topic delay performance overhead

[DanBmh]

When trying to measure the topic delay of camera images for a real time setup, I found that ros2 topic delay is not very well suited for this:

• too few digits and metrics to measure millisecond delays
• using python to calculate the delay introduces a notable overhead

This can be well seen if comparing it against a custom c++ tooling i've created for those reasons:

ros2 topic delay /camera02/pylon_ros2_camera_node/image_raw 
average delay: 0.014 	min: 0.014s max: 0.015s std dev: 0.00016s window: 131
average delay: 0.014	min: 0.014s max: 0.015s std dev: 0.00015s window: 181
average delay: 0.014	min: 0.014s max: 0.015s std dev: 0.00015s window: 231
average delay: 0.014	min: 0.014s max: 0.015s std dev: 0.00014s window: 281
average delay: 0.014	min: 0.014s max: 0.015s std dev: 0.00014s window: 331

ros2 run micro_delay micro_delay --ros-args -p topic:=/camera02/pylon_ros2_camera_node/image_raw
Avg: 0.012300s | Min: 0.012173s | P5: 0.012185s | P50: 0.012308s | P95: 0.012422s | Max: 0.012436s
Avg: 0.012321s | Min: 0.012186s | P5: 0.012194s | P50: 0.012323s | P95: 0.012513s | Max: 0.012531s
Avg: 0.012347s | Min: 0.012180s | P5: 0.012197s | P50: 0.012352s | P95: 0.012506s | Max: 0.012522s
Avg: 0.012358s | Min: 0.012172s | P5: 0.012194s | P50: 0.012374s | P95: 0.012473s | Max: 0.012477s
Avg: 0.012375s | Min: 0.012251s | P5: 0.012266s | P50: 0.012403s | P95: 0.012460s | Max: 0.012463s

The delay measured by my custom c++ tool is around 2ms less than the one reported by ros2 topic delay. I assume the main difference is caused by the better optimized programming language.

One important advantage of the python implementation here is that it works with arbitrary topics, while my c++ version only with sensor_msgs::msg::Image. So I can't provide an adequate replacement, but only inform about this possible performance problem.

The code for the micro_delay tool can be found here: https://gitlab.com/Percipiote/CamStreamer/-/blob/master/extras/delay/micro_delay/src/micro_delay.cpp

[fujitatomoya]

this is known issue for the ros2cli performance. as far as i know, there isn't any plan to address this directly such as using another language.... similar issue, #1043