CPU load with image segmentation example

Robotics & Edge Computing Robotics - Isaac Isaac ROS
1

Hi,

I’m trying to run a default Unet people segmentation pipeline example isaac_ros_unet from here.
As far as I understand all nodes from the pipeline I run (see attached rosgraph and launch file) should use NITROS for communication, so a CPU load should be low. But in real tests the CPU load for composable container is about 300% (see the jtop screenshot below).

So I have a few questions:

  1. Is it a normal CPU load for this pipeline, or I’m doing something wrong?
  2. How can I decrease the CPU load for this and similar pipelines?
  3. How can I check that all nodes are using NITROS?

My setup:

  • Jetson AGX Orin 64GB (Jetpack 5.1.2)
  • LI-JAG-ADP-GMSL2-8CH (P3762_A03 version)
  • 4 x LI OWL cameras

argus_unet_util
argus_unet_util1835×646 87.2 KB

rosgraph
rosgraph5400×3457 726 KB

2

Hi Pavel,

Thank you for your detailed message.

Can you try to check with top or htop if it is only the “component_container” the only process using the CPU?
(jtop provides only the processes that are working with the GPU)

About point 3, I was looking at your graph, and all nodes NITROS accelerated are connected.
You can quickly figure out when you see a topic ending with /nitros connected with another node

3

Hi @Raffaello,

Please find the jtop and htop screenshot below:

htop_and_jtop
htop_and_jtop1850×1055 239 KB

4

Hi @pavel.butov ,

We are currently investigating the high CPU load you reported. I will keep you updated with any news regarding your issue.

Best regards,
Raffaello

5

Hi @Raffaello !

Do you have any updates? This question is very important for us to determine the real benefits of NITROS and to adjust our software architecture accordingly.

6

Hi @pavel.butov

Our engineering team is actively working on resolving the issue you’ve experienced.
As soon as there are any updates, I will promptly notify you.

Thank you for your patience and understanding.