Jetson RAM grows continuously with realtime sensors, but stable with rosbag — expected? (ROS1)

[xuanai131]

Hi, I’m running FAST-LIVO2 on a Jetson Orin (8GB), ubuntu 22.04 and ros Humble, all work well (mapping, odometry result are good) except RAM. With realtime sensors (Livox + camera), RSS grows steadily until OOM. With rosbag playback on the same device, RAM stabilizes around 4–5 GB (I have tried to disable pcd_save_en, dense_map_en and enable local_map.map_sliding_en).

I also tried the original repo with ros1 setup (no local modifications) and see the same RAM growth in realtime.

** Environment
Device: Jetson Orin, 8GB RAM
OS: Docker Ubuntu 20.04 on Jetson host Ubuntu 22.04
ROS: ROS1 noetic
Lidar: Livox MID360

** Behavior:
Realtime: RSS grows continuously (e.g., +100–200MB per ~5s) until OOM.
Rosbag: RAM stabilizes around 4–5GB.
Running with ros2 version directly on jetson give the same behavior.

** Questions:

• Is continuous RAM growth expected for realtime runs on Jetson, and is 8GB insufficient? If so, what minimum RAM do you recommend?
• Why would rosbag playback be stable, but realtime grows? Is there a known DDS/QoS or sensor timing issue that causes heap growth? (Even though this is ROS1.)
• Are there recommended parameters or build options to keep memory bounded on embedded devices (e.g., publish settings, map pruning)?
  If there’s a preferred way to profile memory or collect a diagnostic for you, I can run it.

Thanks!

[natthaphxt]

On ubuntu 22.04, which dataset are you using? I tried using the dataset of r3live and hilti dataset and after running it for a while, it showed error -11 or -6 and wouldn't continue.

[xuanai131]

On ubuntu 22.04, which dataset are you using? I tried using the dataset of r3live and hilti dataset and after running it for a while, it showed error -11 or -6 and wouldn't continue.

I'm using dataset collecting from my own hardware setup (arduicam and mid360 livox lidar).

[natthaphxt]

Oh, okay, thank you for your answer. I'd like to know how much editing you needed on the images, especially regarding resolution and various Yamal file settings.

[fzoric8]

Hi @xuanai131 I had similar problem and I think that problem relates to the rviz visualization.

When I turned off rviz, RAM stayed the same. I suppose that ROS 1 noetic rviz has some problems with memory management during PointCloud collection.

[xuanai131]

Hi @fzoric8, I didn't launched rviz on jetson when testing performance, just visualize odometry on rviz PC to see if it drift.

[xuanai131]

Oh, okay, thank you for your answer. I'd like to know how much editing you needed on the images, especially regarding resolution and various Yamal file settings.

Hi @natthaphxt , I just modified camera intrinsic in camera_pinhole.yaml and camera extrinsic in avia.yaml.
And try to enable map_sliding_en, disable pcd_save_en but still getting RAM growth continuously behavior. Are there any changing to other config params can affect to performance?