What kind of performance numbers are you looking for? Latency, Mem usage, CPU usage? I've recently been working on reducing the CPU overhead for ROS2. Once improvements are made there, people will tackle the RMW layer to get ROS2 to be more deterministic. When both of those are implemented I think we will see a performance increase and will be able to properly predict/measure latency. http://drops.dagstuhl.de/opus/volltexte/2019/10743/pdf/LIPIcs-ECRTS-2019-6.pdf <- this paper mentions the determinism issue that is currently present and gives some numbers on latency.

Your question also suggest you are more interested in performance than the features that DDS has to offer. The main reason to use ROS2 is to benefit from DDS' many QoS settings. This could come at the cost of some performance (though as I mentioned, people are working on this). If you do not really care about using DDS, you could plug a different communication layer at the bottom of ROS2 (assuming there is an rmw implementation), but this would kinda defeat the purpose of using ROS2.

@MCornelis I am mainly interested in message latency.

I'm (at the moment) not aware of more papers that could help you out. I do know that Apex AI has worked on benchmarking ROS2 and that Bosch is involved in creating tracing tools for ROS2. You could try contacting people with knowledge about the subject matter by posting on ROS discourse (requesting papers/ information).

This discourse post https://discourse.ros.org/t/latency-and-throughput-in-ros2/4367 discusses the same paper you have found. You could have a look at the responses, try to revive the post or post a new discussion on the same subject.