ros2 topic hz perform worse than c++ code, also worse than result in another docker container

Bug report

Required Info:

• Operating System:
  • Ubuntu18.04
• Installation type:
  • source
• Version or commit hash:
  • humble
• DDS implementation:
  • FastDDS

Steps to reproduce issue

• run two docker container using the following docker-compose.yml

container1:
    image: image_name
    network_mode: "host"
    cpuset: 2-9
    cpu_shares: 2048
    shm_size: 4G
    container_name: container1
    runtime: nvidia
    privileged: true
    environment:
      - NVIDIA_VISIBLE_DEVICES=all
    logging:
      options:
        max-size: "${DOCKER_LOGS_MAX_SIZE:-100M}"

container2:
    image: image_name
    network_mode: "host"
    privileged: true
    cpuset: 12-17
    cpu_shares: 2048
    shm_size: 2G
    container_name: container2
    environment:
      - NVIDIA_VISIBLE_DEVICES=all
    runtime: nvidia
    logging:
      options:
        max-size: "${DOCKER_LOGS_MAX_SIZE:-100M}"

• publish sensor_msgs::msg::PointCloud2 type message with frequency 10hz in container1
• run subscriber and process data c++ code in container2, record the frequency subscription callback called
• use ros2 topic hz tool follow the frequency of topic in both containers

Expected behavior

• In both containers, ros2 topic hz shows the same frequency which is around 10hz
• Log of c++ code in container2 shows that pointcloud2 messages arrive each 100ms

Actual behavior

• In container1 terminal, ros2 topic hz shows frequency slightly under 10hz
  
• In container2 terminal, ros2 topic hz shows frequency highly under 10hz, verity from 7hz to 9hz
  
• In log of container2, c++ code receive message each 100ms stably
  

Additional information

Since code in both container are quite complicate, I can't paste the code here.
As discussed in the blow question, ros2 topic hz result is lower than c++ code because the qos used is different. But why the hz results in different containers turn out to be different?

https://answers.ros.org/question/350753/ros2-topic-hz-provides-wrong-rate-for-larger-msgs/
Also, since ros2 topic hz result is not reliable, is therea a better way to track the transportaion stability between many nodes?

Start by reproducing ros2 topic hz with the provided two-container Docker Compose setup and compare its readings with the C++ subscription callback. Read the ros2 topic hz entry point and investigate the differing results for the PointCloud2 topic, then document a reliable explanation or measurement approach.