Distributed Testing

It is useful to let several JMeter installations process a test plan for large load and stress tests. The nodes report the results via the Java remote method invocation API (RMI) [6] to a central JMeter report server, which is only responsible for collecting and processing data. The tester starts the master server and the slave client as usual from a single installation.

Enter the slave client IP addresses in the bin/jmeter.properties file under remote_hosts or deliver the information to the command line along with the test plan:

jmeter -n -t script.jmx -R Slave1,Slave2,Slave3

Start the test plan loaded on the master via the master's GUI individually, or distribute and start it on all slave clients using Remote Start All. Remote Stop ends the test activities again. A typical distributed test configuration looks like Figure 3. You should store the data after your graphical analysis for documentation and comparison purposes.

Figure 3: A typical distributed master-slave test with three slaves.

Proxies, Logging, Maven

If you want JMeter to test the websites behind a corporate proxy, the tool needs to receive the proxy server as a parameter:

jmeter -H <My.Proxy.Server> -P 8000 \
  -u <username> -a <password> -N localhost

Anyone who wants to start JMeter without an interface to store the results in a separate file should add the parameter

-n -t my_test.jmx -l log.jtl

To list other command-line parameters, use jmeter -?. Larger test plans may make it necessary to optimize the memory settings for the JRE in the startup file – such as matching the HEAP setting to the JVM.

A plugin for the Maven build tool [7] lets you run the load test in the continuous integration process. Integrate and configure the design in Listing 1 into the pom.xml file.

Avoiding False Measurements

It is common knowledge that the measurement itself often influences the measuring result. Practically no load test can completely avoid measurement-related side effects and measurement inaccuracies. Testers should calibrate their measurements in a warm-up phase and repeat any important measurement to minimize effects. Also, be sure to distribute the load across multiple slaves, so the individual test client does not become a bottleneck.