2015-11-19 11:53 C.mtx.mtd
drwxr-xr-x   - hadoop supergroup          0 2015-11-19 12:10 Centroids.mtx
-rw-r--r--   3 hadoop supergroup        174 2015-11-19 12:10 Centroids.mtx.mtd
drwxr-xr-x   - hadoop supergroup          0 2015-11-19 11:56 X.mtx
-rw-r--r--   3 hadoop supergroup        186 2015-11-19 11:56 X.mtx.mtd
drwxr-xr-x   - hadoop supergroup          0 2015-11-19 11:53 Y.mtx
-rw-r--r--   3 hadoop supergroup        182 2015-11-19 11:53 Y.mtx.mtd
drwxr-xr-x   - hadoop supergroup          0 2015-11-19 11:56 YbyC.mtx
-rw-r--r--   3 hadoop supergroup        182 2015-11-19 11:56 YbyC.mtx.mtd
drwxr-xr-x   - hadoop supergroup          0 2015-11-19 12:10 scratch_space

Now that we have trained our model, next we will test our model. We can do this with the Kmeans-predict.dml script.

[hadoop@host1 ~]$ hadoop jar systemml-1.2.0-SNAPSHOT/SystemML.jar -f systemml-1.2.0-SNAPSHOT/algorithms/Kmeans-predict.dml -config systemml-1.2.0-SNAPSHOT/SystemML-config.xml -nvargs X=X.mtx C=Centroids.mtx prY=PredY.mtx O=stats.txt

In the file system, we can see that the PredY.mtx matrix was created. The stats.txt file lists statistics about the results.

[hadoop@host1 ~]$ hdfs dfs -ls
Found 15 items
drwxr-xr-x   - hadoop supergroup          0 2015-11-19 11:53 C.mtx
-rw-r--r--   3 hadoop supergroup        176 2015-11-19 11:53 C.mtx.mtd
drwxr-xr-x   - hadoop supergroup          0 2015-11-19 12:10 Centroids.mtx
-rw-r--r--   3 hadoop supergroup        174 2015-11-19 12:10 Centroids.mtx.mtd
drwxr-xr-x   - hadoop supergroup          0 2015-11-19 13:20 PredY.mtx
-rw-r--r--   3 hadoop supergroup        182 2015-11-19 13:20 PredY.mtx.mtd
drwxr-xr-x   - hadoop supergroup          0 2015-11-19 11:56 X.mtx
-rw-r--r--   3 hadoop supergroup        186 2015-11-19 11:56 X.mtx.mtd
drwxr-xr-x   - hadoop supergroup          0 2015-11-19 11:53 Y.mtx
-rw-r--r--   3 hadoop supergroup        182 2015-11-19 11:53 Y.mtx.mtd
drwxr-xr-x   - hadoop supergroup          0 2015-11-19 11:56 YbyC.mtx
-rw-r--r--   3 hadoop supergroup        182 2015-11-19 11:56 YbyC.mtx.mtd
drwxr-xr-x   - hadoop supergroup          0 2015-11-19 13:21 scratch_space
-rw-r--r--   3 hadoop supergroup        261 2015-11-19 13:21 stats.txt
-rw-r--r--   3 hadoop supergroup        127 2015-11-19 13:21 stats.txt.mtd

The PredY.mtx matrix consists of a single column of a million rows of doubles, as we can see in the resulting metadata file.

[hadoop@host1 ~]$ hdfs dfs -cat PredY.mtx.mtd
{
    "data_type": "matrix"
    ,"value_type": "double"
    ,"rows": 1000000
    ,"cols": 1
    ,"nnz": 1000000
    ,"format": "text"
    ,"description": { "author": "SystemML" }
}

The statistics generated from testing the method are displayed below.

[hadoop@host1 ~]$ hdfs dfs -cat stats.txt
TSS,,1.1262427174414966E11
WCSS_M,,9.77022617396343E10
WCSS_M_PC,,86.75062686450579
BCSS_M,,1.4922010004515366E10
BCSS_M_PC,,13.249373135494215
WCSS_C,,9.770230517014426E10
WCSS_C_PC,,86.75066542680617
BCSS_C,,1.4921964103415842E10
BCSS_C_PC,,13.249332379537428

Recommended Hadoop Cluster Configuration Settings

Below are some recommended Hadoop configuration file settings that may be of assistance when running SystemML on Hadoop in a clustered environment.

Configuration File	Setting	Value	Description
`mapred-site.xml`	`mapreduce.map.java.opts`	`-Xmx2g -Xms2g -Xmn200m`	Increase memory of child JVMs of Maps, max/min heap size of 2GB, -Xmn specifies young generation which is typically 10% of maximum heap size
	`mapreduce.reduce.java.opts`	`-Xmx2g -Xms2g -Xmn200m`	Increase memory of child JVMs of Reduces, max/min heap size of 2GB, -Xmn specifies young generation which is typically 10% of maximum heap size
	`mapreduce.map.memory.mb`	`3072`	Set to at least 1.5 times the value of the Map max heap size
	`mapreduce.reduce.memory.mb`	`3072`	Set to at least 1.5 times the value of the Reduce max heap size
	`io.sort.mb` (deprecated) / `mapreduce.task.io.sort.mb`	`384`	Memory limit while sorting data
`yarn-site.xml`	`yarn.nodemanager.vmem-pmem-ratio`	`2` to `5`	Maximum ratio of virtual memory to physical memory