Data Flow Scheduler

Note

This is a modified version of the original Data Flow Scheduler with some additional features:

• Configurable activity timeout supervision based on thread CPU time or system time
• n:1 subscription improved
• New examples

ROS2 Data Flow Scheduler

This project was conducted as a master's thesis in collaboration with TTTech Auto AG. The research project aims to enhance the coordination capabilities of ROS2, a widely adopted robotic framework. By enabling cross-process coordination, the project addresses the need for scalable and flexible coordination of executables across multiple ROS2 nodes

Concept

The Data Flow Scheduler employs an innovative concept to enable efficient cross-process coordination within the ROS2 framework. At its core, this concept utilizes a control system responsible for managing and coordinating the execution of callbacks across multiple ROS2 nodes. The system establishes Inter-Process socket Communication (IPC), facilitating the exchange of execution orders and ensuring data flow between nodes.

Installation

• Developed and tested on:

  • WSL2 (Ubuntu 20.04)
  • Ubuntu 22.04

• Tested on:

  • Raspberry Pi 4 B 8GB RAM (Ubuntu Server 22.04)

• Prerequisites:

  • LEMON Graph Library [https://lemon.cs.elte.hu/trac/lemon]

Please refer to the LEMON Graph Library website for installation guidelines and instructions.

Build

To compile the ROS2 package, use colcon build:

colcon build

Alternatively, if you only want to build specific packages, use the --packages-select option:

colcon build --packages-select data_flow_scheduler autoware_reference_system reference_system reference_system_interface

This will build the following packages: data_flow_scheduler, autoware_reference_system, reference_system, and reference_system_interface.

How to use

See examples folder for some examples on how to deploy a node network and the callbacks.

Configuration

Take a look to data_flow_types.h to configure the system.

Some available parameters are:

• Number of parallel activities to dispatch
• Threads priority
• Scheduling policies

Test with custom network

This test deploys a node network to test the latency of the system.

ros2 launch data_flow_scheduler demo_graph_launch.py

Test with Autoware Reference System

This test runs the Autoware Reference System where each node runs in a separate process.

ros2 launch data_flow_scheduler autoware_reference_system_launch.py

Changes in autoware_reference_system

To test the data_flow_scheduler with the reference system nodes the shared pointer to the activities has to be accesssible. Following code has been added for each pointer:

rclcpp::SubscriptionBase::SharedPtr return_subscription_(){
  return subscription_;
}

Also fusion node has been modified to provide subscription for more than 1 topic.

Check the intersection node and the examples folder.

TODO list, future work and limitations

• The DFSched::CallbackInfo, DFSched::NodeInfo, DFSched::TopicInfo and DFSched::TimerInfo needs to be unified.
• Data structures that are holding the graph status can be improved for better perfomance.
• More decision algorithms can be implemented.
• Functions and methods like CallbackHandler::run_callback() needs to be rewritten in a better way
• Dynamic memory during runtime must be avoided
• Thread creation during rumtime must be removed

Tricks

It is possible to disable the longest path in graphcreator.cpp:188 just commenting out the call to compute_longestpath();

This enables a pseudo random choice of the node.

License

The Data Flow Scheduler is released under the Apache 2.0 License. For more details, please refer to the LICENSE file.