DNN Image Encoder - High CPU and VRAM usage

Description

When running DNN Encoder on an NVIDIA Jetson Orin NX, the following observations can be made:

  • 1 CPU core is reserved 100% for the encoder step, even if no images enter the pipeline.
  • 17.2G of virtual memory are allocated to this process.

Why is one CPU core being allocated for the task, when no images are running through the pipeline? What is causing high CPU load?
How to manage the allocated Virtual RAM? What factors contribute to it?

Environment

TensorRT Version: 8.5.2.1+cuda11.4
GPU Type: NVIDIA Tegra Orin
Operating System + Version: Ubuntu 20.04.6 LTS
Container (if container which image + tag): ROS2 Humble from: Isaac ROS Common — isaac_ros_docs documentation

Relevant Files

Screenshot form htop:

Steps To Reproduce

Encoder parameters:

  • Input size: 1936x1216
  • Network size:1280x1280

Hi,

This should be application-related.
Could you share which repo/API/sample you use?

Thanks.