set-up for real-time control in ROS

edit

In our research group we have used OROCOS for several years, and have hired some of the OROCOS developers as consultants to help us tune the real-time performance of our system. I would recommend starting with the easiest to set up system (ROS on a vanilla kernel, components written in Python), and make the following upgrades ONLY as necessary after MEASURING your current performance, and ONLY to the parts of your system that NEED it:

1. Make sure you have your components/nodes that need it running with real-time priority
2. Upgrade your kernel to one with the PREEMPT_RT patches applied
3. Dedicate a fraction of the cores on your CPU (if multicore) to your real-time processes, make sure the non-critical processes are launched on the other cores.
4. Port components to C++ if delays within the components are the bottleneck
5. Port components to Orocos if delays of message transmission between components is the bottleneck
6. Install Xenomai kernel and recompile against it's API's (at this point you are pretty deep into black magic territory; the only people using this are either operating obscenely tight control loops, or implementing safety-critical systems)

Never make any assumptions about where your bottleneck is; ALWAYS measure at every stage of development. Rules of thumb like "avoid dynamic memory allocation" and "only use low-level systems programming languages like C/C++/Fortran" come secondary to this.

Hi, I've mainly used option 2. Theoretically options 2 and 4 should give you the same performance. The benefit of OROCOS is that it makes the integration easier, and makes it possible to switch between different real-time Kernels. OROCOS also provides many other tools to help develop real-time applications, such as methods for real-time modules to communicate with each other.

So your list should include options 5,6, etc. for other real-time Kernels.

e.g. Option 5, ROS on Preempt patched Ubuntu (this is what the PR2 uses)