7. Using RTI Perftest for Large Samples

RTI Perftest can send samples from 28 Bytes to 2,147,483,135 Bytes (2 GBytes - 512 Bytes - 8 Bytes), which corresponds to the maximum payload that RTI Connext DDS is able to send in a single sample.

The size of data is configured by the Command-Line Parameter -dataLen <bytes>. Depending on this parameter, RTI Perftest will automatically configure certain RTI Connext DDS behaviors.

In the case that the sample size is smaller or equal to 63000 Bytes, RTI Perftest will, by default, use types with Bounded-Sequences (bound set to 63000 elements). If the sample size is bigger than 63000 Bytes, RTI Perftest will automatically switch to equivalent types to the ones mentioned previously, but with Unbounded-Sequences.

The reason behind this behavior is that in the case when RTI Perftest uses Unbounded-Sequences, RTI Connext DDS will not pre-allocate the sequences to their maximum size (as opposite as when using bounded sequences). For Unbounded-Members, the code generated by RTI Connext DDS will de-serialize the samples by dynamically allocating and de-allocating memory to accommodate the actual size of the unbounded member. Unbounded-Sequences and strings are also supported with DynamicData (Command-Line parameter -DynamicData).

Apart from the use of Unbounded-Sequences, by setting samples bigger than 63000 Bytes, RTI Perftest will enable the use of Asynchronous Publishing, as set RTI Connext DDS default flow controller.

The case where the user wants to use Unbounded-Sequences, Asynchronous Publishing or a Flow Controller different than the default one but the sample size is smaller than 63000 Bytes is also supported. These behaviors can be achieved by using the Command-Line Parameters -unbounded <managerMemory>, -asynchronous and -flowController, see the Test Parameters section for more details.

7.1. Adjusting the configuration

Find here an example where the sample size is configured to 1GB:

  • Publisher:
bin/<architecture>/<release or debug>/perftest_cpp -pub -dataLen 1073741824
  • Subscriber:
bin/<architecture>/<release or debug>/perftest_cpp -sub -dataLen 1073741824

This is a perfectly valid configuration, accepted by RTI Perftest, however, in certain cases, the communication in this scenario will be limited or non-optimal, due to the large sample size. Therefore some extra tuning might be required:

7.1.1. By using the parameter -sendQueueSize <number>

The default value for the Send Queue in the Writer side is 50. That might be a very high value, and the RTI Connext DDS middleware might not perform in optimal conditions. Therefore, for large data samples it is recommended to reduce the send Queue to lower values.

7.1.2. By using the parameter -pubRate <samples/s>

This parameter can be used in order to limit the frequency in which RTI Perftest publishes samples. This can help reducing the number of packages in the network, helping to achieve better Latency and Throughput numbers.

7.1.3. By using a flow controller -flowController <default,1Gbps,10Gbps>

Since the sample size is bigger than 63000 Bytes, RTI Perftest will enable Asynchronous Publing. By enabling that, you also make use of the default Flow Controller, which might not be optimal. It might be a good practice to specify a Flow Controller tailored for the network where the scenario is.

RTI Perftest provides options to use a flow controller designed for a 10Gbps network and a 1Gbps one. However, by accessing to the perftest_qos_profiles.xml configuration file it is possible to modify these 2 Flow Controllers and tailor them to specific network requirements.

<qos_profile name="BaseProfileQos">
    <participant_qos>
        . . .
        <property>
            <value>
                <element>
                    <name>dds.flow_controller.token_bucket.10Gbps.token_bucket.max_tokens</name>
                    <value>300</value>
                </element>
                <element>
                    <name>dds.flow_controller.token_bucket.10Gbps.token_bucket.tokens_added_per_period</name>
                    <value>200</value>
                </element>
                <element>
                    <name>dds.flow_controller.token_bucket.10Gbps.token_bucket.bytes_per_token</name>
                    <value>65536</value>
                </element>
                <element>
                    <name>dds.flow_controller.token_bucket.10Gbps.token_bucket.period.sec</name>
                    <value>0</value>
                </element>
                <element>
                    <name>dds.flow_controller.token_bucket.10Gbps.token_bucket.period.nanosec</name>
                    <value>10000000</value>
                </element>
            </value>
        </property>
        . . .
    </participant_qos>
</qos_profile>

The specific values for the Flow Controller and the Send Queue will highly depend on the scenario and machines performing the test, but as a general suggestion, these changes are recommended:

  • Publisher:
bin/<architecture>/<release or debug>/perftest_cpp -pub -dataLen 1073741824 -sendQueueSize 1 -flowController 1Gbps
  • Subscriber:
bin/<architecture>/<release or debug>/perftest_cpp -sub -dataLen 1073741824

7.2. Large Samples in Java

When using the RTI Perftest implementation for Java and large data samples, the following error may appear:

Exception in thread "main" java.lang.OutOfMemoryError: Java heap space

The the memory reserved for the heap is not enough in this case, the way how to solve this is by increasing the size we allow Java to reserve. This can be done by using the Command-Line Parameter -Xmx in the scripts used to run the Java examples: bin/Release/perftest_java.sh and bin\Release\perftest_java.bat. The increased amount will depend on the -dataLen parameter and the memory specifications of device where RTI Perftest is running.