High Priority First Flow Controller

Submitted by rajive on Wed, 10/09/2013 - 19:06

Concept

A flow controller can be used to shape the output traffic. There are many types of flow controllers, including a high priority one (DDS_HPF_FLOW_CONTROLLER_SCHED_POLICY) that sends out higher priority samples first. Priority can be assigned on a per sample basis, per channel basis (if using multiple channels), or per writer basis.

Using a flow-controller requires the use of an asynchronous publishing (ASYNCHRONOUS_PUBLISH_MODE_QOS). With asynchronous publishing, a DataWriter's write() call simply queues up the messages. A background thread delivers the actual samples based on the scheduling policy specified by the flow controller.

Large data is defined as samples larger than the message_size_max of the installed transports. Sending large data requires a sample to be fragmented into smaller messages, and reassembled at the other end. In RTI Connext 7.3.0 and below, sending large data reliably required the use of an asynchronous publisher. Starting in release 7.4.0, this is no longer a requirement. Reliable large data can be sent either synchronously or asynchronously with a flow controller.

The default QoS configuration must be modified to use a high priority flow controller and asynchronous publishing. Further optimizations can be made for large data.

Example Description

This Example shows how to configure a High Priority First Flow Controller, and demonstrates it for Large Data being sent reliably. The following changes are made to the default QoS configuration file auto-generated by rtiddsgen.

  1. Increase the transport sizes
    • Send and receive buffers
    • Gather-send buffers to maximize message coalescing/batching for a destination
    • May require the OS to be tuned to match these settings. Stock Ubuntu Linux 12.04 appears to be okay out of the box.
  2. Increase the participant resources to increase the maximum message size
  3. Install a High Priority First Flow Controller on the participant. The example also shows how to tune the flow controller for a certain rate. This can be skipped/customized as needed.
  4. Configure data writer to use the High Priority First flow controller installed with the participant. Assigns a priority of 10 to the DataWriiter (for illustration).
    1. Assuming LARGE DATA will be sent, optimize the RESOURCE_LIMITS and the PROTOCOL. Those settings can be skipped if large data is not being used.
  5. Assuming LARGE DATA will be sent, optimize the settings. These settings can be skipped if large data is not being used.

Note: These changes are documented as XML comments in USER_QOS_PROFILES.xml.

The example uses a modified HelloWorld example to show the High Priority First Flow Controller. The stock examples is modified as follows:

  1. $NDDSHOME/example/CPP/Hello_idl/src/HelloWorld.idl is copied to hpf.idl, and modified to increase the
    HPF_MAX_PAYLOAD_SIZE = 8192*16; // 131MB
    so that we can easily test the the flow controller with asynchronous large data.
  2. A HelloWorld example is generated in C++
    rtiddsgen -example x64Darwin10gcc4.2.1 hpf.idl
  3. The generated hpf_publisher.cxx is modified top add a line tagged with <<HPF>>
    instance->payload.ensure_length(HPF_MAX_PAYLOAD_SIZE, HPF_MAX_PAYLOAD_SIZE); //<<HPF>>
  4. The USER_QOS_PROFILES.xml is modified to configure a High Priority First Flow controller.

References for the current release:

  1. User's Manual: FlowControllers (DDS Extension)
  2. User's Manual: Setting the Maximum Gather-Send Buffer Count for UDPv4 and UDPv6
  3. Best Practices: Tune Your OS for Performance

References for release 5.0.0:

  1. RTI_CoreLibrariesAndUtilities_UsersManual, Section 6.6: FlowControllers (DDS Extension)
  2. RTI_CoreLibrariesAndUtilities_UsersManual, Section 15.6.2: Setting the Maximum Gather-Send Buffer Count for UDPv4 and UDPv6
  3. Best Practice: Tune Your OS for Performance

Browse the files in GitHub

Product:
Languages:
Tags: